diff --git a/.claude/settings.json b/.claude/settings.json
index 9d91475..c4beb6f 100644
--- a/.claude/settings.json
+++ b/.claude/settings.json
@@ -1,41 +1,5 @@
 {
   "sandbox": {
     "enabled": false
-  },
-  "permissions": {
-    "allow": [
-      "Bash(bash setup.sh*)",
-      "Bash(git remote *)",
-      "Bash(pnpm exec tsx setup/index.ts*)",
-      "Bash(pnpm exec tsx scripts/init-first-agent.ts*)",
-      "Bash(pnpm install @chat-adapter/*)",
-      "Bash(pnpm install chat-adapter-imessage*)",
-      "Bash(pnpm install @bitbasti/chat-adapter-webex*)",
-      "Bash(pnpm install @resend/chat-sdk-adapter*)",
-      "Bash(pnpm install @whiskeysockets/baileys*)",
-      "Bash(pnpm install @beeper/chat-adapter-matrix*)",
-      "Bash(pnpm install @nanoco/nanoclaw-dashboard*)",
-      "Bash(pnpm install --frozen-lockfile*)",
-      "Bash(pnpm run build*)",
-      "Bash(curl -fsSL onecli.sh*)",
-      "Bash(onecli *)",
-      "Bash(grep -q *)",
-      "Bash(echo *>> .env)",
-      "Bash(ls *)",
-      "Bash(cat ~/.config/nanoclaw/*)",
-      "Bash(tail *logs/*)",
-      "Bash(launchctl *nanoclaw*)",
-      "Bash(sqlite3 data/*)",
-      "Bash(docker info*)",
-      "Bash(docker logs *)",
-      "Bash(mkdir -p *)",
-      "Bash(cp .env *)",
-      "Bash(rsync -a .claude/skills/*)",
-      "Bash(head *)",
-      "Bash(xattr *)",
-      "Bash(find ~/.npm *)",
-      "Bash(which onecli*)",
-      "Bash(./container/build.sh*)"
-    ]
   }
 }
diff --git a/.claude/skills/add-atomic-chat-tool/SKILL.md b/.claude/skills/add-atomic-chat-tool/SKILL.md
new file mode 100644
index 0000000..6a6d858
--- /dev/null
+++ b/.claude/skills/add-atomic-chat-tool/SKILL.md
@@ -0,0 +1,243 @@
+---
+name: add-atomic-chat-tool
+description: Add Atomic Chat MCP server so the container agent can call local models served by the Atomic Chat desktop app via its OpenAI-compatible API.
+---
+
+# Add Atomic Chat Integration
+
+This skill adds a stdio-based MCP server that exposes models running in the local [Atomic Chat](https://github.com/AtomicBot-ai/Atomic-Chat) desktop app as tools for the container agent. Claude remains the orchestrator but can offload work to local models served by Atomic Chat on `http://127.0.0.1:1337/v1` (OpenAI-compatible).
+
+Tools exposed:
+- `atomic_chat_list_models` — list models currently available in Atomic Chat (`GET /v1/models`)
+- `atomic_chat_generate` — send a prompt to a specified model and return the response (`POST /v1/chat/completions`)
+
+Model management (download, delete) is done through the **Atomic Chat desktop UI** — the app is a fork of Jan and manages its own model library.
+
+The skill ships the MCP server source in this folder and copies it into the agent-runner tree at install time, then wires it up with small edits to `index.ts`, `providers/claude.ts`, and `container-runner.ts`. No branch merge — all edits are additive and idempotent.
+
+## Phase 1: Pre-flight
+
+### Check if already applied
+
+Check if `container/agent-runner/src/atomic-chat-mcp-stdio.ts` exists. If it does, skip to Phase 3 (Configure).
+
+### Check prerequisites
+
+Verify Atomic Chat is installed and its local API server is running. On the host:
+
+```bash
+curl -s http://127.0.0.1:1337/v1/models | head
+```
+
+If the request fails:
+
+1. Install Atomic Chat from the [latest release](https://github.com/AtomicBot-ai/Atomic-Chat/releases) (macOS only for now — `atomic-chat.dmg`).
+2. Open the app.
+3. Open **Settings → Local API Server** and make sure it's enabled on port `1337`.
+4. Go to the **Hub** (or **Models**) tab and download at least one model (e.g. Llama 3.2 3B, Qwen 2.5 Coder 7B).
+5. Load the model once by sending any message in Atomic Chat's UI to warm it up.
+
+## Phase 2: Apply Code Changes
+
+### Copy the MCP server source
+
+```bash
+cp .claude/skills/add-atomic-chat-tool/atomic-chat-mcp-stdio.ts container/agent-runner/src/atomic-chat-mcp-stdio.ts
+```
+
+### Register the MCP server in the agent-runner
+
+Edit `container/agent-runner/src/index.ts`. Find the `mcpServers` object that currently looks like this:
+
+```ts
+  const mcpServers: Record<string, { command: string; args: string[]; env: Record<string, string> }> = {
+    nanoclaw: {
+      command: 'bun',
+      args: ['run', mcpServerPath],
+      env: {},
+    },
+  };
+```
+
+Add an `atomic_chat` entry alongside `nanoclaw`:
+
+```ts
+  const mcpServers: Record<string, { command: string; args: string[]; env: Record<string, string> }> = {
+    nanoclaw: {
+      command: 'bun',
+      args: ['run', mcpServerPath],
+      env: {},
+    },
+    atomic_chat: {
+      command: 'bun',
+      args: ['run', path.join(__dirname, 'atomic-chat-mcp-stdio.ts')],
+      env: {
+        ...(process.env.ATOMIC_CHAT_HOST ? { ATOMIC_CHAT_HOST: process.env.ATOMIC_CHAT_HOST } : {}),
+        ...(process.env.ATOMIC_CHAT_API_KEY ? { ATOMIC_CHAT_API_KEY: process.env.ATOMIC_CHAT_API_KEY } : {}),
+      },
+    },
+  };
+```
+
+### Add the tool glob to the allowlist
+
+Edit `container/agent-runner/src/providers/claude.ts`. Find `'mcp__nanoclaw__*',` in the `TOOL_ALLOWLIST` array and add `'mcp__atomic_chat__*',` on the following line:
+
+```ts
+  'mcp__nanoclaw__*',
+  'mcp__atomic_chat__*',
+];
+```
+
+### Forward host env vars into the container
+
+Edit `src/container-runner.ts` in `buildContainerArgs`. Find the `TZ` env line:
+
+```ts
+  args.push('-e', `TZ=${TIMEZONE}`);
+```
+
+Add ATOMIC_CHAT forwarding right after it:
+
+```ts
+  args.push('-e', `TZ=${TIMEZONE}`);
+
+  // Atomic Chat MCP tool: forward host overrides if set (default is host.docker.internal:1337).
+  if (process.env.ATOMIC_CHAT_HOST) {
+    args.push('-e', `ATOMIC_CHAT_HOST=${process.env.ATOMIC_CHAT_HOST}`);
+  }
+  if (process.env.ATOMIC_CHAT_API_KEY) {
+    args.push('-e', `ATOMIC_CHAT_API_KEY=${process.env.ATOMIC_CHAT_API_KEY}`);
+  }
+```
+
+### Surface `[ATOMIC]` log lines at info level
+
+In the same file, find the stderr logger:
+
+```ts
+  container.stderr?.on('data', (data) => {
+    for (const line of data.toString().trim().split('\n')) {
+      if (line) log.debug(line, { container: agentGroup.folder });
+    }
+  });
+```
+
+Replace it with:
+
+```ts
+  container.stderr?.on('data', (data) => {
+    for (const line of data.toString().trim().split('\n')) {
+      if (!line) continue;
+      if (line.includes('[ATOMIC]')) {
+        log.info(line, { container: agentGroup.folder });
+      } else {
+        log.debug(line, { container: agentGroup.folder });
+      }
+    }
+  });
+```
+
+### Add env-var stubs to `.env.example`
+
+Append to `.env.example`:
+
+```bash
+# Atomic Chat MCP tool (.claude/skills/add-atomic-chat-tool)
+# Override the host where Atomic Chat exposes its OpenAI-compatible API.
+# Default: http://host.docker.internal:1337 (with fallback to localhost)
+# ATOMIC_CHAT_HOST=http://host.docker.internal:1337
+
+# Optional API key. Leave unset for a local Atomic Chat install — it does not require auth.
+# ATOMIC_CHAT_API_KEY=
+```
+
+### Validate code changes
+
+```bash
+pnpm run build
+pnpm exec tsc -p container/agent-runner/tsconfig.json --noEmit
+./container/build.sh
+```
+
+All three must be clean before proceeding.
+
+## Phase 3: Configure
+
+### Set Atomic Chat host (optional)
+
+By default, the MCP server connects to `http://host.docker.internal:1337` (Docker Desktop) with a fallback to `localhost`. To use a custom host, add to `.env`:
+
+```bash
+ATOMIC_CHAT_HOST=http://your-atomic-chat-host:1337
+```
+
+### Set API key (optional)
+
+Atomic Chat does **not require authentication** when running locally — leave this unset. Only set it if you've put Atomic Chat behind a reverse proxy that enforces auth:
+
+```bash
+ATOMIC_CHAT_API_KEY=sk-...
+```
+
+### Restart the service
+
+```bash
+launchctl kickstart -k gui/$(id -u)/com.nanoclaw  # macOS
+# Linux: systemctl --user restart nanoclaw
+```
+
+## Phase 4: Verify
+
+### Test inference
+
+Tell the user:
+
+> Send a message like: "use atomic chat to tell me the capital of France"
+>
+> The agent should use `atomic_chat_list_models` to find available models, then `atomic_chat_generate` to get a response.
+
+### Check logs if needed
+
+```bash
+tail -f logs/nanoclaw.log | grep -i atomic
+```
+
+Look for:
+- `[ATOMIC] Listing models...` — list request started
+- `[ATOMIC] Found N models` — models discovered
+- `[ATOMIC] >>> Generating with <model>` — generation started
+- `[ATOMIC] <<< Done: <model> | Xs | N tokens | M chars` — generation completed
+
+## Troubleshooting
+
+### Agent says "Atomic Chat is not installed" or tries to run a CLI
+
+The agent is looking for a CLI that doesn't exist instead of using the MCP tools. This means:
+1. The MCP server wasn't copied — check `container/agent-runner/src/atomic-chat-mcp-stdio.ts` exists
+2. The MCP server wasn't registered — check `container/agent-runner/src/index.ts` has the `atomic_chat` entry in `mcpServers`
+3. The allowlist wasn't updated — check `container/agent-runner/src/providers/claude.ts` includes `mcp__atomic_chat__*` in `TOOL_ALLOWLIST`
+4. The container wasn't rebuilt — run `./container/build.sh`
+
+### "Failed to connect to Atomic Chat"
+
+1. Verify the host API is reachable: `curl http://127.0.0.1:1337/v1/models`
+2. Confirm the Local API Server is enabled in Atomic Chat's settings
+3. Check Docker can reach the host: `docker run --rm curlimages/curl curl -s http://host.docker.internal:1337/v1/models`
+4. If using a custom host, check `ATOMIC_CHAT_HOST` in `.env`
+
+### `model not found` / 404 on generate
+
+The model ID passed to `atomic_chat_generate` must exactly match one of the IDs returned by `atomic_chat_list_models`. Ask the agent to list models first, then pick one from that list.
+
+### Slow first response
+
+Atomic Chat lazy-loads models into memory on first use. The initial call may take longer while the model warms up. Subsequent calls against the same model are fast.
+
+### Agent doesn't use Atomic Chat tools
+
+The agent may not know about the tools. Try being explicit: "use the atomic_chat_generate tool with llama3.2-3b-instruct to answer: ..."
+
+### Context window or output size issues
+
+Atomic Chat respects each model's native context length. If you hit limits, pass `max_tokens` explicitly when calling `atomic_chat_generate`, or switch to a model with a larger context window in the Atomic Chat UI.
diff --git a/.claude/skills/add-atomic-chat-tool/atomic-chat-mcp-stdio.ts b/.claude/skills/add-atomic-chat-tool/atomic-chat-mcp-stdio.ts
new file mode 100644
index 0000000..0198644
--- /dev/null
+++ b/.claude/skills/add-atomic-chat-tool/atomic-chat-mcp-stdio.ts
@@ -0,0 +1,229 @@
+/**
+ * Atomic Chat MCP Server for NanoClaw
+ * Exposes local Atomic Chat models (OpenAI-compatible, /v1) as tools for the container agent.
+ * Uses host.docker.internal to reach the host's Atomic Chat desktop app from Docker.
+ */
+
+import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
+import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
+import { z } from 'zod';
+
+import fs from 'fs';
+import path from 'path';
+
+const ATOMIC_CHAT_HOST =
+  process.env.ATOMIC_CHAT_HOST || 'http://host.docker.internal:1337';
+const ATOMIC_CHAT_API_KEY = process.env.ATOMIC_CHAT_API_KEY || '';
+const ATOMIC_CHAT_STATUS_FILE = '/workspace/ipc/atomic_chat_status.json';
+
+function log(msg: string): void {
+  console.error(`[ATOMIC] ${msg}`);
+}
+
+function writeStatus(status: string, detail?: string): void {
+  try {
+    const data = { status, detail, timestamp: new Date().toISOString() };
+    const tmpPath = `${ATOMIC_CHAT_STATUS_FILE}.tmp`;
+    fs.mkdirSync(path.dirname(ATOMIC_CHAT_STATUS_FILE), { recursive: true });
+    fs.writeFileSync(tmpPath, JSON.stringify(data));
+    fs.renameSync(tmpPath, ATOMIC_CHAT_STATUS_FILE);
+  } catch {
+    /* best-effort */
+  }
+}
+
+async function atomicFetch(
+  apiPath: string,
+  options?: RequestInit,
+): Promise<Response> {
+  const url = `${ATOMIC_CHAT_HOST}${apiPath}`;
+  const headers: Record<string, string> = {
+    ...((options?.headers as Record<string, string>) || {}),
+  };
+  if (ATOMIC_CHAT_API_KEY) {
+    headers.Authorization = `Bearer ${ATOMIC_CHAT_API_KEY}`;
+  }
+  const finalOptions: RequestInit = { ...options, headers };
+  try {
+    return await fetch(url, finalOptions);
+  } catch (err) {
+    // Fallback to localhost if host.docker.internal fails
+    if (ATOMIC_CHAT_HOST.includes('host.docker.internal')) {
+      const fallbackUrl = url.replace('host.docker.internal', 'localhost');
+      return await fetch(fallbackUrl, finalOptions);
+    }
+    throw err;
+  }
+}
+
+const server = new McpServer({
+  name: 'atomic_chat',
+  version: '1.0.0',
+});
+
+server.tool(
+  'atomic_chat_list_models',
+  'List all models available in the local Atomic Chat desktop app. Use this to see which models are loaded before calling atomic_chat_generate.',
+  {},
+  async () => {
+    log('Listing models...');
+    writeStatus('listing', 'Listing available models');
+    try {
+      const res = await atomicFetch('/v1/models');
+      if (!res.ok) {
+        return {
+          content: [
+            {
+              type: 'text' as const,
+              text: `Atomic Chat API error: ${res.status} ${res.statusText}`,
+            },
+          ],
+          isError: true,
+        };
+      }
+
+      const data = (await res.json()) as {
+        data?: Array<{ id: string; owned_by?: string }>;
+      };
+      const models = data.data || [];
+
+      if (models.length === 0) {
+        return {
+          content: [
+            {
+              type: 'text' as const,
+              text: 'No models available. Open Atomic Chat on the host and download a model from the Hub.',
+            },
+          ],
+        };
+      }
+
+      const list = models
+        .map((m) => `- ${m.id}${m.owned_by ? ` (${m.owned_by})` : ''}`)
+        .join('\n');
+
+      log(`Found ${models.length} models`);
+      return {
+        content: [
+          { type: 'text' as const, text: `Available models:\n${list}` },
+        ],
+      };
+    } catch (err) {
+      return {
+        content: [
+          {
+            type: 'text' as const,
+            text: `Failed to connect to Atomic Chat at ${ATOMIC_CHAT_HOST}: ${err instanceof Error ? err.message : String(err)}`,
+          },
+        ],
+        isError: true,
+      };
+    }
+  },
+);
+
+server.tool(
+  'atomic_chat_generate',
+  'Send a prompt to a local Atomic Chat model and get a response. Good for cheaper/faster tasks like summarization, translation, or general queries. Use atomic_chat_list_models first to see available models.',
+  {
+    model: z
+      .string()
+      .describe(
+        'The model ID as returned by atomic_chat_list_models (e.g. "llama3.2-3b-instruct")',
+      ),
+    prompt: z.string().describe('The prompt to send to the model'),
+    system: z
+      .string()
+      .optional()
+      .describe('Optional system prompt to set model behavior'),
+    temperature: z
+      .number()
+      .optional()
+      .describe('Sampling temperature (0.0–2.0). Defaults to model default.'),
+    max_tokens: z
+      .number()
+      .optional()
+      .describe('Maximum number of tokens to generate in the response.'),
+  },
+  async (args) => {
+    log(`>>> Generating with ${args.model} (${args.prompt.length} chars)...`);
+    writeStatus('generating', `Generating with ${args.model}`);
+    try {
+      const messages: Array<{ role: string; content: string }> = [];
+      if (args.system) {
+        messages.push({ role: 'system', content: args.system });
+      }
+      messages.push({ role: 'user', content: args.prompt });
+
+      const body: Record<string, unknown> = {
+        model: args.model,
+        messages,
+        stream: false,
+      };
+      if (args.temperature !== undefined) body.temperature = args.temperature;
+      if (args.max_tokens !== undefined) body.max_tokens = args.max_tokens;
+
+      const startedAt = Date.now();
+      const res = await atomicFetch('/v1/chat/completions', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(body),
+      });
+
+      if (!res.ok) {
+        const errorText = await res.text();
+        return {
+          content: [
+            {
+              type: 'text' as const,
+              text: `Atomic Chat error (${res.status}): ${errorText}`,
+            },
+          ],
+          isError: true,
+        };
+      }
+
+      const data = (await res.json()) as {
+        choices?: Array<{ message?: { content?: string } }>;
+        usage?: {
+          prompt_tokens?: number;
+          completion_tokens?: number;
+          total_tokens?: number;
+        };
+      };
+
+      const response = data.choices?.[0]?.message?.content ?? '';
+      const elapsedSec = ((Date.now() - startedAt) / 1000).toFixed(1);
+      const completionTokens = data.usage?.completion_tokens;
+
+      const meta = `\n\n[${args.model} | ${elapsedSec}s${
+        completionTokens !== undefined ? ` | ${completionTokens} tokens` : ''
+      }]`;
+
+      log(
+        `<<< Done: ${args.model} | ${elapsedSec}s | ${
+          completionTokens ?? '?'
+        } tokens | ${response.length} chars`,
+      );
+      writeStatus(
+        'done',
+        `${args.model} | ${elapsedSec}s | ${completionTokens ?? '?'} tokens`,
+      );
+
+      return { content: [{ type: 'text' as const, text: response + meta }] };
+    } catch (err) {
+      return {
+        content: [
+          {
+            type: 'text' as const,
+            text: `Failed to call Atomic Chat: ${err instanceof Error ? err.message : String(err)}`,
+          },
+        ],
+        isError: true,
+      };
+    }
+  },
+);
+
+const transport = new StdioServerTransport();
+await server.connect(transport);
diff --git a/.claude/skills/add-codex/SKILL.md b/.claude/skills/add-codex/SKILL.md
new file mode 100644
index 0000000..14b3072
--- /dev/null
+++ b/.claude/skills/add-codex/SKILL.md
@@ -0,0 +1,161 @@
+---
+name: add-codex
+description: Use Codex (CLI + AppServer) as the full agent provider — planning, tool orchestration, native compaction, MCP tools, session resume — in place of the Claude Agent SDK. ChatGPT subscription or OPENAI_API_KEY. Per-group via agent_provider. Distinct from using OpenAI as an MCP tool (where Claude remains the planner).
+---
+
+# Codex agent provider
+
+NanoClaw runs agents in a long-lived **poll loop** inside the container. The backend is selected with **`AGENT_PROVIDER`** (`claude` | `opencode` | `codex` | `mock`).
+
+Trunk ships with only the `claude` provider baked in. This skill copies the Codex provider files in from the `providers` branch, wires them into the host and container barrels, updates the Dockerfile to install the Codex CLI, and rebuilds the image.
+
+The Codex provider runs `codex app-server` as a child process and speaks JSON-RPC over stdio. That gives it native session resume, streaming events, MCP tool access, and `thread/compact/start` compaction — same feature bar as the Claude Agent SDK, without the Anthropic-only lock-in.
+
+## Install
+
+### Pre-flight
+
+If all of the following are already present, skip to **Configuration**:
+
+- `src/providers/codex.ts`
+- `container/agent-runner/src/providers/codex.ts`
+- `container/agent-runner/src/providers/codex-app-server.ts`
+- `container/agent-runner/src/providers/codex.factory.test.ts`
+- `import './codex.js';` line in `src/providers/index.ts`
+- `import './codex.js';` line in `container/agent-runner/src/providers/index.ts`
+- `ARG CODEX_VERSION` and `"@openai/codex@${CODEX_VERSION}"` in the pnpm global-install block in `container/Dockerfile`
+
+Missing pieces — continue below. All steps are idempotent; re-running is safe.
+
+### 1. Fetch the providers branch
+
+```bash
+git fetch origin providers
+```
+
+### 2. Copy the Codex source files
+
+Wholesale copies (owned entirely by this skill — user edits to these files won't survive a re-run, as designed):
+
+```bash
+git show origin/providers:src/providers/codex.ts                                      > src/providers/codex.ts
+git show origin/providers:container/agent-runner/src/providers/codex.ts               > container/agent-runner/src/providers/codex.ts
+git show origin/providers:container/agent-runner/src/providers/codex-app-server.ts    > container/agent-runner/src/providers/codex-app-server.ts
+git show origin/providers:container/agent-runner/src/providers/codex.factory.test.ts  > container/agent-runner/src/providers/codex.factory.test.ts
+```
+
+### 3. Append the self-registration imports
+
+Each barrel gets one line — alphabetical placement keeps diffs small.
+
+`src/providers/index.ts`:
+
+```typescript
+import './codex.js';
+```
+
+`container/agent-runner/src/providers/index.ts`:
+
+```typescript
+import './codex.js';
+```
+
+### 4. Add the Codex CLI to the container Dockerfile
+
+Two edits to `container/Dockerfile`, both idempotent (skip if already present):
+
+**(a)** In the "Pin CLI versions" ARG block (around line 18), add after `ARG CLAUDE_CODE_VERSION=...`:
+
+```dockerfile
+ARG CODEX_VERSION=0.124.0
+```
+
+**(b)** Add a new standalone `RUN` block for the Codex CLI, after the existing per-CLI install blocks (around line 106, right after the `@anthropic-ai/claude-code` block). The Dockerfile splits each global CLI into its own layer for cache granularity — keep that pattern; do not collapse them into a single combined `pnpm install -g` call:
+
+```dockerfile
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    pnpm install -g "@openai/codex@${CODEX_VERSION}"
+```
+
+Note: **no agent-runner package dependency** — Codex is a CLI binary, not a library. Unlike OpenCode, there's nothing to add to `container/agent-runner/package.json`.
+
+### 5. Build
+
+```bash
+pnpm run build                                         # host
+pnpm exec tsc -p container/agent-runner/tsconfig.json --noEmit   # container typecheck
+./container/build.sh                                   # agent image
+```
+
+## Configuration
+
+Codex supports two primary auth paths and one experimental BYO-endpoint path. Pick the one that matches your setup.
+
+### Option A — ChatGPT subscription (recommended for individuals)
+
+On the host (not inside the container), run Codex's OAuth login:
+
+```bash
+codex login
+```
+
+This writes `~/.codex/auth.json` with a subscription token. The host-side Codex provider ([src/providers/codex.ts](../../../src/providers/codex.ts)) copies `auth.json` into a per-session `~/.codex` directory mounted into the container — your host's own Codex CLI is never touched.
+
+No `.env` variables required for this mode.
+
+### Option B — API key (recommended for CI or API billing)
+
+```env
+OPENAI_API_KEY=sk-...
+CODEX_MODEL=gpt-5.4-mini
+```
+
+The host forwards both variables into the container. If both subscription (`auth.json`) and `OPENAI_API_KEY` are present, Codex prefers the subscription.
+
+### Option C — BYO OpenAI-compatible endpoint (experimental)
+
+Codex's built-in `openai` provider honors the `OPENAI_BASE_URL` env var directly. Point it at any OpenAI-compatible endpoint — Groq, Together, self-hosted vLLM, an OpenAI proxy, etc.
+
+```env
+OPENAI_API_KEY=...
+OPENAI_BASE_URL=https://api.groq.com/openai/v1
+CODEX_MODEL=llama-3.3-70b-versatile
+```
+
+Codex also ships first-class local-runner flags — `codex --oss --local-provider ollama` or `--local-provider lmstudio` — that auto-detect a local server. To use those inside NanoClaw, set `CODEX_MODEL` to a model your local runner serves and add the corresponding base URL; see the Codex CLI docs for the full `model_provider = oss` configuration.
+
+**Experimental caveat:** tool-calling quality depends on the model and endpoint. Not every OpenAI-compat provider implements the full function-calling spec, and smaller models (< 30B) often struggle with multi-step tool orchestration. Test before committing.
+
+### Per group / per session
+
+Set `"provider": "codex"` in the group's **`container.json`** (`groups/<folder>/container.json`) — the in-container runner reads `provider` from there, not from the DB. The DB columns **`agent_groups.agent_provider`** and **`sessions.agent_provider`** (session overrides group) only drive host-side provider contribution — per-session `~/.codex` mount, `OPENAI_*` / `CODEX_MODEL` env passthrough — and do not propagate into `container.json` at spawn time. Set both, or just edit `container.json`; if they disagree, the runner uses `container.json` and the host-side resolver falls back through session → group → `container.json` → `'claude'`.
+
+`CODEX_MODEL` applies process-wide via `.env`; if you need different models for different groups, set them via `container_config.env` on the group.
+
+Extra MCP servers still come from **`NANOCLAW_MCP_SERVERS`** / `container_config.mcpServers` on the host. The runner merges them into the same `mcpServers` object passed to all providers.
+
+## Operational notes
+
+- **Spawn-per-query:** Codex's app-server is spawned fresh per query invocation, matching the OpenCode pattern. No long-lived daemon to keep healthy across sessions.
+- **Per-session `~/.codex` isolation:** each group gets its own copy of the host's `auth.json`. The container can rewrite `config.toml` freely on every wake without touching the host's Codex config.
+- **Native compaction:** kicks in automatically at 40K cumulative input tokens between turns, via `thread/compact/start`. If compaction fails, the provider logs and continues uncompacted — no fatal error.
+- **Approvals:** auto-accepted inside the container (the container is the sandbox; same posture as Claude/OpenCode).
+- **Mid-turn input:** Codex turns don't accept mid-turn messages. Follow-up `push()` calls queue and drain between turns, matching the OpenCode pattern. The poll-loop only pushes between turns anyway, so no messages are dropped.
+- **Stale thread recovery:** `isSessionInvalid` matches on stale-thread-ID errors (`thread not found`, `unknown thread`, etc.) so a cold-started app-server can recover cleanly when it sees a stored continuation it no longer has.
+
+## Verify
+
+```bash
+grep -q "./codex.js" container/agent-runner/src/providers/index.ts && echo "container barrel: OK"
+grep -q "./codex.js" src/providers/index.ts && echo "host barrel: OK"
+grep -q "@openai/codex@" container/Dockerfile && echo "Dockerfile install: OK"
+cd container/agent-runner && bun test src/providers/codex.factory.test.ts && cd -
+```
+
+After the image rebuild, set `agent_provider = 'codex'` on a test group and send a message. Successful round-trip looks like:
+
+- `init` event with a stable thread ID as continuation
+- One or more `activity` / `progress` events during the turn
+- `result` event with the model's reply
+
+If the agent hangs or errors, check `~/.codex/auth.json` exists on the host (Option A) or that `OPENAI_API_KEY` is forwarding correctly (Option B) — `docker exec` into a running container and `env | grep -i openai` to confirm.
diff --git a/.claude/skills/add-compact/SKILL.md b/.claude/skills/add-compact/SKILL.md
deleted file mode 100644
index ee9674a..0000000
--- a/.claude/skills/add-compact/SKILL.md
+++ /dev/null
@@ -1,135 +0,0 @@
----
-name: add-compact
-description: Add /compact command for manual context compaction. Solves context rot in long sessions by forwarding the SDK's built-in /compact slash command. Main-group or trusted sender only.
----
-
-# Add /compact Command
-
-Adds a `/compact` session command that compacts conversation history to fight context rot in long-running sessions. Uses the Claude Agent SDK's built-in `/compact` slash command — no synthetic system prompts.
-
-**Session contract:** `/compact` keeps the same logical session alive. The SDK returns a new session ID after compaction (via the `init` system message), which the agent-runner forwards to the orchestrator as `newSessionId`. No destructive reset occurs — the agent retains summarized context.
-
-## Phase 1: Pre-flight
-
-Check if `src/session-commands.ts` exists:
-
-```bash
-test -f src/session-commands.ts && echo "Already applied" || echo "Not applied"
-```
-
-If already applied, skip to Phase 3 (Verify).
-
-## Phase 2: Apply Code Changes
-
-Merge the skill branch:
-
-```bash
-git fetch upstream skill/compact
-git merge upstream/skill/compact
-```
-
-> **Note:** `upstream` is the remote pointing to `qwibitai/nanoclaw`. If using a different remote name, substitute accordingly.
-
-This adds:
-- `src/session-commands.ts` (extract and authorize session commands)
-- `src/session-commands.test.ts` (unit tests for command parsing and auth)
-- Session command interception in `src/index.ts` (both `processGroupMessages` and `startMessageLoop`)
-- Slash command handling in `container/agent-runner/src/index.ts`
-
-### Validate
-
-```bash
-pnpm test
-pnpm run build
-```
-
-### Rebuild container
-
-```bash
-./container/build.sh
-```
-
-### Restart service
-
-```bash
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw  # macOS
-# Linux: systemctl --user restart nanoclaw
-```
-
-## Phase 3: Verify
-
-### Integration Test
-
-1. Start NanoClaw in dev mode: `pnpm run dev`
-2. From the **main group** (self-chat), send exactly: `/compact`
-3. Verify:
-   - The agent acknowledges compaction (e.g., "Conversation compacted.")
-   - The session continues — send a follow-up message and verify the agent responds coherently
-   - A conversation archive is written to `groups/{folder}/conversations/` (by the PreCompact hook)
-   - Container logs show `Compact boundary observed` (confirms SDK actually compacted)
-   - If `compact_boundary` was NOT observed, the response says "compact_boundary was not observed"
-4. From a **non-main group** as a non-admin user, send: `@<assistant> /compact`
-5. Verify:
-   - The bot responds with "Session commands require admin access."
-   - No compaction occurs, no container is spawned for the command
-6. From a **non-main group** as the admin (device owner / `is_from_me`), send: `@<assistant> /compact`
-7. Verify:
-   - Compaction proceeds normally (same behavior as main group)
-8. While an **active container** is running for the main group, send `/compact`
-9. Verify:
-   - The active container is signaled to close (authorized senders only — untrusted senders cannot kill in-flight work)
-   - Compaction proceeds via a new container once the active one exits
-   - The command is not dropped (no cursor race)
-10. Send a normal message, then `/compact`, then another normal message in quick succession (same polling batch):
-11. Verify:
-    - Pre-compact messages are sent to the agent first (check container logs for two `runAgent` calls)
-    - Compaction proceeds after pre-compact messages are processed
-    - Messages **after** `/compact` in the batch are preserved (cursor advances to `/compact`'s timestamp only) and processed on the next poll cycle
-12. From a **non-main group** as a non-admin user, send `@<assistant> /compact`:
-13. Verify:
-    - Denial message is sent ("Session commands require admin access.")
-    - The `/compact` is consumed (cursor advanced) — it does NOT replay on future polls
-    - Other messages in the same batch are also consumed (cursor is a high-water mark — this is an accepted tradeoff for the narrow edge case of denied `/compact` + other messages in the same polling interval)
-    - No container is killed or interrupted
-14. From a **non-main group** (with `requiresTrigger` enabled) as a non-admin user, send bare `/compact` (no trigger prefix):
-15. Verify:
-    - No denial message is sent (trigger policy prevents untrusted bot responses)
-    - The `/compact` is consumed silently
-    - Note: in groups where `requiresTrigger` is `false`, a denial message IS sent because the sender is considered reachable
-16. After compaction, verify **no auto-compaction** behavior — only manual `/compact` triggers it
-
-### Validation on Fresh Clone
-
-```bash
-git clone <your-fork> /tmp/nanoclaw-test
-cd /tmp/nanoclaw-test
-claude  # then run /add-compact
-pnpm run build
-pnpm test
-./container/build.sh
-# Manual: send /compact from main group, verify compaction + continuation
-# Manual: send @<assistant> /compact from non-main as non-admin, verify denial
-# Manual: send @<assistant> /compact from non-main as admin, verify allowed
-# Manual: verify no auto-compaction behavior
-```
-
-## Security Constraints
-
-- **Main-group or trusted/admin sender only.** The main group is the user's private self-chat and is trusted (see `docs/SECURITY.md`). Non-main groups are untrusted — a careless or malicious user could wipe the agent's short-term memory. However, the device owner (`is_from_me`) is always trusted and can compact from any group.
-- **No auto-compaction.** This skill implements manual compaction only. Automatic threshold-based compaction is a separate concern and should be a separate skill.
-- **No config file.** NanoClaw's philosophy is customization through code changes, not configuration sprawl.
-- **Transcript archived before compaction.** The existing `PreCompact` hook in the agent-runner archives the full transcript to `conversations/` before the SDK compacts it.
-- **Session continues after compaction.** This is not a destructive reset. The conversation continues with summarized context.
-
-## What This Does NOT Do
-
-- No automatic compaction threshold (add separately if desired)
-- No `/clear` command (separate skill, separate semantics — `/clear` is a destructive reset)
-- No cross-group compaction (each group's session is isolated)
-- No changes to the container image, Dockerfile, or build script
-
-## Troubleshooting
-
-- **"Session commands require admin access"**: Only the device owner (`is_from_me`) or main-group senders can use `/compact`. Other users are denied.
-- **No compact_boundary in logs**: The SDK may not emit this event in all versions. Check the agent-runner logs for the warning message. Compaction may still have succeeded.
-- **Pre-compact failure**: If messages before `/compact` fail to process, the error message says "Failed to process messages before /compact." The cursor advances past sent output to prevent duplicates; `/compact` remains pending for the next attempt.
diff --git a/.claude/skills/add-emacs/SKILL.md b/.claude/skills/add-emacs/SKILL.md
index 8a4100e..82a5098 100644
--- a/.claude/skills/add-emacs/SKILL.md
+++ b/.claude/skills/add-emacs/SKILL.md
@@ -1,12 +1,11 @@
 ---
 name: add-emacs
-description: Add Emacs as a channel. Opens an interactive chat buffer and org-mode integration so you can talk to NanoClaw from within Emacs (Doom, Spacemacs, or vanilla). Uses a local HTTP bridge — no bot token or external service needed.
+description: Add Emacs as a channel. Opens an interactive chat buffer and org-mode integration so you can talk to NanoClaw from within Emacs (Doom, Spacemacs, or vanilla). Local HTTP bridge — no bot token or external service needed.
 ---
 
 # Add Emacs Channel
 
-This skill adds Emacs support to NanoClaw, then walks through interactive setup.
-Works with Doom Emacs, Spacemacs, and vanilla Emacs 27.1+.
+Adds Emacs support via a local HTTP bridge. Works with Doom Emacs, Spacemacs, and vanilla Emacs 27.1+.
 
 ## What you can do with this
 
@@ -15,95 +14,99 @@ Works with Doom Emacs, Spacemacs, and vanilla Emacs 27.1+.
 - **Meeting notes** — send an org agenda entry; get a summary or action item list back as a child node
 - **Draft writing** — send org prose; receive revisions or continuations in place
 - **Research capture** — ask a question directly in your org notes; the answer lands exactly where you need it
-- **Schedule tasks** — ask Andy to set a reminder or create a scheduled NanoClaw task (e.g. "remind me tomorrow to review the PR")
 
-## Phase 1: Pre-flight
+## Install
 
-### Check if already applied
+NanoClaw doesn't ship channels in trunk. This skill copies the Emacs adapter and the Lisp client in from the `channels` branch. Native HTTP bridge — no Chat SDK, no adapter package.
 
-Check if `src/channels/emacs.ts` exists:
+### Pre-flight (idempotent)
+
+Skip to **Enable** if all of these are already in place:
+
+- `src/channels/emacs.ts` exists
+- `emacs/nanoclaw.el` exists
+- `src/channels/index.ts` contains `import './emacs.js';`
+
+Otherwise continue. Every step below is safe to re-run.
+
+### 1. Fetch the channels branch
 
 ```bash
-test -f src/channels/emacs.ts && echo "already applied" || echo "not applied"
+git fetch origin channels
 ```
 
-If it exists, skip to Phase 3 (Setup). The code changes are already in place.
-
-## Phase 2: Apply Code Changes
-
-### Ensure the upstream remote
+### 2. Copy the adapter and Lisp client
 
 ```bash
-git remote -v
+mkdir -p emacs
+git show origin/channels:src/channels/emacs.ts      > src/channels/emacs.ts
+git show origin/channels:src/channels/emacs.test.ts > src/channels/emacs.test.ts
+git show origin/channels:emacs/nanoclaw.el          > emacs/nanoclaw.el
 ```
 
-If an `upstream` remote pointing to `https://github.com/qwibitai/nanoclaw.git` is missing,
-add it:
+### 3. Append the self-registration import
 
-```bash
-git remote add upstream https://github.com/qwibitai/nanoclaw.git
+Append to `src/channels/index.ts` (skip if the line is already present):
+
+```typescript
+import './emacs.js';
 ```
 
-### Merge the skill branch
-
-```bash
-git fetch upstream skill/emacs
-git merge upstream/skill/emacs
-```
-
-If there are merge conflicts on `pnpm-lock.yaml`, resolve them by accepting the incoming
-version and continuing:
-
-```bash
-git checkout --theirs pnpm-lock.yaml
-git add pnpm-lock.yaml
-git merge --continue
-```
-
-For any other conflict, read the conflicted file and reconcile both sides manually.
-
-This adds:
-- `src/channels/emacs.ts` — `EmacsBridgeChannel` HTTP server (port 8766)
-- `src/channels/emacs.test.ts` — unit tests
-- `emacs/nanoclaw.el` — Emacs Lisp package (`nanoclaw-chat`, `nanoclaw-org-send`)
-- `import './emacs.js'` appended to `src/channels/index.ts`
-
-If the merge reports conflicts, resolve them by reading the conflicted files and understanding the intent of both sides.
-
-### Validate code changes
+### 4. Build
 
 ```bash
 pnpm run build
-pnpm exec vitest run src/channels/emacs.test.ts
 ```
 
-Build must be clean and tests must pass before proceeding.
+No npm package to install — the adapter uses only Node builtins (`http`).
 
-## Phase 3: Setup
+## Enable
 
-### Configure environment (optional)
-
-The channel works out of the box with defaults. Add to `.env` only if you need non-defaults:
+The adapter is gated by `EMACS_ENABLED` so the HTTP port isn't opened on hosts that aren't running Emacs. Add to `.env`:
 
 ```bash
-EMACS_CHANNEL_PORT=8766     # default — change if 8766 is already in use
-EMACS_AUTH_TOKEN=<random>   # optional — locks the endpoint to Emacs only
+EMACS_ENABLED=true
+EMACS_CHANNEL_PORT=8766       # optional — change only if 8766 is taken
+EMACS_AUTH_TOKEN=             # optional — set to a random string to lock the endpoint
+EMACS_PLATFORM_ID=default     # optional — only change if you want a non-default chat id
 ```
 
-If you change or add values, sync to the container environment:
+Generate an auth token (recommended even on single-user machines — prevents other local processes from poking the endpoint):
 
 ```bash
-mkdir -p data/env && cp .env data/env/env
+node -e "console.log(require('crypto').randomBytes(16).toString('hex'))"
 ```
 
-### Configure Emacs
+## Wire the channel
 
-The `nanoclaw.el` package requires only Emacs 27.1+ built-in libraries (`url`, `json`, `org`) — no package manager setup needed.
+Emacs is a single-user, single-chat channel. One host = one messaging group with `platform_id = "default"`.
+
+### If this is your first agent group
+
+Run `/init-first-agent` — pick **Emacs** as the channel, use any short handle as the "user id" (e.g. your OS username), and the skill will create the agent group, wire the channel, and write a welcome message that the agent delivers back to your Emacs buffer.
+
+### Otherwise — wire to an existing agent group
+
+Run the `register` step directly. The `EMACS_PLATFORM_ID` (default `default`) becomes the messaging group's platform id:
+
+```bash
+pnpm exec tsx setup/index.ts --step register -- \
+  --platform-id "default" --name "Emacs" \
+  --folder "<existing-folder>" --channel "emacs" \
+  --session-mode "agent-shared" \
+  --assistant-name "<existing-assistant-name>"
+```
+
+`agent-shared` puts Emacs messages in the same session as any other channel wired to the same agent group — so a conversation you started in Telegram continues in Emacs. Use `shared` to keep an independent Emacs thread with the same workspace, or a new `--folder` for a dedicated Emacs-only agent.
+
+## Configure Emacs
+
+`nanoclaw.el` needs only Emacs 27.1+ builtins (`url`, `json`, `org`) — no package manager.
 
 AskUserQuestion: Which Emacs distribution are you using?
-- **Doom Emacs** - config.el with map! keybindings
-- **Spacemacs** - dotspacemacs/user-config in ~/.spacemacs
-- **Vanilla Emacs / other** - init.el with global-set-key
+- **Doom Emacs** — `config.el` with `map!` keybindings
+- **Spacemacs** — `dotspacemacs/user-config` in `~/.spacemacs`
+- **Vanilla Emacs / other** — `init.el` with `global-set-key`
 
 **Doom Emacs** — add to `~/.config/doom/config.el` (or `~/.doom.d/config.el`):
 
@@ -117,7 +120,7 @@ AskUserQuestion: Which Emacs distribution are you using?
       :desc "Send org"     "o" #'nanoclaw-org-send)
 ```
 
-Then reload: `M-x doom/reload`
+Reload: `M-x doom/reload`
 
 **Spacemacs** — add to `dotspacemacs/user-config` in `~/.spacemacs`:
 
@@ -129,9 +132,9 @@ Then reload: `M-x doom/reload`
 (spacemacs/set-leader-keys "aNo" #'nanoclaw-org-send)
 ```
 
-Then reload: `M-x dotspacemacs/sync-configuration-layers` or restart Emacs.
+Reload: `M-x dotspacemacs/sync-configuration-layers` or restart Emacs.
 
-**Vanilla Emacs** — add to `~/.emacs.d/init.el` (or `~/.emacs`):
+**Vanilla Emacs** — add to `~/.emacs.d/init.el`:
 
 ```elisp
 ;; NanoClaw — personal AI assistant channel
@@ -141,61 +144,75 @@ Then reload: `M-x dotspacemacs/sync-configuration-layers` or restart Emacs.
 (global-set-key (kbd "C-c n o") #'nanoclaw-org-send)
 ```
 
-Then reload: `M-x eval-buffer` or restart Emacs.
+Reload: `M-x eval-buffer` or restart Emacs.
 
-If `EMACS_AUTH_TOKEN` was set, also add (any distribution):
+Replace `~/src/nanoclaw/emacs/nanoclaw.el` with your actual NanoClaw checkout path.
+
+If `EMACS_AUTH_TOKEN` is set, also add (any distribution):
 
 ```elisp
 (setq nanoclaw-auth-token "<your-token>")
 ```
 
-If `EMACS_CHANNEL_PORT` was changed from the default, also add:
+If you changed `EMACS_CHANNEL_PORT` from the default:
 
 ```elisp
 (setq nanoclaw-port <your-port>)
 ```
 
-### Restart NanoClaw
+## Restart NanoClaw
 
 ```bash
 pnpm run build
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw  # macOS
-# Linux: systemctl --user restart nanoclaw
+launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
+# systemctl --user restart nanoclaw                # Linux
 ```
 
-## Phase 4: Verify
+## Verify
 
-### Test the HTTP endpoint
+### HTTP endpoint
 
 ```bash
-curl -s "http://localhost:8766/api/messages?since=0"
+curl -s http://localhost:8766/api/messages?since=0
 ```
 
-Expected: `{"messages":[]}`
-
-If you set `EMACS_AUTH_TOKEN`:
+Expected: `{"messages":[]}`. With an auth token:
 
 ```bash
-curl -s -H "Authorization: Bearer <token>" "http://localhost:8766/api/messages?since=0"
+curl -s -H "Authorization: Bearer <token>" http://localhost:8766/api/messages?since=0
 ```
 
-### Test from Emacs
+### From Emacs
 
 Tell the user:
 
 > 1. Open the chat buffer with your keybinding (`SPC N c`, `SPC a N c`, or `C-c n c`)
-> 2. Type a message and press `RET`
-> 3. A response from Andy should appear within a few seconds
+> 2. Type a message and press `C-c C-c` to send (RET inserts newlines)
+> 3. A response should appear within a few seconds
 >
 > For org-mode: open any `.org` file, position the cursor on a heading, and use `SPC N o` / `SPC a N o` / `C-c n o`
 
-### Check logs if needed
+### Log line
 
-```bash
-tail -f logs/nanoclaw.log
-```
+`tail -f logs/nanoclaw.log` should show `Emacs channel listening` at startup.
 
-Look for `Emacs channel listening` at startup and `Emacs message received` when a message is sent.
+## Channel Info
+
+- **type**: `emacs`
+- **terminology**: Single local buffer. There are no "groups" or separate chats — one host = one chat, addressed by a `platform_id` string (default `default`).
+- **how-to-find-id**: The platform id is whatever you set in `EMACS_PLATFORM_ID` (default `default`). User handles are arbitrary; your OS username or first name is fine (e.g. `emacs:<username>`).
+- **supports-threads**: no
+- **typical-use**: Single developer talking to the assistant from within Emacs, alongside whatever other channel they use (Slack, Telegram, Discord).
+- **default-isolation**: Same agent group as the primary DM, with `session-mode = agent-shared` so a conversation started elsewhere continues in Emacs. Pick a separate folder only if you specifically want an Emacs-only persona.
+
+### Features
+
+- Interactive chat buffer (`nanoclaw-chat`) with markdown → org-mode rendering
+- Org integration (`nanoclaw-org-send`) — sends the current subtree or region; reply lands as a child heading
+- Optional bearer-token auth for the local endpoint
+- Single-user: the adapter exposes exactly one messaging group per host
+
+Not applicable (design): multi-user channels, threads, cold DM initiation, typing indicators, attachments.
 
 ## Troubleshooting
 
@@ -205,66 +222,53 @@ Look for `Emacs channel listening` at startup and `Emacs message received` when
 Error: listen EADDRINUSE: address already in use :::8766
 ```
 
-Either a stale NanoClaw process is running, or 8766 is taken by another app.
-
-Find and kill the stale process:
+Either a stale NanoClaw is running or another app has the port. Kill stale process or change port:
 
 ```bash
 lsof -ti :8766 | xargs kill -9
+# or set EMACS_CHANNEL_PORT in .env and mirror in Emacs config (nanoclaw-port)
 ```
 
-Or change the port in `.env` (`EMACS_CHANNEL_PORT=8767`) and update `nanoclaw-port` in Emacs config.
+### Adapter not starting
+
+If `grep "Emacs channel listening" logs/nanoclaw.log` returns nothing, check that `EMACS_ENABLED=true` is in `.env` and that the adapter import is present:
+
+```bash
+grep -q '^EMACS_ENABLED=true' .env && echo "enabled" || echo "not enabled"
+grep -q "import './emacs.js'" src/channels/index.ts && echo "imported" || echo "not imported"
+```
 
 ### No response from agent
 
-Check:
-1. NanoClaw is running: `launchctl list | grep nanoclaw` (macOS) or `systemctl --user status nanoclaw` (Linux)
-2. Emacs group is registered: `sqlite3 store/messages.db "SELECT * FROM registered_groups WHERE jid = 'emacs:default'"`
-3. Logs show activity: `tail -50 logs/nanoclaw.log`
+1. NanoClaw running: `launchctl list | grep nanoclaw` (macOS) / `systemctl --user status nanoclaw` (Linux)
+2. Messaging group wired: `sqlite3 data/v2.db "SELECT mg.platform_id, ag.folder FROM messaging_groups mg JOIN messaging_group_agents mga ON mg.id = mga.messaging_group_id JOIN agent_groups ag ON ag.id = mga.agent_group_id WHERE mg.channel_type = 'emacs'"`
+3. Logs show inbound: `grep 'channel_type=emacs\|Emacs' logs/nanoclaw.log | tail -20`
 
-If the group is not registered, it will be created automatically on the next NanoClaw restart.
+If no messaging group row exists, run the `register` command above.
 
 ### Auth token mismatch (401 Unauthorized)
 
-Verify the token in Emacs matches `.env`:
-
 ```elisp
-;; M-x describe-variable RET nanoclaw-auth-token RET
+M-x describe-variable RET nanoclaw-auth-token RET
 ```
 
-Must exactly match `EMACS_AUTH_TOKEN` in `.env`.
+Must match `EMACS_AUTH_TOKEN` in `.env`. If you didn't set one server-side, clear it in Emacs too:
+
+```elisp
+(setq nanoclaw-auth-token nil)
+```
 
 ### nanoclaw.el not loading
 
-Check the path is correct:
-
 ```bash
 ls ~/src/nanoclaw/emacs/nanoclaw.el
 ```
 
 If NanoClaw is cloned elsewhere, update the `load`/`load-file` path in your Emacs config.
 
-## After Setup
-
-If running `pnpm run dev` while the service is active:
-
-```bash
-# macOS:
-launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist
-pnpm run dev
-# When done testing:
-launchctl load ~/Library/LaunchAgents/com.nanoclaw.plist
-
-# Linux:
-# systemctl --user stop nanoclaw
-# pnpm run dev
-# systemctl --user start nanoclaw
-```
-
 ## Agent Formatting
 
-The Emacs bridge converts markdown → org-mode automatically. Agents should
-output standard markdown — **not** org-mode syntax. The conversion handles:
+The Emacs bridge converts markdown → org-mode automatically. Agents should output standard markdown, **not** org-mode syntax:
 
 | Markdown | Org-mode |
 |----------|----------|
@@ -274,16 +278,19 @@ output standard markdown — **not** org-mode syntax. The conversion handles:
 | `` `code` `` | `~code~` |
 | ` ```lang ` | `#+begin_src lang` |
 
-If an agent outputs org-mode directly, bold/italic/etc. will be double-converted
-and render incorrectly.
+If an agent outputs org-mode directly, markers get double-converted and render incorrectly.
 
 ## Removal
 
-To remove the Emacs channel:
+```bash
+rm src/channels/emacs.ts src/channels/emacs.test.ts emacs/nanoclaw.el
+# Remove the `import './emacs.js';` line from src/channels/index.ts
+# Remove EMACS_* lines from .env
+pnpm run build
+launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
+# systemctl --user restart nanoclaw                # Linux
 
-1. Delete `src/channels/emacs.ts`, `src/channels/emacs.test.ts`, and `emacs/nanoclaw.el`
-2. Remove `import './emacs.js'` from `src/channels/index.ts`
-3. Remove the NanoClaw block from your Emacs config file
-4. Remove Emacs registration from SQLite: `sqlite3 store/messages.db "DELETE FROM registered_groups WHERE jid = 'emacs:default'"`
-5. Remove `EMACS_CHANNEL_PORT` and `EMACS_AUTH_TOKEN` from `.env` if set
-6. Rebuild: `pnpm run build && launchctl kickstart -k gui/$(id -u)/com.nanoclaw` (macOS) or `pnpm run build && systemctl --user restart nanoclaw` (Linux)
\ No newline at end of file
+# Remove the NanoClaw block from your Emacs config
+# Optionally clean up the messaging group:
+sqlite3 data/v2.db "DELETE FROM messaging_group_agents WHERE messaging_group_id IN (SELECT id FROM messaging_groups WHERE channel_type='emacs'); DELETE FROM messaging_groups WHERE channel_type='emacs';"
+```
diff --git a/.claude/skills/add-gcal-tool/SKILL.md b/.claude/skills/add-gcal-tool/SKILL.md
new file mode 100644
index 0000000..5751933
--- /dev/null
+++ b/.claude/skills/add-gcal-tool/SKILL.md
@@ -0,0 +1,210 @@
+---
+name: add-gcal-tool
+description: Add Google Calendar as an MCP tool (list calendars, list/search/create events, free/busy queries) using OneCLI-managed OAuth. Multi-calendar and multi-account supported. Mirrors /add-gmail-tool's stub pattern — no raw credentials ever reach the container; OneCLI injects real tokens at request time.
+---
+
+# Add Google Calendar Tool (OneCLI-native)
+
+This skill wires [`@cocal/google-calendar-mcp`](https://github.com/cocal-com/google-calendar-mcp) into selected agent groups. The MCP server reads stub credentials containing the `onecli-managed` placeholder; the OneCLI gateway intercepts outbound calls to `calendar.googleapis.com` / `oauth2.googleapis.com` and swaps the bearer for the real OAuth token from its vault.
+
+**Why this package (and not gongrzhe's):** `@gongrzhe/server-calendar-autoauth-mcp` only supports the `primary` calendar and exposes 5 tools (no `list_calendars`). `@cocal/google-calendar-mcp` explicitly supports multi-calendar and multi-account, and is actively maintained.
+
+Tools exposed (surfaced as `mcp__calendar__<name>`, exact set depends on version — run `tools/list` against the MCP server to enumerate): `list-calendars`, `list-events`, `search-events`, `create-event`, `update-event`, `delete-event`, `get-event`, `list-colors`, `get-freebusy`, `get-current-time`, plus multi-account management tools.
+
+**Why this pattern:** v2's invariant is that containers never receive raw API keys (CHANGELOG 2.0.0). Same stub pattern `/add-gmail-tool` uses. This skill is deliberately a sibling, not a combined "Google Workspace" skill — installs independently and removes cleanly.
+
+## Phase 1: Pre-flight
+
+### Verify OneCLI has Google Calendar connected
+
+```bash
+onecli apps get --provider google-calendar
+```
+
+Expected: `"connection": { "status": "connected" }` with scopes including `calendar.readonly` and `calendar.events`.
+
+If not connected, tell the user:
+
+> Open the OneCLI web UI at http://127.0.0.1:10254, go to Apps → Google Calendar, and click Connect. Sign in with the Google account the agent should act as. `calendar.readonly` + `calendar.events` are the minimum useful scopes.
+
+### Verify stub credentials exist
+
+The stub lives at `~/.calendar-mcp/` by convention (shared with `/add-gmail-tool`'s sibling). cocal doesn't default to this path (it uses `~/.config/google-calendar-mcp/tokens.json`) — we override via env vars below so it reads our stubs instead.
+
+```bash
+ls -la ~/.calendar-mcp/gcp-oauth.keys.json ~/.calendar-mcp/credentials.json 2>&1
+```
+
+If both exist with `onecli-managed`:
+
+```bash
+grep -l onecli-managed ~/.calendar-mcp/gcp-oauth.keys.json ~/.calendar-mcp/credentials.json
+```
+
+...skip to Phase 2. If either file has real credentials (no `onecli-managed`), **STOP** — back up and delete before proceeding.
+
+If absent, write them:
+
+```bash
+mkdir -p ~/.calendar-mcp
+cat > ~/.calendar-mcp/gcp-oauth.keys.json <<'EOF'
+{
+  "installed": {
+    "client_id": "onecli-managed.apps.googleusercontent.com",
+    "client_secret": "onecli-managed",
+    "redirect_uris": ["http://localhost:3000/oauth2callback"]
+  }
+}
+EOF
+cat > ~/.calendar-mcp/credentials.json <<'EOF'
+{
+  "access_token": "onecli-managed",
+  "refresh_token": "onecli-managed",
+  "token_type": "Bearer",
+  "expiry_date": 99999999999999,
+  "scope": "https://www.googleapis.com/auth/calendar.readonly https://www.googleapis.com/auth/calendar.events"
+}
+EOF
+chmod 600 ~/.calendar-mcp/*.json
+```
+
+### Verify mount allowlist covers the path
+
+```bash
+cat ~/.config/nanoclaw/mount-allowlist.json
+```
+
+`~/.calendar-mcp` must sit under an `allowedRoots` entry.
+
+### Check agent secret-mode
+
+For each target agent group, confirm OneCLI will inject the Google Calendar token:
+
+```bash
+onecli agents list
+```
+
+`secretMode: all` is sufficient. If `selective`, explicitly assign the Calendar secret.
+
+## Phase 2: Apply Code Changes
+
+### Check if already applied
+
+```bash
+grep -q 'CALENDAR_MCP_VERSION' container/Dockerfile && \
+grep -q "mcp__calendar__\*" container/agent-runner/src/providers/claude.ts && \
+echo "ALREADY APPLIED — skip to Phase 3"
+```
+
+### Add MCP server to Dockerfile
+
+Edit `container/Dockerfile`. Find the pinned-version ARG block and add:
+
+```dockerfile
+ARG CALENDAR_MCP_VERSION=2.6.1
+```
+
+If `/add-gmail-tool` has already been applied, the pnpm global-install block already exists with its `zod-to-json-schema@3.22.5` pin. Just append the calendar package — **the calendar-mcp uses `zod@4.x` and does NOT need that pin**, but it's harmless to share the block:
+
+```dockerfile
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    pnpm install -g \
+        "@gongrzhe/server-gmail-autoauth-mcp@${GMAIL_MCP_VERSION}" \
+        "@cocal/google-calendar-mcp@${CALENDAR_MCP_VERSION}" \
+        "zod-to-json-schema@3.22.5"
+```
+
+If `/add-gmail-tool` hasn't been applied, install Calendar standalone:
+
+```dockerfile
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    pnpm install -g "@cocal/google-calendar-mcp@${CALENDAR_MCP_VERSION}"
+```
+
+### Add tools to allowlist
+
+Edit `container/agent-runner/src/providers/claude.ts`. Add `'mcp__calendar__*'` to `TOOL_ALLOWLIST` after `'mcp__nanoclaw__*'` (or after `'mcp__gmail__*'` if present).
+
+### Rebuild the container image
+
+```bash
+./container/build.sh
+```
+
+## Phase 3: Wire Per-Agent-Group
+
+For each agent group, merge into `groups/<folder>/container.json`:
+
+```jsonc
+{
+  "mcpServers": {
+    "calendar": {
+      "command": "google-calendar-mcp",
+      "args": [],
+      "env": {
+        "GOOGLE_OAUTH_CREDENTIALS": "/workspace/extra/.calendar-mcp/gcp-oauth.keys.json",
+        "GOOGLE_CALENDAR_MCP_TOKEN_PATH": "/workspace/extra/.calendar-mcp/credentials.json"
+      }
+    }
+  },
+  "additionalMounts": [
+    {
+      "hostPath": "/home/<user>/.calendar-mcp",
+      "containerPath": ".calendar-mcp",
+      "readonly": false
+    }
+  ]
+}
+```
+
+Substitute `<user>` with `echo $HOME`. `containerPath` is relative (mount-security rejects absolute paths — additional mounts land at `/workspace/extra/<relative>`).
+
+**Same-group-as-gmail tip:** if this group already has the gmail MCP + `.gmail-mcp` mount, **merge, don't replace** — both entries coexist in `mcpServers` and `additionalMounts`.
+
+## Phase 4: Build and Restart
+
+```bash
+pnpm run build
+systemctl --user restart nanoclaw   # Linux
+# launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
+```
+
+Kill any existing agent containers so they respawn with the new mcpServers config:
+
+```bash
+docker ps -q --filter 'name=nanoclaw-v2-' | xargs -r docker kill
+```
+
+## Phase 5: Verify
+
+### Test from a wired agent
+
+> Send: **"list my calendars"** or **"what's on my work calendar next Monday?"**.
+>
+> First call takes 2–3s while the MCP server starts and OneCLI does the token exchange.
+
+### Check logs if the tool isn't working
+
+```bash
+tail -100 logs/nanoclaw.log | grep -iE 'calendar|mcp'
+```
+
+Common signals:
+- `command not found: google-calendar-mcp` → image not rebuilt.
+- `ENOENT ...credentials.json` → mount missing. Check the mount allowlist.
+- `401 Unauthorized` from `*.googleapis.com` → OneCLI isn't injecting; verify agent's secret mode and that Google Calendar is connected.
+- Agent says "I don't have calendar tools" → `mcp__calendar__*` missing from `TOOL_ALLOWLIST`, or image cache stale (`./container/build.sh` again).
+
+## Removal
+
+1. Delete `"calendar"` from `mcpServers` and the `.calendar-mcp` mount from `additionalMounts` in each group's `container.json`.
+2. Remove `'mcp__calendar__*'` from `TOOL_ALLOWLIST`.
+3. Remove `CALENDAR_MCP_VERSION` ARG and the calendar package from the Dockerfile install block.
+4. `pnpm run build && ./container/build.sh && systemctl --user restart nanoclaw`.
+5. Optional: `rm -rf ~/.calendar-mcp/` and `onecli apps disconnect --provider google-calendar`.
+
+## Credits & references
+
+- **MCP server:** [`@cocal/google-calendar-mcp`](https://github.com/cocal-com/google-calendar-mcp) — MIT-licensed, actively maintained, multi-account and multi-calendar.
+- **Why not gongrzhe:** earlier versions of this skill used `@gongrzhe/server-calendar-autoauth-mcp@1.0.2` which only supports the primary calendar with 5 event-level tools. The cocal server supersedes it.
+- **Skill pattern:** direct sibling of [`/add-gmail-tool`](../add-gmail-tool/SKILL.md); same OneCLI stub mechanism.
diff --git a/.claude/skills/add-gmail-tool/SKILL.md b/.claude/skills/add-gmail-tool/SKILL.md
new file mode 100644
index 0000000..095c285
--- /dev/null
+++ b/.claude/skills/add-gmail-tool/SKILL.md
@@ -0,0 +1,229 @@
+---
+name: add-gmail-tool
+description: Add Gmail as an MCP tool (read, search, send, label, draft) using OneCLI-managed OAuth. The agent gets Gmail tools in every enabled group; OneCLI injects real tokens at request time so no raw credentials are ever in the container or on disk in usable form.
+---
+
+# Add Gmail Tool (OneCLI-native)
+
+This skill wires the [`@gongrzhe/server-gmail-autoauth-mcp`](https://www.npmjs.com/package/@gongrzhe/server-gmail-autoauth-mcp) stdio MCP server into selected agent groups. The MCP server reads stub credentials containing the `onecli-managed` placeholder; the OneCLI gateway intercepts outbound calls to `gmail.googleapis.com` and injects the real OAuth bearer from its vault.
+
+Tools exposed (from `gmail-mcp@1.1.11`, surfaced to the agent as `mcp__gmail__<name>`): `search_emails`, `read_email`, `send_email`, `draft_email`, `delete_email`, `modify_email`, `batch_modify_emails`, `batch_delete_emails`, `download_attachment`, `list_email_labels`, `create_label`, `update_label`, `delete_label`, `get_or_create_label`, `list_filters`, `get_filter`, `create_filter`, `create_filter_from_template`, `delete_filter`.
+
+**Why this pattern:** v2's invariant is that containers never receive raw API keys — OneCLI is the sole credential path (see CHANGELOG v2.0.0). The stub-file pattern satisfies this: the container sees `"onecli-managed"` placeholders, the gateway swaps them in flight.
+
+## Phase 1: Pre-flight
+
+### Verify OneCLI has Gmail connected
+
+```bash
+onecli apps get --provider gmail
+```
+
+Expected: `"connection": { "status": "connected" }` with scopes including `gmail.readonly`, `gmail.modify`, `gmail.send`.
+
+If not connected, tell the user:
+
+> Open the OneCLI web UI at http://127.0.0.1:10254, go to Apps → Gmail, and click Connect. Sign in with the Google account you want the agent to act as.
+
+### Verify stub credentials exist
+
+```bash
+ls -la ~/.gmail-mcp/gcp-oauth.keys.json ~/.gmail-mcp/credentials.json 2>&1
+```
+
+If both exist and contain `"onecli-managed"`:
+
+```bash
+grep -l onecli-managed ~/.gmail-mcp/gcp-oauth.keys.json ~/.gmail-mcp/credentials.json
+```
+
+...skip to Phase 2.
+
+If either file exists but does **not** contain `onecli-managed`, **STOP** and tell the user — these are real OAuth credentials from a previous non-OneCLI install. Back them up, then delete before proceeding. The OneCLI migration normally handles this; if it didn't, something is wrong.
+
+If both files are absent, write them now:
+
+```bash
+mkdir -p ~/.gmail-mcp
+cat > ~/.gmail-mcp/gcp-oauth.keys.json <<'EOF'
+{
+  "installed": {
+    "client_id": "onecli-managed.apps.googleusercontent.com",
+    "client_secret": "onecli-managed",
+    "redirect_uris": ["http://localhost:3000/oauth2callback"]
+  }
+}
+EOF
+cat > ~/.gmail-mcp/credentials.json <<'EOF'
+{
+  "access_token": "onecli-managed",
+  "refresh_token": "onecli-managed",
+  "token_type": "Bearer",
+  "expiry_date": 99999999999999,
+  "scope": "https://www.googleapis.com/auth/gmail.readonly https://www.googleapis.com/auth/gmail.modify https://www.googleapis.com/auth/gmail.send"
+}
+EOF
+chmod 600 ~/.gmail-mcp/gcp-oauth.keys.json ~/.gmail-mcp/credentials.json
+```
+
+### Verify mount allowlist covers the path
+
+```bash
+cat ~/.config/nanoclaw/mount-allowlist.json
+```
+
+`~/.gmail-mcp` must sit under an `allowedRoots` entry (e.g. `/home/<user>`). If it doesn't, tell the user to run `/manage-mounts` first or add their home directory.
+
+### Check agent secret-mode
+
+For each target agent group, confirm OneCLI will inject Gmail secrets into its container. Find the OneCLI agent ID that matches the group's `agentGroupId`:
+
+```bash
+onecli agents list
+```
+
+If that agent's `secretMode` is `all`, you're done — Gmail secrets (identified by OneCLI's Gmail hostPattern) will auto-inject. If it's `selective`, explicitly assign the Gmail secrets:
+
+```bash
+onecli secrets list     # find Gmail secret IDs (OneCLI creates one per connected app)
+onecli agents set-secrets --id <agent-id> --secret-ids <gmail-secret-id>
+```
+
+## Phase 2: Apply Code Changes
+
+### Check if already applied
+
+```bash
+grep -q 'GMAIL_MCP_VERSION' container/Dockerfile && \
+grep -q "mcp__gmail__\*" container/agent-runner/src/providers/claude.ts && \
+echo "ALREADY APPLIED — skip to Phase 3"
+```
+
+### Add MCP server to Dockerfile
+
+Edit `container/Dockerfile`. Find the pinned-version ARG block:
+
+```dockerfile
+ARG CLAUDE_CODE_VERSION=2.1.116
+ARG AGENT_BROWSER_VERSION=latest
+ARG VERCEL_VERSION=latest
+ARG BUN_VERSION=1.3.12
+```
+
+Add a new line:
+
+```dockerfile
+ARG GMAIL_MCP_VERSION=1.1.11
+```
+
+Then find the last pnpm global-install `RUN` block (the one that installs `@anthropic-ai/claude-code`) and add a new block after it, before `# ---- Entrypoint`:
+
+```dockerfile
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    pnpm install -g \
+        "@gongrzhe/server-gmail-autoauth-mcp@${GMAIL_MCP_VERSION}" \
+        "zod-to-json-schema@3.22.5"
+```
+
+Pinned version matters — `minimumReleaseAge` in `pnpm-workspace.yaml` gates trunk installs, and CLAUDE.md requires a fixed ARG version for all Node CLIs installed into the image.
+
+**Why the `zod-to-json-schema` pin:** `@gongrzhe/server-gmail-autoauth-mcp@1.1.11` has loose deps (`zod-to-json-schema: ^3.22.1`, `zod: ^3.22.4`). pnpm resolves `zod-to-json-schema` to the latest 3.25.x, which imports `zod/v3` — a subpath that only exists in `zod>=3.25`. But `zod` resolves to `3.24.x` (highest satisfying `^3.22.4` without breaking peer ranges). Result: `ERR_PACKAGE_PATH_NOT_EXPORTED` at import time. Pinning `zod-to-json-schema` to a pre-v3-subpath version avoids it. Re-check if you bump `GMAIL_MCP_VERSION`.
+
+### Add tools to allowlist
+
+Edit `container/agent-runner/src/providers/claude.ts`. Find `'mcp__nanoclaw__*',` in `TOOL_ALLOWLIST` and add `'mcp__gmail__*',` after it.
+
+### Rebuild the container image
+
+```bash
+./container/build.sh
+```
+
+Must complete cleanly. The new `pnpm install -g` layer is ~60s first time (cached on rebuild).
+
+## Phase 3: Wire Per-Agent-Group
+
+For each agent group that should have Gmail (ask the user — typically their personal DM and CLI agents, sometimes shared household agents), edit `groups/<folder>/container.json` to add the mount and MCP server.
+
+Merge these into the group's `container.json`:
+
+```jsonc
+{
+  "mcpServers": {
+    "gmail": {
+      "command": "gmail-mcp",
+      "args": [],
+      "env": {
+        "GMAIL_OAUTH_PATH": "/workspace/extra/.gmail-mcp/gcp-oauth.keys.json",
+        "GMAIL_CREDENTIALS_PATH": "/workspace/extra/.gmail-mcp/credentials.json"
+      }
+    }
+  },
+  "additionalMounts": [
+    {
+      "hostPath": "/home/<user>/.gmail-mcp",
+      "containerPath": ".gmail-mcp",
+      "readonly": false
+    }
+  ]
+}
+```
+
+Substitute `<user>` with the host user's home (use `echo $HOME`, don't assume `~` will expand — `container-runner.ts` does expand `~` via `expandPath`, but an explicit absolute path is clearer and matches what `/manage-mounts` writes).
+
+**Why the container path is relative:** `mount-security` rejects absolute `containerPath` values. Additional mounts are prefixed with `/workspace/extra/`, so `containerPath: ".gmail-mcp"` lands at `/workspace/extra/.gmail-mcp`. The MCP server's `GMAIL_OAUTH_PATH` / `GMAIL_CREDENTIALS_PATH` env vars point at that absolute location inside the container.
+
+## Phase 4: Build and Restart
+
+```bash
+pnpm run build
+systemctl --user restart nanoclaw  # Linux
+# launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
+```
+
+## Phase 5: Verify
+
+### Test from the wired agent
+
+Tell the user:
+
+> In your `<agent-name>` chat, send: **"list my gmail labels"** or **"search my inbox for invoices from last month"**.
+>
+> The agent should use `mcp__gmail__list_labels` / `mcp__gmail__search`. The first call may take a second or two while the MCP server starts and OneCLI does the token exchange.
+
+### Check logs if the tool isn't working
+
+```bash
+tail -100 logs/nanoclaw.log logs/nanoclaw.error.log | grep -iE 'gmail|mcp'
+# Per-container logs — session-scoped:
+ls data/v2-sessions/*/stderr.log | head
+```
+
+Common signals:
+- `command not found: gmail-mcp` → image wasn't rebuilt or PATH doesn't include `/pnpm` (should — `ENV PATH="$PNPM_HOME:$PATH"` in Dockerfile).
+- `ENOENT: no such file or directory, open '/workspace/extra/.gmail-mcp/credentials.json'` → mount is missing. Check `~/.config/nanoclaw/mount-allowlist.json` includes a parent of `~/.gmail-mcp`.
+- `401 Unauthorized` from `gmail.googleapis.com` → OneCLI isn't injecting. Check the agent's secret mode (`onecli agents secrets --id <agent-id>`) and that the Gmail app is connected (`onecli apps get --provider gmail`).
+- Agent says "I don't have Gmail tools" → `mcp__gmail__*` wasn't added to `TOOL_ALLOWLIST`, or the agent-runner wasn't rebuilt (image cache — run `./container/build.sh` again with `--no-cache` if suspicious).
+
+## Removal
+
+1. Delete the `"gmail"` entry from `mcpServers` and the `.gmail-mcp` entry from `additionalMounts` in each group's `container.json`.
+2. Remove `'mcp__gmail__*'` from `TOOL_ALLOWLIST` in `container/agent-runner/src/providers/claude.ts`.
+3. Remove the `GMAIL_MCP_VERSION` ARG and the `pnpm install -g @gongrzhe/server-gmail-autoauth-mcp` block from `container/Dockerfile`.
+4. `pnpm run build && ./container/build.sh && systemctl --user restart nanoclaw`.
+5. (Optional) `rm -rf ~/.gmail-mcp/` if no other host-side tool needs the stubs.
+6. (Optional) Disconnect Gmail in OneCLI: `onecli apps disconnect --provider gmail`.
+
+## Notes
+
+- **Stub format is OneCLI-prescribed.** The `access_token: "onecli-managed"` pattern with `expiry_date: 99999999999999` tells the Google auth client the token is valid; OneCLI intercepts the outgoing Gmail API call and rewrites `Authorization: Bearer onecli-managed` to the real token. `expiry_date: 0` (refresh-interception) is an alternative the OneCLI docs describe — both work but OneCLI's own `migrate` command writes the far-future variant, which is what this skill assumes.
+- **Scopes are set at OAuth connect time.** If the agent needs scopes beyond what's currently connected (e.g. the user later wants `calendar.readonly` for combined email/calendar workflows), disconnect and reconnect Gmail in the OneCLI web UI with the expanded scope set.
+- **This is tool-only.** Inbound email as a channel (emails trigger the agent) is a separate piece of work — it needs a `src/channels/gmail.ts` adapter that polls the inbox and routes to a messaging group. The pre-v2 qwibitai skill had this; it has not been ported to v2's channel architecture as of v2.0.0.
+
+## Credits & references
+
+- **MCP server:** [`@gongrzhe/server-gmail-autoauth-mcp`](https://github.com/GongRzhe/Gmail-MCP-Server) by GongRzhe — MIT-licensed.
+- **OneCLI credential stubs:** pattern documented at `https://onecli.sh/docs/guides/credential-stubs/gmail.md`.
+- **Skill pattern:** modeled on [`add-atomic-chat-tool`](../add-atomic-chat-tool/SKILL.md) and [`add-vercel`](../add-vercel/SKILL.md).
+- **Addresses:** [issue #1500](https://github.com/qwibitai/nanoclaw/issues/1500) (proxy Gmail/Calendar OAuth tokens through credential proxy) for the Gmail side.
+- **Related PRs:** [#1810](https://github.com/qwibitai/nanoclaw/pull/1810) (pre-install Gmail/Notion MCP) overlaps on the "install the MCP server in the image" idea but bundles many unrelated changes; this skill is the focused OneCLI-native version.
diff --git a/.claude/skills/add-gmail/SKILL.md b/.claude/skills/add-gmail/SKILL.md
deleted file mode 100644
index 6a13291..0000000
--- a/.claude/skills/add-gmail/SKILL.md
+++ /dev/null
@@ -1,236 +0,0 @@
----
-name: add-gmail
-description: Add Gmail integration to NanoClaw. Can be configured as a tool (agent reads/sends emails when triggered from WhatsApp) or as a full channel (emails can trigger the agent, schedule tasks, and receive replies). Guides through GCP OAuth setup and implements the integration.
----
-
-# Add Gmail Integration
-
-This skill adds Gmail support to NanoClaw — either as a tool (read, send, search, draft) or as a full channel that polls the inbox.
-
-## Phase 1: Pre-flight
-
-### Check if already applied
-
-Check if `src/channels/gmail.ts` exists. If it does, skip to Phase 3 (Setup). The code changes are already in place.
-
-### Ask the user
-
-Use `AskUserQuestion`:
-
-AskUserQuestion: Should incoming emails be able to trigger the agent?
-
-- **Yes** — Full channel mode: the agent listens on Gmail and responds to incoming emails automatically
-- **No** — Tool-only: the agent gets full Gmail tools (read, send, search, draft) but won't monitor the inbox. No channel code is added.
-
-## Phase 2: Apply Code Changes
-
-### Ensure channel remote
-
-```bash
-git remote -v
-```
-
-If `gmail` is missing, add it:
-
-```bash
-git remote add gmail https://github.com/qwibitai/nanoclaw-gmail.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch gmail main
-git merge gmail/main || {
-  git checkout --theirs pnpm-lock.yaml
-  git add pnpm-lock.yaml
-  git merge --continue
-}
-```
-
-This merges in:
-- `src/channels/gmail.ts` (GmailChannel class with self-registration via `registerChannel`)
-- `src/channels/gmail.test.ts` (unit tests)
-- `import './gmail.js'` appended to the channel barrel file `src/channels/index.ts`
-- Gmail credentials mount (`~/.gmail-mcp`) in `src/container-runner.ts`
-- Gmail MCP server (`@gongrzhe/server-gmail-autoauth-mcp`) and `mcp__gmail__*` allowed tool in `container/agent-runner/src/index.ts`
-- `googleapis` npm dependency in `package.json`
-
-If the merge reports conflicts, resolve them by reading the conflicted files and understanding the intent of both sides.
-
-### Add email handling instructions (Channel mode only)
-
-If the user chose channel mode, append the following to `groups/main/CLAUDE.md` (before the formatting section):
-
-```markdown
-## Email Notifications
-
-When you receive an email notification (messages starting with `[Email from ...`), inform the user about it but do NOT reply to the email unless specifically asked. You have Gmail tools available — use them only when the user explicitly asks you to reply, forward, or take action on an email.
-```
-
-### Validate code changes
-
-```bash
-pnpm install
-pnpm run build
-pnpm exec vitest run src/channels/gmail.test.ts
-```
-
-All tests must pass (including the new Gmail tests) and build must be clean before proceeding.
-
-## Phase 3: Setup
-
-### Check existing Gmail credentials
-
-```bash
-ls -la ~/.gmail-mcp/ 2>/dev/null || echo "No Gmail config found"
-```
-
-If `credentials.json` already exists with real tokens (not `onecli-managed` values), skip to "Build and restart" below.
-
-### GCP Project Setup
-
-Check if OneCLI is configured:
-
-```bash
-grep -q 'ONECLI_URL=.' .env 2>/dev/null && echo "onecli" || echo "manual"
-```
-
-**If OneCLI:** Tell the user to open `${ONECLI_URL}/connections?connect=gmail` to set up their Gmail connection. The dashboard walks them through creating a Google Cloud OAuth app and authorizing it. Ask them to let you know when done.
-
-Once the user confirms, run:
-
-```bash
-onecli apps get --provider gmail
-```
-
-Check that `config.hasCredentials` is `true` or `connection` is not null. The response `hint` field has instructions and a docs URL for what stub credential files to create under `~/.gmail-mcp/`. Follow the hint — never overwrite existing files that don't contain `onecli-managed` values.
-
-**If manual:** Tell the user:
-
-> I need you to set up Google Cloud OAuth credentials:
->
-> 1. Open https://console.cloud.google.com — create a new project or select existing
-> 2. Go to **APIs & Services > Library**, search "Gmail API", click **Enable**
-> 3. Go to **APIs & Services > Credentials**, click **+ CREATE CREDENTIALS > OAuth client ID**
->    - If prompted for consent screen: choose "External", fill in app name and email, save
->    - Application type: **Desktop app**, name: anything (e.g., "NanoClaw Gmail")
-> 4. Click **DOWNLOAD JSON** and save as `gcp-oauth.keys.json`
->
-> Where did you save the file? (Give me the full path, or paste the file contents here)
-
-If user provides a path, copy it:
-
-```bash
-mkdir -p ~/.gmail-mcp
-cp "/path/user/provided/gcp-oauth.keys.json" ~/.gmail-mcp/gcp-oauth.keys.json
-```
-
-If user pastes JSON content, write it to `~/.gmail-mcp/gcp-oauth.keys.json`.
-
-### OAuth Authorization
-
-Tell the user:
-
-> I'm going to run Gmail authorization. A browser window will open — sign in and grant access. If you see an "app isn't verified" warning, click "Advanced" then "Go to [app name] (unsafe)" — this is normal for personal OAuth apps.
-
-Run the authorization:
-
-```bash
-pnpm dlx @gongrzhe/server-gmail-autoauth-mcp auth
-```
-
-If that fails (some versions don't have an auth subcommand), try `timeout 60 pnpm dlx @gongrzhe/server-gmail-autoauth-mcp || true`. Verify with `ls ~/.gmail-mcp/credentials.json`.
-
-### Build and restart
-
-Clear stale per-group agent-runner copies (they only get re-created if missing, so existing copies won't pick up the new Gmail server):
-
-```bash
-rm -r data/sessions/*/agent-runner-src 2>/dev/null || true
-```
-
-Rebuild the container (agent-runner changed):
-
-```bash
-cd container && ./build.sh
-```
-
-Then compile and restart:
-
-```bash
-pnpm run build
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw  # macOS
-# Linux: systemctl --user restart nanoclaw
-```
-
-## Phase 4: Verify
-
-### Test tool access (both modes)
-
-Tell the user:
-
-> Gmail is connected! Send this in your main channel:
->
-> `@Andy check my recent emails` or `@Andy list my Gmail labels`
-
-### Test channel mode (Channel mode only)
-
-Tell the user to send themselves a test email. The agent should pick it up within a minute. Monitor: `tail -f logs/nanoclaw.log | grep -iE "(gmail|email)"`.
-
-Once verified, offer filter customization via `AskUserQuestion` — by default, only emails in the Primary inbox trigger the agent (Promotions, Social, Updates, and Forums are excluded). The user can keep this default or narrow further by sender, label, or keywords. No code changes needed for filters.
-
-### Check logs if needed
-
-```bash
-tail -f logs/nanoclaw.log
-```
-
-## Troubleshooting
-
-### Gmail connection not responding
-
-Test directly:
-
-```bash
-pnpm dlx @gongrzhe/server-gmail-autoauth-mcp
-```
-
-### OAuth token expired
-
-Re-authorize:
-
-```bash
-rm ~/.gmail-mcp/credentials.json
-pnpm dlx @gongrzhe/server-gmail-autoauth-mcp
-```
-
-### Container can't access Gmail
-
-- Verify `~/.gmail-mcp` is mounted: check `src/container-runner.ts` for the `.gmail-mcp` mount
-- Check container logs: `cat groups/main/logs/container-*.log | tail -50`
-
-### Emails not being detected (Channel mode only)
-
-- By default, the channel polls unread Primary inbox emails (`is:unread category:primary`)
-- Check logs for Gmail polling errors
-
-## Removal
-
-### Tool-only mode
-
-1. Remove `~/.gmail-mcp` mount from `src/container-runner.ts`
-2. Remove `gmail` MCP server and `mcp__gmail__*` from `container/agent-runner/src/index.ts`
-3. Rebuild and restart
-4. Clear stale agent-runner copies: `rm -r data/sessions/*/agent-runner-src 2>/dev/null || true`
-5. Rebuild: `cd container && ./build.sh && cd .. && pnpm run build && launchctl kickstart -k gui/$(id -u)/com.nanoclaw` (macOS) or `systemctl --user restart nanoclaw` (Linux)
-
-### Channel mode
-
-1. Delete `src/channels/gmail.ts` and `src/channels/gmail.test.ts`
-2. Remove `import './gmail.js'` from `src/channels/index.ts`
-3. Remove `~/.gmail-mcp` mount from `src/container-runner.ts`
-4. Remove `gmail` MCP server and `mcp__gmail__*` from `container/agent-runner/src/index.ts`
-5. Uninstall: `pnpm uninstall googleapis`
-6. Rebuild and restart
-7. Clear stale agent-runner copies: `rm -r data/sessions/*/agent-runner-src 2>/dev/null || true`
-8. Rebuild: `cd container && ./build.sh && cd .. && pnpm run build && launchctl kickstart -k gui/$(id -u)/com.nanoclaw` (macOS) or `systemctl --user restart nanoclaw` (Linux)
diff --git a/.claude/skills/add-image-vision/SKILL.md b/.claude/skills/add-image-vision/SKILL.md
deleted file mode 100644
index 4a9da26..0000000
--- a/.claude/skills/add-image-vision/SKILL.md
+++ /dev/null
@@ -1,94 +0,0 @@
----
-name: add-image-vision
-description: Add image vision to NanoClaw agents. Resizes and processes WhatsApp image attachments, then sends them to Claude as multimodal content blocks.
----
-
-# Image Vision Skill
-
-Adds the ability for NanoClaw agents to see and understand images sent via WhatsApp. Images are downloaded, resized with sharp, saved to the group workspace, and passed to the agent as base64-encoded multimodal content blocks.
-
-## Phase 1: Pre-flight
-
-1. Check if `src/image.ts` exists — skip to Phase 3 if already applied
-2. Confirm `sharp` is installable (native bindings require build tools)
-
-**Prerequisite:** WhatsApp must be installed first (`skill/whatsapp` merged). This skill modifies WhatsApp channel files.
-
-## Phase 2: Apply Code Changes
-
-### Ensure WhatsApp fork remote
-
-```bash
-git remote -v
-```
-
-If `whatsapp` is missing, add it:
-
-```bash
-git remote add whatsapp https://github.com/qwibitai/nanoclaw-whatsapp.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch whatsapp skill/image-vision
-git merge whatsapp/skill/image-vision || {
-  git checkout --theirs pnpm-lock.yaml
-  git add pnpm-lock.yaml
-  git merge --continue
-}
-```
-
-This merges in:
-- `src/image.ts` (image download, resize via sharp, base64 encoding)
-- `src/image.test.ts` (8 unit tests)
-- Image attachment handling in `src/channels/whatsapp.ts`
-- Image passing to agent in `src/index.ts` and `src/container-runner.ts`
-- Image content block support in `container/agent-runner/src/index.ts`
-- `sharp` npm dependency in `package.json`
-
-If the merge reports conflicts, resolve them by reading the conflicted files and understanding the intent of both sides.
-
-### Validate code changes
-
-```bash
-pnpm install
-pnpm run build
-pnpm exec vitest run src/image.test.ts
-```
-
-All tests must pass and build must be clean before proceeding.
-
-## Phase 3: Configure
-
-1. Rebuild the container (agent-runner changes need a rebuild):
-   ```bash
-   ./container/build.sh
-   ```
-
-2. Sync agent-runner source to group caches:
-   ```bash
-   for dir in data/sessions/*/agent-runner-src/; do
-     cp container/agent-runner/src/*.ts "$dir"
-   done
-   ```
-
-3. Restart the service:
-   ```bash
-   launchctl kickstart -k gui/$(id -u)/com.nanoclaw
-   ```
-
-## Phase 4: Verify
-
-1. Send an image in a registered WhatsApp group
-2. Check the agent responds with understanding of the image content
-3. Check logs for "Processed image attachment":
-   ```bash
-   tail -50 groups/*/logs/container-*.log
-   ```
-
-## Troubleshooting
-
-- **"Image - download failed"**: Check WhatsApp connection stability. The download may timeout on slow connections.
-- **"Image - processing failed"**: Sharp may not be installed correctly. Run `pnpm ls sharp` to verify.
-- **Agent doesn't mention image content**: Check container logs for "Loaded image" messages. If missing, ensure agent-runner source was synced to group caches.
diff --git a/.claude/skills/add-matrix/SKILL.md b/.claude/skills/add-matrix/SKILL.md
index 7a6ea0f..cf6da75 100644
--- a/.claude/skills/add-matrix/SKILL.md
+++ b/.claude/skills/add-matrix/SKILL.md
@@ -47,7 +47,29 @@ import './matrix.js';
 pnpm install @beeper/chat-adapter-matrix@0.2.0
 ```
 
-### 5. Build
+### 5. Patch matrix-js-sdk ESM imports
+
+The adapter's published dist references `matrix-js-sdk/lib/...` without `.js`
+extensions, which fails under Node 22 strict ESM resolution. Add the missing
+extensions (idempotent — safe to re-run):
+
+```bash
+node -e '
+  const fs = require("fs"), path = require("path");
+  const root = "node_modules/.pnpm";
+  const dir = fs.readdirSync(root).find(d => d.startsWith("@beeper+chat-adapter-matrix@"));
+  if (!dir) { console.log("Matrix adapter not installed"); process.exit(0); }
+  const f = path.join(root, dir, "node_modules/@beeper/chat-adapter-matrix/dist/index.js");
+  fs.writeFileSync(f, fs.readFileSync(f, "utf8").replace(
+    /from "(matrix-js-sdk\/lib\/[^"]+?)(?<!\.js)"/g, "from \"$1.js\""
+  ));
+  console.log("Patched", f);
+'
+```
+
+Re-run this after every `pnpm install` that touches the adapter.
+
+### 6. Build
 
 ```bash
 pnpm run build
@@ -55,25 +77,60 @@ pnpm run build
 
 ## Credentials
 
-1. Register a bot account on your Matrix homeserver (e.g., via Element)
-2. Get the homeserver URL (e.g., `https://matrix.org` or your self-hosted URL)
-3. Get an access token:
-   - In Element: **Settings** > **Help & About** > **Access Token** (advanced)
-   - Or via API: `curl -XPOST 'https://matrix.org/_matrix/client/r0/login' -d '{"type":"m.login.password","user":"botuser","password":"..."}'`
-4. Note the bot's user ID (e.g., `@botuser:matrix.org`)
+The bot needs its own Matrix account — separate from the user's account. This is required because Matrix cannot send DMs to yourself.
 
-### Configure environment
+### Create a bot account
 
-Add to `.env`:
+1. Open [app.element.io](https://app.element.io) in a private/incognito window (or sign out first)
+2. Register a new account for the bot (e.g. `andybot` on matrix.org)
+3. Note the bot's user ID (e.g. `@andybot:matrix.org`)
+
+### Choose an auth method
+
+**Option A: Username + Password (simpler)**
+
+No extra steps — just use the bot account's credentials directly. The adapter logs in automatically.
+
+```bash
+MATRIX_BASE_URL=https://matrix.org
+MATRIX_USERNAME=andybot
+MATRIX_PASSWORD=your-bot-password
+MATRIX_USER_ID=@andybot:matrix.org
+MATRIX_BOT_USERNAME=Andy
+```
+
+**Option B: Access Token (recommended for production)**
+
+Get an access token from Element: sign into the bot account → **Settings** > **Help & About** > **Access Token** (under Advanced). Or via API:
+
+```bash
+curl -XPOST 'https://matrix.org/_matrix/client/r0/login' \
+  -d '{"type":"m.login.password","user":"andybot","password":"..."}'
+```
 
 ```bash
 MATRIX_BASE_URL=https://matrix.org
 MATRIX_ACCESS_TOKEN=your-access-token
-MATRIX_USER_ID=@botuser:matrix.org
-MATRIX_BOT_USERNAME=botuser
+MATRIX_USER_ID=@andybot:matrix.org
+MATRIX_BOT_USERNAME=Andy
 ```
 
-Sync to container: `mkdir -p data/env && cp .env data/env/env`
+### Optional settings
+
+```bash
+MATRIX_INVITE_AUTOJOIN=true                    # Auto-accept room invites (default: true)
+MATRIX_INVITE_AUTOJOIN_ALLOWLIST=@you:matrix.org  # Only accept invites from these users
+MATRIX_RECOVERY_KEY=your-recovery-key          # Enable E2EE cross-signing
+MATRIX_DEVICE_ID=NANOCLAW01                    # Stable device ID across restarts
+```
+
+### Configure environment
+
+Add the chosen env vars to `.env`, then sync:
+
+```bash
+mkdir -p data/env && cp .env data/env/env
+```
 
 ## Next Steps
 
@@ -85,7 +142,7 @@ Otherwise, run `/manage-channels` to wire this channel to an agent group.
 
 - **type**: `matrix`
 - **terminology**: Matrix has "rooms." A room can be a group chat or a direct message. Rooms have internal IDs (like `!abc123:matrix.org`) and optional aliases (like `#general:matrix.org`).
-- **how-to-find-id**: In Element, click the room name > Settings > Advanced — the "Internal room ID" is the platform ID (starts with `!`). Or use a room alias like `#general:matrix.org`.
+- **how-to-find-id**: For DMs, use the bot's `openDM` to resolve the room automatically. For group rooms, in Element click the room name > Settings > Advanced — the "Internal room ID" is the platform ID (starts with `!`). Or use a room alias like `#general:matrix.org`.
 - **supports-threads**: partial (some clients support threads, but not all — treat as no for reliability)
-- **typical-use**: Interactive chat — rooms or direct messages
+- **typical-use**: Interactive chat — rooms or direct messages. Requires a separate bot account (the agent cannot DM users from their own account).
 - **default-isolation**: Same agent group for rooms where you're the primary user. Separate agent group for rooms with different communities or sensitive contexts.
diff --git a/.claude/skills/add-ollama-provider/SKILL.md b/.claude/skills/add-ollama-provider/SKILL.md
new file mode 100644
index 0000000..83f7e5a
--- /dev/null
+++ b/.claude/skills/add-ollama-provider/SKILL.md
@@ -0,0 +1,179 @@
+---
+name: add-ollama-provider
+description: Route a NanoClaw agent group to a local Ollama model instead of the Anthropic API. Ollama speaks the Anthropic API natively (v1/messages), so no provider code changes are needed — just env var overrides and a model setting. Use when the user wants to run their agent locally, cut API costs, or experiment with open-weight models. See docs/ollama.md for background.
+---
+
+# Add Ollama Provider
+
+Routes an agent group to a local Ollama instance instead of the Anthropic API.
+See `docs/ollama.md` for how this works and the tradeoffs involved.
+
+## Prerequisites
+
+1. **Ollama is installed and running** on the host — verify: `curl -s http://localhost:11434/api/tags`
+2. **A model is pulled** — e.g. `ollama pull gemma4` or `ollama pull qwen3-coder`
+3. **The agent group already exists** — run `/init-first-agent` first if needed
+
+## 1. Check source support
+
+The feature requires two fields in `ContainerConfig` (`env` and `blockedHosts`) and their
+corresponding wiring in `container-runner.ts`. Check if already present:
+
+```bash
+grep -c 'blockedHosts' src/container-config.ts src/container-runner.ts
+```
+
+If either count is 0, apply the changes in steps 1a and 1b. Otherwise skip to step 2.
+
+### 1a. Extend ContainerConfig
+
+In `src/container-config.ts`, add to the `ContainerConfig` interface:
+
+```typescript
+env?: Record<string, string>;
+blockedHosts?: string[];
+```
+
+And in `readContainerConfig`, add inside the returned object:
+
+```typescript
+env: raw.env,
+blockedHosts: raw.blockedHosts,
+```
+
+### 1b. Wire into container-runner
+
+In `src/container-runner.ts`, after the `NANOCLAW_MCP_SERVERS` block, add:
+
+```typescript
+// Per-agent-group env overrides — applied last to win over OneCLI values.
+if (containerConfig.env) {
+  for (const [key, value] of Object.entries(containerConfig.env)) {
+    args.push('-e', `${key}=${value}`);
+  }
+}
+
+// Blocked hosts: resolve to 0.0.0.0 so they are unreachable inside the container.
+if (containerConfig.blockedHosts) {
+  for (const host of containerConfig.blockedHosts) {
+    args.push('--add-host', `${host}:0.0.0.0`);
+  }
+}
+```
+
+### 1c. Fix home directory permissions (if not already done)
+
+The container may run as your host uid (not uid 1000). Check the Dockerfile:
+
+```bash
+grep 'chmod.*home/node' container/Dockerfile
+```
+
+If it shows `chmod 755`, change it to `chmod 777` so any uid can write there.
+Then rebuild the container image: `./container/build.sh`
+
+## 2. Identify the setup
+
+Ask the user (plain text, not AskUserQuestion):
+
+1. **Which agent group?** List available groups: `sqlite3 data/v2.db "SELECT folder, name FROM agent_groups;"`
+2. **Which Ollama model?** List available: `curl -s http://localhost:11434/api/tags | grep '"name"'`
+3. **Block Anthropic API?** Recommended yes — prevents accidental spend if config drifts.
+
+Record as `FOLDER`, `MODEL`, and `BLOCK_ANTHROPIC`.
+
+## 3. Configure container.json
+
+Read `groups/<FOLDER>/container.json`. Add (or merge into) an `env` block and optionally `blockedHosts`:
+
+```json
+{
+  "env": {
+    "ANTHROPIC_BASE_URL": "http://host.docker.internal:11434",
+    "ANTHROPIC_API_KEY": "ollama",
+    "NO_PROXY": "host.docker.internal",
+    "no_proxy": "host.docker.internal"
+  },
+  "blockedHosts": ["api.anthropic.com"]
+}
+```
+
+Omit `blockedHosts` if the user declined step 2.
+
+**Why these vars:** `ANTHROPIC_BASE_URL` redirects the Anthropic SDK to Ollama.
+`ANTHROPIC_API_KEY=ollama` satisfies the SDK's key requirement (Ollama ignores it).
+`NO_PROXY` bypasses the OneCLI HTTPS proxy for requests to `host.docker.internal`
+so they reach Ollama directly instead of going through the credential gateway.
+
+## 4. Set the model
+
+Read the agent group's shared Claude settings:
+
+```bash
+# Find the agent group ID
+AG_ID=$(sqlite3 data/v2.db "SELECT id FROM agent_groups WHERE folder='<FOLDER>';")
+SETTINGS=data/v2-sessions/$AG_ID/.claude-shared/settings.json
+```
+
+Add `"model": "<MODEL>"` to that settings file. Create the file if it doesn't exist:
+
+```json
+{
+  "model": "gemma4:latest"
+}
+```
+
+If the file already has content, merge the `model` key in — don't overwrite existing keys.
+
+**Why here and not container.json:** Claude Code reads its model from its own settings
+file, not from env vars. This file is bind-mounted into the container as `~/.claude/settings.json`.
+
+## 5. Build and restart
+
+```bash
+export PATH="/opt/homebrew/bin:$PATH"
+pnpm run build
+launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist
+launchctl load ~/Library/LaunchAgents/com.nanoclaw.plist
+# Linux: systemctl --user restart nanoclaw
+```
+
+## 6. Verify
+
+Send a message to the agent. Then confirm:
+
+```bash
+# Ollama shows the model as active
+curl -s http://localhost:11434/api/ps | grep '"name"'
+
+# Container has the right env vars
+CTR=$(docker ps --filter "name=nanoclaw-v2-<FOLDER>" --format "{{.Names}}" | head -1)
+docker inspect "$CTR" --format '{{json .HostConfig.ExtraHosts}}'
+docker exec "$CTR" env | grep ANTHROPIC
+```
+
+Expected: `api.anthropic.com:0.0.0.0` in ExtraHosts, `ANTHROPIC_BASE_URL=http://host.docker.internal:11434`.
+
+## Reverting to Claude
+
+To switch back to the Anthropic API:
+
+1. Remove the `env` and `blockedHosts` keys from `groups/<FOLDER>/container.json`
+2. Remove `"model"` from the shared settings file
+3. Restart the service
+
+No rebuild needed — both files are read at container spawn time.
+
+## Troubleshooting
+
+**Agent hangs, no response:** Ollama may be loading the model cold (large models take 10–30s).
+Watch `curl -s http://localhost:11434/api/ps` — the model appears once loaded.
+
+**"model not found" error in container logs:** The model name in settings.json doesn't match
+what Ollama has. Run `ollama list` on the host and use the exact name shown.
+
+**Responses claim to be Claude:** The model was trained on data that includes Claude conversations.
+Add a line to `groups/<FOLDER>/CLAUDE.md` telling it what model it runs on.
+
+**Agent responds but Ollama shows no activity:** `NO_PROXY` may not have taken effect for
+`http_proxy` (lowercase). Add both `NO_PROXY` and `no_proxy` to the env block.
diff --git a/.claude/skills/add-opencode/SKILL.md b/.claude/skills/add-opencode/SKILL.md
index 1dd31df..555f0fe 100644
--- a/.claude/skills/add-opencode/SKILL.md
+++ b/.claude/skills/add-opencode/SKILL.md
@@ -60,10 +60,10 @@ import './opencode.js';
 
 ### 4. Add the agent-runner dependency
 
-Pinned. Bump deliberately, not with `bun update`.
+Pinned. Bump deliberately, not with `bun update`. Use `1.4.17` — must match the `opencode-ai` CLI version pinned in step 5. The 1.14.x SDK has a completely different API and is **incompatible** with the current provider code.
 
 ```bash
-cd container/agent-runner && bun add @opencode-ai/sdk@1.4.3 && cd -
+cd container/agent-runner && bun add @opencode-ai/sdk@1.4.17 && cd -
 ```
 
 ### 5. Add `opencode-ai` to the container Dockerfile
@@ -73,9 +73,11 @@ Two edits to `container/Dockerfile`, both idempotent (skip if already present):
 **(a)** In the "Pin CLI versions" ARG block (around line 18), add after `ARG VERCEL_VERSION=latest`:
 
 ```dockerfile
-ARG OPENCODE_VERSION=latest
+ARG OPENCODE_VERSION=1.4.17
 ```
 
+> **Do not use `latest`** — the CLI and SDK must be the same version. `latest` silently upgrades the CLI to 1.14.x which has a breaking session API change (UUID session IDs → `ses_` prefix) incompatible with SDK 1.4.x.
+
 **(b)** In the `pnpm install -g` block (around line 80), append `"opencode-ai@${OPENCODE_VERSION}"` to the list:
 
 ```dockerfile
@@ -94,6 +96,25 @@ pnpm exec tsc -p container/agent-runner/tsconfig.json --noEmit   # container typ
 ./container/build.sh                                   # agent image
 ```
 
+> **Build cache gotcha:** The container buildkit caches COPY steps aggressively. If provider files were already present in the build context before, the new files may not be picked up. If you see "Unknown provider: opencode" after the build, prune the builder and rebuild:
+> ```bash
+> docker builder prune -f && ./container/build.sh
+> ```
+
+### 7. Propagate to existing per-group overlays
+
+Each agent group has a live source overlay at `data/v2-sessions/<group-id>/agent-runner-src/providers/` that **overrides the image at runtime**. This overlay is created when the group is first wired and never auto-updated by image rebuilds. Any group that already existed before this skill ran needs the new files copied in manually.
+
+```bash
+for overlay in data/v2-sessions/*/agent-runner-src/providers/; do
+  [ -d "$overlay" ] || continue
+  cp container/agent-runner/src/providers/opencode.ts "$overlay"
+  cp container/agent-runner/src/providers/mcp-to-opencode.ts "$overlay"
+  cp container/agent-runner/src/providers/index.ts "$overlay"
+  echo "Updated: $overlay"
+done
+```
+
 ## Configuration
 
 ### Host `.env` (typical)
@@ -102,35 +123,62 @@ Set model/provider strings in the form OpenCode expects (often `provider/model-i
 
 These variables are read **on the host** and passed into the container only when the effective provider is `opencode`. They do not switch the provider by themselves; the DB still needs `agent_provider` set (below).
 
-- `OPENCODE_PROVIDER` — OpenCode provider id, e.g. `openrouter`, `anthropic` (if unset, the runner defaults to `anthropic`).
-- `OPENCODE_MODEL` — full model id, e.g. `openrouter/anthropic/claude-sonnet-4`.
-- `OPENCODE_SMALL_MODEL` — optional second model for "small" tasks.
+- `OPENCODE_PROVIDER` — OpenCode provider id, e.g. `openrouter`, `anthropic`, `deepseek`.
+- `OPENCODE_MODEL` — full model id in `provider/model` form, e.g. `deepseek/deepseek-chat`.
+- `OPENCODE_SMALL_MODEL` — optional second model for lighter tasks; defaults to `OPENCODE_MODEL` if unset.
+- `ANTHROPIC_BASE_URL` — **required for non-`anthropic` providers.** The opencode container provider passes this as the `baseURL` for the upstream provider config so requests route through OneCLI's credential proxy or directly to the provider's API. Set it to the provider's API base URL (e.g. `https://api.deepseek.com/v1`, `https://openrouter.ai/api/v1`).
 
-Credentials: OneCLI / credential proxy patterns are unchanged. For non-`anthropic` OpenCode providers, the runner registers a placeholder API key and **`ANTHROPIC_BASE_URL`** (the credential proxy) as `baseURL` so the real key never lives in the container.
+Credentials: register provider API keys in OneCLI with the matching `--host-pattern` (e.g. `api.deepseek.com`, `openrouter.ai`). OneCLI injects them via `HTTPS_PROXY` in the container — the key never lives in `.env` or the container environment.
+
+After adding a secret, **grant the agent access** — agents in `selective` mode only receive secrets they've been explicitly assigned:
+
+```bash
+# Find the agent id and secret id, then:
+onecli agents set-secrets --id <agent-id> --secret-ids <existing-ids>,<new-secret-id>
+```
+
+Always include existing secret IDs in the list — `set-secrets` replaces, not appends.
+
+#### Example: DeepSeek
+
+```env
+OPENCODE_PROVIDER=deepseek
+OPENCODE_MODEL=deepseek/deepseek-chat
+OPENCODE_SMALL_MODEL=deepseek/deepseek-chat
+ANTHROPIC_BASE_URL=https://api.deepseek.com/v1
+```
+
+Register the key:
+```bash
+onecli secrets create --name "DeepSeek" --type generic \
+  --value YOUR_KEY --host-pattern "api.deepseek.com" \
+  --header-name "Authorization" --value-format "Bearer {value}"
+```
 
 #### Example: OpenRouter
 
 ```env
-# OpenCode — host passes these into the container when agent_provider is opencode
 OPENCODE_PROVIDER=openrouter
 OPENCODE_MODEL=openrouter/anthropic/claude-sonnet-4
 OPENCODE_SMALL_MODEL=openrouter/anthropic/claude-haiku-4.5
+ANTHROPIC_BASE_URL=https://openrouter.ai/api/v1
 ```
 
-#### Example: Anthropic via existing proxy env
+Register the key:
+```bash
+onecli secrets create --name "OpenRouter" --type generic \
+  --value YOUR_KEY --host-pattern "openrouter.ai" \
+  --header-name "Authorization" --value-format "Bearer {value}"
+```
 
-When `OPENCODE_PROVIDER` is `anthropic`, OpenCode uses normal Anthropic env inside the container (proxy + placeholder key pattern unchanged).
+#### Example: Anthropic (no ANTHROPIC_BASE_URL needed)
+
+When `OPENCODE_PROVIDER` is `anthropic`, OpenCode uses normal Anthropic env inside the container — the proxy + placeholder key pattern is unchanged and `ANTHROPIC_BASE_URL` is not required.
 
 ```env
 OPENCODE_PROVIDER=anthropic
 OPENCODE_MODEL=anthropic/claude-sonnet-4-20250514
-```
-
-#### Example: only a main model
-
-```env
-OPENCODE_PROVIDER=openrouter
-OPENCODE_MODEL=openrouter/google/gemini-2.5-pro-preview
+OPENCODE_SMALL_MODEL=anthropic/claude-haiku-4-5-20251001
 ```
 
 #### OpenCode Zen (`x-api-key`, not Bearer)
@@ -142,13 +190,9 @@ Zen's HTTP API (e.g. `POST …/zen/v1/messages`) expects the key in the **`x-api
 **Host `.env` (typical Zen shape):**
 
 ```env
-# NanoClaw still resolves AGENT_PROVIDER from agent_groups / sessions; set agent_provider to opencode there.
-# OpenCode SDK: Zen as the upstream provider + models under opencode/…
 OPENCODE_PROVIDER=opencode
 OPENCODE_MODEL=opencode/big-pickle
 OPENCODE_SMALL_MODEL=opencode/big-pickle
-
-# Point the credential proxy at Zen's Anthropic-compatible base URL (host + OneCLI must forward this host).
 ANTHROPIC_BASE_URL=https://opencode.ai/zen/v1
 ```
 
@@ -162,18 +206,16 @@ onecli secrets create --name "OpenCode Zen" --type generic \
   --header-name "x-api-key" --value-format "{value}"
 ```
 
-For comparison, OpenRouter uses `Authorization` + `Bearer {value}`. Zen is different by design.
-
 ### Per group / per session
 
-Schema: **`agent_groups.agent_provider`** and **`sessions.agent_provider`**. Set to `opencode` for groups or sessions that should use OpenCode. The container receives `AGENT_PROVIDER` from the resolved value (session overrides group).
+Set `"provider": "opencode"` in the group's **`container.json`** (`groups/<folder>/container.json`) — the in-container runner reads `provider` from there, not from the DB. The DB columns **`agent_groups.agent_provider`** and **`sessions.agent_provider`** (session overrides group) only drive host-side provider contribution — per-session XDG mount, `OPENCODE_*` env passthrough — and do not propagate into `container.json` at spawn time. Set both, or just edit `container.json`; if they disagree, the runner uses `container.json` and the host-side resolver falls back through session → group → `container.json` → `'claude'`.
 
 Extra MCP servers still come from **`NANOCLAW_MCP_SERVERS`** / `container_config.mcpServers` on the host; the runner merges them into the same `mcpServers` object passed to **both** Claude and OpenCode providers.
 
 ## Operational notes
 
 - OpenCode keeps a local **`opencode serve`** process and SSE subscription; the provider tears down with **`stream.return`** and **SIGKILL** on the server process on **`abort()`** / shared runtime reset to avoid MCP/zombie hangs.
-- Session continuation is opaque (`ses_*` ids); stale sessions are cleared using **`isSessionInvalid`** on OpenCode-specific errors (timeouts, connection resets, not-found patterns) in addition to the poll-loop's existing recovery.
+- Session continuation uses UUID format (SDK 1.4.x / CLI 1.4.x). Stale sessions are cleared by `isSessionInvalid` on OpenCode-specific error patterns. If you see UUID-related errors after an accidental CLI upgrade, clear `session_state` in `outbound.db` and wipe the `opencode-xdg` directory under the session folder.
 - **`NO_PROXY`** for localhost matters when the OpenCode client talks to `127.0.0.1` inside the container while HTTP(S)_PROXY is set (e.g. OneCLI).
 
 ## Verify
diff --git a/.claude/skills/add-pdf-reader/SKILL.md b/.claude/skills/add-pdf-reader/SKILL.md
deleted file mode 100644
index aecc347..0000000
--- a/.claude/skills/add-pdf-reader/SKILL.md
+++ /dev/null
@@ -1,104 +0,0 @@
----
-name: add-pdf-reader
-description: Add PDF reading to NanoClaw agents. Extracts text from PDFs via pdftotext CLI. Handles WhatsApp attachments, URLs, and local files.
----
-
-# Add PDF Reader
-
-Adds PDF reading capability to all container agents using poppler-utils (pdftotext/pdfinfo). PDFs sent as WhatsApp attachments are auto-downloaded to the group workspace.
-
-## Phase 1: Pre-flight
-
-1. Check if `container/skills/pdf-reader/pdf-reader` exists — skip to Phase 3 if already applied
-2. Confirm WhatsApp is installed first (`skill/whatsapp` merged). This skill modifies WhatsApp channel files.
-
-## Phase 2: Apply Code Changes
-
-### Ensure WhatsApp fork remote
-
-```bash
-git remote -v
-```
-
-If `whatsapp` is missing, add it:
-
-```bash
-git remote add whatsapp https://github.com/qwibitai/nanoclaw-whatsapp.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch whatsapp skill/pdf-reader
-git merge whatsapp/skill/pdf-reader || {
-  git checkout --theirs pnpm-lock.yaml
-  git add pnpm-lock.yaml
-  git merge --continue
-}
-```
-
-This merges in:
-- `container/skills/pdf-reader/SKILL.md` (agent-facing documentation)
-- `container/skills/pdf-reader/pdf-reader` (CLI script)
-- `poppler-utils` in `container/Dockerfile`
-- PDF attachment download in `src/channels/whatsapp.ts`
-- PDF tests in `src/channels/whatsapp.test.ts`
-
-If the merge reports conflicts, resolve them by reading the conflicted files and understanding the intent of both sides.
-
-### Validate
-
-```bash
-pnpm run build
-pnpm exec vitest run src/channels/whatsapp.test.ts
-```
-
-### Rebuild container
-
-```bash
-./container/build.sh
-```
-
-### Restart service
-
-```bash
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw  # macOS
-# Linux: systemctl --user restart nanoclaw
-```
-
-## Phase 3: Verify
-
-### Test PDF extraction
-
-Send a PDF file in any registered WhatsApp chat. The agent should:
-1. Download the PDF to `attachments/`
-2. Respond acknowledging the PDF
-3. Be able to extract text when asked
-
-### Test URL fetching
-
-Ask the agent to read a PDF from a URL. It should use `pdf-reader fetch <url>`.
-
-### Check logs if needed
-
-```bash
-tail -f logs/nanoclaw.log | grep -i pdf
-```
-
-Look for:
-- `Downloaded PDF attachment` — successful download
-- `Failed to download PDF attachment` — media download issue
-
-## Troubleshooting
-
-### Agent says pdf-reader command not found
-
-Container needs rebuilding. Run `./container/build.sh` and restart the service.
-
-### PDF text extraction is empty
-
-The PDF may be scanned (image-based). pdftotext only handles text-based PDFs. Consider using the agent-browser to open the PDF visually instead.
-
-### WhatsApp PDF not detected
-
-Verify the message has `documentMessage` with `mimetype: application/pdf`. Some file-sharing apps send PDFs as generic files without the correct mimetype.
diff --git a/.claude/skills/add-reactions/SKILL.md b/.claude/skills/add-reactions/SKILL.md
deleted file mode 100644
index 435bef9..0000000
--- a/.claude/skills/add-reactions/SKILL.md
+++ /dev/null
@@ -1,117 +0,0 @@
----
-name: add-reactions
-description: Add WhatsApp emoji reaction support — receive, send, store, and search reactions.
----
-
-# Add Reactions
-
-This skill adds emoji reaction support to NanoClaw's WhatsApp channel: receive and store reactions, send reactions from the container agent via MCP tool, and query reaction history from SQLite.
-
-## Phase 1: Pre-flight
-
-### Check if already applied
-
-Check if `src/status-tracker.ts` exists:
-
-```bash
-test -f src/status-tracker.ts && echo "Already applied" || echo "Not applied"
-```
-
-If already applied, skip to Phase 3 (Verify).
-
-## Phase 2: Apply Code Changes
-
-### Ensure WhatsApp fork remote
-
-```bash
-git remote -v
-```
-
-If `whatsapp` is missing, add it:
-
-```bash
-git remote add whatsapp https://github.com/qwibitai/nanoclaw-whatsapp.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch whatsapp skill/reactions
-git merge whatsapp/skill/reactions || {
-  git checkout --theirs pnpm-lock.yaml
-  git add pnpm-lock.yaml
-  git merge --continue
-}
-```
-
-This adds:
-- `scripts/migrate-reactions.ts` (database migration for `reactions` table with composite PK and indexes)
-- `src/status-tracker.ts` (forward-only emoji state machine for message lifecycle signaling, with persistence and retry)
-- `src/status-tracker.test.ts` (unit tests for StatusTracker)
-- `container/skills/reactions/SKILL.md` (agent-facing documentation for the `react_to_message` MCP tool)
-- Reaction support in `src/db.ts`, `src/channels/whatsapp.ts`, `src/types.ts`, `src/ipc.ts`, `src/index.ts`, `src/group-queue.ts`, and `container/agent-runner/src/ipc-mcp-stdio.ts`
-
-### Run database migration
-
-```bash
-pnpm exec tsx scripts/migrate-reactions.ts
-```
-
-### Validate code changes
-
-```bash
-pnpm test
-pnpm run build
-```
-
-All tests must pass and build must be clean before proceeding.
-
-## Phase 3: Verify
-
-### Build and restart
-
-```bash
-pnpm run build
-```
-
-Linux:
-```bash
-systemctl --user restart nanoclaw
-```
-
-macOS:
-```bash
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw
-```
-
-### Test receiving reactions
-
-1. Send a message from your phone
-2. React to it with an emoji on WhatsApp
-3. Check the database:
-
-```bash
-sqlite3 store/messages.db "SELECT * FROM reactions ORDER BY timestamp DESC LIMIT 5;"
-```
-
-### Test sending reactions
-
-Ask the agent to react to a message via the `react_to_message` MCP tool. Check your phone — the reaction should appear on the message.
-
-## Troubleshooting
-
-### Reactions not appearing in database
-
-- Check NanoClaw logs for `Failed to process reaction` errors
-- Verify the chat is registered
-- Confirm the service is running
-
-### Migration fails
-
-- Ensure `store/messages.db` exists and is accessible
-- If "table reactions already exists", the migration already ran — skip it
-
-### Agent can't send reactions
-
-- Check IPC logs for `Unauthorized IPC reaction attempt blocked` — the agent can only react in its own group's chat
-- Verify WhatsApp is connected: check logs for connection status
diff --git a/.claude/skills/add-signal/REMOVE.md b/.claude/skills/add-signal/REMOVE.md
new file mode 100644
index 0000000..db37ade
--- /dev/null
+++ b/.claude/skills/add-signal/REMOVE.md
@@ -0,0 +1,13 @@
+# Remove Signal
+
+1. Comment out `import './signal.js'` in `src/channels/index.ts`
+2. Remove `SIGNAL_ACCOUNT` (and any other `SIGNAL_*` vars) from `.env`
+3. Rebuild and restart
+
+If you also want to unlink the Signal account from `signal-cli`:
+
+```bash
+signal-cli -a +1YOURNUMBER removeDevice --deviceId <id>
+```
+
+(Find the device id with `signal-cli -a +1YOURNUMBER listDevices`.)
diff --git a/.claude/skills/add-signal/SKILL.md b/.claude/skills/add-signal/SKILL.md
new file mode 100644
index 0000000..e6d41aa
--- /dev/null
+++ b/.claude/skills/add-signal/SKILL.md
@@ -0,0 +1,148 @@
+---
+name: add-signal
+description: Add Signal channel integration via signal-cli TCP daemon. Native adapter — no Chat SDK bridge.
+---
+
+# Add Signal Channel
+
+Adds Signal messaging support via a native adapter that speaks JSON-RPC to a [signal-cli](https://github.com/AsamK/signal-cli) TCP daemon. No Chat SDK bridge, no npm deps — only Node.js builtins.
+
+## Prerequisites
+
+`signal-cli` installed and a Signal account linked:
+
+- macOS: `brew install signal-cli`
+- Linux: download from [GitHub releases](https://github.com/AsamK/signal-cli/releases)
+- Link your account: `signal-cli -a +1YOURNUMBER link` (follow the QR instructions)
+
+## Install
+
+NanoClaw doesn't ship channels in trunk. This skill copies the Signal adapter and its tests in from the `channels` branch.
+
+### Pre-flight (idempotent)
+
+Skip to **Credentials** if all of these are already in place:
+
+- `src/channels/signal.ts` and `src/channels/signal.test.ts` both exist
+- `src/channels/index.ts` contains `import './signal.js';`
+
+Otherwise continue. Every step below is safe to re-run.
+
+### 1. Fetch the channels branch
+
+```bash
+git fetch origin channels
+```
+
+### 2. Copy the adapter and tests
+
+```bash
+git show origin/channels:src/channels/signal.ts      > src/channels/signal.ts
+git show origin/channels:src/channels/signal.test.ts > src/channels/signal.test.ts
+```
+
+### 3. Append the self-registration import
+
+Append to `src/channels/index.ts` (skip if the line is already present):
+
+```typescript
+import './signal.js';
+```
+
+### 4. Build
+
+```bash
+pnpm run build
+```
+
+No npm packages to install — the adapter uses only Node.js builtins (`node:net`, `node:child_process`, `node:fs`).
+
+## Credentials
+
+Add to `.env`:
+
+```bash
+SIGNAL_ACCOUNT=+1YOURNUMBER
+```
+
+### Optional settings
+
+```bash
+# TCP daemon host and port (default: 127.0.0.1:7583)
+SIGNAL_TCP_HOST=127.0.0.1
+SIGNAL_TCP_PORT=7583
+
+# Path to the signal-cli binary (default: resolved on PATH)
+SIGNAL_CLI_PATH=/usr/local/bin/signal-cli
+
+# Whether NanoClaw manages the daemon lifecycle (default: true).
+# Set to false if you run signal-cli daemon externally.
+SIGNAL_MANAGE_DAEMON=true
+
+# signal-cli data directory (default: ~/.local/share/signal-cli)
+SIGNAL_DATA_DIR=~/.local/share/signal-cli
+```
+
+**Security note:** keep the TCP host on `127.0.0.1`. The daemon has no auth — binding it to a public interface would expose your full Signal account to the network.
+
+Sync to container: `mkdir -p data/env && cp .env data/env/env`
+
+### Restart
+
+```bash
+# macOS
+launchctl kickstart -k gui/$(id -u)/com.nanoclaw
+
+# Linux
+systemctl --user restart nanoclaw
+```
+
+## Next Steps
+
+If you're in the middle of `/setup`, return to the setup flow now.
+
+Otherwise, run `/init-first-agent` to create an agent and wire it to your Signal DM, or `/manage-channels` to wire this channel to an existing agent group. Signal is direct-addressable — your phone number is the platform ID.
+
+## Channel Info
+
+- **type**: `signal`
+- **terminology**: Signal has "chats" (1:1 DMs) and "groups."
+- **how-to-find-id**: DMs use your phone number (e.g. `+15555550123`). Groups use `group:<groupId>` — find group IDs via `signal-cli -a +1YOURNUMBER listGroups`.
+- **supports-threads**: no
+- **typical-use**: Personal assistant via Signal DMs or small group chats
+- **default-isolation**: One agent per Signal account. Multiple chats with the same operator can share an agent group; groups with other people should typically be separate.
+
+### Features
+
+- Markdown formatting — `**bold**`, `*italic*` / `_italic_`, `` `code` ``, ` ```code fence``` `, `~~strike~~`, `||spoiler||` (converted to Signal's offset-based text styles)
+- Quoted replies — `replyTo*` fields populated from Signal quotes
+- Typing indicators — DMs only (Signal doesn't support group typing)
+- Echo suppression — outbound messages are matched on `(platformId, text)` within a 10 s TTL to avoid syncMessage loops
+- Note to Self — messages you send to your own account from another device route to the agent as inbound with `isFromMe: true`
+- Voice attachments — detected but not transcribed by default; the agent receives `[Voice Message]` placeholder text. Run `/add-voice-transcription` for local transcription via parakeet-mlx
+
+Not supported yet: outbound file attachments (logged and dropped), edit/delete messages, reactions.
+
+## Troubleshooting
+
+### Daemon not reachable
+
+```bash
+grep "Signal" logs/nanoclaw.log | tail
+```
+
+If you see `Signal daemon failed to start. Is signal-cli installed and your account linked?`:
+- Confirm `signal-cli` is on PATH (or set `SIGNAL_CLI_PATH`)
+- Confirm the account is linked: `signal-cli -a +1YOURNUMBER listIdentities` should succeed without prompting
+
+If you see `Signal daemon not reachable at 127.0.0.1:7583` and `SIGNAL_MANAGE_DAEMON=false`, start the daemon yourself: `signal-cli -a +1YOURNUMBER daemon --tcp 127.0.0.1:7583`.
+
+### Bot not responding
+
+1. Channel initialized: `grep "Signal channel connected" logs/nanoclaw.log | tail -1`
+2. Channel wired: `sqlite3 data/v2.db "SELECT mg.platform_id, mg.name FROM messaging_groups mg JOIN messaging_group_agents mga ON mg.id = mga.messaging_group_id WHERE mg.channel_type='signal'"`
+3. Service running: `launchctl print gui/$(id -u)/com.nanoclaw` (macOS) / `systemctl --user status nanoclaw` (Linux)
+
+### Lost connection mid-session
+
+If you see `Signal channel lost TCP connection to signal-cli daemon` in the logs, the daemon dropped us. There's no auto-reconnect yet — restart the service to re-establish.
diff --git a/.claude/skills/add-signal/VERIFY.md b/.claude/skills/add-signal/VERIFY.md
new file mode 100644
index 0000000..b1ae851
--- /dev/null
+++ b/.claude/skills/add-signal/VERIFY.md
@@ -0,0 +1,5 @@
+# Verify Signal
+
+Send a message to your own Signal number (Note to Self) from another device, or have someone send your linked number a DM. The bot should respond within a few seconds.
+
+If nothing happens, tail `logs/nanoclaw.log` for `Signal channel connected` and `Signal message received`.
diff --git a/.claude/skills/add-telegram-swarm/SKILL.md b/.claude/skills/add-telegram-swarm/SKILL.md
deleted file mode 100644
index 8f6a4fc..0000000
--- a/.claude/skills/add-telegram-swarm/SKILL.md
+++ /dev/null
@@ -1,384 +0,0 @@
----
-name: add-telegram-swarm
-description: Add Agent Swarm (Teams) support to Telegram. Each subagent gets its own bot identity in the group. Requires Telegram channel to be set up first (use /add-telegram). Triggers on "agent swarm", "agent teams telegram", "telegram swarm", "bot pool".
----
-
-# Add Agent Swarm to Telegram
-
-This skill adds Agent Teams (Swarm) support to an existing Telegram channel. Each subagent in a team gets its own bot identity in the Telegram group, so users can visually distinguish which agent is speaking.
-
-**Prerequisite**: Telegram must already be set up via the `/add-telegram` skill. If `src/telegram.ts` does not exist or `TELEGRAM_BOT_TOKEN` is not configured, tell the user to run `/add-telegram` first.
-
-## How It Works
-
-- The **main bot** receives messages and sends lead agent responses (already set up by `/add-telegram`)
-- **Pool bots** are send-only — each gets a Grammy `Api` instance (no polling)
-- When a subagent calls `send_message` with a `sender` parameter, the host assigns a pool bot and renames it to match the sender's role
-- Messages appear in Telegram from different bot identities
-
-```
-Subagent calls send_message(text: "Found 3 results", sender: "Researcher")
-  → MCP writes IPC file with sender field
-  → Host IPC watcher picks it up
-  → Assigns pool bot #2 to "Researcher" (round-robin, stable per-group)
-  → Renames pool bot #2 to "Researcher" via setMyName
-  → Sends message via pool bot #2's Api instance
-  → Appears in Telegram from "Researcher" bot
-```
-
-## Prerequisites
-
-### 1. Create Pool Bots
-
-Tell the user:
-
-> I need you to create 3-5 Telegram bots to use as the agent pool. These will be renamed dynamically to match agent roles.
->
-> 1. Open Telegram and search for `@BotFather`
-> 2. Send `/newbot` for each bot:
->    - Give them any placeholder name (e.g., "Bot 1", "Bot 2")
->    - Usernames like `myproject_swarm_1_bot`, `myproject_swarm_2_bot`, etc.
-> 3. Copy all the tokens
-> 4. Add all bots to your Telegram group(s) where you want agent teams
-
-Wait for user to provide the tokens.
-
-### 2. Disable Group Privacy for Pool Bots
-
-Tell the user:
-
-> **Important**: Each pool bot needs Group Privacy disabled so it can send messages in groups.
->
-> For each pool bot in `@BotFather`:
-> 1. Send `/mybots` and select the bot
-> 2. Go to **Bot Settings** > **Group Privacy** > **Turn off**
->
-> Then add all pool bots to your Telegram group(s).
-
-## Implementation
-
-### Step 1: Update Configuration
-
-Read `src/config.ts` and add the bot pool config near the other Telegram exports:
-
-```typescript
-export const TELEGRAM_BOT_POOL = (process.env.TELEGRAM_BOT_POOL || '')
-  .split(',')
-  .map((t) => t.trim())
-  .filter(Boolean);
-```
-
-### Step 2: Add Bot Pool to Telegram Module
-
-Read `src/telegram.ts` and add the following:
-
-1. **Update imports** — add `Api` to the Grammy import:
-
-```typescript
-import { Api, Bot } from 'grammy';
-```
-
-2. **Add pool state** after the existing `let bot` declaration:
-
-```typescript
-// Bot pool for agent teams: send-only Api instances (no polling)
-const poolApis: Api[] = [];
-// Maps "{groupFolder}:{senderName}" → pool Api index for stable assignment
-const senderBotMap = new Map<string, number>();
-let nextPoolIndex = 0;
-```
-
-3. **Add pool functions** — place these before the `isTelegramConnected` function:
-
-```typescript
-/**
- * Initialize send-only Api instances for the bot pool.
- * Each pool bot can send messages but doesn't poll for updates.
- */
-export async function initBotPool(tokens: string[]): Promise<void> {
-  for (const token of tokens) {
-    try {
-      const api = new Api(token);
-      const me = await api.getMe();
-      poolApis.push(api);
-      logger.info(
-        { username: me.username, id: me.id, poolSize: poolApis.length },
-        'Pool bot initialized',
-      );
-    } catch (err) {
-      logger.error({ err }, 'Failed to initialize pool bot');
-    }
-  }
-  if (poolApis.length > 0) {
-    logger.info({ count: poolApis.length }, 'Telegram bot pool ready');
-  }
-}
-
-/**
- * Send a message via a pool bot assigned to the given sender name.
- * Assigns bots round-robin on first use; subsequent messages from the
- * same sender in the same group always use the same bot.
- * On first assignment, renames the bot to match the sender's role.
- */
-export async function sendPoolMessage(
-  chatId: string,
-  text: string,
-  sender: string,
-  groupFolder: string,
-): Promise<void> {
-  if (poolApis.length === 0) {
-    // No pool bots — fall back to main bot
-    await sendTelegramMessage(chatId, text);
-    return;
-  }
-
-  const key = `${groupFolder}:${sender}`;
-  let idx = senderBotMap.get(key);
-  if (idx === undefined) {
-    idx = nextPoolIndex % poolApis.length;
-    nextPoolIndex++;
-    senderBotMap.set(key, idx);
-    // Rename the bot to match the sender's role, then wait for Telegram to propagate
-    try {
-      await poolApis[idx].setMyName(sender);
-      await new Promise((r) => setTimeout(r, 2000));
-      logger.info({ sender, groupFolder, poolIndex: idx }, 'Assigned and renamed pool bot');
-    } catch (err) {
-      logger.warn({ sender, err }, 'Failed to rename pool bot (sending anyway)');
-    }
-  }
-
-  const api = poolApis[idx];
-  try {
-    const numericId = chatId.replace(/^tg:/, '');
-    const MAX_LENGTH = 4096;
-    if (text.length <= MAX_LENGTH) {
-      await api.sendMessage(numericId, text);
-    } else {
-      for (let i = 0; i < text.length; i += MAX_LENGTH) {
-        await api.sendMessage(numericId, text.slice(i, i + MAX_LENGTH));
-      }
-    }
-    logger.info({ chatId, sender, poolIndex: idx, length: text.length }, 'Pool message sent');
-  } catch (err) {
-    logger.error({ chatId, sender, err }, 'Failed to send pool message');
-  }
-}
-```
-
-### Step 3: Add sender Parameter to MCP Tool
-
-Read `container/agent-runner/src/ipc-mcp-stdio.ts` and update the `send_message` tool to accept an optional `sender` parameter:
-
-Change the tool's schema from:
-```typescript
-{ text: z.string().describe('The message text to send') },
-```
-
-To:
-```typescript
-{
-  text: z.string().describe('The message text to send'),
-  sender: z.string().optional().describe('Your role/identity name (e.g. "Researcher"). When set, messages appear from a dedicated bot in Telegram.'),
-},
-```
-
-And update the handler to include `sender` in the IPC data:
-
-```typescript
-async (args) => {
-    const data: Record<string, string | undefined> = {
-      type: 'message',
-      chatJid,
-      text: args.text,
-      sender: args.sender || undefined,
-      groupFolder,
-      timestamp: new Date().toISOString(),
-    };
-
-    writeIpcFile(MESSAGES_DIR, data);
-
-    return { content: [{ type: 'text' as const, text: 'Message sent.' }] };
-  },
-```
-
-### Step 4: Update Host IPC Routing
-
-Read `src/ipc.ts` and make these changes:
-
-1. **Add imports** — add `sendPoolMessage` and `initBotPool` from the Telegram swarm module, and `TELEGRAM_BOT_POOL` from config.
-
-2. **Update IPC message routing** — in `src/ipc.ts`, find where the `sendMessage` dependency is called to deliver IPC messages (inside `processIpcFiles`). The `sendMessage` is passed in via the `IpcDeps` parameter. Wrap it to route Telegram swarm messages through the bot pool:
-
-```typescript
-if (data.sender && data.chatJid.startsWith('tg:')) {
-  await sendPoolMessage(
-    data.chatJid,
-    data.text,
-    data.sender,
-    sourceGroup,
-  );
-} else {
-  await deps.sendMessage(data.chatJid, data.text);
-}
-```
-
-Note: The assistant name prefix is handled by `formatOutbound()` in the router — Telegram channels have `prefixAssistantName = false` so no prefix is added for `tg:` JIDs.
-
-3. **Initialize pool in `main()` in `src/index.ts`** — after creating the Telegram channel, add:
-
-```typescript
-if (TELEGRAM_BOT_POOL.length > 0) {
-  await initBotPool(TELEGRAM_BOT_POOL);
-}
-```
-
-### Step 5: Update CLAUDE.md Files
-
-#### 5a. Add global message formatting rules
-
-Read `groups/global/CLAUDE.md` and add a Message Formatting section:
-
-```markdown
-## Message Formatting
-
-NEVER use markdown. Only use WhatsApp/Telegram formatting:
-- *single asterisks* for bold (NEVER **double asterisks**)
-- _underscores_ for italic
-- • bullet points
-- ```triple backticks``` for code
-
-No ## headings. No [links](url). No **double stars**.
-```
-
-#### 5b. Update existing group CLAUDE.md headings
-
-In any group CLAUDE.md that has a "WhatsApp Formatting" section (e.g. `groups/main/CLAUDE.md`), rename the heading to reflect multi-channel support:
-
-```
-## WhatsApp Formatting (and other messaging apps)
-```
-
-#### 5c. Add Agent Teams instructions to Telegram groups
-
-For each Telegram group that will use agent teams, create or update its `groups/{folder}/CLAUDE.md` with these instructions. Read the existing CLAUDE.md first (or `groups/global/CLAUDE.md` as a base) and add the Agent Teams section:
-
-```markdown
-## Agent Teams
-
-When creating a team to tackle a complex task, follow these rules:
-
-### CRITICAL: Follow the user's prompt exactly
-
-Create *exactly* the team the user asked for — same number of agents, same roles, same names. Do NOT add extra agents, rename roles, or use generic names like "Researcher 1". If the user says "a marine biologist, a physicist, and Alexander Hamilton", create exactly those three agents with those exact names.
-
-### Team member instructions
-
-Each team member MUST be instructed to:
-
-1. *Share progress in the group* via `mcp__nanoclaw__send_message` with a `sender` parameter matching their exact role/character name (e.g., `sender: "Marine Biologist"` or `sender: "Alexander Hamilton"`). This makes their messages appear from a dedicated bot in the Telegram group.
-2. *Also communicate with teammates* via `SendMessage` as normal for coordination.
-3. Keep group messages *short* — 2-4 sentences max per message. Break longer content into multiple `send_message` calls. No walls of text.
-4. Use the `sender` parameter consistently — always the same name so the bot identity stays stable.
-5. NEVER use markdown formatting. Use ONLY WhatsApp/Telegram formatting: single *asterisks* for bold (NOT **double**), _underscores_ for italic, • for bullets, ```backticks``` for code. No ## headings, no [links](url), no **double asterisks**.
-
-### Example team creation prompt
-
-When creating a teammate, include instructions like:
-
-\```
-You are the Marine Biologist. When you have findings or updates for the user, send them to the group using mcp__nanoclaw__send_message with sender set to "Marine Biologist". Keep each message short (2-4 sentences max). Use emojis for strong reactions. ONLY use single *asterisks* for bold (never **double**), _underscores_ for italic, • for bullets. No markdown. Also communicate with teammates via SendMessage.
-\```
-
-### Lead agent behavior
-
-As the lead agent who created the team:
-
-- You do NOT need to react to or relay every teammate message. The user sees those directly from the teammate bots.
-- Send your own messages only to comment, share thoughts, synthesize, or direct the team.
-- When processing an internal update from a teammate that doesn't need a user-facing response, wrap your *entire* output in `<internal>` tags.
-- Focus on high-level coordination and the final synthesis.
-```
-
-### Step 6: Update Environment
-
-Add pool tokens to `.env`:
-
-```bash
-TELEGRAM_BOT_POOL=TOKEN1,TOKEN2,TOKEN3,...
-```
-
-**Important**: Sync to all required locations:
-
-```bash
-cp .env data/env/env
-```
-
-Also add `TELEGRAM_BOT_POOL` to the launchd plist (`~/Library/LaunchAgents/com.nanoclaw.plist`) in the `EnvironmentVariables` dict if using launchd.
-
-### Step 7: Rebuild and Restart
-
-```bash
-pnpm run build
-./container/build.sh  # Required — MCP tool changed
-# macOS:
-launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist
-launchctl load ~/Library/LaunchAgents/com.nanoclaw.plist
-# Linux:
-# systemctl --user restart nanoclaw
-```
-
-Must use `unload/load` (macOS) or `restart` (Linux) because the service env vars changed.
-
-### Step 8: Test
-
-Tell the user:
-
-> Send a message in your Telegram group asking for a multi-agent task, e.g.:
-> "Assemble a team of a researcher and a coder to build me a hello world app"
->
-> You should see:
-> - The lead agent (main bot) acknowledging and creating the team
-> - Each subagent messaging from a different bot, renamed to their role
-> - Short, scannable messages from each agent
->
-> Check logs: `tail -f logs/nanoclaw.log | grep -i pool`
-
-## Architecture Notes
-
-- Pool bots use Grammy's `Api` class — lightweight, no polling, just send
-- Bot names are set via `setMyName` — changes are global to the bot, not per-chat
-- A 2-second delay after `setMyName` allows Telegram to propagate the name change before the first message
-- Sender→bot mapping is stable within a group (keyed as `{groupFolder}:{senderName}`)
-- Mapping resets on service restart — pool bots get reassigned fresh
-- If pool runs out, bots are reused (round-robin wraps)
-
-## Troubleshooting
-
-### Pool bots not sending messages
-
-1. Verify tokens: `curl -s "https://api.telegram.org/botTOKEN/getMe"`
-2. Check pool initialized: `grep "Pool bot" logs/nanoclaw.log`
-3. Ensure all pool bots are members of the Telegram group
-4. Check Group Privacy is disabled for each pool bot
-
-### Bot names not updating
-
-Telegram caches bot names client-side. The 2-second delay after `setMyName` helps, but users may need to restart their Telegram client to see updated names immediately.
-
-### Subagents not using send_message
-
-Check the group's `CLAUDE.md` has the Agent Teams instructions. The lead agent reads this when creating teammates and must include the `send_message` + `sender` instructions in each teammate's prompt.
-
-## Removal
-
-To remove Agent Swarm support while keeping basic Telegram:
-
-1. Remove `TELEGRAM_BOT_POOL` from `src/config.ts`
-2. Remove pool code from `src/telegram.ts` (`poolApis`, `senderBotMap`, `initBotPool`, `sendPoolMessage`)
-3. Remove pool routing from IPC handler in `src/index.ts` (revert to plain `sendMessage`)
-4. Remove `initBotPool` call from `main()`
-5. Remove `sender` param from MCP tool in `container/agent-runner/src/ipc-mcp-stdio.ts`
-6. Remove Agent Teams section from group CLAUDE.md files
-7. Remove `TELEGRAM_BOT_POOL` from `.env`, `data/env/env`, and launchd plist/systemd unit
-8. Rebuild: `pnpm run build && ./container/build.sh && launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist && launchctl load ~/Library/LaunchAgents/com.nanoclaw.plist` (macOS) or `pnpm run build && ./container/build.sh && systemctl --user restart nanoclaw` (Linux)
diff --git a/.claude/skills/add-voice-transcription/SKILL.md b/.claude/skills/add-voice-transcription/SKILL.md
deleted file mode 100644
index cae1e47..0000000
--- a/.claude/skills/add-voice-transcription/SKILL.md
+++ /dev/null
@@ -1,148 +0,0 @@
----
-name: add-voice-transcription
-description: Add voice message transcription to NanoClaw using OpenAI's Whisper API. Automatically transcribes WhatsApp voice notes so the agent can read and respond to them.
----
-
-# Add Voice Transcription
-
-This skill adds automatic voice message transcription to NanoClaw's WhatsApp channel using OpenAI's Whisper API. When a voice note arrives, it is downloaded, transcribed, and delivered to the agent as `[Voice: <transcript>]`.
-
-## Phase 1: Pre-flight
-
-### Check if already applied
-
-Check if `src/transcription.ts` exists. If it does, skip to Phase 3 (Configure). The code changes are already in place.
-
-### Ask the user
-
-Use `AskUserQuestion` to collect information:
-
-AskUserQuestion: Do you have an OpenAI API key for Whisper transcription?
-
-If yes, collect it now. If no, direct them to create one at https://platform.openai.com/api-keys.
-
-## Phase 2: Apply Code Changes
-
-**Prerequisite:** WhatsApp must be installed first (`skill/whatsapp` merged). This skill modifies WhatsApp channel files.
-
-### Ensure WhatsApp fork remote
-
-```bash
-git remote -v
-```
-
-If `whatsapp` is missing, add it:
-
-```bash
-git remote add whatsapp https://github.com/qwibitai/nanoclaw-whatsapp.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch whatsapp skill/voice-transcription
-git merge whatsapp/skill/voice-transcription || {
-  git checkout --theirs pnpm-lock.yaml
-  git add pnpm-lock.yaml
-  git merge --continue
-}
-```
-
-This merges in:
-- `src/transcription.ts` (voice transcription module using OpenAI Whisper)
-- Voice handling in `src/channels/whatsapp.ts` (isVoiceMessage check, transcribeAudioMessage call)
-- Transcription tests in `src/channels/whatsapp.test.ts`
-- `openai` npm dependency in `package.json`
-- `OPENAI_API_KEY` in `.env.example`
-
-If the merge reports conflicts, resolve them by reading the conflicted files and understanding the intent of both sides.
-
-### Validate code changes
-
-```bash
-pnpm install
-pnpm run build
-pnpm exec vitest run src/channels/whatsapp.test.ts
-```
-
-All tests must pass and build must be clean before proceeding.
-
-## Phase 3: Configure
-
-### Get OpenAI API key (if needed)
-
-If the user doesn't have an API key:
-
-> I need you to create an OpenAI API key:
->
-> 1. Go to https://platform.openai.com/api-keys
-> 2. Click "Create new secret key"
-> 3. Give it a name (e.g., "NanoClaw Transcription")
-> 4. Copy the key (starts with `sk-`)
->
-> Cost: ~$0.006 per minute of audio (~$0.003 per typical 30-second voice note)
-
-Wait for the user to provide the key.
-
-### Add to environment
-
-Add to `.env`:
-
-```bash
-OPENAI_API_KEY=<their-key>
-```
-
-Sync to container environment:
-
-```bash
-mkdir -p data/env && cp .env data/env/env
-```
-
-The container reads environment from `data/env/env`, not `.env` directly.
-
-### Build and restart
-
-```bash
-pnpm run build
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw  # macOS
-# Linux: systemctl --user restart nanoclaw
-```
-
-## Phase 4: Verify
-
-### Test with a voice note
-
-Tell the user:
-
-> Send a voice note in any registered WhatsApp chat. The agent should receive it as `[Voice: <transcript>]` and respond to its content.
-
-### Check logs if needed
-
-```bash
-tail -f logs/nanoclaw.log | grep -i voice
-```
-
-Look for:
-- `Transcribed voice message` — successful transcription with character count
-- `OPENAI_API_KEY not set` — key missing from `.env`
-- `OpenAI transcription failed` — API error (check key validity, billing)
-- `Failed to download audio message` — media download issue
-
-## Troubleshooting
-
-### Voice notes show "[Voice Message - transcription unavailable]"
-
-1. Check `OPENAI_API_KEY` is set in `.env` AND synced to `data/env/env`
-2. Verify key works: `curl -s https://api.openai.com/v1/models -H "Authorization: Bearer $OPENAI_API_KEY" | head -c 200`
-3. Check OpenAI billing — Whisper requires a funded account
-
-### Voice notes show "[Voice Message - transcription failed]"
-
-Check logs for the specific error. Common causes:
-- Network timeout — transient, will work on next message
-- Invalid API key — regenerate at https://platform.openai.com/api-keys
-- Rate limiting — wait and retry
-
-### Agent doesn't respond to voice notes
-
-Verify the chat is registered and the agent is running. Voice transcription only runs for registered groups.
diff --git a/.claude/skills/add-wechat/REMOVE.md b/.claude/skills/add-wechat/REMOVE.md
new file mode 100644
index 0000000..366739e
--- /dev/null
+++ b/.claude/skills/add-wechat/REMOVE.md
@@ -0,0 +1,49 @@
+# Remove WeChat Channel
+
+Undo `/add-wechat`.
+
+### 1. Remove credentials
+
+Delete WeChat lines from `.env`:
+
+```bash
+sed -i.bak '/^WECHAT_ENABLED=/d' .env && rm -f .env.bak
+cp .env data/env/env
+```
+
+### 2. Remove adapter and import
+
+```bash
+rm -f src/channels/wechat.ts
+sed -i.bak "/import '\.\/wechat\.js';/d" src/channels/index.ts && rm -f src/channels/index.ts.bak
+```
+
+### 3. Uninstall the package
+
+```bash
+pnpm remove wechat-ilink-client
+```
+
+### 4. Remove saved auth + sync state
+
+```bash
+rm -rf data/wechat
+```
+
+### 5. Remove DB wiring
+
+```sql
+-- Remove any sessions first (foreign key)
+DELETE FROM sessions WHERE messaging_group_id IN (SELECT id FROM messaging_groups WHERE channel_type = 'wechat');
+DELETE FROM messaging_group_agents WHERE messaging_group_id IN (SELECT id FROM messaging_groups WHERE channel_type = 'wechat');
+DELETE FROM messaging_groups WHERE channel_type = 'wechat';
+```
+
+### 6. Rebuild and restart
+
+```bash
+pnpm run build
+systemctl --user restart nanoclaw   # Linux
+# or
+launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
+```
diff --git a/.claude/skills/add-wechat/SKILL.md b/.claude/skills/add-wechat/SKILL.md
new file mode 100644
index 0000000..49746ce
--- /dev/null
+++ b/.claude/skills/add-wechat/SKILL.md
@@ -0,0 +1,170 @@
+---
+name: add-wechat
+description: Add WeChat (personal) channel integration via Tencent's official iLink Bot API. Uses long-polling and QR scan — no webhook, no ToS risk, no paid token.
+---
+
+# Add WeChat Channel
+
+Adds WeChat support via **iLink Bot API** — the first-party Tencent API for personal WeChat bots (different from WeCom / Official Account).
+
+**Why this is different from wechaty/PadLocal:**
+
+- Official Tencent API — no ToS violation, no ban risk
+- Free — no PadLocal token required
+- No public webhook URL needed — uses long-poll
+- Works with any personal WeChat account
+
+## Prerequisites
+
+- A **personal WeChat account** with the mobile app installed
+- A phone to scan the QR code for login
+- Node.js >= 20 (already required by NanoClaw)
+
+## Install
+
+NanoClaw doesn't ship channels in trunk. This skill copies the WeChat adapter in from the `channels` branch.
+
+### Pre-flight (idempotent)
+
+Skip to **Credentials** if all of these are already in place:
+
+- `src/channels/wechat.ts` exists
+- `src/channels/index.ts` contains `import './wechat.js';`
+- `wechat-ilink-client` is listed in `package.json` dependencies
+
+Otherwise continue. Every step below is safe to re-run.
+
+### 1. Fetch the channels branch
+
+```bash
+git fetch origin channels
+```
+
+### 2. Copy the adapter
+
+```bash
+git show origin/channels:src/channels/wechat.ts > src/channels/wechat.ts
+```
+
+### 3. Append the self-registration import
+
+Append to `src/channels/index.ts` (skip if the line is already present):
+
+```typescript
+import './wechat.js';
+```
+
+### 4. Install the library (pinned)
+
+```bash
+pnpm install wechat-ilink-client@0.1.0
+```
+
+### 5. Build
+
+```bash
+pnpm run build
+```
+
+## Credentials
+
+Unlike most channels, WeChat requires **no pre-configured API keys**. Auth happens via QR code scan from your phone.
+
+### 1. Enable the channel
+
+Add to `.env`:
+
+```bash
+WECHAT_ENABLED=true
+```
+
+Sync to container: `mkdir -p data/env && cp .env data/env/env`
+
+### 2. Start the service and scan the QR
+
+Restart NanoClaw:
+
+```bash
+systemctl --user restart nanoclaw   # Linux
+# or
+launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
+```
+
+The adapter will print a **QR URL** to the logs and save it to `data/wechat/qr.txt`:
+
+```bash
+tail -f logs/nanoclaw.log | grep WeChat
+# or
+cat data/wechat/qr.txt
+```
+
+Open the URL in a browser (it renders a QR code), then:
+
+1. Open WeChat on your phone
+2. Use its built-in QR scanner (top-right "+" → Scan)
+3. Approve the authorization on your phone
+4. Auth credentials are saved to `data/wechat/auth.json` — do not commit this file
+
+The bot is now connected as your WeChat account.
+
+## Wire your first DM
+
+A successful QR login alone isn't enough — the adapter still needs to be wired to an agent group before it can respond.
+
+### 1. Trigger the first inbound message
+
+Have a different WeChat account send a message to the bot account. This auto-creates a `messaging_groups` row with the sender's `platform_id`.
+
+### 2. Run the wire script
+
+```bash
+pnpm exec tsx .claude/skills/add-wechat/scripts/wire-dm.ts
+```
+
+Interactive flow: the script lists all unwired WeChat messaging groups, asks which agent group to wire it to, and creates the `messaging_group_agents` row with sensible defaults (sender policy `request_approval`, session mode `shared`).
+
+With `request_approval`, the next DM from the stranger fires an approval card to the admin — admin taps Approve/Deny, approved users are added as members and their queued message replays through the agent.
+
+Non-interactive:
+
+```bash
+pnpm exec tsx .claude/skills/add-wechat/scripts/wire-dm.ts \
+  --platform-id wechat:wxid_xxxxx \
+  --agent-group ag-xxxxx \
+  --non-interactive
+```
+
+Flags:
+
+- `--platform-id <id>` — wire a specific messaging group (default: most recent unwired)
+- `--agent-group <id>` — target agent group (default: prompt; or solo admin group in non-interactive)
+- `--sender-policy public|strict|request_approval` — default `request_approval` (fires an admin approval card on unknown-sender DMs)
+- `--session-mode shared|per-thread` — default `shared`
+
+### 3. Test
+
+Have the sender message the bot again — the agent should respond.
+
+## Operational notes
+
+- **Only one instance can use a given token at a time.** Don't run multiple NanoClaw instances pointing to the same `data/wechat/auth.json`.
+- **Re-login on session expiry:** if you see `WeChat: session expired` in logs, delete `data/wechat/auth.json` and restart — you'll be asked to re-scan.
+- **Sync cursor persistence:** `data/wechat/sync-buf.txt` holds the long-poll cursor. Deleting it replays recent history on next start; don't delete it in normal operation.
+- **Account safety:** this uses the official Tencent API, so account bans for bot automation aren't a risk. That said, don't spam — normal rate limits still apply.
+
+## Next Steps
+
+If you're in the middle of `/setup`, return to the setup flow now.
+
+Otherwise, restart the service to pick up the new channel and wiring.
+
+## Channel Info
+
+- **type**: `wechat`
+- **terminology**: WeChat has "contacts" (DMs) and "group chats" (rooms). Each DM or group is a separate messaging group.
+- **how-to-find-id**: Send a message to the bot from the target account; the adapter auto-creates a messaging group and logs `WeChat inbound platformId=wechat:<id>`. Use `wechat:<user_id>` for DMs, `wechat:<group_id>` for rooms.
+- **admin-user-id**: The operator's WeChat user_id (for `init-first-agent.ts --admin-user-id`) is saved to `data/wechat/auth.json` as `operatorUserId` after the QR scan. Read it with `cat data/wechat/auth.json | jq -r .operatorUserId` and prefix with `wechat:` (i.e. `wechat:<operatorUserId>`).
+- **supports-threads**: no (WeChat has no reply threads)
+- **typical-use**: Long-poll — the adapter holds a persistent connection to Tencent's iLink API and receives messages in real time. No webhook URL needed.
+- **default-isolation**: `shared` session mode per messaging group (DM or room). Use `strict` sender policy if you want only specific users to reach the agent; `public` opens it to anyone who messages the bot.
+- **post-install-wiring**: Use the `wire-dm.ts` helper (see the "Wire your first DM" section above) if running this skill standalone. If running as part of `bash nanoclaw.sh`, `init-first-agent.ts` handles wiring — just pass the `platform-id` and `admin-user-id` captured above.
diff --git a/.claude/skills/add-wechat/scripts/wire-dm.ts b/.claude/skills/add-wechat/scripts/wire-dm.ts
new file mode 100644
index 0000000..f94c88d
--- /dev/null
+++ b/.claude/skills/add-wechat/scripts/wire-dm.ts
@@ -0,0 +1,172 @@
+#!/usr/bin/env pnpm exec tsx
+/**
+ * Wire a WeChat DM (or group) to an agent group.
+ *
+ * After /add-wechat installs the adapter and the user scans the QR login,
+ * the first inbound message from another WeChat account auto-creates a
+ * `messaging_groups` row. This script finds that row, asks the operator
+ * which agent group to wire it to, and inserts the `messaging_group_agents`
+ * join row with sensible defaults — the "post-login wiring" step /add-wechat
+ * otherwise requires manual SQL for.
+ *
+ * Usage:
+ *   pnpm exec tsx .claude/skills/add-wechat/scripts/wire-dm.ts
+ *
+ * Flags:
+ *   --platform-id <id>      Wire a specific messaging group (default: most recent unwired)
+ *   --agent-group <id>      Target agent group (default: interactive pick; or solo admin group)
+ *   --sender-policy <p>     public | strict (default: public)
+ *   --session-mode <m>      shared | per-thread (default: shared)
+ *   --non-interactive       Fail instead of prompting
+ */
+import Database from 'better-sqlite3';
+import path from 'node:path';
+import readline from 'node:readline';
+
+const DB_PATH = process.env.NANOCLAW_DB_PATH ?? path.join(process.cwd(), 'data', 'v2.db');
+
+type SenderPolicy = 'public' | 'strict' | 'request_approval';
+
+interface Args {
+  platformId?: string;
+  agentGroupId?: string;
+  senderPolicy: SenderPolicy;
+  sessionMode: 'shared' | 'per-thread';
+  interactive: boolean;
+}
+
+function parseArgs(argv: string[]): Args {
+  const args: Args = {
+    // Default matches the router's auto-create (`request_approval`) so the
+    // admin gets an approval card on the next unknown-sender DM rather than
+    // a silent allow. Pass `--sender-policy public` to open the channel to
+    // anyone, or `strict` to require explicit membership.
+    senderPolicy: 'request_approval',
+    sessionMode: 'shared',
+    interactive: true,
+  };
+  for (let i = 0; i < argv.length; i++) {
+    const flag = argv[i];
+    const val = argv[i + 1];
+    switch (flag) {
+      case '--platform-id': args.platformId = val; i++; break;
+      case '--agent-group': args.agentGroupId = val; i++; break;
+      case '--sender-policy':
+        if (val !== 'public' && val !== 'strict' && val !== 'request_approval') {
+          throw new Error(`bad --sender-policy: ${val} (use public | strict | request_approval)`);
+        }
+        args.senderPolicy = val; i++; break;
+      case '--session-mode':
+        if (val !== 'shared' && val !== 'per-thread') throw new Error(`bad --session-mode: ${val}`);
+        args.sessionMode = val; i++; break;
+      case '--non-interactive': args.interactive = false; break;
+      case '--help': case '-h':
+        console.log('See .claude/skills/add-wechat/scripts/wire-dm.ts header for usage.');
+        process.exit(0);
+    }
+  }
+  return args;
+}
+
+async function prompt(q: string): Promise<string> {
+  const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
+  return new Promise((resolve) => rl.question(q, (a) => { rl.close(); resolve(a.trim()); }));
+}
+
+function generateId(prefix: string): string {
+  return `${prefix}-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+}
+
+async function main(): Promise<void> {
+  const args = parseArgs(process.argv.slice(2));
+  const db = new Database(DB_PATH);
+  db.pragma('journal_mode = WAL');
+
+  // 1. Pick the messaging group
+  let platformId = args.platformId;
+  if (!platformId) {
+    const rows = db.prepare(`
+      SELECT mg.id, mg.platform_id, mg.name, mg.is_group, mg.created_at
+      FROM messaging_groups mg
+      LEFT JOIN messaging_group_agents mga ON mga.messaging_group_id = mg.id
+      WHERE mg.channel_type = 'wechat' AND mga.id IS NULL
+      ORDER BY mg.created_at DESC
+    `).all() as Array<{ id: string; platform_id: string; name: string | null; is_group: number; created_at: string }>;
+
+    if (rows.length === 0) {
+      console.error('No unwired WeChat messaging groups found.');
+      console.error('Send a message to the bot first (from another WeChat account), then re-run.');
+      process.exit(1);
+    }
+
+    if (rows.length === 1 || !args.interactive) {
+      platformId = rows[0].platform_id;
+      console.log(`Using most recent unwired group: ${platformId} (${rows[0].is_group ? 'group' : 'DM'})`);
+    } else {
+      console.log('Unwired WeChat messaging groups:');
+      rows.forEach((r, i) => {
+        console.log(`  ${i + 1}. ${r.platform_id}  (${r.is_group ? 'group' : 'DM'}, ${r.created_at})`);
+      });
+      const pick = await prompt('Pick one [1]: ');
+      const idx = pick === '' ? 0 : parseInt(pick, 10) - 1;
+      if (Number.isNaN(idx) || idx < 0 || idx >= rows.length) throw new Error('invalid choice');
+      platformId = rows[idx].platform_id;
+    }
+  }
+
+  const mg = db.prepare(
+    'SELECT id, platform_id, is_group FROM messaging_groups WHERE channel_type = ? AND platform_id = ?'
+  ).get('wechat', platformId) as { id: string; platform_id: string; is_group: number } | undefined;
+  if (!mg) throw new Error(`no wechat messaging_group with platform_id = ${platformId}`);
+
+  // 2. Pick the agent group
+  let agentGroupId = args.agentGroupId;
+  if (!agentGroupId) {
+    const agents = db.prepare('SELECT id, name, is_admin FROM agent_groups ORDER BY is_admin DESC, created_at ASC')
+      .all() as Array<{ id: string; name: string; is_admin: number }>;
+    if (agents.length === 0) throw new Error('no agent groups exist — create one first');
+
+    const adminAgents = agents.filter((a) => a.is_admin === 1);
+    if (adminAgents.length === 1 && !args.interactive) {
+      agentGroupId = adminAgents[0].id;
+      console.log(`Auto-selected sole admin agent group: ${adminAgents[0].name} (${agentGroupId})`);
+    } else if (args.interactive) {
+      console.log('Agent groups:');
+      agents.forEach((a, i) => {
+        console.log(`  ${i + 1}. ${a.name} (${a.id})${a.is_admin ? ' [admin]' : ''}`);
+      });
+      const pick = await prompt('Pick one [1]: ');
+      const idx = pick === '' ? 0 : parseInt(pick, 10) - 1;
+      if (Number.isNaN(idx) || idx < 0 || idx >= agents.length) throw new Error('invalid choice');
+      agentGroupId = agents[idx].id;
+    } else {
+      throw new Error('multiple agent groups exist; pass --agent-group <id>');
+    }
+  }
+
+  const ag = db.prepare('SELECT id, name FROM agent_groups WHERE id = ?').get(agentGroupId) as
+    { id: string; name: string } | undefined;
+  if (!ag) throw new Error(`no agent_group with id = ${agentGroupId}`);
+
+  // 3. Update sender policy + wire
+  const tx = db.transaction(() => {
+    db.prepare('UPDATE messaging_groups SET unknown_sender_policy = ? WHERE id = ?')
+      .run(args.senderPolicy, mg.id);
+
+    db.prepare(`
+      INSERT INTO messaging_group_agents
+        (id, messaging_group_id, agent_group_id, trigger_rules, response_scope, session_mode, priority, created_at)
+      VALUES (?, ?, ?, '', 'all', ?, 10, datetime('now'))
+    `).run(generateId('mga'), mg.id, ag.id, args.sessionMode);
+  });
+  tx();
+
+  console.log('');
+  console.log(`WIRED platform_id=${mg.platform_id} agent_group=${ag.name} policy=${args.senderPolicy} mode=${args.sessionMode}`);
+  db.close();
+}
+
+main().catch((err) => {
+  console.error('FAILED:', err.message);
+  process.exit(1);
+});
diff --git a/.claude/skills/channel-formatting/SKILL.md b/.claude/skills/channel-formatting/SKILL.md
deleted file mode 100644
index 8d27ffc..0000000
--- a/.claude/skills/channel-formatting/SKILL.md
+++ /dev/null
@@ -1,137 +0,0 @@
----
-name: channel-formatting
-description: Convert Claude's Markdown output to each channel's native text syntax before delivery. Adds zero-dependency formatting for WhatsApp, Telegram, and Slack (marker substitution). Also ships a Signal rich-text helper (parseSignalStyles) used by the Signal skill.
----
-
-# Channel Formatting
-
-This skill wires channel-aware Markdown conversion into the outbound pipeline so Claude's
-responses render natively on each platform — no more literal `**asterisks**` in WhatsApp or
-Telegram.
-
-| Channel | Transformation |
-|---------|---------------|
-| WhatsApp | `**bold**` → `*bold*`, `*italic*` → `_italic_`, headings → bold, links → `text (url)` |
-| Telegram | same as WhatsApp, but `[text](url)` links are preserved (Markdown v1 renders them natively) |
-| Slack | same as WhatsApp, but links become `<url\|text>` |
-| Discord | passthrough (Discord already renders Markdown) |
-| Signal | passthrough for `parseTextStyles`; `parseSignalStyles` in `src/text-styles.ts` produces plain text + native `textStyle` ranges for use by the Signal skill |
-
-Code blocks (fenced and inline) are always protected — their content is never transformed.
-
-## Phase 1: Pre-flight
-
-### Check if already applied
-
-```bash
-test -f src/text-styles.ts && echo "already applied" || echo "not yet applied"
-```
-
-If `already applied`, skip to Phase 3 (Verify).
-
-## Phase 2: Apply Code Changes
-
-### Ensure the upstream remote
-
-```bash
-git remote -v
-```
-
-If an `upstream` remote pointing to `https://github.com/qwibitai/nanoclaw.git` is missing,
-add it:
-
-```bash
-git remote add upstream https://github.com/qwibitai/nanoclaw.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch upstream skill/channel-formatting
-git merge upstream/skill/channel-formatting
-```
-
-If there are merge conflicts on `pnpm-lock.yaml`, resolve them by accepting the incoming
-version and continuing:
-
-```bash
-git checkout --theirs pnpm-lock.yaml
-git add pnpm-lock.yaml
-git merge --continue
-```
-
-For any other conflict, read the conflicted file and reconcile both sides manually.
-
-This merge adds:
-
-- `src/text-styles.ts` — `parseTextStyles(text, channel)` for marker substitution and
-  `parseSignalStyles(text)` for Signal native rich text
-- `src/router.ts` — `formatOutbound` gains an optional `channel` parameter; when provided
-  it calls `parseTextStyles` after stripping `<internal>` tags
-- `src/index.ts` — both outbound `sendMessage` paths pass `channel.name` to `formatOutbound`
-- `src/formatting.test.ts` — test coverage for both functions across all channels
-
-### Validate
-
-```bash
-pnpm install
-pnpm run build
-pnpm exec vitest run src/formatting.test.ts
-```
-
-All 73 tests should pass and the build should be clean before continuing.
-
-## Phase 3: Verify
-
-### Rebuild and restart
-
-```bash
-pnpm run build
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw   # macOS
-# Linux: systemctl --user restart nanoclaw
-```
-
-### Spot-check formatting
-
-Send a message through any registered WhatsApp or Telegram chat that will trigger a
-response from Claude. Ask something that will produce formatted output, such as:
-
-> Summarise the three main advantages of TypeScript using bullet points and **bold** headings.
-
-Confirm that the response arrives with native bold (`*text*`) rather than raw double
-asterisks.
-
-### Check logs if needed
-
-```bash
-tail -f logs/nanoclaw.log
-```
-
-## Signal Skill Integration
-
-If you have the Signal skill installed, `src/channels/signal.ts` can import
-`parseSignalStyles` from the newly present `src/text-styles.ts`:
-
-```typescript
-import { parseSignalStyles, SignalTextStyle } from '../text-styles.js';
-```
-
-`parseSignalStyles` returns `{ text: string, textStyle: SignalTextStyle[] }` where
-`textStyle` is an array of `{ style, start, length }` objects suitable for the
-`signal-cli` JSON-RPC `textStyles` parameter (format: `"start:length:STYLE"`).
-
-## Removal
-
-```bash
-# Remove the new file
-rm src/text-styles.ts
-
-# Revert router.ts to remove the channel param
-git diff upstream/main src/router.ts   # review changes
-git checkout upstream/main -- src/router.ts
-
-# Revert the index.ts sendMessage call sites to plain formatOutbound(rawText)
-# (edit manually or: git checkout upstream/main -- src/index.ts)
-
-pnpm run build
-```
\ No newline at end of file
diff --git a/.claude/skills/init-first-agent/SKILL.md b/.claude/skills/init-first-agent/SKILL.md
index be78845..6b110d3 100644
--- a/.claude/skills/init-first-agent/SKILL.md
+++ b/.claude/skills/init-first-agent/SKILL.md
@@ -87,18 +87,17 @@ The script:
 2. Creates the `agent_groups` row and calls `initGroupFilesystem` at `groups/dm-with-<name>/`.
 3. Reuses or creates the DM `messaging_groups` row.
 4. Wires them via `messaging_group_agents` (which auto-creates the companion `agent_destinations` row).
-5. Resolves the session (creates `inbound.db` / `outbound.db`).
-6. Writes a `kind: 'chat'`, `sender: 'system'` welcome message into `inbound.db`.
+5. Hands the welcome message to the running service via its CLI socket (`data/cli.sock`), targeting the DM messaging group. The service routes it into the DM session, which wakes the container synchronously. If the socket isn't reachable (service down), falls back to a direct `inbound.db` write that the next host sweep picks up.
 
 Show the script's output to the user.
 
 ## 5. Verify
 
-Host sweep runs every ~60s. Within one sweep window the container wakes, the agent processes the system message, and the reply flows through `outbound.db` to the channel.
+The welcome DM is queued synchronously; the only wait is container cold-start (~60s on first launch) before the agent processes the message and the reply flows through `outbound.db` to the channel.
 
 Do not tail the log or poll in a sleep loop. Ask the user in plain text:
 
-> The welcome DM should arrive within ~60 seconds. Let me know when you've received it (or if it doesn't arrive within two minutes).
+> The welcome DM should arrive shortly. Let me know when you've received it (or if it doesn't arrive within two minutes).
 
 Wait for the user's reply. If they confirm receipt, the skill is done.
 
diff --git a/.claude/skills/migrate-nanoclaw/diagnostics.md b/.claude/skills/migrate-nanoclaw/diagnostics.md
index a308b7a..170fd7a 100644
--- a/.claude/skills/migrate-nanoclaw/diagnostics.md
+++ b/.claude/skills/migrate-nanoclaw/diagnostics.md
@@ -45,8 +45,7 @@ rm /tmp/nanoclaw-diagnostics.json
 **No**: `rm /tmp/nanoclaw-diagnostics.json`
 
 **Never ask again**:
-1. Replace contents of `.claude/skills/setup/diagnostics.md` with `# Diagnostics — opted out`
-2. Replace contents of `.claude/skills/update-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
-3. Replace contents of `.claude/skills/migrate-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
-4. Remove the diagnostics sections from each corresponding SKILL.md
-5. `rm /tmp/nanoclaw-diagnostics.json`
+1. Replace contents of `.claude/skills/update-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
+2. Replace contents of `.claude/skills/migrate-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
+3. Remove the diagnostics sections from each corresponding SKILL.md
+4. `rm /tmp/nanoclaw-diagnostics.json`
diff --git a/.claude/skills/new-setup/SKILL.md b/.claude/skills/new-setup/SKILL.md
deleted file mode 100644
index 6b95695..0000000
--- a/.claude/skills/new-setup/SKILL.md
+++ /dev/null
@@ -1,137 +0,0 @@
----
-name: new-setup
-description: Shortest path from zero to a working two-way agent chat, for any user regardless of technical background — ends at a running NanoClaw instance with at least one CLI-reachable agent.
-allowed-tools: Bash(bash setup.sh) Bash(bash setup/probe.sh) Bash(pnpm exec tsx setup/index.ts *) Bash(pnpm run chat *) Bash(brew install *) Bash(curl -fsSL https://get.docker.com | sh) Bash(sudo usermod -aG docker *) Bash(open -a Docker) Bash(sudo systemctl start docker)
----
-
-# NanoClaw bare-minimum setup
-
-Purpose of this skill is to take any user — technical or not — from zero to a two-way chat with an agent in the fewest steps possible. Done means a running NanoClaw instance with at least one agent reachable via the CLI channel.
-
-Only run the steps strictly required for the NanoClaw process to start and respond to the user end-to-end. Everything else is deferred to post-setup skills.
-
-Before each step, narrate to the user in your own words what's about to happen — one short, friendly sentence, no jargon. Don't read a scripted line; use the step context below to speak naturally.
-
-Each step is invoked as `pnpm exec tsx setup/index.ts --step <name>` and emits a structured status block Claude parses to decide what to do next.
-
-Start with a probe: a single upfront scan that snapshots every prerequisite and dependency. The rest of the flow reads this snapshot to decide what to run, skip, or ask about — no per-step re-checking. The probe is pure bash (`setup/probe.sh`) with no external deps so it runs correctly before Node has been installed.
-
-## Current state
-
-!`bash setup/probe.sh`
-
-## Flow
-
-Parse the probe block above. For each step below, consult the named probe fields and skip, ask, or run accordingly. The probe always returns a real snapshot — there is no "node not installed" fallback; `HOST_DEPS=missing` is how you know Node/pnpm haven't been bootstrapped yet.
-
-## Ordering and parallelism
-
-Run steps sequentially by default: invoke the step, wait for its status block, act on the result, move to the next.
-
-One permitted parallelism:
-
-- **Step 2 (container image build) and step 3 (OneCLI install)** are independent — they may start together in the background.
-- **Step 4 (auth) must NOT start until step 3 has completed.** Auth writes the secret into the OneCLI vault; if OneCLI isn't installed and healthy yet, the user gets asked for a credential the system can't store. Do not open an `AskUserQuestion` for step 4 while OneCLI is still installing.
-- Step 2's image build may continue running past step 4 — the image isn't consumed until step 6 (first CLI agent). Join before step 6.
-
-### 1. Node bootstrap
-
-Check probe results and skip if `HOST_DEPS=ok` — Node, pnpm, `node_modules`, and `better-sqlite3`'s native binding are already in place.
-
-If `HOST_DEPS=missing` and `node --version` fails (Node isn't installed at all), install Node 22 **before** running `bash setup.sh`, otherwise the first bootstrap run is guaranteed to fail:
-
-- macOS: `brew install node@22`
-- Linux / WSL: `curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash - && sudo apt-get install -y nodejs`
-
-Then run `bash setup.sh`. If Node is already present and only `HOST_DEPS=missing`, run `bash setup.sh` directly — deps just haven't been installed yet.
-
-Parse the status block:
-
-- `NODE_OK=false` → Node install didn't take effect (PATH issue, keg-only formula, etc.). Investigate `logs/setup.log`, resolve, re-run.
-- `DEPS_OK=false` or `NATIVE_OK=false` → Read `logs/setup.log`, fix, re-run.
-
-> **Loose command:** `bash setup.sh`. Justification: pre-Node bootstrap. Can't call the Node-based dispatcher before Node and `pnpm install` are in place.
-
-### 2. Docker
-
-Check probe results and skip if `DOCKER=running` AND `IMAGE_PRESENT=true`.
-
-**Runtime:**
-- `DOCKER=not_found` → Docker itself is missing — install it so agent containers have an isolated place to run.
-  - macOS: `brew install --cask docker && open -a Docker`
-  - Linux: `curl -fsSL https://get.docker.com | sh && sudo usermod -aG docker $USER` (tell user they may need to log out/in for group membership)
-- `DOCKER=installed_not_running` → Docker is installed but the daemon is down — start it.
-  - macOS: `open -a Docker`
-  - Linux: `sudo systemctl start docker`
-
-Wait ~15s after either, then proceed.
-
-> **Loose commands:** Docker install/start. Justification: platform-specific package-manager invocations. Wrapping them in a `--step` would just move the same branching into TypeScript with no added value.
-
-**Image (run if `IMAGE_PRESENT=false`):** build the agent container image — takes a few minutes the first time, one-off cost.
-
-`pnpm exec tsx setup/index.ts --step container -- --runtime docker`
-
-### 3. OneCLI
-
-Check probe results and skip if `ONECLI_STATUS=healthy`.
-
-OneCLI is the local vault that holds API keys and only releases them to agents when they need them.
-
-`pnpm exec tsx setup/index.ts --step onecli`
-
-### 4. Anthropic credential
-
-Check probe results and skip if `ANTHROPIC_SECRET=true`.
-
-The credential never travels through chat — the user generates it, registers it with OneCLI themselves, and the skill verifies.
-
-**4a. Pick the source.** `AskUserQuestion`:
-
-1. **Claude subscription (Pro/Max)** — "Generate a token via `claude setup-token` in another terminal."
-2. **Anthropic API key** — "Use a pay-per-use key from console.anthropic.com/settings/keys."
-
-**4b. Wait for the user to obtain the credential.** For subscription, have them run `claude setup-token` in another terminal. For API key, point them to the console URL above. Either way, they keep the token — just confirm when they have it.
-
-**4c. Pick the registration path.** `AskUserQuestion` — substitute `${ONECLI_URL}` from the probe (or `.env`):
-
-1. **Dashboard** — "Open ${ONECLI_URL} in a browser; add a secret of type `anthropic`, value = the token, host-pattern `api.anthropic.com`."
-2. **CLI** — "Run in another terminal: `onecli secrets create --name Anthropic --type anthropic --value YOUR_TOKEN --host-pattern api.anthropic.com`"
-
-Wait for the user's confirmation. If their reply happens to include a token (starts with `sk-ant-`), register it for them: `pnpm exec tsx setup/index.ts --step auth -- --create --value <TOKEN>`.
-
-**4d. Verify.**
-
-`pnpm exec tsx setup/index.ts --step auth -- --check`
-
-If `ANTHROPIC_OK=false`, the secret isn't there yet — ask them to retry, then re-check.
-
-### 5. Service
-
-Check probe results and skip if `SERVICE_STATUS=running`.
-
-Start the NanoClaw background service — it relays messages between the user and the agent.
-
-`pnpm exec tsx setup/index.ts --step service`
-
-### 6. First CLI agent
-
-If step 2's container build is still running in the background, join it here before proceeding — the agent needs the image.
-
-Create the first agent and wire it to the CLI channel. Ask the user "What should I call you?" first — default the offered value to `INFERRED_DISPLAY_NAME` from the probe.
-
-`pnpm exec tsx setup/index.ts --step cli-agent -- --display-name "<name>"`
-
-### 7. First chat
-
-Everything's ready — send the first message to the agent.
-
-`pnpm run chat hi`
-
-The agent should reply within ~60s (first container spin-up is slowest). If no reply, tail `logs/nanoclaw.log`.
-
-> **Loose command:** `pnpm run chat hi`. Justification: this is the command the user will keep using after setup. Hiding it behind a `--step` would force them to memorize a second way to do the same thing.
-
-## If anything fails
-
-Any step that reports `STATUS: failed` in its status block: read `logs/setup.log`, diagnose, fix the underlying cause, re-run the same `--step`. Don't bypass errors to keep moving.
diff --git a/.claude/skills/setup/SKILL.md b/.claude/skills/setup/SKILL.md
index 3b21e14..5ce96e3 100644
--- a/.claude/skills/setup/SKILL.md
+++ b/.claude/skills/setup/SKILL.md
@@ -1,349 +1,10 @@
 ---
 name: setup
-description: Run initial NanoClaw setup. Use when user wants to install dependencies, authenticate messaging channels, register their main channel, or start the background services. Triggers on "setup", "install", "configure nanoclaw", or first-time setup requests.
+description: Run initial NanoClaw setup. Use when user wants to install NanoClaw, configure it, or go through first-time setup. Triggers on "setup", "install", "configure nanoclaw", or first-time setup requests.
 ---
 
 # NanoClaw Setup
 
-Welcome the user to NanoClaw. Introduce yourself — you'll be walking them through the entire setup process step by step, from installing dependencies to getting their first message through. Keep it warm and brief (2-3 sentences).
+Tell the user to run `bash nanoclaw.sh` in their terminal. That script handles the full end-to-end setup — dependencies, container image, OneCLI vault, Anthropic credential, service, first agent, and optional channel wiring.
 
-Then explain that setup involves running many shell commands (installing packages, building containers, starting services), and recommend pre-approving the standard setup commands so they don't have to confirm each one individually.
-
-Use `AskUserQuestion` with these options:
-
-1. **Pre-approve (recommended)** — description: "Pre-approve standard setup commands so you don't have to confirm each one. You can review the list first if you'd like."
-2. **No thanks** — description: "I'll approve each command individually as it comes up."
-3. **Show me the list first** — description: "Show me exactly which commands will be pre-approved before I decide."
-
-If they pick option 1: read `.claude/skills/setup/setup-permissions.json`, then read the project settings file at `.claude/settings.json` (create it if it doesn't exist with `{}`), and directly edit it to add/merge the permissions into the `permissions.allow` array. Do NOT use the `update-config` skill.
-
-If they pick option 3: read and display `.claude/skills/setup/setup-permissions.json`, then re-ask with just options 1 and 2.
-
-If they decline, continue — they'll approve commands individually.
-
----
-
-**Internal guidance (do not show to user):**
-
-- Run setup steps automatically. Only pause when user action is required (channel authentication, configuration choices).
-- Setup uses `bash setup.sh` for bootstrap, then `npx tsx setup/index.ts --step <name>` for all other steps. Steps emit structured status blocks to stdout. Verbose logs go to `logs/setup.log`.
-- **Principle:** When something is broken or missing, fix it. Don't tell the user to go fix it themselves unless it genuinely requires their manual action (e.g. authenticating a channel, pasting a secret token). If a dependency is missing, install it. If a service won't start, diagnose and repair.
-- **UX Note:** Use `AskUserQuestion` for multiple-choice questions only (e.g. "which credential method?"). Do NOT use it when free-text input is needed (e.g. phone numbers, tokens, paths) — just ask the question in plain text and wait for the user's reply.
-- **Timeouts:** Use 5m timeouts for install and build steps.
-- **Waiting on user:** When the user needs to do something (change a setting, get a token, open a browser, etc.), stop and wait. Give clear instructions, then say "Let me know when done or if you need help." Do NOT continue to the next step. If they ask for help, give more detail, ask where they got stuck, and try to assist.
-
-## 0. Git Upstream
-
-Ensure `upstream` remote points to `qwibitai/nanoclaw`. If missing, add it silently:
-
-```bash
-git remote -v
-git remote add upstream https://github.com/qwibitai/nanoclaw.git 2>/dev/null || true
-```
-
-## 1. Bootstrap (Node.js + Dependencies)
-
-Run `bash setup.sh` and parse the status block.
-
-- If NODE_OK=false → Node.js is missing or too old. Use `AskUserQuestion: Would you like me to install Node.js 22?` If confirmed:
-  - macOS: `brew install node@22` (if brew available) or install nvm then `nvm install 22`
-  - Linux: `curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash - && sudo apt-get install -y nodejs`, or nvm
-  - After installing Node, re-run `bash setup.sh`
-- If DEPS_OK=false → Read `logs/setup.log`. Try: delete `node_modules`, re-run `bash setup.sh`. If native module build fails, install build tools (`xcode-select --install` on macOS, `build-essential` on Linux), then retry.
-- If NATIVE_OK=false → better-sqlite3 failed to load. Install build tools and re-run.
-- Record PLATFORM and IS_WSL for later steps.
-
-## 2. Check Environment
-
-Run `pnpm exec tsx setup/index.ts --step environment` and parse the status block.
-
-- If HAS_AUTH=true → WhatsApp is already configured, note for step 5
-- If HAS_REGISTERED_GROUPS=true → note existing config, offer to skip or reconfigure
-- Record DOCKER value for step 3
-
-### OpenClaw Migration Detection
-
-If OPENCLAW_PATH is not `none` from the environment check above, AskUserQuestion:
-
-1. **Migrate now** — "Import identity, credentials, and settings from OpenClaw before continuing setup."
-2. **Fresh start** — "Skip migration and set up NanoClaw from scratch."
-3. **Migrate later** — "Continue setup now, run `/migrate-from-openclaw` anytime later."
-
-If "Migrate now": invoke `/migrate-from-openclaw`, then return here and continue at step 2a (Timezone).
-
-## 2a. Timezone
-
-Run `pnpm exec tsx setup/index.ts --step timezone` and parse the status block.
-
-- If NEEDS_USER_INPUT=true → The system timezone could not be autodetected (e.g. POSIX-style TZ like `IST-2`). AskUserQuestion: "What is your timezone?" with common options (America/New_York, Europe/London, Asia/Jerusalem, Asia/Tokyo) and an "Other" escape. Then re-run: `pnpm exec tsx setup/index.ts --step timezone -- --tz <their-answer>`.
-- If STATUS=success and RESOLVED_TZ is `UTC` or `Etc/UTC` → confirm with the user: "Your system timezone is UTC — is that correct, or are you on a remote server?" If wrong, ask for their actual timezone and re-run with `--tz`.
-- If STATUS=success → Timezone is configured. Note RESOLVED_TZ for reference.
-
-## 3. Container Runtime (Docker)
-
-### 3a. Install Docker
-
-- DOCKER=running → continue to step 4
-- DOCKER=installed_not_running → start Docker: `open -a Docker` (macOS) or `sudo systemctl start docker` (Linux). Wait 15s, re-check with `docker info`.
-- DOCKER=not_found → Use `AskUserQuestion: Docker is required for running agents. Would you like me to install it?` If confirmed:
-  - macOS: install via `brew install --cask docker`, then `open -a Docker` and wait for it to start. If brew not available, direct to Docker Desktop download at https://docker.com/products/docker-desktop
-  - Linux: install with `curl -fsSL https://get.docker.com | sh && sudo usermod -aG docker $USER`. Note: user may need to log out/in for group membership.
-
-### 3b. CJK fonts
-
-Agent containers skip CJK fonts by default (~200MB saved). Without them, Chromium-rendered screenshots and PDFs show tofu for Chinese/Japanese/Korean.
-
-- **User writing to you in Chinese, Japanese, or Korean** → enable without asking. Mention it briefly.
-- **Resolved timezone from step 2a is a CJK region** (`Asia/Tokyo`, `Asia/Shanghai`, `Asia/Hong_Kong`, `Asia/Taipei`, `Asia/Seoul`) or other signal short of active CJK use → ask: "Enable CJK fonts? Adds ~200MB, lets the agent render CJK in screenshots and PDFs."
-- **Otherwise** → skip.
-
-To enable, write `INSTALL_CJK_FONTS=true` to `.env`:
-
-```bash
-grep -q '^INSTALL_CJK_FONTS=' .env && sed -i.bak 's/^INSTALL_CJK_FONTS=.*/INSTALL_CJK_FONTS=true/' .env && rm -f .env.bak || echo 'INSTALL_CJK_FONTS=true' >> .env
-```
-
-The next step's build picks it up automatically.
-
-### 3c. Build and test
-
-Run `pnpm exec tsx setup/index.ts --step container -- --runtime docker` and parse the status block.
-
-**If BUILD_OK=false:** Read `logs/setup.log` tail for the build error.
-- Cache issue (stale layers): `docker builder prune -f`. Retry.
-- Dockerfile syntax or missing files: diagnose from the log and fix, then retry.
-
-**If TEST_OK=false but BUILD_OK=true:** The image built but won't run. Check logs — common cause is runtime not fully started. Wait a moment and retry the test.
-
-## 4. Credential System
-
-### 4a. OneCLI
-
-Install OneCLI and its CLI tool:
-
-```bash
-curl -fsSL onecli.sh/install | sh
-curl -fsSL onecli.sh/cli/install | sh
-```
-
-Verify both installed: `onecli version`. If the command is not found, the CLI was likely installed to `~/.local/bin/`. Add it to PATH for the current session and persist it:
-
-```bash
-export PATH="$HOME/.local/bin:$PATH"
-# Persist for future sessions (append to shell profile if not already present)
-grep -q '.local/bin' ~/.bashrc 2>/dev/null || echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc
-grep -q '.local/bin' ~/.zshrc 2>/dev/null || echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc
-```
-
-Then re-verify with `onecli version`.
-
-Point the CLI at the local OneCLI instance, the ONECLI_URL was output from the install script above:
-```bash
-onecli config set api-host ${ONECLI_URL}
-```
-
-Ensure `.env` has the OneCLI URL (create the file if it doesn't exist):
-```bash
-grep -q 'ONECLI_URL' .env 2>/dev/null || echo 'ONECLI_URL=${ONECLI_URL}' >> .env
-```
-
-Check if a secret already exists:
-```bash
-onecli secrets list
-```
-
-If an Anthropic secret is listed, confirm with user: keep or reconfigure? If keeping, skip to step 5.
-
-AskUserQuestion: Do you want to use your **Claude subscription** (Pro/Max) or an **Anthropic API key**?
-
-1. **Claude subscription (Pro/Max)** — description: "Uses your existing Claude Pro or Max subscription. You'll run `claude setup-token` in another terminal to get your token."
-2. **Anthropic API key** — description: "Pay-per-use API key from console.anthropic.com."
-
-#### Subscription path
-
-Tell the user:
-
-> Run `claude setup-token` in another terminal. It will output a token — copy it but don't paste it here.
-
-Then stop and wait for the user to confirm they have the token. Do NOT proceed until they respond.
-
-Once they confirm, they register it with OneCLI. AskUserQuestion with two options:
-
-1. **Dashboard** — description: "Best if you have a browser on this machine. Open ${ONECLI_URL} and add the secret in the UI. Use type 'anthropic' and paste your token as the value."
-2. **CLI** — description: "Best for remote/headless servers. Run: `onecli secrets create --name Anthropic --type anthropic --value YOUR_TOKEN --host-pattern api.anthropic.com`"
-
-#### API key path
-
-Tell the user to get an API key from https://console.anthropic.com/settings/keys if they don't have one.
-
-Then AskUserQuestion with two options:
-
-1. **Dashboard** — description: "Best if you have a browser on this machine. Open ${ONECLI_URL} and add the secret in the UI."
-2. **CLI** — description: "Best for remote/headless servers. Run: `onecli secrets create --name Anthropic --type anthropic --value YOUR_KEY --host-pattern api.anthropic.com`"
-
-#### After either path
-
-Ask them to let you know when done.
-
-**If the user's response happens to contain a token or key** (starts with `sk-ant-`): handle it gracefully — run the `onecli secrets create` command with that value on their behalf.
-
-**After user confirms:** verify with `onecli secrets list` that an Anthropic secret exists. If not, ask again.
-
-## 5. Set Up Channels
-
-Show the full list of available channels in plain text (do NOT use AskUserQuestion — it limits to 4 options). Ask which one they want to start with. They can add more later with `/customize`.
-
-Channels where the agent gets its own identity (name and avatar) are marked as recommended.
-
-1. Discord *(recommended — agent gets own identity)*
-2. Slack *(recommended — agent gets own identity)*
-3. Telegram *(recommended — agent gets own identity)*
-4. Microsoft Teams *(recommended — agent gets own identity)*
-5. Webex *(recommended — agent gets own identity)*
-6. WhatsApp
-7. WhatsApp Cloud API
-8. iMessage
-9. GitHub
-10. Linear
-11. Google Chat
-12. Resend (email)
-13. Matrix
-
-**Delegate to the selected channel's skill.** Each channel skill handles its own package installation, authentication, registration, and configuration.
-
-Invoke the matching skill:
-
-- **Discord:** Invoke `/add-discord`
-- **Slack:** Invoke `/add-slack`
-- **Telegram:** Invoke `/add-telegram`
-- **GitHub:** Invoke `/add-github`
-- **Linear:** Invoke `/add-linear`
-- **Microsoft Teams:** Invoke `/add-teams`
-- **Google Chat:** Invoke `/add-gchat`
-- **WhatsApp Cloud API:** Invoke `/add-whatsapp-cloud`
-- **WhatsApp Baileys:** Invoke `/add-whatsapp`
-- **Resend:** Invoke `/add-resend`
-- **Matrix:** Invoke `/add-matrix`
-- **Webex:** Invoke `/add-webex`
-- **iMessage:** Invoke `/add-imessage`
-
-The skill will:
-1. Install the Chat SDK adapter package
-2. Uncomment the channel import in `src/channels/index.ts`
-3. Collect credentials/tokens and write to `.env`
-4. Build and verify
-
-**After the channel skill completes**, install dependencies and rebuild — channel merges may introduce new packages:
-
-```bash
-pnpm install && pnpm run build
-```
-
-If the build fails, read the error output and fix it (usually a missing dependency). Then continue to step 5a.
-
-## 6. Mount Allowlist
-
-Set empty mount allowlist (agents only access their own workspace). Users can configure mounts later with `/manage-mounts`.
-
-```bash
-pnpm exec tsx setup/index.ts --step mounts -- --empty
-```
-
-## 7. Start Service
-
-If service already running: unload first.
-- macOS: `launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist`
-- Linux: `systemctl --user stop nanoclaw` (or `systemctl stop nanoclaw` if root)
-
-Run `pnpm exec tsx setup/index.ts --step service` and parse the status block.
-
-**If FALLBACK=wsl_no_systemd:** WSL without systemd detected. Tell user they can either enable systemd in WSL (`echo -e "[boot]\nsystemd=true" | sudo tee /etc/wsl.conf` then restart WSL) or use the generated `start-nanoclaw.sh` wrapper.
-
-**If DOCKER_GROUP_STALE=true:** The user was added to the docker group after their session started — the systemd service can't reach the Docker socket. Ask user to run these two commands:
-
-1. Immediate fix: `sudo setfacl -m u:$(whoami):rw /var/run/docker.sock`
-2. Persistent fix (re-applies after every Docker restart):
-```bash
-sudo mkdir -p /etc/systemd/system/docker.service.d
-sudo tee /etc/systemd/system/docker.service.d/socket-acl.conf << 'EOF'
-[Service]
-ExecStartPost=/usr/bin/setfacl -m u:USERNAME:rw /var/run/docker.sock
-EOF
-sudo systemctl daemon-reload
-```
-Replace `USERNAME` with the actual username (from `whoami`). Run the two `sudo` commands separately — the `tee` heredoc first, then `daemon-reload`. After user confirms setfacl ran, re-run the service step.
-
-**If SERVICE_LOADED=false:**
-- Read `logs/setup.log` for the error.
-- macOS: check `launchctl list | grep nanoclaw`. If PID=`-` and status non-zero, read `logs/nanoclaw.error.log`.
-- Linux: check `systemctl --user status nanoclaw`.
-- Re-run the service step after fixing.
-
-## 7a. Wire Channels to Agent Groups
-
-The service is now running, so polling-based adapters (Telegram) can observe inbound messages — required for pairing.
-
-Invoke `/manage-channels` to wire the installed channels to agent groups. This step:
-1. Creates the agent group(s) and assigns a name to the assistant
-2. Resolves each channel's platform-specific ID (Telegram via pairing code; other channels via the platform's own ID lookup)
-3. Decides the isolation level — whether channels share an agent, session, or are fully separate
-
-The `/manage-channels` skill reads each channel's `## Channel Info` section from its SKILL.md for platform-specific guidance (terminology, how to find IDs, recommended isolation).
-
-**This step is required.** Without it, channels are installed but not wired — messages will be silently dropped because the router has no agent group to route to.
-
-## 7b. Dashboard & Web Applications
-
-AskUserQuestion: Do you want to create a dashboard and build web applications?
-
-1. **Yes (recommended)** — description: "Get a NanoClaw dashboard to monitor your agents and build custom websites however you want. Deploys to Vercel."
-2. **Not now** — description: "You can add this later with `/add-vercel`."
-
-If yes: invoke `/add-vercel`.
-
-## 8. Verify
-
-Run `pnpm exec tsx setup/index.ts --step verify` and parse the status block.
-
-**If STATUS=failed, fix each:**
-- SERVICE=stopped → `pnpm run build`, then restart: `launchctl kickstart -k gui/$(id -u)/com.nanoclaw` (macOS) or `systemctl --user restart nanoclaw` (Linux) or `bash start-nanoclaw.sh` (WSL nohup)
-- SERVICE=not_found → re-run step 7
-- CREDENTIALS=missing → re-run step 4 (check `onecli secrets list`)
-- CHANNEL_AUTH shows `not_found` for any channel → re-invoke that channel's skill (e.g. `/add-telegram`)
-- REGISTERED_GROUPS=0 → re-invoke `/manage-channels` from step 7a
-Tell user to test: send a message in their registered chat. Show: `tail -f logs/nanoclaw.log`
-
-## Troubleshooting
-
-**Service not starting:** Check `logs/nanoclaw.error.log`. Common: wrong Node path (re-run step 7), credential system not running (check `curl ${ONECLI_URL}/api/health`), missing channel credentials (re-invoke channel skill).
-
-**Container agent fails ("Claude Code process exited with code 1"):** Ensure Docker is running — `open -a Docker` (macOS) or `sudo systemctl start docker` (Linux). Check container logs in `groups/main/logs/container-*.log`.
-
-**No response to messages:** Check trigger pattern. Main channel doesn't need prefix. Check DB: `pnpm exec tsx setup/index.ts --step verify`. Check `logs/nanoclaw.log`.
-
-**Channel not connecting:** Verify the channel's credentials are set in `.env`. Channels auto-enable when their credentials are present. For WhatsApp: check `store/auth/creds.json` exists. For token-based channels: check token values in `.env`. Restart the service after any `.env` change.
-
-**Unload service:** macOS: `launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist` | Linux: `systemctl --user stop nanoclaw`
-
-
-## 9. Diagnostics
-
-1. Use the Read tool to read `.claude/skills/setup/diagnostics.md`.
-2. Follow every step in that file before completing setup.
-
-## 10. Fork Setup
-
-Only run this after the user has confirmed 2-way messaging works.
-
-Check `git remote -v`. If `origin` points to `qwibitai/nanoclaw` (not a fork), ask in plain text:
-
-> We recommend forking NanoClaw so you can push your customizations and pull updates easily. Would you like to set up a fork now?
-
-If yes: instruct the user to fork `qwibitai/nanoclaw` on GitHub (they need to do this in their browser), then ask for their GitHub username. Run:
-```bash
-git remote rename origin upstream
-git remote add origin https://github.com/<their-username>/nanoclaw.git
-git push --force origin main
-```
-
-If no: skip — upstream is already configured from step 0.
+If they hit an error partway through, it will offer Claude-assisted recovery inline — no need to come back here.
diff --git a/.claude/skills/setup/diagnostics.md b/.claude/skills/setup/diagnostics.md
deleted file mode 100644
index 26d79b1..0000000
--- a/.claude/skills/setup/diagnostics.md
+++ /dev/null
@@ -1,49 +0,0 @@
-# Diagnostics
-
-Gather system info:
-
-```bash
-node -p "require('./package.json').version"
-uname -s
-uname -m
-node -p "process.versions.node.split('.')[0]"
-```
-
-Check if the user migrated from OpenClaw during this setup session (i.e. `/migrate-from-openclaw` was invoked). If you're unsure (e.g. after context compaction), check for `migration-state.md` in the project root — it exists during and sometimes after migration.
-
-Write `/tmp/nanoclaw-diagnostics.json`. No paths, usernames, hostnames, or IP addresses.
-
-```json
-{
-  "api_key": "phc_fx1Hhx9ucz8GuaJC8LVZWO8u03yXZZJJ6ObS4yplnaP",
-  "event": "setup_complete",
-  "distinct_id": "<uuid>",
-  "properties": {
-    "success": true,
-    "nanoclaw_version": "1.2.21",
-    "os_platform": "darwin",
-    "arch": "arm64",
-    "node_major_version": 22,
-    "channels_selected": ["telegram", "whatsapp"],
-    "migrated_from_openclaw": false,
-    "error_count": 0,
-    "failed_step": null
-  }
-}
-```
-
-Show the entire JSON to the user and ask via AskUserQuestion: **Yes** / **No** / **Never ask again**
-
-**Yes**:
-```bash
-curl -s -X POST https://us.i.posthog.com/capture/ -H 'Content-Type: application/json' -d @/tmp/nanoclaw-diagnostics.json
-rm /tmp/nanoclaw-diagnostics.json
-```
-
-**No**: `rm /tmp/nanoclaw-diagnostics.json`
-
-**Never ask again**:
-1. Replace contents of `.claude/skills/setup/diagnostics.md` with `# Diagnostics — opted out`
-2. Replace contents of `.claude/skills/update-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
-3. Remove the `## 9. Diagnostics` section from `.claude/skills/setup/SKILL.md` and the `## Diagnostics` section from `.claude/skills/update-nanoclaw/SKILL.md`
-4. `rm /tmp/nanoclaw-diagnostics.json`
diff --git a/.claude/skills/setup/setup-permissions.json b/.claude/skills/setup/setup-permissions.json
deleted file mode 100644
index a263d4c..0000000
--- a/.claude/skills/setup/setup-permissions.json
+++ /dev/null
@@ -1,34 +0,0 @@
-[
-  "Bash(bash setup.sh*)",
-  "Bash(git remote *)",
-  "Bash(npx tsx setup/index.ts*)",
-  "Bash(npx tsx scripts/init-first-agent.ts*)",
-  "Bash(npm install @chat-adapter/*)",
-  "Bash(npm install chat-adapter-imessage*)",
-  "Bash(npm install @bitbasti/chat-adapter-webex*)",
-  "Bash(npm install @resend/chat-sdk-adapter*)",
-  "Bash(npm install @whiskeysockets/baileys*)",
-  "Bash(npm install @beeper/chat-adapter-matrix*)",
-  "Bash(npm install @nanoco/nanoclaw-dashboard*)",
-  "Bash(npm ci*)",
-  "Bash(npm run build*)",
-  "Bash(curl -fsSL onecli.sh*)",
-  "Bash(onecli *)",
-  "Bash(grep -q *)",
-  "Bash(echo *>> .env)",
-  "Bash(ls *)",
-  "Bash(cat ~/.config/nanoclaw/*)",
-  "Bash(tail *logs/*)",
-  "Bash(launchctl *nanoclaw*)",
-  "Bash(sqlite3 data/*)",
-  "Bash(docker info*)",
-  "Bash(docker logs *)",
-  "Bash(mkdir -p *)",
-  "Bash(cp .env *)",
-  "Bash(rsync -a .claude/skills/*)",
-  "Bash(head *)",
-  "Bash(xattr *)",
-  "Bash(find ~/.npm *)",
-  "Bash(which onecli*)",
-  "Bash(./container/build.sh*)"
-]
diff --git a/.claude/skills/update-nanoclaw/diagnostics.md b/.claude/skills/update-nanoclaw/diagnostics.md
index 8b06aa4..551842e 100644
--- a/.claude/skills/update-nanoclaw/diagnostics.md
+++ b/.claude/skills/update-nanoclaw/diagnostics.md
@@ -43,7 +43,6 @@ rm /tmp/nanoclaw-diagnostics.json
 **No**: `rm /tmp/nanoclaw-diagnostics.json`
 
 **Never ask again**:
-1. Replace contents of `.claude/skills/setup/diagnostics.md` with `# Diagnostics — opted out`
-2. Replace contents of `.claude/skills/update-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
-3. Remove the `## 9. Diagnostics` section from `.claude/skills/setup/SKILL.md` and the `## Diagnostics` section from `.claude/skills/update-nanoclaw/SKILL.md`
-4. `rm /tmp/nanoclaw-diagnostics.json`
+1. Replace contents of `.claude/skills/update-nanoclaw/diagnostics.md` with `# Diagnostics — opted out`
+2. Remove the `## Diagnostics` section from `.claude/skills/update-nanoclaw/SKILL.md`
+3. `rm /tmp/nanoclaw-diagnostics.json`
diff --git a/.claude/skills/use-local-whisper/SKILL.md b/.claude/skills/use-local-whisper/SKILL.md
deleted file mode 100644
index 664cafa..0000000
--- a/.claude/skills/use-local-whisper/SKILL.md
+++ /dev/null
@@ -1,152 +0,0 @@
----
-name: use-local-whisper
-description: Use when the user wants local voice transcription instead of OpenAI Whisper API. Switches to whisper.cpp running on Apple Silicon. WhatsApp only for now. Requires voice-transcription skill to be applied first.
----
-
-# Use Local Whisper
-
-Switches voice transcription from OpenAI's Whisper API to local whisper.cpp. Runs entirely on-device — no API key, no network, no cost.
-
-**Channel support:** Currently WhatsApp only. The transcription module (`src/transcription.ts`) uses Baileys types for audio download. Other channels (Telegram, Discord, etc.) would need their own audio-download logic before this skill can serve them.
-
-**Note:** The Homebrew package is `whisper-cpp`, but the CLI binary it installs is `whisper-cli`.
-
-## Prerequisites
-
-- `voice-transcription` skill must be applied first (WhatsApp channel)
-- macOS with Apple Silicon (M1+) recommended
-- `whisper-cpp` installed: `brew install whisper-cpp` (provides the `whisper-cli` binary)
-- `ffmpeg` installed: `brew install ffmpeg`
-- A GGML model file downloaded to `data/models/`
-
-## Phase 1: Pre-flight
-
-### Check if already applied
-
-Check if `src/transcription.ts` already uses `whisper-cli`:
-
-```bash
-grep 'whisper-cli' src/transcription.ts && echo "Already applied" || echo "Not applied"
-```
-
-If already applied, skip to Phase 3 (Verify).
-
-### Check dependencies are installed
-
-```bash
-whisper-cli --help >/dev/null 2>&1 && echo "WHISPER_OK" || echo "WHISPER_MISSING"
-ffmpeg -version >/dev/null 2>&1 && echo "FFMPEG_OK" || echo "FFMPEG_MISSING"
-```
-
-If missing, install via Homebrew:
-```bash
-brew install whisper-cpp ffmpeg
-```
-
-### Check for model file
-
-```bash
-ls data/models/ggml-*.bin 2>/dev/null || echo "NO_MODEL"
-```
-
-If no model exists, download the base model (148MB, good balance of speed and accuracy):
-```bash
-mkdir -p data/models
-curl -L -o data/models/ggml-base.bin "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin"
-```
-
-For better accuracy at the cost of speed, use `ggml-small.bin` (466MB) or `ggml-medium.bin` (1.5GB).
-
-## Phase 2: Apply Code Changes
-
-### Ensure WhatsApp fork remote
-
-```bash
-git remote -v
-```
-
-If `whatsapp` is missing, add it:
-
-```bash
-git remote add whatsapp https://github.com/qwibitai/nanoclaw-whatsapp.git
-```
-
-### Merge the skill branch
-
-```bash
-git fetch whatsapp skill/local-whisper
-git merge whatsapp/skill/local-whisper || {
-  git checkout --theirs pnpm-lock.yaml
-  git add pnpm-lock.yaml
-  git merge --continue
-}
-```
-
-This modifies `src/transcription.ts` to use the `whisper-cli` binary instead of the OpenAI API.
-
-### Validate
-
-```bash
-pnpm run build
-```
-
-## Phase 3: Verify
-
-### Ensure launchd PATH includes Homebrew
-
-The NanoClaw launchd service runs with a restricted PATH. `whisper-cli` and `ffmpeg` are in `/opt/homebrew/bin/` (Apple Silicon) or `/usr/local/bin/` (Intel), which may not be in the plist's PATH.
-
-Check the current PATH:
-```bash
-grep -A1 'PATH' ~/Library/LaunchAgents/com.nanoclaw.plist
-```
-
-If `/opt/homebrew/bin` is missing, add it to the `<string>` value inside the `PATH` key in the plist. Then reload:
-```bash
-launchctl unload ~/Library/LaunchAgents/com.nanoclaw.plist
-launchctl load ~/Library/LaunchAgents/com.nanoclaw.plist
-```
-
-### Build and restart
-
-```bash
-pnpm run build
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw
-```
-
-### Test
-
-Send a voice note in any registered group. The agent should receive it as `[Voice: <transcript>]`.
-
-### Check logs
-
-```bash
-tail -f logs/nanoclaw.log | grep -i -E "voice|transcri|whisper"
-```
-
-Look for:
-- `Transcribed voice message` — successful transcription
-- `whisper.cpp transcription failed` — check model path, ffmpeg, or PATH
-
-## Configuration
-
-Environment variables (optional, set in `.env`):
-
-| Variable | Default | Description |
-|----------|---------|-------------|
-| `WHISPER_BIN` | `whisper-cli` | Path to whisper.cpp binary |
-| `WHISPER_MODEL` | `data/models/ggml-base.bin` | Path to GGML model file |
-
-## Troubleshooting
-
-**"whisper.cpp transcription failed"**: Ensure both `whisper-cli` and `ffmpeg` are in PATH. The launchd service uses a restricted PATH — see Phase 3 above. Test manually:
-```bash
-ffmpeg -f lavfi -i anullsrc=r=16000:cl=mono -t 1 -f wav /tmp/test.wav -y
-whisper-cli -m data/models/ggml-base.bin -f /tmp/test.wav --no-timestamps -nt
-```
-
-**Transcription works in dev but not as service**: The launchd plist PATH likely doesn't include `/opt/homebrew/bin`. See "Ensure launchd PATH includes Homebrew" in Phase 3.
-
-**Slow transcription**: The base model processes ~30s of audio in <1s on M1+. If slower, check CPU usage — another process may be competing.
-
-**Wrong language**: whisper.cpp auto-detects language. To force a language, you can set `WHISPER_LANG` and modify `src/transcription.ts` to pass `-l $WHISPER_LANG`.
diff --git a/.github/workflows/label-pr.yml b/.github/workflows/label-pr.yml
index bec9d3e..ebfe3f3 100644
--- a/.github/workflows/label-pr.yml
+++ b/.github/workflows/label-pr.yml
@@ -1,7 +1,12 @@
 name: Label PR
 
+# SECURITY: this workflow runs with write access to the base repo on fork PRs,
+# because `pull_request_target` executes in the context of the base branch.
+# Keep it metadata-only — do NOT add actions/checkout or any step that
+# executes PR-supplied content (install scripts, build commands, etc.).
+# See https://securitylab.github.com/resources/github-actions-preventing-pwn-requests/
 on:
-  pull_request:
+  pull_request_target:
     types: [opened, edited]
 
 jobs:
diff --git a/.gitignore b/.gitignore
index 1035745..8a57c51 100644
--- a/.gitignore
+++ b/.gitignore
@@ -11,18 +11,19 @@ store/
 data/
 logs/
 
-# Groups - only track base structure and specific CLAUDE.md files
+# Groups - per-installation state, not tracked
 groups/*
-!groups/main/
-!groups/global/
-groups/main/*
-groups/global/*
-!groups/main/CLAUDE.md
-!groups/global/CLAUDE.md
+
+# Composer-managed CLAUDE.md artifacts (regenerated every spawn) and
+# per-group memory (CLAUDE.local.md) must never be committed.
+**/CLAUDE.local.md
+**/.claude-shared.md
+**/.claude-fragments/
 
 # Secrets
 *.keys.json
 .env
+.env*
 
 # Temp files
 .tmp-*
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2503be7..ab2fd5a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,21 @@ All notable changes to NanoClaw will be documented in this file.
 
 For detailed release notes, see the [full changelog on the documentation site](https://docs.nanoclaw.dev/changelog).
 
+## [2.0.0] - 2026-04-22
+
+Major version. NanoClaw v2 is a substantial architectural rewrite. Existing forks should run `/migrate-nanoclaw` (clean-base replay of customizations) or `/update-nanoclaw` (selective cherry-pick) before resuming work.
+
+- [BREAKING] **New entity model.** Users, roles (owner/admin), messaging groups, and agent groups are now tracked as separate entities, wired via `messaging_group_agents`. Privilege is user-level instead of channel-level, so the old "main channel = admin" concept is retired. See [docs/architecture.md](docs/architecture.md) and [docs/isolation-model.md](docs/isolation-model.md).
+- [BREAKING] **Two-DB session split.** Each session now has `inbound.db` (host writes, container reads) and `outbound.db` (container writes, host reads) with exactly one writer each. Replaces the single shared session DB and eliminates cross-mount SQLite contention. See [docs/db-session.md](docs/db-session.md).
+- [BREAKING] **Install flow replaced.** `bash nanoclaw.sh` is the new default: a scripted installer that hands off to Claude Code for error recovery and guided decisions. The `/setup` Claude-guided skill still works as an alternative.
+- [BREAKING] **Channels moved to the `channels` branch.** Trunk no longer ships Discord, Slack, Telegram, WhatsApp, iMessage, Teams, Linear, GitHub, WeChat, Matrix, Google Chat, Webex, Resend, or WhatsApp Cloud. Install them per fork via `/add-<channel>` skills, which copy from the `channels` branch. `/update-nanoclaw` will re-install the channels your fork had.
+- [BREAKING] **Alternative providers moved to the `providers` branch.** OpenCode, Codex, and Ollama install via `/add-opencode`, `/add-codex`, `/add-ollama-provider`. Claude remains the default provider baked into trunk.
+- [BREAKING] **Three-level channel isolation.** Wire channels to their own agent (separate agent groups), share an agent with independent conversations (`session_mode: 'shared'`), or merge channels into one shared session (`session_mode: 'agent-shared'`). Chosen per channel via `/manage-channels`.
+- [BREAKING] **Apple Container removed from default setup.** Still available as an opt-in via `/convert-to-apple-container`.
+- **Shared-source agent-runner.** Per-group `agent-runner-src/` overlays are gone; all groups mount the same agent-runner read-only. Per-group customization flows through composed `CLAUDE.md` (shared base + per-group fragments).
+- **Agent-runner runtime moved from Node to Bun.** Container image is self-contained; no host-side impact. Host remains on Node + pnpm.
+- **OneCLI Agent Vault is the sole credential path.** Containers never receive raw API keys; credentials are injected at request time.
+
 ## [1.2.36] - 2026-03-26
 
 - [BREAKING] Replaced pino logger with built-in logger. WhatsApp users must re-merge the WhatsApp fork to pick up the Baileys logger compatibility fix: `git fetch whatsapp main && git merge whatsapp/main`. If the `whatsapp` remote is not configured: `git remote add whatsapp https://github.com/qwibitai/nanoclaw-whatsapp.git`.
diff --git a/CLAUDE.md b/CLAUDE.md
index ec06a64..7115c4c 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -1,3 +1,18 @@
+# ⚠️ STOP — READ THIS FIRST IF YOU ARE CLAUDE ⚠️
+
+**If you are reading this because you just ran `git pull`, `git merge`, `git fetch && git merge`, or any equivalent to bring in upstream changes — and you see merge conflicts or a large diff involving this file — HALT IMMEDIATELY.**
+
+This is NanoClaw v2, a ground-up rewrite with breaking changes throughout. It cannot be merged into an existing v1 install. Attempting to resolve the conflicts by hand, run builds, or "fix" anything will corrupt the user's install and burn tokens for no result.
+
+**Do this instead:**
+1. Abort the merge: `git merge --abort` (or `git reset --hard ORIG_HEAD` if the merge already completed).
+2. Tell the user, verbatim: *"This is the v2 rewrite — it can't be merged into your existing install. Run `bash migrate-v2.sh` to migrate instead."*
+3. Wait for the user to confirm before doing anything else. Do not attempt the migration yourself.
+
+If you are a fresh install (you ran `git clone`, not `git pull`) and there are no conflicts, ignore this banner and continue below.
+
+---
+
 # NanoClaw
 
 Personal Claude assistant. See [README.md](README.md) for philosophy and setup. Architecture lives in `docs/`.
@@ -49,7 +64,9 @@ Exactly one writer per file — no cross-mount lock contention. Heartbeat is a f
 | `src/session-manager.ts` | Resolves sessions; opens `inbound.db` / `outbound.db`; manages heartbeat path |
 | `src/container-runner.ts` | Spawns per-agent-group Docker containers with session DB + outbox mounts, OneCLI `ensureAgent` |
 | `src/container-runtime.ts` | Runtime selection (Docker vs Apple containers), orphan cleanup |
-| `src/access.ts` | `pickApprover`, `pickApprovalDelivery`, admin resolution for `NANOCLAW_ADMIN_USER_IDS` |
+| `src/modules/permissions/access.ts` | `canAccessAgentGroup` — owner / global admin / scoped admin / member resolution against `user_roles` + `agent_group_members` |
+| `src/modules/approvals/primitive.ts` | `pickApprover`, `pickApprovalDelivery`, `requestApproval`, approval-handler registry |
+| `src/command-gate.ts` | Router-side admin command gate — queries `user_roles` directly (no env var, no container-side check) |
 | `src/onecli-approvals.ts` | OneCLI credentialed-action approval bridge |
 | `src/user-dm.ts` | Cold-DM resolution + `user_dms` cache |
 | `src/group-init.ts` | Per-agent-group filesystem scaffold (CLAUDE.md, skills, agent-runner-src overlay) |
@@ -74,7 +91,7 @@ Each `/add-<name>` skill is idempotent: `git fetch origin <branch>` → copy mod
 
 One tier of agent self-modification today:
 
-1. **`install_packages` / `add_mcp_server` / `request_rebuild`** — changes to the per-agent-group container config only (apt/npm deps, wire an existing MCP server). Admin approval, rebuild, container restart. `container/agent-runner/src/mcp-tools/self-mod.ts`.
+1. **`install_packages` / `add_mcp_server`** — changes to the per-agent-group container config only (apt/npm deps, wire an existing MCP server). Single admin approval per request; on approve, the handler in `src/modules/self-mod/apply.ts` rebuilds the image when needed (`install_packages` only) and restarts the container. `container/agent-runner/src/mcp-tools/self-mod.ts`.
 
 A second tier (direct source-level self-edits via a draft/activate flow) is planned but not yet implemented.
 
@@ -82,6 +99,41 @@ A second tier (direct source-level self-edits via a draft/activate flow) is plan
 
 API keys, OAuth tokens, and auth credentials are managed by the OneCLI gateway. Secrets are injected into per-agent containers at request time — none are passed in env vars or through chat context. `src/onecli-approvals.ts`, `ensureAgent()` in `container-runner.ts`. Run `onecli --help`.
 
+### Gotcha: auto-created agents start in `selective` secret mode
+
+When the host first spawns a session for a new agent group, `container-runner.ts:385` calls `onecli.ensureAgent({ name, identifier })`. The OneCLI `POST /api/agents` endpoint creates the agent in **`selective`** secret mode — meaning **no secrets are assigned to it by default**, even if the secrets exist in the vault and have host patterns that would otherwise match.
+
+Symptom: container starts, the proxy + CA cert are wired correctly, but the agent gets `401 Unauthorized` (or similar) from APIs whose credentials *are* in the vault. The credential just isn't in this agent's allow-list.
+
+The SDK does not expose `setSecretMode` — the only fix is the CLI (or the web UI at `http://127.0.0.1:10254`).
+
+```bash
+# Find the agent (identifier is the agent group id)
+onecli agents list
+
+# Flip to "all" so every vault secret with a matching host pattern gets injected
+onecli agents set-secret-mode --id <agent-id> --mode all
+
+# Or, stay selective and assign specific secrets
+onecli secrets list                                    # find secret ids
+onecli agents set-secrets --id <agent-id> --secret-ids <id1>,<id2>
+
+# Inspect what an agent currently has
+onecli agents secrets --id <agent-id>                  # secrets assigned to this agent
+onecli secrets list                                    # all vault secrets (with host patterns)
+```
+
+If you've just enabled `mode all`, no container restart is needed — the gateway looks up secrets per request, so the next API call from the running container will see the new credentials.
+
+### Requiring approval for credential use
+
+Approval-gating credentialed actions is a **two-sided** flow:
+
+- **Server-side** (OneCLI gateway): decides *when* to hold a request and emit a pending approval. As of `onecli@1.3.0`, the CLI does **not** expose this — `rules create --action` only accepts `block` or `rate_limit`, and `secrets create` has no approval flag. Approval policies must be configured via the OneCLI web UI at `http://127.0.0.1:10254`. If/when the CLI grows an `approve` action, this section needs updating.
+- **Host-side** (nanoclaw): receives pending approvals and routes them to a human. `src/modules/approvals/onecli-approvals.ts` registers a callback via `onecli.configureManualApproval(cb)` (long-polls `GET /api/approvals/pending`). The callback uses `pickApprover` + `pickApprovalDelivery` from `src/modules/approvals/primitive.ts` to DM an approver. Approvers are resolved from the `user_roles` table — preference order: scoped admins for the agent group → global admins → owners. There is no env var like `NANOCLAW_ADMIN_USER_IDS`; roles are persisted in the central DB only.
+
+If approvals are configured server-side but the host callback isn't running (or throws), every credentialed call hangs until the gateway times out. Conversely, if the gateway has no rule asking for approval, the host callback never fires regardless of how it's wired.
+
 ## Skills
 
 Four types of skills. See [CONTRIBUTING.md](CONTRIBUTING.md) for the full taxonomy.
@@ -157,7 +209,6 @@ This project uses pnpm with `minimumReleaseAge: 4320` (3 days) in `pnpm-workspac
 | [docs/agent-runner-details.md](docs/agent-runner-details.md) | Agent-runner internals + MCP tool interface |
 | [docs/isolation-model.md](docs/isolation-model.md) | Three-level channel isolation model |
 | [docs/setup-wiring.md](docs/setup-wiring.md) | What's wired, what's open in the setup flow |
-| [docs/checklist.md](docs/checklist.md) | Rolling status checklist across all subsystems |
 | [docs/architecture-diagram.md](docs/architecture-diagram.md) | Diagram version of the architecture |
 | [docs/build-and-runtime.md](docs/build-and-runtime.md) | Runtime split (Node host + Bun container), lockfiles, image build surface, CI, key invariants |
 
diff --git a/README.md b/README.md
index 874a8d7..de61f6d 100644
--- a/README.md
+++ b/README.md
@@ -12,7 +12,7 @@
   <a href="README_zh.md">中文</a>&nbsp; • &nbsp;
   <a href="README_ja.md">日本語</a>&nbsp; • &nbsp;
   <a href="https://discord.gg/VDdww8qS42"><img src="https://img.shields.io/discord/1470188214710046894?label=Discord&logo=discord&v=2" alt="Discord" valign="middle"></a>&nbsp; • &nbsp;
-  <a href="repo-tokens"><img src="repo-tokens/badge.svg" alt="34.9k tokens, 17% of context window" valign="middle"></a>
+  <a href="repo-tokens"><img src="repo-tokens/badge.svg" alt="repo tokens" valign="middle"></a>
 </p>
 
 ---
@@ -26,55 +26,38 @@ NanoClaw provides that same core functionality, but in a codebase small enough t
 ## Quick Start
 
 ```bash
-gh repo fork qwibitai/nanoclaw --clone
-cd nanoclaw
-claude
+git clone https://github.com/qwibitai/nanoclaw.git nanoclaw-v2
+cd nanoclaw-v2
+bash nanoclaw.sh
 ```
 
-<details>
-<summary>Without GitHub CLI</summary>
-
-1. Fork [qwibitai/nanoclaw](https://github.com/qwibitai/nanoclaw) on GitHub (click the Fork button)
-2. `git clone https://github.com/<your-username>/nanoclaw.git`
-3. `cd nanoclaw`
-4. `claude`
-
-</details>
-
-Then run `/setup`. Claude Code handles everything: dependencies, authentication, container setup and service configuration.
-
-> **Note:** Commands prefixed with `/` (like `/setup`, `/add-whatsapp`) are [Claude Code skills](https://code.claude.com/docs/en/skills). Type them inside the `claude` CLI prompt, not in your regular terminal. If you don't have Claude Code installed, get it at [claude.com/product/claude-code](https://claude.com/product/claude-code).
+`nanoclaw.sh` walks you from a fresh machine to a named agent you can message. It installs Node, pnpm, and Docker if missing, registers your Anthropic credential with OneCLI, builds the agent container, and pairs your first channel (Telegram, Discord, WhatsApp, or a local CLI). If a step fails, Claude Code is invoked automatically to diagnose and resume from where it broke.
 
 ## Philosophy
 
 **Small enough to understand.** One process, a few source files and no microservices. If you want to understand the full NanoClaw codebase, just ask Claude Code to walk you through it.
 
-**Secure by isolation.** Agents run in Linux containers (Apple Container on macOS, or Docker) and they can only see what's explicitly mounted. Bash access is safe because commands run inside the container, not on your host.
+**Secure by isolation.** Agents run in Linux containers and they can only see what's explicitly mounted. Bash access is safe because commands run inside the container, not on your host.
 
 **Built for the individual user.** NanoClaw isn't a monolithic framework; it's software that fits each user's exact needs. Instead of becoming bloatware, NanoClaw is designed to be bespoke. You make your own fork and have Claude Code modify it to match your needs.
 
 **Customization = code changes.** No configuration sprawl. Want different behavior? Modify the code. The codebase is small enough that it's safe to make changes.
 
-**AI-native.**
-- No installation wizard; Claude Code guides setup.
-- No monitoring dashboard; ask Claude what's happening.
-- No debugging tools; describe the problem and Claude fixes it.
+**AI-native, hybrid by design.** The install and onboarding flow is an optimized scripted path, fast and deterministic. When a step needs judgment, whether a failed install, a guided decision, or a customization, control hands off to Claude Code seamlessly. Beyond setup there's no monitoring dashboard or debugging UI either: describe the problem in chat and Claude Code handles it.
 
-**Skills over features.** Instead of adding features (e.g. support for Telegram) to the codebase, contributors submit [claude code skills](https://code.claude.com/docs/en/skills) like `/add-telegram` that transform your fork. You end up with clean code that does exactly what you need.
+**Skills over features.** Trunk ships the registry and infrastructure, not specific channel adapters or alternative agent providers. Channels (Discord, Slack, Telegram, WhatsApp, …) live on a long-lived `channels` branch; alternative providers (OpenCode, Ollama) live on `providers`. You run `/add-telegram`, `/add-opencode`, etc. and the skill copies exactly the module(s) you need into your fork. No feature you didn't ask for.
 
-**Best harness, best model.** NanoClaw runs on the Claude Agent SDK, which means you're running Claude Code directly. Claude Code is highly capable and its coding and problem-solving capabilities allow it to modify and expand NanoClaw and tailor it to each user.
+**Best harness, best model.** NanoClaw natively uses Claude Code via Anthropic's official Claude Agent SDK, so you get the latest Claude models and Claude Code's full toolset, including the ability to modify and expand your own NanoClaw fork. Other providers are drop-in options: `/add-codex` for OpenAI's Codex (ChatGPT subscription or API key), `/add-opencode` for OpenRouter, Google, DeepSeek and more via OpenCode, and `/add-ollama-provider` for local open-weight models. Provider is configurable per agent group.
 
 ## What It Supports
 
-- **Multi-channel messaging** - Talk to your assistant from WhatsApp, Telegram, Discord, Slack, or Gmail. Add channels with skills like `/add-whatsapp` or `/add-telegram`. Run one or many at the same time.
-- **Isolated group context** - Each group has its own `CLAUDE.md` memory, isolated filesystem, and runs in its own container sandbox with only that filesystem mounted to it.
-- **Main channel** - Your private channel (self-chat) for admin control; every group is completely isolated
-- **Scheduled tasks** - Recurring jobs that run Claude and can message you back
-- **Web access** - Search and fetch content from the Web
-- **Container isolation** - Agents are sandboxed in Docker (macOS/Linux), [Docker Sandboxes](docs/docker-sandboxes.md) (micro VM isolation), or Apple Container (macOS)
-- **Credential security** - Agents never hold raw API keys. Outbound requests route through [OneCLI's Agent Vault](https://github.com/onecli/onecli), which injects credentials at request time and enforces per-agent policies and rate limits.
-- **Agent Swarms** - Spin up teams of specialized agents that collaborate on complex tasks
-- **Optional integrations** - Add Gmail (`/add-gmail`) and more via skills
+- **Multi-channel messaging** — WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, Linear, GitHub, WeChat, and email via Resend. Installed on demand with `/add-<channel>` skills. Run one or many at the same time.
+- **Flexible isolation** — connect each channel to its own agent for full privacy, share one agent across many channels for unified memory with separate conversations, or fold multiple channels into a single shared session so one conversation spans many surfaces. Pick per channel via `/manage-channels`. See [docs/isolation-model.md](docs/isolation-model.md).
+- **Per-agent workspace** — each agent group has its own `CLAUDE.md`, its own memory, its own container, and only the mounts you allow. Nothing crosses the boundary unless you wire it to.
+- **Scheduled tasks** — recurring jobs that run Claude and can message you back
+- **Web access** — search and fetch content from the web
+- **Container isolation** — agents are sandboxed in Docker (macOS/Linux/WSL2), with optional [Docker Sandboxes](docs/docker-sandboxes.md) micro-VM isolation or Apple Container as a macOS-native opt-in
+- **Credential security** — agents never hold raw API keys. Outbound requests route through [OneCLI's Agent Vault](https://github.com/onecli/onecli), which injects credentials at request time and enforces per-agent policies and rate limits.
 
 ## Usage
 
@@ -86,7 +69,7 @@ Talk to your assistant with the trigger word (default: `@Andy`):
 @Andy every Monday at 8am, compile news on AI developments from Hacker News and TechCrunch and message me a briefing
 ```
 
-From the main channel (your self-chat), you can manage groups and tasks:
+From a channel you own or administer, you can manage groups and tasks:
 ```
 @Andy list all scheduled tasks across groups
 @Andy pause the Monday briefing task
@@ -110,54 +93,58 @@ The codebase is small enough that Claude can safely modify it.
 
 **Don't add features. Add skills.**
 
-If you want to add Telegram support, don't create a PR that adds Telegram to the core codebase. Instead, fork NanoClaw, make the code changes on a branch, and open a PR. We'll create a `skill/telegram` branch from your PR that other users can merge into their fork.
+If you want to add a new channel or agent provider, don't add it to trunk. New channel adapters land on the `channels` branch; new agent providers land on `providers`. Users install them in their own fork with `/add-<name>` skills, which copy the relevant module(s) into the standard paths, wire the registration, and pin dependencies.
 
-Users then run `/add-telegram` on their fork and get clean code that does exactly what they need, not a bloated system trying to support every use case.
+This keeps trunk as pure registry and infra, and every fork stays lean — users get the channels and providers they asked for and nothing else.
 
 ### RFS (Request for Skills)
 
 Skills we'd like to see:
 
 **Communication Channels**
-- `/add-signal` - Add Signal as a channel
+- `/add-signal` — Add Signal as a channel
 
 ## Requirements
 
-- macOS, Linux, or Windows (via WSL2)
-- Node.js 20+
-- [Claude Code](https://claude.ai/download)
-- [Apple Container](https://github.com/apple/container) (macOS) or [Docker](https://docker.com/products/docker-desktop) (macOS/Linux)
+- macOS or Linux (Windows via WSL2)
+- Node.js 20+ and pnpm 10+ (the installer will install both if missing)
+- [Docker Desktop](https://docker.com/products/docker-desktop) (macOS/Windows) or Docker Engine (Linux)
+- [Claude Code](https://claude.ai/download) for `/customize`, `/debug`, error recovery during setup, and all `/add-<channel>` skills
 
 ## Architecture
 
 ```
-Channels --> SQLite --> Polling loop --> Container (Claude Agent SDK) --> Response
+messaging apps → host process (router) → inbound.db → container (Bun, Claude Agent SDK) → outbound.db → host process (delivery) → messaging apps
 ```
 
-Single Node.js process. Channels are added via skills and self-register at startup — the orchestrator connects whichever ones have credentials present. Agents execute in isolated Linux containers with filesystem isolation. Only mounted directories are accessible. Per-group message queue with concurrency control. IPC via filesystem.
+A single Node host orchestrates per-session agent containers. When a message arrives, the host routes it via the entity model (user → messaging group → agent group → session), writes it to the session's `inbound.db`, and wakes the container. The agent-runner inside the container polls `inbound.db`, runs Claude, and writes responses to `outbound.db`. The host polls `outbound.db` and delivers back through the channel adapter.
 
-For the full architecture details, see the [documentation site](https://docs.nanoclaw.dev/concepts/architecture).
+Two SQLite files per session, each with exactly one writer — no cross-mount contention, no IPC, no stdin piping. Channels and alternative providers self-register at startup; trunk ships the registry and the Chat SDK bridge, while the adapters themselves are skill-installed per fork.
+
+For the full architecture writeup see [docs/architecture.md](docs/architecture.md); for the three-level isolation model see [docs/isolation-model.md](docs/isolation-model.md).
 
 Key files:
-- `src/index.ts` - Orchestrator: state, message loop, agent invocation
-- `src/channels/registry.ts` - Channel registry (self-registration at startup)
-- `src/ipc.ts` - IPC watcher and task processing
-- `src/router.ts` - Message formatting and outbound routing
-- `src/group-queue.ts` - Per-group queue with global concurrency limit
-- `src/container-runner.ts` - Spawns streaming agent containers
-- `src/task-scheduler.ts` - Runs scheduled tasks
-- `src/db.ts` - SQLite operations (messages, groups, sessions, state)
-- `groups/*/CLAUDE.md` - Per-group memory
+- `src/index.ts` — entry point: DB init, channel adapters, delivery polls, sweep
+- `src/router.ts` — inbound routing: messaging group → agent group → session → `inbound.db`
+- `src/delivery.ts` — polls `outbound.db`, delivers via adapter, handles system actions
+- `src/host-sweep.ts` — 60s sweep: stale detection, due-message wake, recurrence
+- `src/session-manager.ts` — resolves sessions, opens `inbound.db` / `outbound.db`
+- `src/container-runner.ts` — spawns per-agent-group containers, OneCLI credential injection
+- `src/db/` — central DB (users, roles, agent groups, messaging groups, wiring, migrations)
+- `src/channels/` — channel adapter infra (adapters installed via `/add-<channel>` skills)
+- `src/providers/` — host-side provider config (`claude` baked in; others via skills)
+- `container/agent-runner/` — Bun agent-runner: poll loop, MCP tools, provider abstraction
+- `groups/<folder>/` — per-agent-group filesystem (`CLAUDE.md`, skills, container config)
 
 ## FAQ
 
 **Why Docker?**
 
-Docker provides cross-platform support (macOS, Linux and even Windows via WSL2) and a mature ecosystem. On macOS, you can optionally switch to Apple Container via `/convert-to-apple-container` for a lighter-weight native runtime. For additional isolation, [Docker Sandboxes](docs/docker-sandboxes.md) run each container inside a micro VM.
+Docker provides cross-platform support (macOS, Linux and Windows via WSL2) and a mature ecosystem. On macOS, you can optionally switch to Apple Container via `/convert-to-apple-container` for a lighter-weight native runtime. For additional isolation, [Docker Sandboxes](docs/docker-sandboxes.md) run each container inside a micro VM.
 
 **Can I run this on Linux or Windows?**
 
-Yes. Docker is the default runtime and works on macOS, Linux, and Windows (via WSL2). Just run `/setup`.
+Yes. Docker is the default runtime and works on macOS, Linux, and Windows (via WSL2). Just run `bash nanoclaw.sh`.
 
 **Is this secure?**
 
@@ -169,33 +156,28 @@ We don't want configuration sprawl. Every user should customize NanoClaw so that
 
 **Can I use third-party or open-source models?**
 
-Yes. NanoClaw supports any Claude API-compatible model endpoint. Set these environment variables in your `.env` file:
+Yes. The supported path is `/add-opencode` (OpenRouter, OpenAI, Google, DeepSeek, and more via OpenCode config) or `/add-ollama-provider` (local open-weight models via Ollama). Both are configurable per agent group, so different agents can run on different backends in the same install.
+
+For one-off experiments, any Claude API-compatible endpoint also works via `.env`:
 
 ```bash
 ANTHROPIC_BASE_URL=https://your-api-endpoint.com
 ANTHROPIC_AUTH_TOKEN=your-token-here
 ```
 
-This allows you to use:
-- Local models via [Ollama](https://ollama.ai) with an API proxy
-- Open-source models hosted on [Together AI](https://together.ai), [Fireworks](https://fireworks.ai), etc.
-- Custom model deployments with Anthropic-compatible APIs
-
-Note: The model must support the Anthropic API format for best compatibility.
-
 **How do I debug issues?**
 
 Ask Claude Code. "Why isn't the scheduler running?" "What's in the recent logs?" "Why did this message not get a response?" That's the AI-native approach that underlies NanoClaw.
 
 **Why isn't the setup working for me?**
 
-If you have issues, during setup, Claude will try to dynamically fix them. If that doesn't work, run `claude`, then run `/debug`. If Claude finds an issue that is likely affecting other users, open a PR to modify the setup SKILL.md.
+If a step fails, `nanoclaw.sh` hands off to Claude Code to diagnose and resume. If that doesn't resolve it, run `claude`, then `/debug`. If Claude identifies an issue likely to affect other users, open a PR against the relevant setup step or skill.
 
 **What changes will be accepted into the codebase?**
 
 Only security fixes, bug fixes, and clear improvements will be accepted to the base configuration. That's all.
 
-Everything else (new capabilities, OS compatibility, hardware support, enhancements) should be contributed as skills.
+Everything else (new capabilities, OS compatibility, hardware support, enhancements) should be contributed as skills on the `channels` or `providers` branch.
 
 This keeps the base system minimal and lets every user customize their installation without inheriting features they don't want.
 
diff --git a/README_ja.md b/README_ja.md
index 5c3f648..947db95 100644
--- a/README_ja.md
+++ b/README_ja.md
@@ -8,90 +8,56 @@
 
 <p align="center">
   <a href="https://nanoclaw.dev">nanoclaw.dev</a>&nbsp; • &nbsp;
+  <a href="https://docs.nanoclaw.dev">ドキュメント</a>&nbsp; • &nbsp;
   <a href="README.md">English</a>&nbsp; • &nbsp;
   <a href="README_zh.md">中文</a>&nbsp; • &nbsp;
   <a href="https://discord.gg/VDdww8qS42"><img src="https://img.shields.io/discord/1470188214710046894?label=Discord&logo=discord&v=2" alt="Discord" valign="middle"></a>&nbsp; • &nbsp;
-  <a href="repo-tokens"><img src="repo-tokens/badge.svg" alt="34.9k tokens, 17% of context window" valign="middle"></a>
+  <a href="repo-tokens"><img src="repo-tokens/badge.svg" alt="repo tokens" valign="middle"></a>
 </p>
 
 ---
 
-<h2 align="center">🐳 Dockerサンドボックスで動作</h2>
-<p align="center">各エージェントはマイクロVM内の独立したコンテナで実行されます。<br>ハイパーバイザーレベルの分離。ミリ秒で起動。複雑なセットアップ不要。</p>
-
-**macOS (Apple Silicon)**
-```bash
-curl -fsSL https://nanoclaw.dev/install-docker-sandboxes.sh | bash
-```
-
-**Windows (WSL)**
-```bash
-curl -fsSL https://nanoclaw.dev/install-docker-sandboxes-windows.sh | bash
-```
-
-> 現在、macOS（Apple Silicon）とWindows（x86）に対応しています。Linux対応は近日公開予定。
-
-<p align="center"><a href="https://nanoclaw.dev/blog/nanoclaw-docker-sandboxes">発表記事を読む →</a>&nbsp; · &nbsp;<a href="docs/docker-sandboxes.md">手動セットアップガイド →</a></p>
-
----
-
 ## NanoClawを作った理由
 
-[OpenClaw](https://github.com/openclaw/openclaw)は素晴らしいプロジェクトですが、理解しきれない複雑なソフトウェアに自分の生活へのフルアクセスを与えたまま安心して眠れるとは思えませんでした。OpenClawは約50万行のコード、53の設定ファイル、70以上の依存関係を持っています。セキュリティはアプリケーションレベル（許可リスト、ペアリングコード）であり、真のOS レベルの分離ではありません。すべてが共有メモリを持つ1つのNodeプロセスで動作します。
+[OpenClaw](https://github.com/openclaw/openclaw)は素晴らしいプロジェクトですが、自分が理解しきれない複雑なソフトウェアに生活へのフルアクセスを与えたまま安心して眠れるとは思えませんでした。OpenClawは約50万行のコード、53の設定ファイル、70以上の依存関係を持っています。セキュリティはアプリケーションレベル（許可リスト、ペアリングコード）であり、真のOSレベルの分離ではありません。すべてが共有メモリを持つ1つのNodeプロセスで動作します。
 
-NanoClawは同じコア機能を提供しますが、理解できる規模のコードベースで実現しています：1つのプロセスと少数のファイル。Claudeエージェントは単なるパーミッションチェックの背後ではなく、ファイルシステム分離された独自のLinuxコンテナで実行されます。
+NanoClawは同じコア機能を提供しますが、理解できる規模のコードベースで実現しています。1つのプロセスと少数のファイル。Claudeエージェントは単なるパーミッションチェックの背後ではなく、ファイルシステム分離された独自のLinuxコンテナで実行されます。
 
 ## クイックスタート
 
 ```bash
-gh repo fork qwibitai/nanoclaw --clone
-cd nanoclaw
-claude
+git clone https://github.com/qwibitai/nanoclaw.git nanoclaw-v2
+cd nanoclaw-v2
+bash nanoclaw.sh
 ```
 
-<details>
-<summary>GitHub CLIなしの場合</summary>
-
-1. GitHub上で[qwibitai/nanoclaw](https://github.com/qwibitai/nanoclaw)をフォーク（Forkボタンをクリック）
-2. `git clone https://github.com/<あなたのユーザー名>/nanoclaw.git`
-3. `cd nanoclaw`
-4. `claude`
-
-</details>
-
-その後、`/setup`を実行します。Claude Codeがすべてを処理します：依存関係、認証、コンテナセットアップ、サービス設定。
-
-> **注意:** `/`で始まるコマンド（`/setup`、`/add-whatsapp`など）は[Claude Codeスキル](https://code.claude.com/docs/en/skills)です。通常のターミナルではなく、`claude` CLIプロンプト内で入力してください。Claude Codeをインストールしていない場合は、[claude.com/product/claude-code](https://claude.com/product/claude-code)から入手してください。
+`nanoclaw.sh`は、まっさらなマシンから、メッセージを送れる名前付きエージェントが動く状態までを一気通貫で案内します。NodeやpnpmやDockerが無ければインストールし、AnthropicクレデンシャルをOneCLIに登録し、エージェントコンテナをビルドし、最初のチャネル（Telegram、Discord、WhatsApp、またはローカルCLI）とペアリングします。途中でステップが失敗すれば自動的にClaude Codeが呼び出され、原因を診断して中断箇所から再開します。
 
 ## 設計思想
 
-**理解できる規模。** 1つのプロセス、少数のソースファイル、マイクロサービスなし。NanoClawのコードベース全体を理解したい場合は、Claude Codeに説明を求めるだけです。
+**理解できる規模。** 1つのプロセス、少数のソースファイル、マイクロサービスなし。NanoClawのコードベース全体を把握したいなら、Claude Codeに説明を求めれば十分です。
 
-**分離によるセキュリティ。** エージェントはLinuxコンテナ（macOSではApple Container、またはDocker）で実行され、明示的にマウントされたものだけが見えます。コマンドはホストではなくコンテナ内で実行されるため、Bashアクセスは安全です。
+**分離によるセキュリティ。** エージェントはLinuxコンテナで実行され、明示的にマウントされたものだけが見えます。コマンドはホストではなくコンテナ内で実行されるため、Bashアクセスも安全です。
 
-**個人ユーザー向け。** NanoClawはモノリシックなフレームワークではなく、各ユーザーのニーズに正確にフィットするソフトウェアです。肥大化するのではなく、オーダーメイドになるよう設計されています。自分のフォークを作成し、Claude Codeにニーズに合わせて変更させます。
+**個人ユーザー向け。** NanoClawはモノリシックなフレームワークではなく、各ユーザーのニーズに正確にフィットするソフトウェアです。肥大化するのではなく、オーダーメイドであるよう設計されています。自分のフォークを作り、Claude Codeにニーズに合わせて変更させます。
 
-**カスタマイズ＝コード変更。** 設定ファイルの肥大化なし。動作を変えたい？コードを変更するだけ。コードベースは変更しても安全な規模です。
+**カスタマイズ＝コード変更。** 設定の肥大化はありません。動作を変えたいならコードを変える。コードベースは変更しても安全な規模です。
 
-**AIネイティブ。**
-- インストールウィザードなし — Claude Codeがセットアップを案内。
-- モニタリングダッシュボードなし — Claudeに状況を聞くだけ。
-- デバッグツールなし — 問題を説明すればClaudeが修正。
+**AIネイティブ、設計としてハイブリッド。** インストールとオンボーディングは最適化されたスクリプトのパスで、速く決定的です。判断が必要なところ（インストール失敗、対話的な決定、カスタマイズ）では、制御はシームレスにClaude Codeへ渡されます。セットアップ以降も、監視ダッシュボードやデバッグUIは用意しません。問題をチャットで説明すれば、Claude Codeが処理します。
 
-**機能追加ではなくスキル。** コードベースに機能（例：Telegram対応）を追加する代わりに、コントリビューターは`/add-telegram`のような[Claude Codeスキル](https://code.claude.com/docs/en/skills)を提出し、あなたのフォークを変換します。あなたが必要なものだけを正確に実行するクリーンなコードが手に入ります。
+**機能ではなくスキル。** トランクにはレジストリとインフラのみを同梱し、個別のチャネルアダプターや代替プロバイダーは含めません。チャネル（Discord、Slack、Telegram、WhatsAppなど）は長期運用される`channels`ブランチに、代替プロバイダー（OpenCode、Ollama）は`providers`ブランチに置かれます。`/add-telegram`や`/add-opencode`などを実行すると、スキルが必要なモジュールだけを正確にフォークへコピーします。要求していない機能は一切入りません。
 
-**最高のハーネス、最高のモデル。** NanoClawはClaude Agent SDK上で動作します。つまり、Claude Codeを直接実行しているということです。Claude Codeは高い能力を持ち、そのコーディングと問題解決能力によってNanoClawを変更・拡張し、各ユーザーに合わせてカスタマイズできます。
+**最高のハーネス、最高のモデル。** NanoClawはAnthropic公式のClaude Agent SDK経由でネイティブにClaude Codeを使用します。最新のClaudeモデルとClaude Codeの全ツールセット（自分のNanoClawフォークを変更・拡張する能力を含む）が手に入ります。他プロバイダーはドロップイン・オプションです。OpenAIのCodex（ChatGPTサブスクリプションまたはAPIキー）向けには`/add-codex`、OpenCode経由のOpenRouter、Google、DeepSeekなどには`/add-opencode`、ローカルのオープンウェイトモデルには`/add-ollama-provider`。プロバイダーはエージェントグループごとに設定可能です。
 
 ## サポート機能
 
-- **マルチチャネルメッセージング** - WhatsApp、Telegram、Discord、Slack、Gmailからアシスタントと会話。`/add-whatsapp`や`/add-telegram`などのスキルでチャネルを追加。1つでも複数でも同時に実行可能。
-- **グループごとの分離コンテキスト** - 各グループは独自の`CLAUDE.md`メモリ、分離されたファイルシステムを持ち、そのファイルシステムのみがマウントされた専用コンテナサンドボックスで実行。
-- **メインチャネル** - 管理制御用のプライベートチャネル（セルフチャット）。各グループは完全に分離。
-- **スケジュールタスク** - Claudeを実行し、メッセージを返せる定期ジョブ。
-- **Webアクセス** - Webからのコンテンツ検索・取得。
-- **コンテナ分離** - エージェントは[Dockerサンドボックス](https://nanoclaw.dev/blog/nanoclaw-docker-sandboxes)（マイクロVM分離）、Apple Container（macOS）、またはDocker（macOS/Linux）でサンドボックス化。
-- **エージェントスウォーム** - 複雑なタスクで協力する専門エージェントチームを起動。
-- **オプション連携** - Gmail（`/add-gmail`）などをスキルで追加。
+- **マルチチャネルメッセージング** — WhatsApp、Telegram、Discord、Slack、Microsoft Teams、iMessage、Matrix、Google Chat、Webex、Linear、GitHub、WeChat、Resend経由のメール。`/add-<channel>`スキルでオンデマンドにインストール。1つでも複数でも同時に実行可能。
+- **柔軟な分離モデル** — チャネルごとに専用エージェントを割り当てて完全プライバシーを確保することも、複数チャネルで1つのエージェントを共有して会話は分離しつつメモリを統一することも、複数チャネルを1つの共有セッションにまとめて会話を横断させることもできます。`/manage-channels`でチャネル単位に選択。[docs/isolation-model.md](docs/isolation-model.md)参照。
+- **エージェントごとのワークスペース** — 各エージェントグループは独自の`CLAUDE.md`、独自のメモリ、独自のコンテナ、そしてあなたが許可したマウントのみを持ちます。明示的に配線しない限り、境界を越えるものはありません。
+- **スケジュールタスク** — Claudeを実行し、結果を返信できる定期ジョブ。
+- **Webアクセス** — Webからの検索とコンテンツ取得。
+- **コンテナ分離** — エージェントはDockerでサンドボックス化されます（macOS/Linux/WSL2）。[Docker Sandboxes](docs/docker-sandboxes.md)によるマイクロVM分離や、macOSネイティブのオプトインとしてApple Containerも選択可能です。
+- **クレデンシャルのセキュリティ** — エージェントは生のAPIキーを保持しません。アウトバウンドリクエストは[OneCLI Agent Vault](https://github.com/onecli/onecli)を経由し、リクエスト時に認証情報を注入して、エージェントごとのポリシーとレート制限を適用します。
 
 ## 使い方
 
@@ -103,7 +69,7 @@ claude
 @Andy 毎週月曜の朝8時に、Hacker NewsとTechCrunchからAI関連のニュースをまとめてブリーフィングを送って
 ```
 
-メインチャネル（セルフチャット）から、グループやタスクを管理できます：
+所有または管理しているチャネルからは、グループやタスクを管理できます：
 ```
 @Andy 全グループのスケジュールタスクを一覧表示して
 @Andy 月曜のブリーフィングタスクを一時停止して
@@ -112,14 +78,14 @@ claude
 
 ## カスタマイズ
 
-NanoClawは設定ファイルを使いません。変更するには、Claude Codeに伝えるだけです：
+NanoClawは設定ファイルを使いません。変更したいときは、Claude Codeにやりたいことを伝えるだけです：
 
 - 「トリガーワードを@Bobに変更して」
 - 「今後はレスポンスをもっと短く直接的にして」
 - 「おはようと言ったらカスタム挨拶を追加して」
 - 「会話の要約を毎週保存して」
 
-または`/customize`を実行してガイド付きの変更を行えます。
+または`/customize`を実行すればガイド付きで変更できます。
 
 コードベースは十分に小さいため、Claudeが安全に変更できます。
 
@@ -127,105 +93,101 @@ NanoClawは設定ファイルを使いません。変更するには、Claude Co
 
 **機能を追加するのではなく、スキルを追加してください。**
 
-Telegram対応を追加したい場合、コアコードベースにTelegramを追加するPRを作成しないでください。代わりに、NanoClawをフォークし、ブランチでコード変更を行い、PRを開いてください。あなたのPRから`skill/telegram`ブランチを作成し、他のユーザーが自分のフォークにマージできるようにします。
+新しいチャネルやエージェントプロバイダーを追加したい場合、トランクには追加しないでください。新しいチャネルアダプターは`channels`ブランチに、新しいエージェントプロバイダーは`providers`ブランチに追加します。ユーザーはそれぞれのフォークで`/add-<name>`スキルを実行し、スキルが必要なモジュールを標準パスへコピーし、登録を配線し、依存関係をピン留めします。
 
-ユーザーは自分のフォークで`/add-telegram`を実行するだけで、あらゆるユースケースに対応しようとする肥大化したシステムではなく、必要なものだけを正確に実行するクリーンなコードが手に入ります。
+こうすることでトランクは純粋なレジストリ／インフラのまま保たれ、どのフォークもスリムなままです。ユーザーは求めたチャネルとプロバイダーだけを受け取り、それ以外は入りません。
 
 ### RFS（スキル募集）
 
-私たちが求めているスキル：
+私たちが見たいスキル：
 
 **コミュニケーションチャネル**
-- `/add-signal` - Signalをチャネルとして追加
-
-**セッション管理**
-- `/clear` - 会話をコンパクト化する`/clear`コマンドの追加（同一セッション内で重要な情報を保持しながらコンテキストを要約）。Claude Agent SDKを通じてプログラム的にコンパクト化をトリガーする方法の解明が必要。
+- `/add-signal` — Signalをチャネルとして追加
 
 ## 必要条件
 
-- macOSまたはLinux
-- Node.js 20以上
-- [Claude Code](https://claude.ai/download)
-- [Apple Container](https://github.com/apple/container)（macOS）または[Docker](https://docker.com/products/docker-desktop)（macOS/Linux）
+- macOSまたはLinux（WindowsはWSL2経由）
+- Node.js 20以上とpnpm 10以上（インストーラーが未インストールなら両方をインストールします）
+- [Docker Desktop](https://docker.com/products/docker-desktop)（macOS/Windows）または Docker Engine（Linux）
+- [Claude Code](https://claude.ai/download)（`/customize`、`/debug`、セットアップ時のエラー復旧、全ての`/add-<channel>`スキルで使用）
 
 ## アーキテクチャ
 
 ```
-チャネル --> SQLite --> ポーリングループ --> コンテナ（Claude Agent SDK） --> レスポンス
+メッセージングアプリ → ホストプロセス（ルーター） → inbound.db → コンテナ（Bun、Claude Agent SDK） → outbound.db → ホストプロセス（配信） → メッセージングアプリ
 ```
 
-単一のNode.jsプロセス。チャネルはスキルで追加され、起動時に自己登録します — オーケストレーターは認証情報が存在するチャネルを接続します。エージェントはファイルシステム分離された独立したLinuxコンテナで実行されます。マウントされたディレクトリのみアクセス可能。グループごとのメッセージキューと同時実行制御。ファイルシステム経由のIPC。
+単一のNodeホストがセッションごとのエージェントコンテナをオーケストレーションします。メッセージが到着すると、ホストはエンティティモデル（ユーザー → メッセージンググループ → エージェントグループ → セッション）に沿ってルーティングし、セッションの`inbound.db`に書き込み、コンテナを起こします。コンテナ内部のagent-runnerは`inbound.db`をポーリングしてClaudeを実行し、レスポンスを`outbound.db`に書き込みます。ホストは`outbound.db`をポーリングし、チャネルアダプターを通じて配信します。
 
-詳細なアーキテクチャについては、[docs/SPEC.md](docs/SPEC.md)を参照してください。
+セッションごとに2つのSQLiteファイル、各ファイルにライターは1つだけ — クロスマウントの競合なし、IPCなし、stdinパイプなし。チャネルと代替プロバイダーは起動時に自己登録します。トランクはレジストリとChat SDKブリッジを同梱し、アダプター本体はフォークごとにスキルでインストールされます。
+
+詳しいアーキテクチャ説明は[docs/architecture.md](docs/architecture.md)を、3階層の分離モデルについては[docs/isolation-model.md](docs/isolation-model.md)を参照してください。
 
 主要ファイル：
-- `src/index.ts` - オーケストレーター：状態、メッセージループ、エージェント呼び出し
-- `src/channels/registry.ts` - チャネルレジストリ（起動時の自己登録）
-- `src/ipc.ts` - IPCウォッチャーとタスク処理
-- `src/router.ts` - メッセージフォーマットとアウトバウンドルーティング
-- `src/group-queue.ts` - グローバル同時実行制限付きのグループごとのキュー
-- `src/container-runner.ts` - ストリーミングエージェントコンテナの起動
-- `src/task-scheduler.ts` - スケジュールタスクの実行
-- `src/db.ts` - SQLite操作（メッセージ、グループ、セッション、状態）
-- `groups/*/CLAUDE.md` - グループごとのメモリ
+- `src/index.ts` — エントリーポイント：DB初期化、チャネルアダプター、配信ポーリング、sweep
+- `src/router.ts` — インバウンドルーティング：メッセージンググループ → エージェントグループ → セッション → `inbound.db`
+- `src/delivery.ts` — `outbound.db`をポーリングし、アダプター経由で配信、システムアクションを処理
+- `src/host-sweep.ts` — 60秒ごとのsweep：ストール検出、期限到来メッセージの起動、繰り返し
+- `src/session-manager.ts` — セッションの解決、`inbound.db`と`outbound.db`のオープン
+- `src/container-runner.ts` — エージェントグループごとのコンテナ起動、OneCLIによるクレデンシャル注入
+- `src/db/` — セントラルDB（ユーザー、ロール、エージェントグループ、メッセージンググループ、配線、マイグレーション）
+- `src/channels/` — チャネルアダプターのインフラ（アダプターは`/add-<channel>`スキルでインストール）
+- `src/providers/` — ホスト側プロバイダー設定（`claude`はバンドル、その他はスキル経由）
+- `container/agent-runner/` — Bun製agent-runner：ポーリングループ、MCPツール、プロバイダー抽象化
+- `groups/<folder>/` — エージェントグループごとのファイルシステム（`CLAUDE.md`、スキル、コンテナ設定）
 
 ## FAQ
 
 **なぜDockerなのか？**
 
-Dockerはクロスプラットフォーム対応（macOS、Linux、さらにWSL2経由のWindows）と成熟したエコシステムを提供します。macOSでは、`/convert-to-apple-container`でオプションとしてApple Containerに切り替え、より軽量なネイティブランタイムを使用できます。
+Dockerはクロスプラットフォーム対応（macOS、Linux、WSL2経由のWindows）と成熟したエコシステムを提供します。macOSでは、`/convert-to-apple-container`でオプションとしてApple Containerに切り替え、より軽量なネイティブランタイムを使えます。さらに強い分離が必要なら、[Docker Sandboxes](docs/docker-sandboxes.md)が各コンテナをマイクロVM内で動作させます。
 
-**Linuxで実行できますか？**
+**LinuxやWindowsで実行できますか？**
 
-はい。DockerがデフォルトのランタイムでmacOSとLinuxの両方で動作します。`/setup`を実行するだけです。
+はい。Dockerがデフォルトのランタイムで、macOS、Linux、Windows（WSL2経由）で動作します。`bash nanoclaw.sh`を実行するだけです。
 
 **セキュリティは大丈夫ですか？**
 
-エージェントはアプリケーションレベルのパーミッションチェックの背後ではなく、コンテナで実行されます。明示的にマウントされたディレクトリのみアクセスできます。実行するものをレビューすべきですが、コードベースは十分に小さいため実際にレビュー可能です。完全なセキュリティモデルについては[docs/SECURITY.md](docs/SECURITY.md)を参照してください。
+エージェントはアプリケーションレベルのパーミッションチェックではなく、コンテナ内で実行されます。明示的にマウントされたディレクトリのみアクセス可能です。クレデンシャルはコンテナに渡されず、アウトバウンドAPIリクエストは[OneCLI Agent Vault](https://github.com/onecli/onecli)を経由し、プロキシレベルで認証を注入し、レートリミットやアクセスポリシーをサポートします。実行するものはレビューすべきですが、コードベースは実際にレビュー可能な規模です。完全なセキュリティモデルについては[セキュリティドキュメント](https://docs.nanoclaw.dev/concepts/security)を参照してください。
 
 **なぜ設定ファイルがないのか？**
 
-設定の肥大化を避けたいからです。すべてのユーザーがNanoClawをカスタマイズし、汎用的なシステムを設定するのではなく、コードが必要なことを正確に実行するようにすべきです。設定ファイルが欲しい場合は、Claudeに追加するよう伝えることができます。
+設定の肥大化を避けたいからです。すべてのユーザーがNanoClawをカスタマイズし、汎用的なシステムを設定するのではなくコードが自分の望み通りに動くようにすべきです。設定ファイルが欲しければClaudeに追加するよう伝えれば実現できます。
 
 **サードパーティやオープンソースモデルを使えますか？**
 
-はい。NanoClawはClaude API互換のモデルエンドポイントに対応しています。`.env`ファイルで以下の環境変数を設定してください：
+はい。推奨される方法は`/add-opencode`（OpenCode設定経由でOpenRouter、OpenAI、Google、DeepSeekなど）か`/add-ollama-provider`（Ollama経由でローカルのオープンウェイトモデル）です。どちらもエージェントグループごとに設定可能なので、同じインストール内で異なるエージェントが異なるバックエンドで動作できます。
+
+一時的な実験用には、Claude API互換のエンドポイントも`.env`で利用できます：
 
 ```bash
 ANTHROPIC_BASE_URL=https://your-api-endpoint.com
 ANTHROPIC_AUTH_TOKEN=your-token-here
 ```
 
-以下が使用可能です：
-- [Ollama](https://ollama.ai)とAPIプロキシ経由のローカルモデル
-- [Together AI](https://together.ai)、[Fireworks](https://fireworks.ai)等でホストされたオープンソースモデル
-- Anthropic互換APIのカスタムモデルデプロイメント
-
-注意：最高の互換性のため、モデルはAnthropic APIフォーマットに対応している必要があります。
-
 **問題のデバッグ方法は？**
 
 Claude Codeに聞いてください。「スケジューラーが動いていないのはなぜ？」「最近のログには何がある？」「このメッセージに返信がなかったのはなぜ？」これがNanoClawの基盤となるAIネイティブなアプローチです。
 
 **セットアップがうまくいかない場合は？**
 
-問題がある場合、セットアップ中にClaudeが動的に修正を試みます。それでもうまくいかない場合は、`claude`を実行してから`/debug`を実行してください。Claudeが他のユーザーにも影響する可能性のある問題を見つけた場合は、セットアップのSKILL.mdを修正するPRを開いてください。
+ステップが失敗した場合、`nanoclaw.sh`は診断と再開のためにClaude Codeへ制御を渡します。それでも解決しなければ、`claude`を実行して`/debug`を呼び出してください。他のユーザーにも影響しそうな問題をClaudeが特定した場合は、該当のセットアップステップまたはスキルにPRを送ってください。
 
 **どのような変更がコードベースに受け入れられますか？**
 
-セキュリティ修正、バグ修正、明確な改善のみが基本設定に受け入れられます。それだけです。
+ベース設定に受け入れられるのは、セキュリティ修正、バグ修正、明確な改善のみです。それだけです。
 
-それ以外のすべて（新機能、OS互換性、ハードウェアサポート、機能拡張）はスキルとしてコントリビューションすべきです。
+それ以外（新機能、OS互換性、ハードウェアサポート、拡張など）は、`channels`または`providers`ブランチのスキルとしてコントリビュートしてください。
 
-これにより、基本システムを最小限に保ち、すべてのユーザーが不要な機能を継承することなく、自分のインストールをカスタマイズできます。
+これにより、ベースシステムを最小限に保ち、全ユーザーが不要な機能を継承することなく自分のインストールをカスタマイズできます。
 
 ## コミュニティ
 
-質問やアイデアは？[Discordに参加](https://discord.gg/VDdww8qS42)してください。
+質問やアイデアがありますか？[Discordに参加](https://discord.gg/VDdww8qS42)してください。
 
 ## 変更履歴
 
-破壊的変更と移行ノートについては[CHANGELOG.md](CHANGELOG.md)を参照してください。
+破壊的変更については[CHANGELOG.md](CHANGELOG.md)を、完全なリリース履歴はドキュメントサイトの[full release history](https://docs.nanoclaw.dev/changelog)を参照してください。
 
 ## ライセンス
 
diff --git a/README_zh.md b/README_zh.md
index 714bd87..9db5c28 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -3,90 +3,87 @@
 </p>
 
 <p align="center">
-  NanoClaw —— 您的专属 Claude 助手，在容器中安全运行。它轻巧易懂，并能根据您的个人需求灵活定制。
+  一个将智能体安全运行在独立容器中的 AI 助手。轻量、易于理解，并可根据您的需求完全定制。
 </p>
 
 <p align="center">
   <a href="https://nanoclaw.dev">nanoclaw.dev</a>&nbsp; • &nbsp;
+  <a href="https://docs.nanoclaw.dev">文档</a>&nbsp; • &nbsp;
   <a href="README.md">English</a>&nbsp; • &nbsp;
   <a href="README_ja.md">日本語</a>&nbsp; • &nbsp;
   <a href="https://discord.gg/VDdww8qS42"><img src="https://img.shields.io/discord/1470188214710046894?label=Discord&logo=discord&v=2" alt="Discord" valign="middle"></a>&nbsp; • &nbsp;
-  <a href="repo-tokens"><img src="repo-tokens/badge.svg" alt="34.9k tokens, 17% of context window" valign="middle"></a>
+  <a href="repo-tokens"><img src="repo-tokens/badge.svg" alt="repo tokens" valign="middle"></a>
 </p>
-通过 Claude Code，NanoClaw 可以动态重写自身代码，根据您的需求定制功能。
 
-**新功能：** 首个支持 [Agent Swarms（智能体集群）](https://code.claude.com/docs/en/agent-teams) 的 AI 助手。可轻松组建智能体团队，在您的聊天中高效协作。
+---
 
-## 我为什么创建这个项目
+## 我为什么创建 NanoClaw
 
-[OpenClaw](https://github.com/openclaw/openclaw) 是一个令人印象深刻的项目，但我无法安心使用一个我不了解却能访问我个人隐私的软件。OpenClaw 有近 50 万行代码、53 个配置文件和 70+ 个依赖项。其安全性是应用级别的（通过白名单、配对码实现），而非操作系统级别的隔离。所有东西都在一个共享内存的 Node 进程中运行。
+[OpenClaw](https://github.com/openclaw/openclaw) 是一个令人印象深刻的项目，但我无法安心使用一个我不了解、却能访问我个人隐私的复杂软件。OpenClaw 有近 50 万行代码、53 个配置文件和 70+ 个依赖项。其安全性是应用级别的（白名单、配对码），而非真正的操作系统级隔离。所有东西都在一个共享内存的 Node 进程中运行。
 
-NanoClaw 用一个您能快速理解的代码库，为您提供了同样的核心功能。只有一个进程，少数几个文件。智能体（Agent）运行在具有文件系统隔离的真实 Linux 容器中，而不是依赖于权限检查。
+NanoClaw 用一个您能轻松理解的代码库提供了同样的核心功能：一个进程，少数几个文件。Claude 智能体运行在具有文件系统隔离的独立 Linux 容器中，而不是仅靠权限检查。
 
 ## 快速开始
 
 ```bash
-git clone https://github.com/qwibitai/nanoclaw.git
-cd nanoclaw
-claude
+git clone https://github.com/qwibitai/nanoclaw.git nanoclaw-v2
+cd nanoclaw-v2
+bash nanoclaw.sh
 ```
 
-然后运行 `/setup`。Claude Code 会处理一切：依赖安装、身份验证、容器设置、服务配置。
-
-> **注意：** 以 `/` 开头的命令（如 `/setup`、`/add-whatsapp`）是 [Claude Code 技能](https://code.claude.com/docs/en/skills)。请在 `claude` CLI 提示符中输入，而非在普通终端中。
+`nanoclaw.sh` 会把您从一台全新机器一直带到一个可以直接发消息的命名智能体。它会在缺失时安装 Node、pnpm 和 Docker，向 OneCLI 注册您的 Anthropic 凭据，构建智能体容器，并配对您的第一个渠道（Telegram、Discord、WhatsApp 或本地 CLI）。如果某一步失败，会自动调用 Claude Code 进行诊断并从中断处继续。
 
 ## 设计哲学
 
-**小巧易懂：** 单一进程，少量源文件。无微服务、无消息队列、无复杂抽象层。让 Claude Code 引导您轻松上手。
+**小到可以理解。** 单一进程，少量源文件，无微服务。如果您想了解完整的 NanoClaw 代码库，直接让 Claude Code 给您讲一遍就行。
 
-**通过隔离保障安全:** 智能体运行在 Linux 容器（在 macOS 上是 Apple Container，或 Docker）中。它们只能看到被明确挂载的内容。即便通过 Bash 访问也十分安全，因为所有命令都在容器内执行，不会直接操作您的宿主机。
+**通过隔离实现安全。** 智能体运行在 Linux 容器中，只能看到明确挂载的内容。Bash 访问是安全的，因为命令在容器内执行，而不是在您的宿主机上。
 
-**为单一用户打造:** 这不是一个框架，是一个完全符合您个人需求的、可工作的软件。您可以 Fork 本项目，然后让 Claude Code 根据您的精确需求进行修改和适配。
+**为个人用户打造。** NanoClaw 不是一个单体框架，而是能精确匹配每个用户需求的软件。它被设计成量身定制的，而不是臃肿膨胀。您创建自己的 fork，让 Claude Code 按您的需求修改它。
 
-**定制即代码修改:** 没有繁杂的配置文件。想要不同的行为？直接修改代码。代码库足够小，这样做是安全的。
+**定制 = 修改代码。** 没有配置膨胀。想要不同的行为？改代码。代码库小到改动是安全的。
 
-**AI 原生:** 无安装向导（由 Claude Code 指导安装）。无需监控仪表盘，直接询问 Claude 即可了解系统状况。无调试工具（描述问题，Claude 会修复它）。
+**AI 原生，混合式设计。** 安装与上手流程走的是经过优化的脚本路径，快速且确定。当某一步需要判断（安装失败、引导决策、定制化）时，控制权会无缝地交给 Claude Code。安装之后也不提供监控仪表盘或调试 UI：您在聊天中描述问题，Claude Code 来处理。
 
-**技能（Skills）优于功能（Features）:** 贡献者不应该向代码库添加新功能（例如支持 Telegram）。相反，他们应该贡献像 `/add-telegram` 这样的 [Claude Code 技能](https://code.claude.com/docs/en/skills)，这些技能可以改造您的 fork。最终，您得到的是只做您需要事情的整洁代码。
+**技能优于功能。** 主干只发布注册表和基础设施，不包含具体的渠道适配器或替代智能体提供者。各个渠道（Discord、Slack、Telegram、WhatsApp……）放在长期存在的 `channels` 分支上；替代提供者（OpenCode、Ollama）放在 `providers` 分支上。您运行 `/add-telegram`、`/add-opencode` 等，技能会把您所需要的模块精确地复制到您的 fork 里。不会出现您没要求的功能。
 
-**最好的工具套件，最好的模型:** 本项目运行在 Claude Agent SDK 之上，这意味着您直接运行的就是 Claude Code。Claude Code 高度强大，其编码和问题解决能力使其能够修改和扩展 NanoClaw，为每个用户量身定制。
+**最强的 harness，最强的模型。** NanoClaw 通过 Anthropic 官方的 Claude Agent SDK 原生使用 Claude Code，所以您能用上最新的 Claude 模型以及 Claude Code 的完整工具集——包括修改和扩展自己的 NanoClaw fork 的能力。其他提供者是可插拔选项：`/add-codex` 对应 OpenAI 的 Codex（ChatGPT 订阅或 API key），`/add-opencode` 通过 OpenCode 接入 OpenRouter、Google、DeepSeek 等，`/add-ollama-provider` 用于本地开源权重模型。提供者可按智能体组单独配置。
 
 ## 功能支持
 
-- **多渠道消息** - 通过 WhatsApp、Telegram、Discord、Slack 或 Gmail 与您的助手对话。使用 `/add-whatsapp` 或 `/add-telegram` 等技能添加渠道，可同时运行一个或多个。
-- **隔离的群组上下文** - 每个群组都拥有独立的 `CLAUDE.md` 记忆和隔离的文件系统。它们在各自的容器沙箱中运行，且仅挂载所需的文件系统。
-- **主频道** - 您的私有频道（self-chat），用于管理控制；其他所有群组都完全隔离
-- **计划任务** - 运行 Claude 的周期性作业，并可以给您回发消息
-- **网络访问** - 搜索和抓取网页内容
-- **容器隔离** - 智能体在 Apple Container (macOS) 或 Docker (macOS/Linux) 的沙箱中运行
-- **智能体集群（Agent Swarms）** - 启动多个专业智能体团队，协作完成复杂任务（首个支持此功能的个人 AI 助手）
-- **可选集成** - 通过技能添加 Gmail (`/add-gmail`) 等更多功能
+- **多渠道消息** — WhatsApp、Telegram、Discord、Slack、Microsoft Teams、iMessage、Matrix、Google Chat、Webex、Linear、GitHub、WeChat，以及通过 Resend 的邮件。按需通过 `/add-<channel>` 技能安装。可同时运行一个或多个。
+- **灵活的隔离模式** — 可为每个渠道配一个独立智能体以获得完全隐私，也可让一个智能体在多个渠道上共享、统一记忆但会话独立，或者把多个渠道合并到一个共享会话里，让一场对话横跨多个入口。通过 `/manage-channels` 按渠道选择。详见 [docs/isolation-model.md](docs/isolation-model.md)。
+- **每个智能体的独立工作区** — 每个智能体组都有自己的 `CLAUDE.md`、自己的记忆、自己的容器，以及您允许的挂载点。除非您明确接线，否则不会有东西越过边界。
+- **计划任务** — 运行 Claude 的周期性作业，可以给您回发消息。
+- **网络访问** — 搜索和抓取网页内容。
+- **容器隔离** — 智能体在 Docker（macOS/Linux/WSL2）中沙箱化运行，可选 [Docker Sandboxes](docs/docker-sandboxes.md) 的微虚拟机隔离，或在 macOS 上选用 Apple Container 作为原生运行时。
+- **凭据安全** — 智能体不持有原始 API key。出站请求经由 [OneCLI 的 Agent Vault](https://github.com/onecli/onecli)，在请求时注入凭据，并按每个智能体执行策略和速率限制。
 
 ## 使用方法
 
-使用触发词（默认为 `@Andy`）与您的助手对话：
+用触发词（默认为 `@Andy`）与您的助手对话：
 
 ```
-@Andy 每周一到周五早上9点，给我发一份销售渠道的概览（需要访问我的 Obsidian vault 文件夹）
-@Andy 每周五回顾过去一周的 git 历史，如果与 README 有出入，就更新它
-@Andy 每周一早上8点，从 Hacker News 和 TechCrunch 收集关于 AI 发展的资讯，然后发给我一份简报
+@Andy 每个工作日早上 9 点给我发一份销售渠道概览（可以访问我的 Obsidian vault 文件夹）
+@Andy 每周五回顾过去一周的 git 历史，如果与 README 有出入就更新它
+@Andy 每周一早上 8 点，从 Hacker News 和 TechCrunch 收集 AI 相关资讯，给我发一份简报
 ```
 
-在主频道（您的self-chat）中，可以管理群组和任务：
+在您拥有或管理的渠道里，还可以管理群组和任务：
 ```
-@Andy 列出所有群组的计划任务
+@Andy 列出所有群组里的计划任务
 @Andy 暂停周一简报任务
 @Andy 加入"家庭聊天"群组
 ```
 
 ## 定制
 
-没有需要学习的配置文件。直接告诉 Claude Code 您想要什么：
+NanoClaw 不用配置文件。想改就直接告诉 Claude Code：
 
 - "把触发词改成 @Bob"
-- "记住以后回答要更简短直接"
-- "当我说早上好的时候，加一个自定义的问候"
-- "每周存储一次对话摘要"
+- "以后回答请更简短、更直接"
+- "我说早上好的时候加一个自定义问候"
+- "每周保存一次会话摘要"
 
 或者运行 `/customize` 进行引导式修改。
 
@@ -94,107 +91,103 @@ claude
 
 ## 贡献
 
-**不要添加功能，而是添加技能。**
+**不要加功能，要加技能。**
 
-如果您想添加 Telegram 支持，不要创建一个 PR 同时添加 Telegram 和 WhatsApp。而是贡献一个技能文件 (`.claude/skills/add-telegram/SKILL.md`)，教 Claude Code 如何改造一个 NanoClaw 安装以使用 Telegram。
+如果您想添加新的渠道或智能体提供者，不要把它加到主干上。新的渠道适配器进入 `channels` 分支；新的智能体提供者进入 `providers` 分支。用户在自己的 fork 上运行 `/add-<name>` 技能，由技能把相关模块复制到标准路径、接好注册、固定依赖版本。
 
-然后用户在自己的 fork 上运行 `/add-telegram`，就能得到只做他们需要事情的整洁代码，而不是一个试图支持所有用例的臃肿系统。
+这样主干始终保持为纯粹的注册表和基础设施，每个 fork 也都保持精简——用户只获得他们要求的渠道和提供者，其它什么也不会混进来。
 
-### RFS (技能征集)
+### RFS（技能征集）
 
 我们希望看到的技能：
 
 **通信渠道**
-- `/add-signal` - 添加 Signal 作为渠道
-
-**会话管理**
-- `/clear` - 添加一个 `/clear` 命令，用于压缩会话（在同一会话中总结上下文，同时保留关键信息）。这需要研究如何通过 Claude Agent SDK 以编程方式触发压缩。
+- `/add-signal` — 添加 Signal 作为渠道
 
 ## 系统要求
 
-- macOS 或 Linux
-- Node.js 20+
-- [Claude Code](https://claude.ai/download)
-- [Apple Container](https://github.com/apple/container) (macOS) 或 [Docker](https://docker.com/products/docker-desktop) (macOS/Linux)
+- macOS 或 Linux（Windows 通过 WSL2）
+- Node.js 20+ 和 pnpm 10+（安装脚本会在缺失时自动安装）
+- [Docker Desktop](https://docker.com/products/docker-desktop)（macOS/Windows）或 Docker Engine（Linux）
+- [Claude Code](https://claude.ai/download)，用于 `/customize`、`/debug`、安装过程中的错误恢复以及所有 `/add-<channel>` 技能
 
 ## 架构
 
 ```
-渠道 --> SQLite --> 轮询循环 --> 容器 (Claude Agent SDK) --> 响应
+消息应用 → 主机进程（路由器） → inbound.db → 容器（Bun、Claude Agent SDK） → outbound.db → 主机进程（投递） → 消息应用
 ```
 
-单一 Node.js 进程。渠道通过技能添加，启动时自注册 — 编排器连接具有凭据的渠道。智能体在具有文件系统隔离的 Linux 容器中执行。每个群组的消息队列带有并发控制。通过文件系统进行 IPC。
+单一 Node 主机编排每个会话的智能体容器。当一条消息到来时，主机按实体模型（用户 → 消息组 → 智能体组 → 会话）进行路由，写入该会话的 `inbound.db`，并唤醒容器。容器内部的 agent-runner 轮询 `inbound.db`，调用 Claude，并把响应写入 `outbound.db`。主机轮询 `outbound.db`，通过渠道适配器投递回去。
 
-完整架构详情请见 [docs/SPEC.md](docs/SPEC.md)。
+每个会话两个 SQLite 文件，每个文件只有一个写入者——没有跨挂载的锁争用，没有 IPC，没有 stdin 管道。渠道和替代提供者在启动时自注册；主干提供注册表和 Chat SDK 桥接，而适配器本身在每个 fork 里通过技能安装。
+
+完整架构说明见 [docs/architecture.md](docs/architecture.md)；三级隔离模型见 [docs/isolation-model.md](docs/isolation-model.md)。
 
 关键文件：
-- `src/index.ts` - 编排器：状态管理、消息循环、智能体调用
-- `src/channels/registry.ts` - 渠道注册表（启动时自注册）
-- `src/ipc.ts` - IPC 监听与任务处理
-- `src/router.ts` - 消息格式化与出站路由
-- `src/group-queue.ts` - 带全局并发限制的群组队列
-- `src/container-runner.ts` - 生成流式智能体容器
-- `src/task-scheduler.ts` - 运行计划任务
-- `src/db.ts` - SQLite 操作（消息、群组、会话、状态）
-- `groups/*/CLAUDE.md` - 各群组的记忆
+- `src/index.ts` — 入口：数据库初始化、渠道适配器、投递轮询、sweep
+- `src/router.ts` — 入站路由：消息组 → 智能体组 → 会话 → `inbound.db`
+- `src/delivery.ts` — 轮询 `outbound.db`，通过适配器投递，处理系统动作
+- `src/host-sweep.ts` — 60 秒 sweep：失效检测、到期消息唤醒、循环任务
+- `src/session-manager.ts` — 解析会话，打开 `inbound.db` / `outbound.db`
+- `src/container-runner.ts` — 为每个智能体组启动容器，OneCLI 凭据注入
+- `src/db/` — 中心数据库（用户、角色、智能体组、消息组、接线、迁移）
+- `src/channels/` — 渠道适配器基础设施（适配器通过 `/add-<channel>` 技能安装）
+- `src/providers/` — 主机侧提供者配置（`claude` 内置，其他通过技能安装）
+- `container/agent-runner/` — Bun 版 agent-runner：轮询循环、MCP 工具、提供者抽象
+- `groups/<folder>/` — 每个智能体组的文件系统（`CLAUDE.md`、技能、容器配置）
 
 ## FAQ
 
-**为什么是 Docker？**
+**为什么用 Docker？**
 
-Docker 提供跨平台支持（macOS 和 Linux）和成熟的生态系统。在 macOS 上，您可以选择通过运行 `/convert-to-apple-container` 切换到 Apple Container，以获得更轻量级的原生运行时体验。
+Docker 提供跨平台支持（macOS、Linux、Windows via WSL2）和成熟的生态。在 macOS 上，您可以选择通过 `/convert-to-apple-container` 切换到 Apple Container，以获得更轻量的原生运行时。如需更强隔离，[Docker Sandboxes](docs/docker-sandboxes.md) 会把每个容器放到一台微虚拟机里运行。
 
-**我可以在 Linux 上运行吗？**
+**我可以在 Linux 或 Windows 上运行吗？**
 
-可以。Docker 是默认的容器运行时，在 macOS 和 Linux 上都可以使用。只需运行 `/setup`。
+可以。Docker 是默认运行时，可在 macOS、Linux 以及 Windows（通过 WSL2）上工作。运行 `bash nanoclaw.sh` 就行。
 
 **这个项目安全吗？**
 
-智能体在容器中运行，而不是在应用级别的权限检查之后。它们只能访问被明确挂载的目录。您仍然应该审查您运行的代码，但这个代码库小到您真的可以做到。完整的安全模型请见 [docs/SECURITY.md](docs/SECURITY.md)。
+智能体运行在容器里，而不是躲在应用级权限检查之后。它们只能访问明确挂载的目录。凭据不会进入容器——出站 API 请求通过 [OneCLI 的 Agent Vault](https://github.com/onecli/onecli) 在代理层注入认证，并支持速率限制和访问策略。您仍然应该审查自己要运行的代码，但代码库小到您真的能做到。完整的安全模型见 [安全文档](https://docs.nanoclaw.dev/concepts/security)。
 
 **为什么没有配置文件？**
 
-我们不希望配置泛滥。每个用户都应该定制它，让代码完全符合他们的需求，而不是去配置一个通用的系统。如果您喜欢用配置文件，告诉 Claude 让它加上。
+我们不想让配置泛滥。每位用户都应该定制 NanoClaw，让代码精确地做他们想要的事，而不是去配置一个通用系统。如果您更喜欢有配置文件，可以让 Claude 给您加。
 
 **我可以使用第三方或开源模型吗？**
 
-可以。NanoClaw 支持任何 API 兼容的模型端点。在 `.env` 文件中设置以下环境变量：
+可以。推荐做法是 `/add-opencode`（通过 OpenCode 配置接入 OpenRouter、OpenAI、Google、DeepSeek 等）或 `/add-ollama-provider`（通过 Ollama 使用本地开源权重模型）。两者都可以按智能体组单独配置，所以同一套安装里不同的智能体可以运行在不同的后端上。
+
+对于一次性实验，任何 Claude API 兼容的端点也可以通过 `.env` 使用：
 
 ```bash
 ANTHROPIC_BASE_URL=https://your-api-endpoint.com
 ANTHROPIC_AUTH_TOKEN=your-token-here
 ```
 
-这使您能够使用：
-- 通过 [Ollama](https://ollama.ai) 配合 API 代理运行的本地模型
-- 托管在 [Together AI](https://together.ai)、[Fireworks](https://fireworks.ai) 等平台上的开源模型
-- 兼容 Anthropic API 格式的自定义模型部署
-
-注意：为获得最佳兼容性，模型需支持 Anthropic API 格式。
-
 **我该如何调试问题？**
 
-问 Claude Code。"为什么计划任务没有运行？" "最近的日志里有什么？" "为什么这条消息没有得到回应？" 这就是 AI 原生的方法。
+问 Claude Code。"为什么计划任务没运行？""最近的日志里有什么？""为什么这条消息没有得到回复？"这就是 NanoClaw 底层的 AI 原生方式。
 
-**为什么我的安装不成功？**
+**为什么安装对我不成功？**
 
-如果遇到问题，安装过程中 Claude 会尝试动态修复。如果问题仍然存在，运行 `claude`，然后运行 `/debug`。如果 Claude 发现一个可能影响其他用户的问题，请开一个 PR 来修改 setup SKILL.md。
+如果某一步失败，`nanoclaw.sh` 会把控制权交给 Claude Code 进行诊断并从中断处继续。如果还是没解决，运行 `claude`，然后 `/debug`。如果 Claude 发现一个可能影响其他用户的问题，请对相关的安装步骤或技能提 PR。
 
-**什么样的代码更改会被接受？**
+**什么样的更改会被接受进代码库？**
 
-安全修复、bug 修复，以及对基础配置的明确改进。仅此而已。
+进入基础配置的只会是：安全修复、bug 修复、明显的改进。仅此而已。
 
-其他一切（新功能、操作系统兼容性、硬件支持、增强功能）都应该作为技能来贡献。
+其他一切（新能力、操作系统兼容、硬件支持、增强）都应作为技能贡献到 `channels` 或 `providers` 分支。
 
-这使得基础系统保持最小化，并让每个用户可以定制他们的安装，而无需继承他们不想要的功能。
+这样基础系统保持最小化，每位用户都可以定制自己的安装，而不必继承他们不想要的功能。
 
 ## 社区
 
-有任何疑问或建议？欢迎[加入 Discord 社区](https://discord.gg/VDdww8qS42)与我们交流。
+有问题或想法？欢迎[加入 Discord](https://discord.gg/VDdww8qS42)。
 
 ## 更新日志
 
-破坏性变更和迁移说明请见 [CHANGELOG.md](CHANGELOG.md)。
+破坏性变更见 [CHANGELOG.md](CHANGELOG.md)，完整发布历史见文档站的 [full release history](https://docs.nanoclaw.dev/changelog)。
 
 ## 许可证
 
diff --git a/REFACTOR.md b/REFACTOR.md
deleted file mode 100644
index 2ae7f81..0000000
--- a/REFACTOR.md
+++ /dev/null
@@ -1,175 +0,0 @@
-# NanoClaw Refactor — Forward-Looking Reference
-
-Consolidates what's still relevant from `REFACTOR_PLAN.md` and `REFACTOR_EXECUTION.md`: open decisions, remaining work, operational patterns worth keeping. Historical PR timeline and phase framing have been dropped — the work is in the commit history.
-
----
-
-## Architecture (still authoritative)
-
-### Module tiers
-
-Three categories, distinguished by shipping model and dependency direction:
-
-| Tier | Where it lives | Loaded by default? | Removal cost |
-|------|----------------|--------------------|--------------|
-| **Core** | `src/**` (outside `src/modules/`, `src/channels/`, `src/providers/`) | always | N/A — can't remove |
-| **Default modules** | `src/modules/<name>/` on main | yes — imported by `src/modules/index.ts` | edit core imports (intentional friction) |
-| **Optional modules** | `src/modules/<name>/` on main (for now — see open q #7) | yes, via barrel import | delete files + barrel line + revert `MODULE-HOOK` edits |
-| **Channel adapters** | `src/channels/<name>.ts` on `channels` branch | no — cherry-pick via `/add-<name>` | delete files + barrel line |
-| **Providers** | on `providers` branch | no — cherry-pick via `/add-<provider>` | delete files + barrel line |
-
-Default modules today: `typing`, `mount-security`, `approvals`, `cli`.
-Optional modules: `interactive`, `scheduling`, `permissions`, `agent-to-agent`, `self-mod`.
-
-Dependency rule: **core ← default modules ← optional modules**. Optional modules must not depend on each other. Known transitional violation (flagged): `src/db/messaging-groups.ts` auto-wires `agent_destinations` when agent-to-agent is installed.
-
-### The four registries
-
-Full contract in [`docs/module-contract.md`](docs/module-contract.md). Summary:
-
-1. **Delivery action handlers** — `delivery.ts`; modules call `registerDeliveryAction(name, fn)`.
-2. **Router inbound gate** — `router.ts`; single setter (`setSenderResolver` + `setAccessGate`). Default: allow-all.
-3. **Response dispatcher** — `response-registry.ts`; modules call `registerResponseHandler(fn)`. First to return `true` claims.
-4. **Container MCP tool self-registration** — `container/agent-runner/src/mcp-tools/server.ts`; modules call `registerTools([...])` at import.
-
-Anything else single-consumer uses either a `sqlite_master`-guarded inline read or a `MODULE-HOOK:<name>:start/end` skill edit.
-
-### Module distribution (pending)
-
-- **`main`** — core + default modules + default channel (`cli`). Ships clean.
-- **`channels`** — fully loaded runnable branch with all channel adapters; skills cherry-pick from it.
-- **`providers`** — same pattern for agent providers (OpenCode).
-- **`modules` branch** — proposed but NOT created yet. See "Remaining work" below.
-
----
-
-## Remaining work
-
-### Phase 5: merge `v2` → `main`
-
-Cut-over the refactor. Pre-reqs (already met): green build, green tests, green service boot, clean `channels` / `providers` syncs.
-
-Open logistics:
-- Release versioning: bump to `1.3.0` at merge time or cut a `v2-rc` tag first for internal testing? Non-blocking — decide at merge.
-- Coordinate with anyone still running the old `main` (v1.2.53) — breaking change for them.
-- Announce the new layout + the one shell command that changed (`pnpm run chat` is new default).
-
-### `modules` branch — create, skip, or defer?
-
-The original plan (PR #10) was to fork a `modules` branch and populate it with the 5 optional modules, so future `/add-<module>` skills pull via `git show origin/modules:path`. Three paths:
-
-- **(a) Create it now.** Matches the `channels`/`providers` pattern for consistency. Extra surface to maintain: every core change must be merged into `modules` at phase boundaries (same cadence as channels/providers). Pays off if we ever want to make a module *truly* optional (not shipped on main).
-- **(b) Skip it.** Leave all 5 optional modules shipped on main. No `modules` branch, no install skills, no cherry-picking. Simpler but loses the "opt-in" property for users who want a leaner install.
-- **(c) Defer.** Ship main without the modules branch; create it later if someone actually wants to slim their install. No-cost option for now.
-
-Recommendation leans toward (c) — we've already paid the architectural cost (tier boundary, dependency rule, registries) without needing the branch today.
-
-### Per-module follow-ups (tracked as open questions below)
-
-Each has a specific landing zone when we get to it:
-- #11–13 (admin mechanism, providers registry, container-runner audit) — scope a focused cleanup pass.
-- #14 (CLAUDE.md review) — single dedicated PR touching every module.
-- #15 (A2A / destinations rethink) — requires design, not just cleanup.
-- #17–18 (self-mod rethink, per-group source) — requires design.
-- #19 (system vs user CLAUDE.md) — requires install-skill tooling.
-
----
-
-## Operational patterns (keep using these)
-
-### Standing checks for every PR
-
-Non-negotiable; a unit test suite alone doesn't catch circular-import TDZ bugs:
-
-1. `pnpm run build` clean.
-2. `pnpm test` + `bun test` (in `container/agent-runner/`) all green.
-3. **Service actually starts.** `gtimeout 5 node dist/index.js` (or `launchctl kickstart`) must reach `NanoClaw running`. Unit tests import individual files; only `main()` exercises the module-init order.
-4. Expected boot log lines present (at least: `Central DB ready`, `Delivery polls started`, `Host sweep started`, `NanoClaw running`, plus any module lifecycle line like `OneCLI approval handler started` or `CLI channel listening`).
-
-### Module architecture rule (TDZ bug, PR #3)
-
-Any registry state a module writes to at import time must live in a file with **no back-edge to `src/index.ts`** — transitively. `src/index.ts` imports `src/modules/index.js` for side effects; if a module calls `registerX()` at top level and `X` lives in `src/index.ts`, the ES module loader hits a TDZ reference on the const declaration. Fix: registry state lives in its own dependency-free file (e.g. `src/response-registry.ts`). Any new registry follows the same pattern.
-
-### Branch sync procedure
-
-After every `v2` (or future `main`) sync into `channels` / `providers` / future `modules`:
-
-1. **File-presence diff.** Enumerate files that existed pre-sync but are missing post-sync:
-   ```
-   git ls-tree -r <pre-sync>  | awk '{print $4}' | sort > /tmp/pre.txt
-   git ls-tree -r <post-sync> | awk '{print $4}' | sort > /tmp/post.txt
-   comm -23 /tmp/pre.txt /tmp/post.txt
-   ```
-   Classify each missing file:
-   - **Intentional** (core deleted it) → leave deleted.
-   - **Branch-owned** (channels branch still needs it) → restore from pre-sync HEAD.
-
-   This has caught real losses on both `channels` (17 adapter files plus 3 setup scripts after PR #2's channel move) and `providers` (opencode files after PR #2).
-
-2. **Cross-file consistency.** When restoring a file, check whether something *else* that also changed references it (e.g. `setup/index.ts`'s `STEPS` map).
-
-3. **Run the standing checks** against the synced branch (not just v2).
-
-### Prettier drift pattern
-
-The `format:fix` pre-commit hook sometimes reformats peer files *after* the commit completes, leaving cosmetic-only diffs in the working tree. Discard with `git checkout -- <files>`. Do not re-commit the drift — it's trivial whitespace and noise.
-
----
-
-## Open questions (curated)
-
-### Design / architecture
-
-1. **`NANOCLAW_ADMIN_USER_IDS` as the admin mechanism.** Host queries `user_roles` at container wake, collapses into env var, container compares sender IDs. Conflates identity-at-send with privilege-at-wake and forces the container to care about namespaced user IDs. Revisit during a container-runner audit.
-
-2. **Host-side `src/providers/` registry.** One real consumer (OpenCode). A registry is probably overkill — the install skill could just edit `container-runner.ts` via `MODULE-HOOK`. Fold into the container-runner audit.
-
-3. **Container-runner audit.** `src/container-runner.ts` has accreted wake/spawn/kill, mount assembly, OneCLI credential application, admin-ID env var, idle timers, image rebuild. Some pieces should pull apart or move into modules. Not blocking. Related to #1 and #2.
-
-4. **Revisit destinations + A2A capability holistically.** The destination projection invariant, dual-purpose routing+ACL table, channel vs agent destination shapes, `createMessagingGroupAgent` auto-wire coupling — more machinery than the feature warrants. Phase 3 moved it out of core intact; a redesign is warranted but scoped post-refactor.
-
-5. **Self-mod approach rethink.** Three separate MCP tools + three delivery actions + three approval handlers for what's essentially "mutate container.config.json and rebuild." Also: post-rebuild latency (host sweep waits up to 60s), and agents sometimes send redundant `add_mcp_server` + `request_rebuild` pairs. Consider collapsing into a single "apply this container-config diff" approval primitive.
-
-6. **Per-agent-group source / per-group base image.** Self-mod today layers packages/MCP on a shared base. As groups diverge (different base images, provider configs, runtime toolchains), the shared-base assumption won't scale. Scope post-refactor.
-
-### Distribution / operational
-
-7. **Providers on a consolidated `modules` branch?** Staying separate for now. Revisit if a second optional provider appears.
-
-8. **Per-group module enablement.** Modules are currently project-wide. If one agent group wants approvals and another doesn't, we'd need per-group feature flags. Flag if asked.
-
-9. **Module removal UX.** We do not drop tables on uninstall. Is that the right default? (Alternative: `/remove-<module>` optionally runs a down migration. YAGNI until requested.)
-
-10. **Cross-module ordering for the response dispatcher.** Registration order determines who claims a given `questionId`. IDs are disjoint in practice (`q-…` vs `appr-…`), so first-match-wins is safe. If a third response-consuming module arrives, we may need keyed dispatch.
-
-11. **Versioned module migrations.** Reinstalls are idempotent (migrator skips anything already in `schema_version`). If a module ships a *new* migration in a later version, the install skill must append the new file + barrel entry without touching prior ones. Simplest rule: install skills are additive; content changes to an already-applied migration are a hard error.
-
-12. **Telegram pairing imports from permissions (channels branch).** `src/channels/telegram.ts` reaches into `src/modules/permissions/db/*` for `grantRole`/`hasAnyOwner`/`upsertUser` in the pairing-bootstrap branch. Cross-branch tier violation. Fix: extract those writes into a pairing helper (e.g. `src/channels/telegram-pairing-accept.ts` or `setup/pair-telegram.ts`). Non-blocking.
-
-### Core slotting (files not explicitly discussed)
-
-13. **`state-sqlite.ts`, `webhook-server.ts`, `timezone.ts`.** state-sqlite is likely core (host tracker). Webhook-server likely core (channel infra). Timezone likely core utility. Confirm if any of them prove to be module-shaped during future audits.
-
-14. **Chat SDK bridge location.** `src/channels/chat-sdk-bridge.ts` is channel infra that bridges adapters on the `channels` branch. Stays in `src/channels/` for now.
-
-15. **OneCLI credential injection.** Lives in `container-runner.ts`. Every agent call uses it, no clean optional boundary. Stays core. Related: `onecli-approvals.ts` is bundled inside the `approvals` default module on the assumption OneCLI stays in core. If OneCLI later moves to its own module, `onecli-approvals` follows.
-
-### Documentation
-
-16. **CLAUDE.md content per module.** Every module ships with project.md + agent.md. Need a dedicated review pass: (a) write the missing agent-to-agent snippets, (b) audit other modules for accuracy/tone, (c) confirm `agent.md` files are actually tailored for the agent vs. copy-pastes of `project.md`.
-
-17. **Split system CLAUDE.md from user CLAUDE.md.** Project `CLAUDE.md` and `groups/global/CLAUDE.md` mix system-authored content (module contracts, install-skill appends) with user customizations. Updates currently risk clobbering user intent. Look at a system-owned region (or separate file) that skills rewrite freely plus a user-owned one that's never touched. Related to #16.
-
----
-
-## Where the canonical references live
-
-- **Module contract** — [`docs/module-contract.md`](docs/module-contract.md)
-- **Architecture overview** — [`docs/architecture.md`](docs/architecture.md)
-- **DB layout** — [`docs/db.md`](docs/db.md), [`docs/db-central.md`](docs/db-central.md), [`docs/db-session.md`](docs/db-session.md)
-- **Agent-runner internals** — [`docs/agent-runner-details.md`](docs/agent-runner-details.md)
-- **Channel isolation model** — [`docs/isolation-model.md`](docs/isolation-model.md)
-- **Build + runtime split** — [`docs/build-and-runtime.md`](docs/build-and-runtime.md)
-- **Top-level** — [`CLAUDE.md`](CLAUDE.md)
-
-This doc (`REFACTOR.md`) is transient — prune when open questions close; retire entirely once the refactor is fully behind us and the operational patterns have been absorbed into `CLAUDE.md` or `docs/`.
diff --git a/container/CLAUDE.md b/container/CLAUDE.md
new file mode 100644
index 0000000..baf911a
--- /dev/null
+++ b/container/CLAUDE.md
@@ -0,0 +1,21 @@
+You are a NanoClaw agent. Your name, destinations, and message-sending rules are provided in the runtime system prompt at the top of each turn.
+
+## Communication
+
+Be concise — every message costs the reader's attention. Prefer outcomes over play-by-play; when the work is done, the final message should be about the result, not a transcript of what you did.
+
+## Workspace
+
+Files you create are saved in `/workspace/agent/`. Use this for notes, research, or anything that should persist across turns in this group.
+
+The file `CLAUDE.local.md` in your workspace is your per-group memory. Record things there that you'll want to remember in future sessions — user preferences, project context, recurring facts. Keep entries short and structured.
+
+## Memory
+
+When the user shares any substantive information with you, it must be stored somewhere you can retrieve it when relevant. If it's information that is pertinent to every single conversation turn it should be put into CLAUDE.local.md. Otherwise, create a system for storing the information depending on its type - e.g. create a file of people that the user mentions so you can keep track or a file of projects. For every file you create, add a concise reference in your CLAUDE.local.md so you'll be able to find it in future conversations. 
+
+A core part of your job and the main thing that defines how useful you are to the user is how well you do in creating these systems for organizing information. These are your systems that help you do your job well. Evolve them over time as needed.
+
+## Conversation history
+
+The `conversations/` folder in your workspace holds searchable transcripts of past sessions with this group. Use it to recall prior context when a request references something that happened before. For structured long-lived data, prefer dedicated files (`customers.md`, `preferences.md`, etc.); split any file over ~500 lines into a folder with an index.
diff --git a/container/Dockerfile b/container/Dockerfile
index 12d2bf6..4b4cf22 100644
--- a/container/Dockerfile
+++ b/container/Dockerfile
@@ -3,8 +3,12 @@
 # Runs Claude Agent SDK in isolated Linux VM with browser automation.
 #
 # Runtime split:
-#   - agent-runner (our TypeScript code): Bun
+#   - agent-runner (our TypeScript code): Bun, mounted RO at /app/src by host
 #   - globally-installed Node CLIs (claude-code, agent-browser, vercel): pnpm + Node
+#
+# Source is never baked in — /app/src is provided by a shared read-only
+# bind mount at runtime (see src/container-runner.ts). Source-only changes
+# never require an image rebuild.
 
 FROM node:22-slim
 
@@ -15,7 +19,7 @@ ARG INSTALL_CJK_FONTS=false
 # Pin CLI versions for reproducibility. Bump deliberately — unpinned installs
 # mean every rebuild silently picks up the latest and can break in lockstep
 # across all users.
-ARG CLAUDE_CODE_VERSION=2.1.112
+ARG CLAUDE_CODE_VERSION=2.1.116
 ARG AGENT_BROWSER_VERSION=latest
 ARG VERCEL_VERSION=latest
 ARG BUN_VERSION=1.3.12
@@ -66,44 +70,48 @@ RUN curl -fsSL https://bun.sh/install | bash -s "bun-v${BUN_VERSION}" && \
     install -m 0755 /root/.bun/bin/bun /usr/local/bin/bun && \
     rm -rf /root/.bun
 
-# ---- pnpm + global Node CLIs -------------------------------------------------
-ENV PNPM_HOME="/pnpm"
-ENV PATH="$PNPM_HOME:$PATH"
-RUN corepack enable
-
-# agent-browser has a postinstall build script — pnpm skips these by default.
-# Allowlist it via .npmrc so the install doesn't silently produce a broken
-# package. Pinned versions so every rebuild is reproducible.
-RUN --mount=type=cache,target=/root/.cache/pnpm \
-    echo "only-built-dependencies[]=agent-browser" > /root/.npmrc && \
-    pnpm install -g \
-        "@anthropic-ai/claude-code@${CLAUDE_CODE_VERSION}" \
-        "agent-browser@${AGENT_BROWSER_VERSION}" \
-        "vercel@${VERCEL_VERSION}"
-
-# ---- agent-runner ------------------------------------------------------------
+# ---- agent-runner deps -------------------------------------------------------
+# Deps are cached independently of CLI versions. Source is NOT baked in —
+# it's provided by the shared RO mount at runtime.
 WORKDIR /app
 
-# Copy manifest + lockfile first so the install layer caches independently of
-# source edits.
 COPY agent-runner/package.json agent-runner/bun.lock ./
 
 RUN --mount=type=cache,target=/root/.bun/install/cache \
     bun install --frozen-lockfile
 
-# Source. Bun runs TS directly — no tsc build step. The host remounts this
-# path at runtime via `src/container-runner.ts` so source edits on the host
-# take effect without rebuilding the image; the baked copy is the fallback.
-COPY agent-runner/ ./
+# ---- pnpm + global Node CLIs -------------------------------------------------
+# Most stable first, most frequently bumped last. Bumping claude-code
+# (the most common change) only invalidates one layer.
+#
+# only-built-dependencies gates pnpm's supply-chain policy:
+#   - agent-browser has a postinstall build step.
+#   - @anthropic-ai/claude-code's postinstall downloads the native Claude
+#     binary (linux-arm64 variant on our image). Without the allowlist
+#     the SDK fails at spawn time with "native binary not found".
+ENV PNPM_HOME="/pnpm"
+ENV PATH="$PNPM_HOME:$PATH"
+RUN corepack enable
+
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    echo "only-built-dependencies[]=agent-browser" > /root/.npmrc && \
+    echo "only-built-dependencies[]=@anthropic-ai/claude-code" >> /root/.npmrc && \
+    pnpm install -g "vercel@${VERCEL_VERSION}"
+
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    pnpm install -g "agent-browser@${AGENT_BROWSER_VERSION}"
+
+RUN --mount=type=cache,target=/root/.cache/pnpm \
+    pnpm install -g "@anthropic-ai/claude-code@${CLAUDE_CODE_VERSION}"
 
 # ---- Entrypoint --------------------------------------------------------------
 COPY entrypoint.sh /app/entrypoint.sh
 RUN chmod +x /app/entrypoint.sh
 
 # ---- Workspace + permissions -------------------------------------------------
-RUN mkdir -p /workspace/group /workspace/global /workspace/extra && \
+RUN mkdir -p /workspace/group /workspace/extra && \
     chown -R node:node /workspace && \
-    chmod 755 /home/node
+    chmod 777 /home/node
 
 USER node
 WORKDIR /workspace/group
diff --git a/container/agent-runner/bun.lock b/container/agent-runner/bun.lock
index 99fe840..3c08828 100644
--- a/container/agent-runner/bun.lock
+++ b/container/agent-runner/bun.lock
@@ -5,7 +5,7 @@
     "": {
       "name": "nanoclaw-agent-runner",
       "dependencies": {
-        "@anthropic-ai/claude-agent-sdk": "^0.2.92",
+        "@anthropic-ai/claude-agent-sdk": "^0.2.116",
         "@modelcontextprotocol/sdk": "^1.12.1",
         "cron-parser": "^5.0.0",
         "zod": "^4.0.0",
@@ -18,7 +18,23 @@
     },
   },
   "packages": {
-    "@anthropic-ai/claude-agent-sdk": ["@anthropic-ai/claude-agent-sdk@0.2.112", "", { "dependencies": { "@anthropic-ai/sdk": "^0.81.0", "@modelcontextprotocol/sdk": "^1.29.0" }, "optionalDependencies": { "@img/sharp-darwin-arm64": "^0.34.2", "@img/sharp-darwin-x64": "^0.34.2", "@img/sharp-linux-arm": "^0.34.2", "@img/sharp-linux-arm64": "^0.34.2", "@img/sharp-linux-x64": "^0.34.2", "@img/sharp-linuxmusl-arm64": "^0.34.2", "@img/sharp-linuxmusl-x64": "^0.34.2", "@img/sharp-win32-arm64": "^0.34.2", "@img/sharp-win32-x64": "^0.34.2" }, "peerDependencies": { "zod": "^4.0.0" } }, "sha512-vMFoiDKlOive8p3tphpV1gQaaytOipwGJ+uw9mvvaLQUODSC2+fCdRDAY25i2Tsv+lOtxzXBKctmaDuWqZY7ig=="],
+    "@anthropic-ai/claude-agent-sdk": ["@anthropic-ai/claude-agent-sdk@0.2.116", "", { "dependencies": { "@anthropic-ai/sdk": "^0.81.0", "@modelcontextprotocol/sdk": "^1.29.0" }, "optionalDependencies": { "@anthropic-ai/claude-agent-sdk-darwin-arm64": "0.2.116", "@anthropic-ai/claude-agent-sdk-darwin-x64": "0.2.116", "@anthropic-ai/claude-agent-sdk-linux-arm64": "0.2.116", "@anthropic-ai/claude-agent-sdk-linux-arm64-musl": "0.2.116", "@anthropic-ai/claude-agent-sdk-linux-x64": "0.2.116", "@anthropic-ai/claude-agent-sdk-linux-x64-musl": "0.2.116", "@anthropic-ai/claude-agent-sdk-win32-arm64": "0.2.116", "@anthropic-ai/claude-agent-sdk-win32-x64": "0.2.116" }, "peerDependencies": { "zod": "^4.0.0" } }, "sha512-5NKpgaOZkzNCGCvLxJZUVGimf5IcYmpQ2x2XrR9ilK+2UkWrnnwcUfIWo8bBz9e7lSYcUf9XleGigq2eOOF7aw=="],
+
+    "@anthropic-ai/claude-agent-sdk-darwin-arm64": ["@anthropic-ai/claude-agent-sdk-darwin-arm64@0.2.116", "", { "os": "darwin", "cpu": "arm64" }, "sha512-mG19ovtXCpETmd5KmTU1JO2iIHZBG09IP8DmgZjLA3wLmTzpgn9Au9veRaeJeXb1EqiHiFZU+z+mNB79+w5v9g=="],
+
+    "@anthropic-ai/claude-agent-sdk-darwin-x64": ["@anthropic-ai/claude-agent-sdk-darwin-x64@0.2.116", "", { "os": "darwin", "cpu": "x64" }, "sha512-qC25N0HRM8IXbM4Qi4svH9f51Y6DciDvjLV+oNYnxkdPgDG8p/+b7vQirN7qPxytIQb2TPdoFgUeCsSe7lrQyw=="],
+
+    "@anthropic-ai/claude-agent-sdk-linux-arm64": ["@anthropic-ai/claude-agent-sdk-linux-arm64@0.2.116", "", { "os": "linux", "cpu": "arm64" }, "sha512-MQIcJhhPM+RPJ7kMQdOQarkJ2FlJqOiu953c08YyJOoWdHykd3DIiHws3mf1Mwl/dfFeIyshOVpNND3hyIy5Dg=="],
+
+    "@anthropic-ai/claude-agent-sdk-linux-arm64-musl": ["@anthropic-ai/claude-agent-sdk-linux-arm64-musl@0.2.116", "", { "os": "linux", "cpu": "arm64" }, "sha512-Dg/T3NkSp35ODiwdhj0KquvC6Xu+DMbyWFNkfepA3bz4oF2SVSgyOPYwVmfoJerzEUnYDldP4YhOxRrhbt0vXA=="],
+
+    "@anthropic-ai/claude-agent-sdk-linux-x64": ["@anthropic-ai/claude-agent-sdk-linux-x64@0.2.116", "", { "os": "linux", "cpu": "x64" }, "sha512-Bww1fzQB+vcF0tRhmCAlwSsN4wR2HgX7pBT9AWuwzJj6DKsVC23N54Ea80lsnM7dTUtUTrGYMTwVUHTWqfYnfQ=="],
+
+    "@anthropic-ai/claude-agent-sdk-linux-x64-musl": ["@anthropic-ai/claude-agent-sdk-linux-x64-musl@0.2.116", "", { "os": "linux", "cpu": "x64" }, "sha512-LMYxUMa1nK4N9BPRJdcGBAvl9rjTI4ZHo+kfAKrJ3MlfB6VFF1tRIubwsWOaOtkuNazMdAYovsZJg4bdzOBBTQ=="],
+
+    "@anthropic-ai/claude-agent-sdk-win32-arm64": ["@anthropic-ai/claude-agent-sdk-win32-arm64@0.2.116", "", { "os": "win32", "cpu": "arm64" }, "sha512-h0YO1vkTIeUtffQhONrYbNC1pXmk1yjb1xxMEw7bAwucqtFoFpLDWe+q4+RhxaQr8ZOj6LtRE/U3dzPWHOlshA=="],
+
+    "@anthropic-ai/claude-agent-sdk-win32-x64": ["@anthropic-ai/claude-agent-sdk-win32-x64@0.2.116", "", { "os": "win32", "cpu": "x64" }, "sha512-3lllmtDFHgpW0ZM3iNvxsEjblrgRzF9Qm1lxTOtunP3hIn+pA/IkWMtKlN1ixxWiaBguLVQkJ90V6JHsvJJIvw=="],
 
     "@anthropic-ai/sdk": ["@anthropic-ai/sdk@0.81.0", "", { "dependencies": { "json-schema-to-ts": "^3.1.1" }, "peerDependencies": { "zod": "^3.25.0 || ^4.0.0" }, "optionalPeers": ["zod"], "bin": { "anthropic-ai-sdk": "bin/cli" } }, "sha512-D4K5PvEV6wPiRtVlVsJHIUhHAmOZ6IT/I9rKlTf84gR7GyyAurPJK7z9BOf/AZqC5d1DhYQGJNKRmV+q8dGhgw=="],
 
@@ -26,38 +42,6 @@
 
     "@hono/node-server": ["@hono/node-server@1.19.14", "", { "peerDependencies": { "hono": "^4" } }, "sha512-GwtvgtXxnWsucXvbQXkRgqksiH2Qed37H9xHZocE5sA3N8O8O8/8FA3uclQXxXVzc9XBZuEOMK7+r02FmSpHtw=="],
 
-    "@img/sharp-darwin-arm64": ["@img/sharp-darwin-arm64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-darwin-arm64": "1.2.4" }, "os": "darwin", "cpu": "arm64" }, "sha512-imtQ3WMJXbMY4fxb/Ndp6HBTNVtWCUI0WdobyheGf5+ad6xX8VIDO8u2xE4qc/fr08CKG/7dDseFtn6M6g/r3w=="],
-
-    "@img/sharp-darwin-x64": ["@img/sharp-darwin-x64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-darwin-x64": "1.2.4" }, "os": "darwin", "cpu": "x64" }, "sha512-YNEFAF/4KQ/PeW0N+r+aVVsoIY0/qxxikF2SWdp+NRkmMB7y9LBZAVqQ4yhGCm/H3H270OSykqmQMKLBhBJDEw=="],
-
-    "@img/sharp-libvips-darwin-arm64": ["@img/sharp-libvips-darwin-arm64@1.2.4", "", { "os": "darwin", "cpu": "arm64" }, "sha512-zqjjo7RatFfFoP0MkQ51jfuFZBnVE2pRiaydKJ1G/rHZvnsrHAOcQALIi9sA5co5xenQdTugCvtb1cuf78Vf4g=="],
-
-    "@img/sharp-libvips-darwin-x64": ["@img/sharp-libvips-darwin-x64@1.2.4", "", { "os": "darwin", "cpu": "x64" }, "sha512-1IOd5xfVhlGwX+zXv2N93k0yMONvUlANylbJw1eTah8K/Jtpi15KC+WSiaX/nBmbm2HxRM1gZ0nSdjSsrZbGKg=="],
-
-    "@img/sharp-libvips-linux-arm": ["@img/sharp-libvips-linux-arm@1.2.4", "", { "os": "linux", "cpu": "arm" }, "sha512-bFI7xcKFELdiNCVov8e44Ia4u2byA+l3XtsAj+Q8tfCwO6BQ8iDojYdvoPMqsKDkuoOo+X6HZA0s0q11ANMQ8A=="],
-
-    "@img/sharp-libvips-linux-arm64": ["@img/sharp-libvips-linux-arm64@1.2.4", "", { "os": "linux", "cpu": "arm64" }, "sha512-excjX8DfsIcJ10x1Kzr4RcWe1edC9PquDRRPx3YVCvQv+U5p7Yin2s32ftzikXojb1PIFc/9Mt28/y+iRklkrw=="],
-
-    "@img/sharp-libvips-linux-x64": ["@img/sharp-libvips-linux-x64@1.2.4", "", { "os": "linux", "cpu": "x64" }, "sha512-tJxiiLsmHc9Ax1bz3oaOYBURTXGIRDODBqhveVHonrHJ9/+k89qbLl0bcJns+e4t4rvaNBxaEZsFtSfAdquPrw=="],
-
-    "@img/sharp-libvips-linuxmusl-arm64": ["@img/sharp-libvips-linuxmusl-arm64@1.2.4", "", { "os": "linux", "cpu": "arm64" }, "sha512-FVQHuwx1IIuNow9QAbYUzJ+En8KcVm9Lk5+uGUQJHaZmMECZmOlix9HnH7n1TRkXMS0pGxIJokIVB9SuqZGGXw=="],
-
-    "@img/sharp-libvips-linuxmusl-x64": ["@img/sharp-libvips-linuxmusl-x64@1.2.4", "", { "os": "linux", "cpu": "x64" }, "sha512-+LpyBk7L44ZIXwz/VYfglaX/okxezESc6UxDSoyo2Ks6Jxc4Y7sGjpgU9s4PMgqgjj1gZCylTieNamqA1MF7Dg=="],
-
-    "@img/sharp-linux-arm": ["@img/sharp-linux-arm@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-linux-arm": "1.2.4" }, "os": "linux", "cpu": "arm" }, "sha512-9dLqsvwtg1uuXBGZKsxem9595+ujv0sJ6Vi8wcTANSFpwV/GONat5eCkzQo/1O6zRIkh0m/8+5BjrRr7jDUSZw=="],
-
-    "@img/sharp-linux-arm64": ["@img/sharp-linux-arm64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-linux-arm64": "1.2.4" }, "os": "linux", "cpu": "arm64" }, "sha512-bKQzaJRY/bkPOXyKx5EVup7qkaojECG6NLYswgktOZjaXecSAeCWiZwwiFf3/Y+O1HrauiE3FVsGxFg8c24rZg=="],
-
-    "@img/sharp-linux-x64": ["@img/sharp-linux-x64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-linux-x64": "1.2.4" }, "os": "linux", "cpu": "x64" }, "sha512-MEzd8HPKxVxVenwAa+JRPwEC7QFjoPWuS5NZnBt6B3pu7EG2Ge0id1oLHZpPJdn3OQK+BQDiw9zStiHBTJQQQQ=="],
-
-    "@img/sharp-linuxmusl-arm64": ["@img/sharp-linuxmusl-arm64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-linuxmusl-arm64": "1.2.4" }, "os": "linux", "cpu": "arm64" }, "sha512-fprJR6GtRsMt6Kyfq44IsChVZeGN97gTD331weR1ex1c1rypDEABN6Tm2xa1wE6lYb5DdEnk03NZPqA7Id21yg=="],
-
-    "@img/sharp-linuxmusl-x64": ["@img/sharp-linuxmusl-x64@0.34.5", "", { "optionalDependencies": { "@img/sharp-libvips-linuxmusl-x64": "1.2.4" }, "os": "linux", "cpu": "x64" }, "sha512-Jg8wNT1MUzIvhBFxViqrEhWDGzqymo3sV7z7ZsaWbZNDLXRJZoRGrjulp60YYtV4wfY8VIKcWidjojlLcWrd8Q=="],
-
-    "@img/sharp-win32-arm64": ["@img/sharp-win32-arm64@0.34.5", "", { "os": "win32", "cpu": "arm64" }, "sha512-WQ3AgWCWYSb2yt+IG8mnC6Jdk9Whs7O0gxphblsLvdhSpSTtmu69ZG1Gkb6NuvxsNACwiPV6cNSZNzt0KPsw7g=="],
-
-    "@img/sharp-win32-x64": ["@img/sharp-win32-x64@0.34.5", "", { "os": "win32", "cpu": "x64" }, "sha512-+29YMsqY2/9eFEiW93eqWnuLcWcufowXewwSNIT6UwZdUUCrM3oFjMWH/Z6/TMmb4hlFenmfAVbpWeup2jryCw=="],
-
     "@modelcontextprotocol/sdk": ["@modelcontextprotocol/sdk@1.29.0", "", { "dependencies": { "@hono/node-server": "^1.19.9", "ajv": "^8.17.1", "ajv-formats": "^3.0.1", "content-type": "^1.0.5", "cors": "^2.8.5", "cross-spawn": "^7.0.5", "eventsource": "^3.0.2", "eventsource-parser": "^3.0.0", "express": "^5.2.1", "express-rate-limit": "^8.2.1", "hono": "^4.11.4", "jose": "^6.1.3", "json-schema-typed": "^8.0.2", "pkce-challenge": "^5.0.0", "raw-body": "^3.0.0", "zod": "^3.25 || ^4.0", "zod-to-json-schema": "^3.25.1" }, "peerDependencies": { "@cfworker/json-schema": "^4.1.1" }, "optionalPeers": ["@cfworker/json-schema"] }, "sha512-zo37mZA9hJWpULgkRpowewez1y6ML5GsXJPY8FI0tBBCd77HEvza4jDqRKOXgHNn867PVGCyTdzqpz0izu5ZjQ=="],
 
     "@types/bun": ["@types/bun@1.3.12", "", { "dependencies": { "bun-types": "1.3.12" } }, "sha512-DBv81elK+/VSwXHDlnH3Qduw+KxkTIWi7TXkAeh24zpi5l0B2kUg9Ga3tb4nJaPcOFswflgi/yAvMVBPrxMB+A=="],
diff --git a/container/agent-runner/package.json b/container/agent-runner/package.json
index 06eb394..e9af0b1 100644
--- a/container/agent-runner/package.json
+++ b/container/agent-runner/package.json
@@ -9,7 +9,7 @@
     "test": "bun test"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.2.92",
+    "@anthropic-ai/claude-agent-sdk": "^0.2.116",
     "@modelcontextprotocol/sdk": "^1.12.1",
     "cron-parser": "^5.0.0",
     "zod": "^4.0.0"
diff --git a/container/agent-runner/scripts/sdk-signal-probe.ts b/container/agent-runner/scripts/sdk-signal-probe.ts
new file mode 100644
index 0000000..a4a3c98
--- /dev/null
+++ b/container/agent-runner/scripts/sdk-signal-probe.ts
@@ -0,0 +1,169 @@
+#!/usr/bin/env bun
+/**
+ * SDK signal probe: run a prompt, log every signal the Agent SDK emits —
+ * async-iterator events + hook callbacks + CLI stderr — with absolute
+ * and relative timing.
+ *
+ * Usage:
+ *   bun run scripts/sdk-signal-probe.ts "<prompt>"                 # simple string mode
+ *   bun run scripts/sdk-signal-probe.ts --stream "<prompt>"        # streaming-input mode
+ *   bun run scripts/sdk-signal-probe.ts --stream "<p>" \
+ *     --push "5000:<text>" --push "15000:<text>" --timeout 60000   # multi-push
+ *
+ * Streaming mode (`--stream`) passes an AsyncIterable prompt to `query()`,
+ * which keeps the CLI subprocess alive past the first result (per SDK
+ * deep dive). Required for post-result pushes, agent teams, background
+ * task notifications.
+ */
+import { query } from '@anthropic-ai/claude-agent-sdk';
+
+const args = process.argv.slice(2);
+const prompts: string[] = [];
+const pushes: Array<{ atMs: number; text: string }> = [];
+let streamMode = false;
+let timeoutMs: number | undefined;
+
+for (let i = 0; i < args.length; i++) {
+  const a = args[i];
+  if (a === '--stream') streamMode = true;
+  else if (a === '--push') {
+    const val = args[++i] ?? '';
+    const ix = val.indexOf(':');
+    if (ix === -1) throw new Error(`bad --push (want MS:text): ${val}`);
+    pushes.push({ atMs: parseInt(val.slice(0, ix), 10), text: val.slice(ix + 1) });
+  } else if (a === '--timeout') timeoutMs = parseInt(args[++i] ?? '0', 10);
+  else if (a === '--prompt') prompts.push(args[++i] ?? '');
+  else prompts.push(a);
+}
+
+const prompt = prompts.join(' ');
+if (!prompt) {
+  console.error('usage: sdk-signal-probe.ts [--stream] "<prompt>" [--push MS:<text>]... [--timeout MS]');
+  process.exit(1);
+}
+
+const T0 = Date.now();
+let LAST = T0;
+
+function log(source: string, type: string, payload: unknown = {}): void {
+  const now = Date.now();
+  const entry = { t_ms: now - T0, d_ms: now - LAST, source, type, payload };
+  LAST = now;
+  console.log(JSON.stringify(entry));
+}
+
+function hookLogger(eventName: string) {
+  return async (input: unknown, toolUseID: string | undefined): Promise<any> => {
+    log('hook', eventName, { toolUseID, input });
+    // Stuck-tool simulation: if env flag is set and this is a PreToolUse for Bash,
+    // never resolve — simulates a tool that hangs forever.
+    if (process.env.PROBE_HANG === 'true' && eventName === 'PreToolUse') {
+      const toolName = (input as any)?.tool_name ?? (input as any)?.name;
+      if (toolName === 'Bash') {
+        log('meta', 'pre_tool_use_hanging', { toolUseID, toolName });
+        await new Promise(() => {
+          /* never resolves */
+        });
+      }
+    }
+    return { continue: true };
+  };
+}
+
+const HOOK_EVENTS = [
+  'PreToolUse',
+  'PostToolUse',
+  'PostToolUseFailure',
+  'Notification',
+  'UserPromptSubmit',
+  'SessionStart',
+  'SessionEnd',
+  'Stop',
+  'SubagentStart',
+  'SubagentStop',
+  'PreCompact',
+  'PermissionRequest',
+] as const;
+
+const hooks: Record<string, unknown[]> = {};
+for (const ev of HOOK_EVENTS) hooks[ev] = [{ hooks: [hookLogger(ev)] }];
+
+// Build prompt — string (single-turn) or AsyncIterable (streaming-input)
+let promptInput: any;
+
+if (streamMode) {
+  const sessionId = `probe-${Date.now()}`;
+  async function* streamPrompt() {
+    // Initial user message
+    yield {
+      type: 'user' as const,
+      message: { role: 'user' as const, content: prompt },
+      parent_tool_use_id: null,
+      session_id: sessionId,
+    };
+    // Schedule subsequent pushes
+    const startT = Date.now();
+    const sorted = [...pushes].sort((a, b) => a.atMs - b.atMs);
+    for (const p of sorted) {
+      const waitMs = Math.max(0, p.atMs - (Date.now() - startT));
+      if (waitMs > 0) await new Promise((r) => setTimeout(r, waitMs));
+      log('meta', 'push_message', { atMs: p.atMs, text: p.text });
+      yield {
+        type: 'user' as const,
+        message: { role: 'user' as const, content: p.text },
+        parent_tool_use_id: null,
+        session_id: sessionId,
+      };
+    }
+    // Keep stream open for tail events; iterator ends when we return
+    // (no more work expected). For post-result-idle scenarios, wait here.
+    await new Promise((r) => setTimeout(r, 5000));
+  }
+  promptInput = streamPrompt();
+} else {
+  promptInput = prompt;
+}
+
+log('meta', 'probe_start', { prompt, streamMode, pushes, timeoutMs });
+
+const q = query({
+  prompt: promptInput,
+  options: {
+    includePartialMessages: true,
+    hooks: hooks as any,
+    stderr: (data: string) => log('stderr', 'chunk', { data }),
+    settingSources: [],
+    permissionMode: 'bypassPermissions',
+    allowDangerouslySkipPermissions: true,
+  },
+});
+
+// Absolute time cap — exit cleanly so the log flushes
+if (timeoutMs) {
+  setTimeout(() => {
+    log('meta', 'timeout_hit', { timeoutMs });
+    setTimeout(() => process.exit(0), 250);
+  }, timeoutMs);
+}
+
+try {
+  for await (const event of q) {
+    const snapshot: any = { ...event };
+    try {
+      const raw = JSON.stringify(snapshot);
+      if (raw.length > 2000) {
+        snapshot._truncated_bytes = raw.length;
+        if (snapshot.message?.content) {
+          const c = JSON.stringify(snapshot.message.content);
+          snapshot.message = { ...snapshot.message, content: c.slice(0, 500) + `…<+${c.length - 500}b>` };
+        }
+      }
+    } catch {
+      /* best-effort */
+    }
+    log('event', snapshot.type ?? 'unknown', { subtype: snapshot.subtype, event: snapshot });
+  }
+  log('meta', 'iterator_done');
+} catch (err: any) {
+  log('meta', 'iterator_error', { message: err?.message, stack: err?.stack?.split('\n').slice(0, 5) });
+}
diff --git a/container/agent-runner/src/config.ts b/container/agent-runner/src/config.ts
new file mode 100644
index 0000000..3a022ab
--- /dev/null
+++ b/container/agent-runner/src/config.ts
@@ -0,0 +1,55 @@
+/**
+ * Runner config — reads /workspace/agent/container.json at startup.
+ *
+ * This file is mounted read-only inside the container. The host writes it;
+ * the runner only reads. All NanoClaw-specific configuration lives here
+ * instead of environment variables.
+ */
+import fs from 'fs';
+
+const CONFIG_PATH = '/workspace/agent/container.json';
+
+export interface RunnerConfig {
+  provider: string;
+  assistantName: string;
+  groupName: string;
+  agentGroupId: string;
+  maxMessagesPerPrompt: number;
+  mcpServers: Record<string, { command: string; args: string[]; env: Record<string, string> }>;
+}
+
+const DEFAULT_MAX_MESSAGES = 10;
+
+let _config: RunnerConfig | null = null;
+
+/**
+ * Load config from container.json. Called once at startup.
+ * Falls back to sensible defaults for any missing field.
+ */
+export function loadConfig(): RunnerConfig {
+  if (_config) return _config;
+
+  let raw: Record<string, unknown> = {};
+  try {
+    raw = JSON.parse(fs.readFileSync(CONFIG_PATH, 'utf8'));
+  } catch {
+    console.error(`[config] Failed to read ${CONFIG_PATH}, using defaults`);
+  }
+
+  _config = {
+    provider: (raw.provider as string) || 'claude',
+    assistantName: (raw.assistantName as string) || '',
+    groupName: (raw.groupName as string) || '',
+    agentGroupId: (raw.agentGroupId as string) || '',
+    maxMessagesPerPrompt: (raw.maxMessagesPerPrompt as number) || DEFAULT_MAX_MESSAGES,
+    mcpServers: (raw.mcpServers as RunnerConfig['mcpServers']) || {},
+  };
+
+  return _config;
+}
+
+/** Get the loaded config. Throws if loadConfig() hasn't been called. */
+export function getConfig(): RunnerConfig {
+  if (!_config) throw new Error('Config not loaded — call loadConfig() first');
+  return _config;
+}
diff --git a/container/agent-runner/src/db/connection.ts b/container/agent-runner/src/db/connection.ts
index 9bf2551..3f0e73b 100644
--- a/container/agent-runner/src/db/connection.ts
+++ b/container/agent-runner/src/db/connection.ts
@@ -31,8 +31,7 @@ let _heartbeatPath: string = DEFAULT_HEARTBEAT_PATH;
 /** Inbound DB — container opens read-only (host is the sole writer). */
 export function getInboundDb(): Database {
   if (!_inbound) {
-    const dbPath = process.env.SESSION_INBOUND_DB_PATH || DEFAULT_INBOUND_PATH;
-    _inbound = new Database(dbPath, { readonly: true });
+    _inbound = new Database(DEFAULT_INBOUND_PATH, { readonly: true });
     _inbound.exec('PRAGMA busy_timeout = 5000');
   }
   return _inbound;
@@ -41,8 +40,7 @@ export function getInboundDb(): Database {
 /** Outbound DB — container owns this file (sole writer). */
 export function getOutboundDb(): Database {
   if (!_outbound) {
-    const dbPath = process.env.SESSION_OUTBOUND_DB_PATH || DEFAULT_OUTBOUND_PATH;
-    _outbound = new Database(dbPath);
+    _outbound = new Database(DEFAULT_OUTBOUND_PATH);
     _outbound.exec('PRAGMA journal_mode = DELETE');
     _outbound.exec('PRAGMA busy_timeout = 5000');
     _outbound.exec('PRAGMA foreign_keys = ON');
@@ -64,17 +62,65 @@ export function getOutboundDb(): Database {
     if (!cols.has('updated_at')) {
       _outbound.exec(`ALTER TABLE session_state ADD COLUMN updated_at TEXT NOT NULL DEFAULT ''`);
     }
+    // container_state: tracks the current tool in flight (if any) so the host
+    // sweep can widen its stuck tolerance when Bash is running with a user-
+    // declared long timeout. Forward-compat for older outbound.db files.
+    _outbound.exec(`
+      CREATE TABLE IF NOT EXISTS container_state (
+        id                       INTEGER PRIMARY KEY CHECK (id = 1),
+        current_tool             TEXT,
+        tool_declared_timeout_ms INTEGER,
+        tool_started_at          TEXT,
+        updated_at               TEXT NOT NULL
+      );
+    `);
   }
   return _outbound;
 }
 
+/**
+ * Record that a tool is starting. `declaredTimeoutMs` is the tool's own
+ * timeout hint when one is available (Bash exposes it in the tool_use input);
+ * omit for tools with no declared timeout.
+ */
+export function setContainerToolInFlight(tool: string, declaredTimeoutMs: number | null): void {
+  const now = new Date().toISOString();
+  getOutboundDb()
+    .prepare(
+      `INSERT INTO container_state (id, current_tool, tool_declared_timeout_ms, tool_started_at, updated_at)
+       VALUES (1, ?, ?, ?, ?)
+       ON CONFLICT(id) DO UPDATE SET
+         current_tool = excluded.current_tool,
+         tool_declared_timeout_ms = excluded.tool_declared_timeout_ms,
+         tool_started_at = excluded.tool_started_at,
+         updated_at = excluded.updated_at`,
+    )
+    .run(tool, declaredTimeoutMs, now, now);
+}
+
+/** Clear the in-flight tool — called on PostToolUse / PostToolUseFailure. */
+export function clearContainerToolInFlight(): void {
+  const now = new Date().toISOString();
+  getOutboundDb()
+    .prepare(
+      `INSERT INTO container_state (id, current_tool, tool_declared_timeout_ms, tool_started_at, updated_at)
+       VALUES (1, NULL, NULL, NULL, ?)
+       ON CONFLICT(id) DO UPDATE SET
+         current_tool = NULL,
+         tool_declared_timeout_ms = NULL,
+         tool_started_at = NULL,
+         updated_at = excluded.updated_at`,
+    )
+    .run(now);
+}
+
 /**
  * Touch the heartbeat file — replaces the old touchProcessing() DB writes.
  * The host checks this file's mtime for stale container detection.
  * A file touch is cheaper and avoids cross-boundary DB write contention.
  */
 export function touchHeartbeat(): void {
-  const p = process.env.SESSION_HEARTBEAT_PATH || _heartbeatPath;
+  const p = _heartbeatPath;
   const now = new Date();
   try {
     fs.utimesSync(p, now, now);
@@ -109,7 +155,9 @@ export function initTestSessionDb(): { inbound: Database; outbound: Database } {
       status         TEXT DEFAULT 'pending',
       process_after  TEXT,
       recurrence     TEXT,
+      series_id      TEXT,
       tries          INTEGER DEFAULT 0,
+      trigger        INTEGER NOT NULL DEFAULT 1,
       platform_id    TEXT,
       channel_type   TEXT,
       thread_id      TEXT,
@@ -157,6 +205,13 @@ export function initTestSessionDb(): { inbound: Database; outbound: Database } {
       value      TEXT NOT NULL,
       updated_at TEXT NOT NULL
     );
+    CREATE TABLE container_state (
+      id                       INTEGER PRIMARY KEY CHECK (id = 1),
+      current_tool             TEXT,
+      tool_declared_timeout_ms INTEGER,
+      tool_started_at          TEXT,
+      updated_at               TEXT NOT NULL
+    );
   `);
 
   return { inbound: _inbound, outbound: _outbound };
diff --git a/container/agent-runner/src/db/messages-in.ts b/container/agent-runner/src/db/messages-in.ts
index da1a8dd..4ecf818 100644
--- a/container/agent-runner/src/db/messages-in.ts
+++ b/container/agent-runner/src/db/messages-in.ts
@@ -7,6 +7,7 @@
  * The container never writes to inbound.db — all status tracking goes through
  * processing_ack. The host reads processing_ack to sync message lifecycle.
  */
+import { getConfig } from '../config.js';
 import { getInboundDb, getOutboundDb } from './connection.js';
 
 export interface MessageInRow {
@@ -18,16 +19,35 @@ export interface MessageInRow {
   process_after: string | null;
   recurrence: string | null;
   tries: number;
+  /** 1 = wake-eligible (default); 0 = accumulated context only */
+  trigger: number;
   platform_id: string | null;
   channel_type: string | null;
   thread_id: string | null;
   content: string;
 }
 
+// Cap on how many messages reach the agent in one prompt. Read from
+// container.json; falls back to 10.
+function getMaxMessagesPerPrompt(): number {
+  try {
+    return getConfig().maxMessagesPerPrompt;
+  } catch {
+    // Config not loaded yet (e.g. test harness) — use default
+    return 10;
+  }
+}
+
 /**
  * Fetch pending messages that are due for processing.
  * Reads from inbound.db (read-only), filters against processing_ack in outbound.db
  * to skip messages already picked up by this or a previous container run.
+ *
+ * Returns the most recent `MAX_MESSAGES_PER_PROMPT` pending rows in
+ * chronological order, regardless of their `trigger` flag: accumulated
+ * context (trigger=0) rides along with the wake-eligible rows so the agent
+ * sees the prior context it missed. Host's countDueMessages gates waking on
+ * trigger=1 separately (see src/db/session-db.ts).
  */
 export function getPendingMessages(): MessageInRow[] {
   const inbound = getInboundDb();
@@ -38,9 +58,10 @@ export function getPendingMessages(): MessageInRow[] {
       `SELECT * FROM messages_in
        WHERE status = 'pending'
          AND (process_after IS NULL OR datetime(process_after) <= datetime('now'))
-       ORDER BY timestamp ASC`,
+       ORDER BY seq DESC
+       LIMIT ?`,
     )
-    .all() as MessageInRow[];
+    .all(getMaxMessagesPerPrompt()) as MessageInRow[];
 
   if (pending.length === 0) return [];
 
@@ -51,7 +72,9 @@ export function getPendingMessages(): MessageInRow[] {
     ),
   );
 
-  return pending.filter((m) => !ackedIds.has(m.id));
+  // Reverse: we fetched DESC to take the most recent N, but the agent
+  // should see them in chronological order (oldest first).
+  return pending.filter((m) => !ackedIds.has(m.id)).reverse();
 }
 
 /** Mark messages as processing — writes to processing_ack in outbound.db. */
diff --git a/container/agent-runner/src/db/session-state.test.ts b/container/agent-runner/src/db/session-state.test.ts
new file mode 100644
index 0000000..b5aa269
--- /dev/null
+++ b/container/agent-runner/src/db/session-state.test.ts
@@ -0,0 +1,100 @@
+import { beforeEach, describe, expect, test } from 'bun:test';
+
+import { getOutboundDb, initTestSessionDb } from './connection.js';
+import {
+  clearContinuation,
+  getContinuation,
+  migrateLegacyContinuation,
+  setContinuation,
+} from './session-state.js';
+
+beforeEach(() => {
+  initTestSessionDb();
+});
+
+function seedLegacy(value: string): void {
+  getOutboundDb()
+    .prepare('INSERT INTO session_state (key, value, updated_at) VALUES (?, ?, ?)')
+    .run('sdk_session_id', value, new Date().toISOString());
+}
+
+describe('session-state — per-provider continuations', () => {
+  test('set/get round-trip, case-insensitive provider key', () => {
+    setContinuation('claude', 'claude-conv-1');
+    expect(getContinuation('claude')).toBe('claude-conv-1');
+    expect(getContinuation('Claude')).toBe('claude-conv-1');
+    expect(getContinuation('CLAUDE')).toBe('claude-conv-1');
+  });
+
+  test('providers are isolated — switching reads the right slot', () => {
+    setContinuation('claude', 'claude-conv-1');
+    setContinuation('codex', 'codex-thread-xyz');
+
+    expect(getContinuation('claude')).toBe('claude-conv-1');
+    expect(getContinuation('codex')).toBe('codex-thread-xyz');
+  });
+
+  test('clearContinuation only affects the specified provider', () => {
+    setContinuation('claude', 'keep-me');
+    setContinuation('codex', 'drop-me');
+
+    clearContinuation('codex');
+
+    expect(getContinuation('claude')).toBe('keep-me');
+    expect(getContinuation('codex')).toBeUndefined();
+  });
+
+  test('unknown provider returns undefined', () => {
+    expect(getContinuation('never-used')).toBeUndefined();
+  });
+});
+
+describe('session-state — legacy migration', () => {
+  test('adopts legacy value into current provider when current is empty', () => {
+    seedLegacy('old-session-id');
+
+    const adopted = migrateLegacyContinuation('claude');
+
+    expect(adopted).toBe('old-session-id');
+    expect(getContinuation('claude')).toBe('old-session-id');
+  });
+
+  test('always deletes legacy row regardless of migration outcome', () => {
+    seedLegacy('old-session-id');
+    setContinuation('claude', 'existing');
+
+    migrateLegacyContinuation('claude');
+
+    // After migration the legacy key must be gone, whether or not it was adopted.
+    // A subsequent migration for a different provider must not see it.
+    const resultAfterSecondCall = migrateLegacyContinuation('codex');
+    expect(resultAfterSecondCall).toBeUndefined();
+  });
+
+  test('prefers existing current-provider slot over legacy', () => {
+    seedLegacy('legacy-value');
+    setContinuation('claude', 'claude-value');
+
+    const result = migrateLegacyContinuation('claude');
+
+    expect(result).toBe('claude-value');
+    expect(getContinuation('claude')).toBe('claude-value');
+  });
+
+  test('no legacy row — returns current provider value (possibly undefined)', () => {
+    expect(migrateLegacyContinuation('claude')).toBeUndefined();
+
+    setContinuation('codex', 'codex-value');
+    expect(migrateLegacyContinuation('codex')).toBe('codex-value');
+  });
+
+  test('migration is idempotent on a second call (legacy already gone)', () => {
+    seedLegacy('once');
+
+    const first = migrateLegacyContinuation('claude');
+    expect(first).toBe('once');
+
+    const second = migrateLegacyContinuation('claude');
+    expect(second).toBe('once');
+  });
+});
diff --git a/container/agent-runner/src/db/session-state.ts b/container/agent-runner/src/db/session-state.ts
index a199ae1..9e12309 100644
--- a/container/agent-runner/src/db/session-state.ts
+++ b/container/agent-runner/src/db/session-state.ts
@@ -2,12 +2,20 @@
  * Persistent key/value state for the container. Lives in outbound.db
  * (container-owned, already scoped per channel/thread).
  *
- * Primary use: remember the SDK session ID so the agent's conversation
- * resumes across container restarts. Cleared by /clear.
+ * Primary use: remember each provider's opaque continuation id so the
+ * agent's conversation resumes across container restarts. Keyed per
+ * provider because continuations are provider-private — a Claude
+ * conversation id means nothing to Codex and vice versa. Switching
+ * providers is therefore lossless: each provider's last thread stays
+ * on file and resumes cleanly if the user flips back.
  */
 import { getOutboundDb } from './connection.js';
 
-const SDK_SESSION_KEY = 'sdk_session_id';
+const LEGACY_KEY = 'sdk_session_id';
+
+function continuationKey(providerName: string): string {
+  return `continuation:${providerName.toLowerCase()}`;
+}
 
 function getValue(key: string): string | undefined {
   const row = getOutboundDb()
@@ -18,9 +26,7 @@ function getValue(key: string): string | undefined {
 
 function setValue(key: string, value: string): void {
   getOutboundDb()
-    .prepare(
-      'INSERT OR REPLACE INTO session_state (key, value, updated_at) VALUES (?, ?, ?)',
-    )
+    .prepare('INSERT OR REPLACE INTO session_state (key, value, updated_at) VALUES (?, ?, ?)')
     .run(key, value, new Date().toISOString());
 }
 
@@ -28,14 +34,46 @@ function deleteValue(key: string): void {
   getOutboundDb().prepare('DELETE FROM session_state WHERE key = ?').run(key);
 }
 
-export function getStoredSessionId(): string | undefined {
-  return getValue(SDK_SESSION_KEY);
+/**
+ * One-time migration of the pre-per-provider continuation row.
+ *
+ * Before this was keyed per provider, continuations lived under the
+ * single key `sdk_session_id`. On container start, if that legacy row
+ * exists and the current provider has no continuation of its own, adopt
+ * the legacy value into the current provider's slot (best-guess — the
+ * legacy row was written by whatever provider ran last). The legacy row
+ * is always deleted so future provider flips never re-read a stale id
+ * through the wrong lens.
+ *
+ * Returns the continuation the caller should use at startup (either the
+ * current provider's existing value, the adopted legacy value, or
+ * undefined).
+ */
+export function migrateLegacyContinuation(providerName: string): string | undefined {
+  const legacy = getValue(LEGACY_KEY);
+  const currentKey = continuationKey(providerName);
+  const current = getValue(currentKey);
+
+  if (legacy === undefined) return current;
+
+  // Always drop the legacy row so no future provider reads it.
+  deleteValue(LEGACY_KEY);
+
+  // Prefer the current provider's own slot if one already exists.
+  if (current !== undefined) return current;
+
+  setValue(currentKey, legacy);
+  return legacy;
 }
 
-export function setStoredSessionId(sessionId: string): void {
-  setValue(SDK_SESSION_KEY, sessionId);
+export function getContinuation(providerName: string): string | undefined {
+  return getValue(continuationKey(providerName));
 }
 
-export function clearStoredSessionId(): void {
-  deleteValue(SDK_SESSION_KEY);
+export function setContinuation(providerName: string, id: string): void {
+  setValue(continuationKey(providerName), id);
+}
+
+export function clearContinuation(providerName: string): void {
+  deleteValue(continuationKey(providerName));
 }
diff --git a/container/agent-runner/src/destinations.ts b/container/agent-runner/src/destinations.ts
index d525cf1..013bd3b 100644
--- a/container/agent-runner/src/destinations.ts
+++ b/container/agent-runner/src/destinations.ts
@@ -72,8 +72,26 @@ export function findByRouting(
   return row ? rowToEntry(row) : undefined;
 }
 
-/** Generate the system-prompt addendum describing destinations and syntax. */
-export function buildSystemPromptAddendum(): string {
+/**
+ * Generate the system-prompt addendum: agent identity + destination map.
+ *
+ * Identity is injected here (not in the shared CLAUDE.md) because it's
+ * per-agent-group and changes when the operator renames an agent, while
+ * the shared base is identical across all agents.
+ */
+export function buildSystemPromptAddendum(assistantName?: string): string {
+  const sections: string[] = [];
+
+  if (assistantName) {
+    sections.push(['# You are ' + assistantName, '', `Your name is **${assistantName}**. Use it when the channel asks who you are, when introducing yourself, and when signing any message that explicitly calls for a signature.`].join('\n'));
+  }
+
+  sections.push(buildDestinationsSection());
+
+  return sections.join('\n\n');
+}
+
+function buildDestinationsSection(): string {
   const all = getAllDestinations();
 
   if (all.length === 0) {
diff --git a/container/agent-runner/src/formatter.test.ts b/container/agent-runner/src/formatter.test.ts
new file mode 100644
index 0000000..e34156c
--- /dev/null
+++ b/container/agent-runner/src/formatter.test.ts
@@ -0,0 +1,167 @@
+/**
+ * v1-parity tests for formatter behavior.
+ *
+ * Port of src/v1/formatting.test.ts (at commit 27c5220, parent of the v1
+ * deletion commit 86becf8). Covers: context timezone header, reply_to +
+ * quoted_message rendering, XML escaping, and stripInternalTags.
+ *
+ * Timestamp-format assertions use `formatLocalTime()` output format, which
+ * is host locale-dependent for decorators (month abbr, "," separator) but
+ * stable for the numeric parts we assert on (hour, minute, year).
+ */
+import { describe, it, expect, beforeEach, afterEach } from 'bun:test';
+
+import { initTestSessionDb, closeSessionDb, getInboundDb } from './db/connection.js';
+import { getPendingMessages } from './db/messages-in.js';
+import { formatMessages, stripInternalTags } from './formatter.js';
+import { TIMEZONE } from './timezone.js';
+
+beforeEach(() => {
+  initTestSessionDb();
+});
+
+afterEach(() => {
+  closeSessionDb();
+});
+
+function insertMessage(
+  id: string,
+  kind: string,
+  content: object,
+  opts?: { timestamp?: string },
+) {
+  const timestamp = opts?.timestamp ?? new Date().toISOString();
+  getInboundDb()
+    .prepare(
+      `INSERT INTO messages_in (id, kind, timestamp, status, content)
+       VALUES (?, ?, ?, 'pending', ?)`,
+    )
+    .run(id, kind, timestamp, JSON.stringify(content));
+}
+
+describe('context timezone header', () => {
+  it('prepends <context timezone="..."/> to formatted output', () => {
+    insertMessage('m1', 'chat', { sender: 'Alice', text: 'hello' });
+    const result = formatMessages(getPendingMessages());
+    expect(result).toContain(`<context timezone="${TIMEZONE}"`);
+  });
+
+  it('includes the header even when the message list is empty', () => {
+    const result = formatMessages([]);
+    expect(result).toContain(`<context timezone="${TIMEZONE}"`);
+  });
+
+  it('header comes before the <messages> block', () => {
+    insertMessage('m1', 'chat', { sender: 'Alice', text: 'one' });
+    insertMessage('m2', 'chat', { sender: 'Bob', text: 'two' });
+    const result = formatMessages(getPendingMessages());
+    const ctxIdx = result.indexOf('<context');
+    const msgsIdx = result.indexOf('<messages>');
+    expect(ctxIdx).toBeGreaterThanOrEqual(0);
+    expect(msgsIdx).toBeGreaterThan(ctxIdx);
+  });
+});
+
+describe('timestamp formatting', () => {
+  it('renders time via formatLocalTime (user TZ)', () => {
+    // 2026-06-15T12:00:00Z — timezone-agnostic assertions (year is stable)
+    insertMessage('m1', 'chat', { sender: 'Alice', text: 'hi' }, { timestamp: '2026-06-15T12:00:00.000Z' });
+    const result = formatMessages(getPendingMessages());
+    // formatLocalTime's format in en-US contains the year and a month abbrev
+    expect(result).toContain('2026');
+    expect(result).toMatch(/Jun/);
+  });
+
+  it('uses 12-hour AM/PM format', () => {
+    // 15:30 UTC — some hour will show with AM or PM depending on TZ
+    insertMessage('m1', 'chat', { sender: 'Alice', text: 'hi' }, { timestamp: '2026-06-15T15:30:00.000Z' });
+    const result = formatMessages(getPendingMessages());
+    expect(result).toMatch(/(AM|PM)/);
+  });
+});
+
+describe('reply_to + quoted_message rendering', () => {
+  it('renders reply_to attribute and quoted_message when all fields present', () => {
+    insertMessage('m1', 'chat', {
+      sender: 'Alice',
+      text: 'Yes, on my way!',
+      replyTo: { id: '42', sender: 'Bob', text: 'Are you coming tonight?' },
+    });
+    const result = formatMessages(getPendingMessages());
+    expect(result).toContain('reply_to="42"');
+    expect(result).toContain('<quoted_message from="Bob">Are you coming tonight?</quoted_message>');
+    expect(result).toContain('Yes, on my way!</message>');
+  });
+
+  it('omits reply_to and quoted_message when no reply context', () => {
+    insertMessage('m1', 'chat', { sender: 'Alice', text: 'plain' });
+    const result = formatMessages(getPendingMessages());
+    expect(result).not.toContain('reply_to');
+    expect(result).not.toContain('quoted_message');
+  });
+
+  it('renders reply_to but omits quoted_message when original content is missing', () => {
+    insertMessage('m1', 'chat', {
+      sender: 'Alice',
+      text: 'ack',
+      replyTo: { id: '42', sender: 'Bob' }, // no text
+    });
+    const result = formatMessages(getPendingMessages());
+    expect(result).toContain('reply_to="42"');
+    expect(result).not.toContain('quoted_message');
+  });
+
+  it('XML-escapes reply context', () => {
+    insertMessage('m1', 'chat', {
+      sender: 'Alice',
+      text: 'reply',
+      replyTo: { id: '1', sender: 'A & B', text: '<script>alert("xss")</script>' },
+    });
+    const result = formatMessages(getPendingMessages());
+    expect(result).toContain('from="A &amp; B"');
+    expect(result).toContain('&lt;script&gt;');
+    expect(result).toContain('&quot;xss&quot;');
+  });
+});
+
+describe('XML escaping', () => {
+  it('escapes <, >, &, " in sender and body', () => {
+    insertMessage('m1', 'chat', {
+      sender: 'A & B <Co>',
+      text: '<script>alert("xss")</script>',
+    });
+    const result = formatMessages(getPendingMessages());
+    expect(result).toContain('sender="A &amp; B &lt;Co&gt;"');
+    expect(result).toContain('&lt;script&gt;alert(&quot;xss&quot;)&lt;/script&gt;');
+  });
+});
+
+describe('stripInternalTags', () => {
+  it('strips single-line internal tags and trims', () => {
+    expect(stripInternalTags('hello <internal>secret</internal> world')).toBe('hello  world');
+  });
+
+  it('strips multi-line internal tags', () => {
+    expect(stripInternalTags('hello <internal>\nsecret\nstuff\n</internal> world')).toBe(
+      'hello  world',
+    );
+  });
+
+  it('strips multiple internal tag blocks', () => {
+    expect(stripInternalTags('<internal>a</internal>hello<internal>b</internal>')).toBe('hello');
+  });
+
+  it('returns empty string when input is only internal tags', () => {
+    expect(stripInternalTags('<internal>only this</internal>')).toBe('');
+  });
+
+  it('returns input unchanged when there are no internal tags', () => {
+    expect(stripInternalTags('hello world')).toBe('hello world');
+  });
+
+  it('preserves content that surrounds internal tags', () => {
+    expect(stripInternalTags('<internal>thinking</internal>The answer is 42')).toBe(
+      'The answer is 42',
+    );
+  });
+});
diff --git a/container/agent-runner/src/formatter.ts b/container/agent-runner/src/formatter.ts
index fbf1ed9..c0475b2 100644
--- a/container/agent-runner/src/formatter.ts
+++ b/container/agent-runner/src/formatter.ts
@@ -1,5 +1,6 @@
 import { findByRouting } from './destinations.js';
 import type { MessageInRow } from './db/messages-in.js';
+import { TIMEZONE, formatLocalTime } from './timezone.js';
 
 /**
  * Command categories for messages starting with '/'.
@@ -11,7 +12,7 @@ import type { MessageInRow } from './db/messages-in.js';
 export type CommandCategory = 'admin' | 'filtered' | 'passthrough' | 'none';
 
 const ADMIN_COMMANDS = new Set(['/remote-control', '/clear', '/compact', '/context', '/cost', '/files']);
-const FILTERED_COMMANDS = new Set(['/help', '/login', '/logout', '/doctor', '/config']);
+const FILTERED_COMMANDS = new Set(['/help', '/login', '/logout', '/doctor', '/config', '/start']);
 
 export interface CommandInfo {
   category: CommandCategory;
@@ -54,6 +55,17 @@ export function categorizeMessage(msg: MessageInRow): CommandInfo {
   return { category: 'passthrough', command, text, senderId };
 }
 
+/**
+ * Narrow check for /clear — the only command the runner handles directly.
+ * All other command gating (filtered, admin) is done by the host router
+ * before messages reach the container.
+ */
+export function isClearCommand(msg: MessageInRow): boolean {
+  const content = parseContent(msg.content);
+  const text = (content.text || '').trim();
+  return text.toLowerCase().startsWith('/clear');
+}
+
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
 function extractSenderId(msg: MessageInRow, content: any): string | null {
   const raw: string | null = content?.senderId || content?.author?.userId || null;
@@ -92,10 +104,19 @@ export function extractRouting(messages: MessageInRow[]): RoutingContext {
 
 /**
  * Format a batch of messages_in rows into a prompt string.
+ *
+ * Prepends a `<context timezone="<IANA>" />` header so the agent always knows
+ * what timezone it's in — every timestamp it sees in message bodies is the
+ * user's local time, and every time it produces (schedules, suggests) should
+ * be interpreted as local time in that same zone. This header is v1 behavior
+ * (src/v1/router.ts:20-22); dropping it led to misinterpretations where the
+ * agent scheduled tasks for the wrong hour.
+ *
  * Strips routing fields — the agent never sees platform_id, channel_type, thread_id.
  */
 export function formatMessages(messages: MessageInRow[]): string {
-  if (messages.length === 0) return '';
+  const header = `<context timezone="${escapeXml(TIMEZONE)}" />\n`;
+  if (messages.length === 0) return header;
 
   // Group by kind
   const chatMessages = messages.filter((m) => m.kind === 'chat' || m.kind === 'chat-sdk');
@@ -118,7 +139,7 @@ export function formatMessages(messages: MessageInRow[]): string {
     parts.push(...systemMessages.map(formatSystemMessage));
   }
 
-  return parts.join('\n\n');
+  return header + parts.join('\n\n');
 }
 
 function formatChatMessages(messages: MessageInRow[]): string {
@@ -137,9 +158,10 @@ function formatChatMessages(messages: MessageInRow[]): string {
 function formatSingleChat(msg: MessageInRow): string {
   const content = parseContent(msg.content);
   const sender = content.sender || content.author?.fullName || content.author?.userName || 'Unknown';
-  const time = formatTime(msg.timestamp);
+  const time = formatLocalTime(msg.timestamp, TIMEZONE);
   const text = content.text || '';
   const idAttr = msg.seq != null ? ` id="${msg.seq}"` : '';
+  const replyAttr = content.replyTo?.id ? ` reply_to="${escapeXml(String(content.replyTo.id))}"` : '';
   const replyPrefix = formatReplyContext(content.replyTo);
   const attachmentsSuffix = formatAttachments(content.attachments);
 
@@ -154,7 +176,7 @@ function formatSingleChat(msg: MessageInRow): string {
       ? ` from="unknown:${escapeXml(msg.channel_type || '')}:${escapeXml(msg.platform_id || '')}"`
       : '';
 
-  return `<message${idAttr}${fromAttr} sender="${escapeXml(sender)}" time="${time}">${replyPrefix}${escapeXml(text)}${attachmentsSuffix}</message>`;
+  return `<message${idAttr}${fromAttr} sender="${escapeXml(sender)}" time="${escapeXml(time)}"${replyAttr}>${replyPrefix}${escapeXml(text)}${attachmentsSuffix}</message>`;
 }
 
 function formatTaskMessage(msg: MessageInRow): string {
@@ -179,13 +201,22 @@ function formatSystemMessage(msg: MessageInRow): string {
   return `[SYSTEM RESPONSE]\n\nAction: ${content.action || 'unknown'}\nStatus: ${content.status || 'unknown'}\nResult: ${JSON.stringify(content.result || null)}`;
 }
 
+/**
+ * Render the quoted original inside the <message> body.
+ *
+ * Matches v1 format (src/v1/router.ts:10-18): `<quoted_message from="X">Y</quoted_message>`.
+ * Requires BOTH sender and text — if only id is present the reply_to attribute
+ * on the parent <message> carries the link without an inline preview.
+ *
+ * No truncation here (v1 didn't truncate).
+ */
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
 function formatReplyContext(replyTo: any): string {
   if (!replyTo) return '';
-  const sender = replyTo.sender || 'Unknown';
-  const text = replyTo.text || '';
-  const preview = text.length > 100 ? text.slice(0, 100) + '…' : text;
-  return `\n<reply-to sender="${escapeXml(sender)}">${escapeXml(preview)}</reply-to>\n`;
+  const sender = replyTo.sender;
+  const text = replyTo.text;
+  if (!sender || !text) return '';
+  return `\n  <quoted_message from="${escapeXml(sender)}">${escapeXml(text)}</quoted_message>\n`;
 }
 
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
@@ -213,15 +244,15 @@ function parseContent(json: string): any {
   }
 }
 
-function formatTime(timestamp: string): string {
-  try {
-    const d = new Date(timestamp);
-    return `${d.getHours().toString().padStart(2, '0')}:${d.getMinutes().toString().padStart(2, '0')}`;
-  } catch {
-    return timestamp;
-  }
-}
-
 function escapeXml(str: string): string {
   return str.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;').replace(/"/g, '&quot;');
 }
+
+/**
+ * Strip `<internal>...</internal>` blocks from agent output, then trim.
+ * Ported from v1 (src/v1/router.ts:25-27). Used to remove the agent's
+ * own scratchpad/reasoning before a reply goes out over a channel.
+ */
+export function stripInternalTags(text: string): string {
+  return text.replace(/<internal>[\s\S]*?<\/internal>/g, '').trim();
+}
diff --git a/container/agent-runner/src/index.ts b/container/agent-runner/src/index.ts
index a0b0dc8..90c690f 100644
--- a/container/agent-runner/src/index.ts
+++ b/container/agent-runner/src/index.ts
@@ -4,14 +4,8 @@
  * Runs inside a container. All IO goes through the session DB.
  * No stdin, no stdout markers, no IPC files.
  *
- * Config:
- *   - SESSION_INBOUND_DB_PATH:  path to host-owned inbound DB (default: /workspace/inbound.db)
- *   - SESSION_OUTBOUND_DB_PATH: path to container-owned outbound DB (default: /workspace/outbound.db)
- *   - SESSION_HEARTBEAT_PATH:   heartbeat file path (default: /workspace/.heartbeat)
- *   - AGENT_PROVIDER: any registered provider name (default: claude). The
- *     set of registered providers is whatever `providers/index.ts` imports.
- *   - NANOCLAW_ASSISTANT_NAME: assistant name for transcript archiving
- *   - NANOCLAW_ADMIN_USER_IDS: comma-separated user IDs allowed to run admin commands
+ * Config is read from /workspace/agent/container.json (mounted RO).
+ * Only TZ and OneCLI networking vars come from env.
  *
  * Mount structure:
  *   /workspace/
@@ -19,14 +13,19 @@
  *     outbound.db       ← container-owned session DB
  *     .heartbeat        ← container touches for liveness detection
  *     outbox/           ← outbound files
- *     agent/            ← agent group folder (CLAUDE.md, skills, working files)
- *     .claude/          ← Claude SDK session data
+ *     agent/            ← agent group folder (CLAUDE.md, container.json, working files)
+ *       container.json  ← per-group config (RO nested mount)
+ *     global/           ← shared global memory (RO)
+ *   /app/src/           ← shared agent-runner source (RO)
+ *   /app/skills/        ← shared skills (RO)
+ *   /home/node/.claude/ ← Claude SDK state + skill symlinks (RW)
  */
 
 import fs from 'fs';
 import path from 'path';
 import { fileURLToPath } from 'url';
 
+import { loadConfig } from './config.js';
 import { buildSystemPromptAddendum } from './destinations.js';
 // Providers barrel — each enabled provider self-registers on import.
 // Provider skills append imports to providers/index.ts.
@@ -41,22 +40,18 @@ function log(msg: string): void {
 const CWD = '/workspace/agent';
 
 async function main(): Promise<void> {
-  const providerName = (process.env.AGENT_PROVIDER || 'claude').toLowerCase() as ProviderName;
-  const assistantName = process.env.NANOCLAW_ASSISTANT_NAME;
-  const adminUserIds = new Set(
-    (process.env.NANOCLAW_ADMIN_USER_IDS || '')
-      .split(',')
-      .map((s) => s.trim())
-      .filter(Boolean),
-  );
+  const config = loadConfig();
+  const providerName = config.provider.toLowerCase() as ProviderName;
 
   log(`Starting v2 agent-runner (provider: ${providerName})`);
 
-  // Destinations addendum is the only runtime-generated context we inject.
-  // Global CLAUDE.md is loaded by Claude Code from /workspace/agent/CLAUDE.md
-  // (which imports /workspace/global/CLAUDE.md via @-syntax) — no need to
-  // read it manually anymore.
-  const instructions = buildSystemPromptAddendum();
+  // Runtime-generated system-prompt addendum: agent identity (name) plus
+  // the live destinations map. Everything else (capabilities, per-module
+  // instructions, per-channel formatting) is loaded by Claude Code from
+  // /workspace/agent/CLAUDE.md — the composed entry imports the shared
+  // base (/app/CLAUDE.md) and each enabled module's fragment. Per-group
+  // memory lives in /workspace/agent/CLAUDE.local.md (auto-loaded).
+  const instructions = buildSystemPromptAddendum(config.assistantName || undefined);
 
   // Discover additional directories mounted at /workspace/extra/*
   const additionalDirectories: string[] = [];
@@ -77,34 +72,22 @@ async function main(): Promise<void> {
   const __dirname = path.dirname(fileURLToPath(import.meta.url));
   const mcpServerPath = path.join(__dirname, 'mcp-tools', 'index.ts');
 
-  // Build MCP servers config: nanoclaw built-in + any additional from host
+  // Build MCP servers config: nanoclaw built-in + any from container.json
   const mcpServers: Record<string, { command: string; args: string[]; env: Record<string, string> }> = {
     nanoclaw: {
       command: 'bun',
       args: ['run', mcpServerPath],
-      env: {
-        SESSION_INBOUND_DB_PATH: process.env.SESSION_INBOUND_DB_PATH || '/workspace/inbound.db',
-        SESSION_OUTBOUND_DB_PATH: process.env.SESSION_OUTBOUND_DB_PATH || '/workspace/outbound.db',
-        SESSION_HEARTBEAT_PATH: process.env.SESSION_HEARTBEAT_PATH || '/workspace/.heartbeat',
-      },
+      env: {},
     },
   };
 
-  // Merge additional MCP servers from host configuration
-  if (process.env.NANOCLAW_MCP_SERVERS) {
-    try {
-      const additional = JSON.parse(process.env.NANOCLAW_MCP_SERVERS) as Record<string, { command: string; args: string[]; env: Record<string, string> }>;
-      for (const [name, config] of Object.entries(additional)) {
-        mcpServers[name] = config;
-        log(`Additional MCP server: ${name} (${config.command})`);
-      }
-    } catch (e) {
-      log(`Failed to parse NANOCLAW_MCP_SERVERS: ${e}`);
-    }
+  for (const [name, serverConfig] of Object.entries(config.mcpServers)) {
+    mcpServers[name] = serverConfig;
+    log(`Additional MCP server: ${name} (${serverConfig.command})`);
   }
 
   const provider = createProvider(providerName, {
-    assistantName,
+    assistantName: config.assistantName || undefined,
     mcpServers,
     env: { ...process.env },
     additionalDirectories: additionalDirectories.length > 0 ? additionalDirectories : undefined,
@@ -112,9 +95,9 @@ async function main(): Promise<void> {
 
   await runPollLoop({
     provider,
+    providerName,
     cwd: CWD,
     systemContext: { instructions },
-    adminUserIds,
   });
 }
 
diff --git a/container/agent-runner/src/integration.test.ts b/container/agent-runner/src/integration.test.ts
index 4a8b091..3447c38 100644
--- a/container/agent-runner/src/integration.test.ts
+++ b/container/agent-runner/src/integration.test.ts
@@ -98,6 +98,7 @@ async function runPollLoopWithTimeout(provider: MockProvider, signal: AbortSigna
   return Promise.race([
     runPollLoop({
       provider,
+      providerName: 'mock',
       cwd: '/tmp',
     }),
     new Promise<void>((_, reject) => {
diff --git a/container/agent-runner/src/mcp-tools/agents.instructions.md b/container/agent-runner/src/mcp-tools/agents.instructions.md
new file mode 100644
index 0000000..8ada129
--- /dev/null
+++ b/container/agent-runner/src/mcp-tools/agents.instructions.md
@@ -0,0 +1,26 @@
+## Companion and collaborator agents (`create_agent`)
+
+`mcp__nanoclaw__create_agent({ name, instructions })` spins up a new long-lived agent and wires it as a destination — bidirectional, so you can send it tasks and it can message you back.
+
+### How it works
+
+- Creates a new agent with its own container, workspace, and session. Your `instructions` string seeds the agent's `CLAUDE.local.md` — its starting role and personality.
+- The agent's `name` becomes a destination on both sides: you address it via `send_message({ to: "<name>", ... })`, and its replies arrive as inbound messages with `from="<name>"`.
+- Each agent has its own persistent workspace under `groups/<folder>/` — memory, conversation history, and notes all survive across sessions. This is a full standalone agent, not a stateless sub-query.
+- **Fire-and-forget:** the call returns immediately without waiting for the agent to confirm it's ready. Messages you send will queue until it's up.
+
+### When to use
+
+- **Companions** — a long-running presence that accumulates context over time: a `Researcher` tracking an ongoing inquiry, a `Calendar` agent managing scheduling, an assistant that knows your preferences and history.
+- **Collaborators** — a parallel specialist that works independently and reports back: a `Builder` handling code edits while you stay in conversation, a `Reviewer` running checks in the background.
+
+The right frame is: does this agent need its own memory and context that builds over time, or does it need to work independently without blocking your turn? Either is a good reason to spawn one.
+
+### When NOT to use
+
+- **One-off lookups or short tasks** — use the SDK `Agent` tool instead. It's stateless, spins up and completes in one shot, and leaves no persistent footprint.
+- **Work that finishes before the user's next message** — agents persist indefinitely. Don't create one for something you could do inline.
+
+### Writing good `instructions`
+
+Cover: the agent's role, who it takes tasks from (you, by name), how it should report back (on completion only? with milestones for long work?), and any domain-specific rules. Don't restate NanoClaw base behavior — the shared base is already loaded on the agent's end.
\ No newline at end of file
diff --git a/container/agent-runner/src/mcp-tools/core.instructions.md b/container/agent-runner/src/mcp-tools/core.instructions.md
new file mode 100644
index 0000000..d9995bf
--- /dev/null
+++ b/container/agent-runner/src/mcp-tools/core.instructions.md
@@ -0,0 +1,27 @@
+## Sending messages
+
+Your final response is delivered via the `## Sending messages` rules in your runtime system prompt (single-destination: just write; multi-destination: use `<message to="name">...</message>` blocks). See that section for the current destination list.
+
+### Mid-turn updates (`send_message`)
+
+Use the `mcp__nanoclaw__send_message` tool to send a message while you're still working (before your final output). If you have one destination, `to` is optional; with multiple, specify it. Pace your updates to the length of the work:
+
+- **Short turn (≤2 quick tool calls):** Don't narrate. Output any response.
+- **Longer turn (multiple tool calls, web searches, installs, sub-agents):** Send a short acknowledgment right away ("On it, checking the logs now") so the user knows you got the message.
+- **Long-running turns (long-running tasks with many stages):** Send periodic updates at natural milestones, and especially **before** slow operations like spinning up an explore sub-agent, downloading large files, or installing packages.
+
+**Never narrate micro-steps.** "I'm going to read the file now… okay, I'm reading it… now I'm parsing it…" is noise. Updates should mark meaningful transitions, not every tool call.
+
+**Outcomes, not play-by-play.** When the turn is done, the final message should be about the result, not a transcript of what you did.
+
+### Sending files (`send_file`)
+
+Use `mcp__nanoclaw__send_file({ path, text?, filename?, to? })` to deliver a file from your workspace. `path` is absolute or relative to `/workspace/agent/`; `filename` overrides the display name shown in chat (defaults to the file's basename); `text` is an optional accompanying message. Use this for artifacts you produce (charts, PDFs, generated images, reports) rather than dumping contents into chat.
+
+### Reacting to messages (`add_reaction`)
+
+Use `mcp__nanoclaw__add_reaction({ messageId, emoji })` to react to a specific inbound message by its `#N` id — pass `messageId` as an integer (e.g. `22`, not `"22"`). Good for lightweight acknowledgment (`eyes` = seen, `white_check_mark` = done) when a full reply would be noise. `emoji` is the shortcode name (e.g. `thumbs_up`, `heart`), not the raw character.
+
+### Internal thoughts
+
+Wrap reasoning in `<internal>...</internal>` tags to mark it as scratchpad — logged but not sent.
diff --git a/container/agent-runner/src/mcp-tools/interactive.instructions.md b/container/agent-runner/src/mcp-tools/interactive.instructions.md
new file mode 100644
index 0000000..f6601bd
--- /dev/null
+++ b/container/agent-runner/src/mcp-tools/interactive.instructions.md
@@ -0,0 +1,22 @@
+## Interactive prompts
+
+The two tools here solve different problems: `ask_user_question` forces a decision and waits for it; `send_card` displays structured content and moves on.
+
+### Asking a multiple-choice question (`ask_user_question`)
+
+`mcp__nanoclaw__ask_user_question({ title, question, options, timeout? })` presents the user with a set of choices and **blocks your turn** until they tap one or the timeout expires (default: 300 seconds). Returns their chosen value.
+
+`options` can be plain strings or `{ label, selectedLabel?, value? }` objects:
+- `label` — the button text shown before selection
+- `selectedLabel` — the text shown on the button *after* selection (useful for confirmations, e.g. `"✓ Confirmed"`)
+- `value` — the string returned to you when that option is chosen (defaults to `label`)
+
+Use this when you genuinely cannot proceed without a decision. For free-text input, send a normal message and wait for their reply — don't reach for this tool.
+
+### Structured cards (`send_card`)
+
+`mcp__nanoclaw__send_card({ card, fallbackText? })` renders a structured card and **returns immediately** — it does not pause your turn or collect a response.
+
+`card` supports: `title`, `description`, `children` (nested text or content blocks), and `actions` (buttons). `fallbackText` is sent as a plain message on platforms without card support.
+
+Use this for presenting information in a cleaner format than prose: summaries, options the user can read (but you're not waiting on), or results with contextual buttons. If you need the user to actually *choose* something and return a value, use `ask_user_question` instead.
\ No newline at end of file
diff --git a/container/agent-runner/src/mcp-tools/scheduling.instructions.md b/container/agent-runner/src/mcp-tools/scheduling.instructions.md
new file mode 100644
index 0000000..9b6b829
--- /dev/null
+++ b/container/agent-runner/src/mcp-tools/scheduling.instructions.md
@@ -0,0 +1,40 @@
+## Task scheduling (`schedule_task`)
+
+For any recurring task, use `schedule_task`. This is the scheduling path — tasks persist across sessions and restarts, and support the pre-task `script` hook described below.
+
+To inspect or change existing tasks, use `list_tasks` (returns one row per series with the stable id) and `update_task` / `cancel_task` / `pause_task` / `resume_task`. Prefer `update_task` over cancel + reschedule.
+
+Frequent recurring scheduled tasks — more than a few times a day — consume API credits and can risk account restrictions. You can add a `script` that runs first, and you will only be called when the check passes.
+
+### How it works
+
+1. Provide a bash `script` alongside the `prompt` when scheduling
+2. When the task fires, the script runs first
+3. Script returns: `{ "wakeAgent": true/false, "data": {...} }`
+4. If `wakeAgent: false` — nothing happens, task waits for next run
+5. If `wakeAgent: true` — claude receives the script's data + prompt and handles
+
+### Always test your script first
+
+Before scheduling, run the script directly to verify it works:
+
+```bash
+bash -c 'node --input-type=module -e "
+  const r = await fetch(\"https://api.github.com/repos/owner/repo/pulls?state=open\");
+  const prs = await r.json();
+  console.log(JSON.stringify({ wakeAgent: prs.length > 0, data: prs.slice(0, 5) }));
+"'
+```
+
+### When NOT to use scripts
+
+If a task requires your judgment every time (daily briefings, reminders, reports), skip the script — just use a regular prompt. Do not attempt to do things like sentiment analysis or advanced nlp in scripts.
+
+### Frequent task guidance
+
+If a user wants a task to run more than a few times a day and a script can't be used:
+
+- Explain that each time the task fires it uses API credits and risks rate limits
+- Suggest adjusting the task requirements in a way that will allow you to use a script
+- If the user needs an LLM to evaluate data, suggest using an API key with direct Anthropic API calls inside the script
+- Help the user find the minimum viable frequency
diff --git a/container/agent-runner/src/mcp-tools/scheduling.ts b/container/agent-runner/src/mcp-tools/scheduling.ts
index 168808c..00e41bb 100644
--- a/container/agent-runner/src/mcp-tools/scheduling.ts
+++ b/container/agent-runner/src/mcp-tools/scheduling.ts
@@ -8,6 +8,7 @@
 import { getInboundDb } from '../db/connection.js';
 import { writeMessageOut } from '../db/messages-out.js';
 import { getSessionRouting } from '../db/session-routing.js';
+import { TIMEZONE, parseZonedToUtc } from '../timezone.js';
 import { registerTools } from './server.js';
 import type { McpToolDefinition } from './types.js';
 
@@ -35,13 +36,21 @@ export const scheduleTask: McpToolDefinition = {
   tool: {
     name: 'schedule_task',
     description:
-      'Schedule a one-shot or recurring task. The task will be processed at the specified time. Use cron expressions for recurring tasks.',
+      `Schedule a one-shot or recurring task. The user's timezone is declared in the <context timezone="..."/> header of your prompt — interpret the user's "9pm" etc. in that zone. Cron expressions are interpreted in the user's timezone too.`,
     inputSchema: {
       type: 'object' as const,
       properties: {
         prompt: { type: 'string', description: 'Task instructions/prompt' },
-        processAfter: { type: 'string', description: 'ISO timestamp for first run (e.g., 2024-01-15T09:00:00Z)' },
-        recurrence: { type: 'string', description: 'Cron expression for recurring tasks (e.g., "0 9 * * 1-5" for weekdays at 9am)' },
+        processAfter: {
+          type: 'string',
+          description:
+            `ISO 8601 timestamp for the first run. Accepts either UTC (ending in "Z" or "+00:00") or a naive local timestamp (no offset) which is interpreted in the user's timezone (e.g. "2026-01-15T21:00:00" = 9pm user-local). Prefer naive local.`,
+        },
+        recurrence: {
+          type: 'string',
+          description:
+            'Cron expression for recurring tasks (e.g., "0 9 * * 1-5" = weekdays at 9am user-local). Evaluated in the user\'s timezone.',
+        },
         script: { type: 'string', description: 'Optional pre-agent script to run before processing' },
       },
       required: ['prompt', 'processAfter'],
@@ -49,8 +58,17 @@ export const scheduleTask: McpToolDefinition = {
   },
   async handler(args) {
     const prompt = args.prompt as string;
-    const processAfter = args.processAfter as string;
-    if (!prompt || !processAfter) return err('prompt and processAfter are required');
+    const processAfterIn = args.processAfter as string;
+    if (!prompt || !processAfterIn) return err('prompt and processAfter are required');
+
+    let processAfter: string;
+    try {
+      const d = parseZonedToUtc(processAfterIn, TIMEZONE);
+      if (Number.isNaN(d.getTime())) return err(`invalid processAfter: ${processAfterIn}`);
+      processAfter = d.toISOString();
+    } catch {
+      return err(`invalid processAfter: ${processAfterIn}`);
+    }
 
     const id = generateId();
     const r = routing();
@@ -233,7 +251,11 @@ export const updateTask: McpToolDefinition = {
           type: 'string',
           description: 'New cron expression (optional). Pass empty string to clear and make the task one-shot.',
         },
-        processAfter: { type: 'string', description: 'New ISO timestamp for the next run (optional)' },
+        processAfter: {
+          type: 'string',
+          description:
+            `New ISO 8601 timestamp for the next run (optional). Accepts either UTC (ending in "Z" / "+00:00") or a naive local timestamp interpreted in the user's timezone.`,
+        },
         script: {
           type: 'string',
           description: 'New pre-agent script (optional). Pass empty string to clear.',
@@ -248,7 +270,15 @@ export const updateTask: McpToolDefinition = {
 
     const update: Record<string, unknown> = { taskId };
     if (typeof args.prompt === 'string') update.prompt = args.prompt;
-    if (typeof args.processAfter === 'string') update.processAfter = args.processAfter;
+    if (typeof args.processAfter === 'string') {
+      try {
+        const d = parseZonedToUtc(args.processAfter, TIMEZONE);
+        if (Number.isNaN(d.getTime())) return err(`invalid processAfter: ${args.processAfter}`);
+        update.processAfter = d.toISOString();
+      } catch {
+        return err(`invalid processAfter: ${args.processAfter}`);
+      }
+    }
     // Empty string clears recurrence/script; undefined leaves them as-is.
     if (typeof args.recurrence === 'string') update.recurrence = args.recurrence === '' ? null : args.recurrence;
     if (typeof args.script === 'string') update.script = args.script === '' ? null : args.script;
diff --git a/container/agent-runner/src/mcp-tools/self-mod.instructions.md b/container/agent-runner/src/mcp-tools/self-mod.instructions.md
new file mode 100644
index 0000000..15057e0
--- /dev/null
+++ b/container/agent-runner/src/mcp-tools/self-mod.instructions.md
@@ -0,0 +1,25 @@
+## Installing packages & tools
+
+To install packages that persist, use the self-modification tools:
+
+**`install_packages`** — request system (apt) or global npm packages. Requires admin approval.
+
+Example flow:
+```
+install_packages({ apt: ["ffmpeg"], npm: ["@xenova/transformers"], reason: "Audio transcription" })
+# → Admin gets an approval card → approves
+```
+
+**When to use this vs workspace `pnpm install`:**
+- `pnpm install` if you only need it temporarily to do one task. Will not be available in subsequent truns.
+- `install_packages` persists for all future turns. Use especially if the user specifically asks you to add a capability
+
+### MCP servers (`add_mcp_server`)
+
+Use **`add_mcp_server`** to add an MCP server to your configuration. Browse available servers at https://mcp.so — it's a curated directory of high-quality MCP servers. Most Node.js servers run via `pnpm dlx`, e.g.:
+
+```
+add_mcp_server({ name: "memory", command: "pnpm", args: ["dlx", "@modelcontextprotocol/server-memory"] })
+```
+
+Do not ask the user to give you credentials. Credentials are managed by the user in the OneCLI agent vault. Add a "placeholder" string instead of the credential, and ask the user to add the credential to the vault. You can make a test request before the secret is added and the vault proxy will respond with the local url of the vault dashboard on the user's machine and a link to a form for adding that specific credential.
diff --git a/container/agent-runner/src/mcp-tools/self-mod.ts b/container/agent-runner/src/mcp-tools/self-mod.ts
index 775ec3b..3e2a2d8 100644
--- a/container/agent-runner/src/mcp-tools/self-mod.ts
+++ b/container/agent-runner/src/mcp-tools/self-mod.ts
@@ -1,9 +1,13 @@
 /**
- * Self-modification MCP tools: install_packages, add_mcp_server, request_rebuild.
+ * Self-modification MCP tools: install_packages, add_mcp_server.
  *
- * All three are fire-and-forget — the tool writes a system action row and
- * returns immediately. The host processes the request (including admin
- * approval) and notifies the agent via a chat message when complete.
+ * Both are fire-and-forget — the tool writes a system action row and returns
+ * immediately. The host processes the request (including admin approval)
+ * and notifies the agent via a chat message when complete. Admin approval
+ * is approval to apply the change: `install_packages` auto-rebuilds the
+ * per-agent image and restarts the container; `add_mcp_server` just
+ * updates `container.json` and restarts (bun runs TS directly — no build
+ * step needed for a pure MCP wiring change).
  *
  * Package names are sanitized here at the tool boundary AND re-validated on
  * the host side (defense in depth).
@@ -36,7 +40,7 @@ export const installPackages: McpToolDefinition = {
   tool: {
     name: 'install_packages',
     description:
-      'Install apt and/or npm packages into YOUR per-agent container image. Requires admin approval; fire-and-forget. After approval, call `request_rebuild` to apply.',
+      'Install apt and/or npm packages into YOUR per-agent container image. Requires admin approval; fire-and-forget. On approval, the image is rebuilt and the container is restarted automatically.',
     inputSchema: {
       type: 'object' as const,
       properties: {
@@ -113,32 +117,4 @@ export const addMcpServer: McpToolDefinition = {
   },
 };
 
-export const requestRebuild: McpToolDefinition = {
-  tool: {
-    name: 'request_rebuild',
-    description:
-      'Rebuild YOUR container image to pick up approved `install_packages` / `add_mcp_server` changes. Requires admin approval; fire-and-forget.',
-    inputSchema: {
-      type: 'object' as const,
-      properties: {
-        reason: { type: 'string', description: 'Why the rebuild is needed' },
-      },
-    },
-  },
-  async handler(args) {
-    const requestId = generateId();
-    writeMessageOut({
-      id: requestId,
-      kind: 'system',
-      content: JSON.stringify({
-        action: 'request_rebuild',
-        reason: (args.reason as string) || '',
-      }),
-    });
-
-    log(`request_rebuild: ${requestId}`);
-    return ok(`Rebuild request submitted. You will be notified when admin approves or rejects.`);
-  },
-};
-
-registerTools([installPackages, addMcpServer, requestRebuild]);
+registerTools([installPackages, addMcpServer]);
diff --git a/container/agent-runner/src/poll-loop.test.ts b/container/agent-runner/src/poll-loop.test.ts
index de5fb68..356108f 100644
--- a/container/agent-runner/src/poll-loop.test.ts
+++ b/container/agent-runner/src/poll-loop.test.ts
@@ -14,13 +14,13 @@ afterEach(() => {
   closeSessionDb();
 });
 
-function insertMessage(id: string, kind: string, content: object, opts?: { processAfter?: string }) {
+function insertMessage(id: string, kind: string, content: object, opts?: { processAfter?: string; trigger?: 0 | 1 }) {
   getInboundDb()
     .prepare(
-      `INSERT INTO messages_in (id, kind, timestamp, status, process_after, content)
-     VALUES (?, ?, datetime('now'), 'pending', ?, ?)`,
+      `INSERT INTO messages_in (id, kind, timestamp, status, process_after, trigger, content)
+     VALUES (?, ?, datetime('now'), 'pending', ?, ?, ?)`,
     )
-    .run(id, kind, opts?.processAfter ?? null, JSON.stringify(content));
+    .run(id, kind, opts?.processAfter ?? null, opts?.trigger ?? 1, JSON.stringify(content));
 }
 
 describe('formatter', () => {
@@ -84,6 +84,51 @@ describe('formatter', () => {
   });
 });
 
+describe('accumulate gate (trigger column)', () => {
+  it('getPendingMessages returns both trigger=0 and trigger=1 rows', () => {
+    // trigger=0 rides along as context, trigger=1 is the wake-eligible row.
+    // The poll loop's gate depends on this data contract.
+    insertMessage('m1', 'chat', { sender: 'A', text: 'chit chat' }, { trigger: 0 });
+    insertMessage('m2', 'chat', { sender: 'B', text: 'actual mention' }, { trigger: 1 });
+    const messages = getPendingMessages();
+    expect(messages).toHaveLength(2);
+    const byId = Object.fromEntries(messages.map((m) => [m.id, m]));
+    expect(byId.m1.trigger).toBe(0);
+    expect(byId.m2.trigger).toBe(1);
+  });
+
+  it('trigger=0-only batch: gate predicate `some(trigger===1)` is false', () => {
+    insertMessage('m1', 'chat', { sender: 'A', text: 'noise' }, { trigger: 0 });
+    insertMessage('m2', 'chat', { sender: 'B', text: 'more noise' }, { trigger: 0 });
+    const messages = getPendingMessages();
+    // This is the exact predicate the poll loop uses to skip accumulate-only
+    // batches — gate should be false, so the loop sleeps without waking the agent.
+    expect(messages.some((m) => m.trigger === 1)).toBe(false);
+  });
+
+  it('mixed batch: gate is true → loop proceeds, accumulated rows ride along', () => {
+    insertMessage('m1', 'chat', { sender: 'A', text: 'earlier chatter' }, { trigger: 0 });
+    insertMessage('m2', 'chat', { sender: 'B', text: 'the real mention' }, { trigger: 1 });
+    const messages = getPendingMessages();
+    expect(messages.some((m) => m.trigger === 1)).toBe(true);
+    // Both messages are present for the formatter → agent sees the prior context.
+    expect(messages.map((m) => m.id).sort()).toEqual(['m1', 'm2']);
+  });
+
+  it('trigger column defaults to 1 for legacy inserts without explicit value', () => {
+    // The schema default is 1 (see src/db/schema.ts INBOUND_SCHEMA) — existing
+    // rows / tests without the column set are effectively wake-eligible.
+    getInboundDb()
+      .prepare(
+        `INSERT INTO messages_in (id, kind, timestamp, status, content)
+         VALUES ('m1', 'chat', datetime('now'), 'pending', '{"text":"hi"}')`,
+      )
+      .run();
+    const [msg] = getPendingMessages();
+    expect(msg.trigger).toBe(1);
+  });
+});
+
 describe('routing', () => {
   it('should extract routing from messages', () => {
     getInboundDb()
diff --git a/container/agent-runner/src/poll-loop.ts b/container/agent-runner/src/poll-loop.ts
index cc26286..bd48db2 100644
--- a/container/agent-runner/src/poll-loop.ts
+++ b/container/agent-runner/src/poll-loop.ts
@@ -2,13 +2,16 @@ import { findByName, getAllDestinations, type DestinationEntry } from './destina
 import { getPendingMessages, markProcessing, markCompleted, type MessageInRow } from './db/messages-in.js';
 import { writeMessageOut } from './db/messages-out.js';
 import { touchHeartbeat, clearStaleProcessingAcks } from './db/connection.js';
-import { getStoredSessionId, setStoredSessionId, clearStoredSessionId } from './db/session-state.js';
-import { formatMessages, extractRouting, categorizeMessage, type RoutingContext } from './formatter.js';
+import {
+  clearContinuation,
+  migrateLegacyContinuation,
+  setContinuation,
+} from './db/session-state.js';
+import { formatMessages, extractRouting, categorizeMessage, isClearCommand, stripInternalTags, type RoutingContext } from './formatter.js';
 import type { AgentProvider, AgentQuery, ProviderEvent } from './providers/types.js';
 
 const POLL_INTERVAL_MS = 1000;
 const ACTIVE_POLL_INTERVAL_MS = 500;
-const IDLE_END_MS = 20_000; // End stream after 20s with no SDK events
 
 function log(msg: string): void {
   console.error(`[poll-loop] ${msg}`);
@@ -20,16 +23,16 @@ function generateId(): string {
 
 export interface PollLoopConfig {
   provider: AgentProvider;
+  /**
+   * Name of the provider (e.g. "claude", "codex", "opencode"). Used to key
+   * the stored continuation per-provider so flipping providers doesn't
+   * resurrect a stale id from a different backend.
+   */
+  providerName: string;
   cwd: string;
   systemContext?: {
     instructions?: string;
   };
-  /**
-   * Set of user IDs allowed to run admin commands (e.g. /clear) in this
-   * agent group. Host populates from owners + global admins + scoped admins
-   * at container wake time, so role changes take effect on next spawn.
-   */
-  adminUserIds?: Set<string>;
 }
 
 /**
@@ -46,8 +49,9 @@ export async function runPollLoop(config: PollLoopConfig): Promise<void> {
   // Resume the agent's prior session from a previous container run if one
   // was persisted. The continuation is opaque to the poll-loop — the
   // provider decides how to use it (Claude resumes a .jsonl transcript,
-  // other providers may reload a thread ID, etc.).
-  let continuation: string | undefined = getStoredSessionId();
+  // other providers may reload a thread ID, etc.). Keyed per-provider so
+  // a Codex thread id never gets handed to Claude or vice versa.
+  let continuation: string | undefined = migrateLegacyContinuation(config.providerName);
 
   if (continuation) {
     log(`Resuming agent session ${continuation}`);
@@ -73,79 +77,54 @@ export async function runPollLoop(config: PollLoopConfig): Promise<void> {
       continue;
     }
 
+    // Accumulate gate: if the batch contains only trigger=0 rows
+    // (context-only, router-stored under ignored_message_policy='accumulate'),
+    // don't wake the agent. Leave them `pending` — they'll ride along the
+    // next time a real trigger=1 message lands via this same getPendingMessages
+    // query. Without this gate, a warm container keeps processing
+    // (and potentially responding to) every accumulate-only batch, defeating
+    // the "store as context, don't engage" contract. Host-side countDueMessages
+    // gates the same way for wake-from-cold (see src/db/session-db.ts).
+    if (!messages.some((m) => m.trigger === 1)) {
+      await sleep(POLL_INTERVAL_MS);
+      continue;
+    }
+
     const ids = messages.map((m) => m.id);
     markProcessing(ids);
 
     const routing = extractRouting(messages);
 
-    // Handle commands: categorize chat messages
-    const adminUserIds = config.adminUserIds ?? new Set<string>();
-    const normalMessages = [];
+    // Command handling: the host router gates filtered and unauthorized
+    // admin commands before they reach the container. The only command
+    // the runner handles directly is /clear (session reset).
+    const normalMessages: MessageInRow[] = [];
     const commandIds: string[] = [];
 
     for (const msg of messages) {
-      if (msg.kind !== 'chat' && msg.kind !== 'chat-sdk') {
-        normalMessages.push(msg);
-        continue;
-      }
-
-      const cmdInfo = categorizeMessage(msg);
-
-      if (cmdInfo.category === 'filtered') {
-        // Silently drop — mark completed, don't process
-        log(`Filtered command: ${cmdInfo.command} (msg: ${msg.id})`);
+      if ((msg.kind === 'chat' || msg.kind === 'chat-sdk') && isClearCommand(msg)) {
+        log('Clearing session (resetting continuation)');
+        continuation = undefined;
+        clearContinuation(config.providerName);
+        writeMessageOut({
+          id: generateId(),
+          kind: 'chat',
+          platform_id: routing.platformId,
+          channel_type: routing.channelType,
+          thread_id: routing.threadId,
+          content: JSON.stringify({ text: 'Session cleared.' }),
+        });
         commandIds.push(msg.id);
         continue;
       }
-
-      if (cmdInfo.category === 'admin') {
-        if (!cmdInfo.senderId || !adminUserIds.has(cmdInfo.senderId)) {
-          log(`Admin command denied: ${cmdInfo.command} from ${cmdInfo.senderId} (msg: ${msg.id})`);
-          writeMessageOut({
-            id: generateId(),
-            kind: 'chat',
-            platform_id: routing.platformId,
-            channel_type: routing.channelType,
-            thread_id: routing.threadId,
-            content: JSON.stringify({ text: `Permission denied: ${cmdInfo.command} requires admin access.` }),
-          });
-          commandIds.push(msg.id);
-          continue;
-        }
-        // Handle admin commands directly
-        if (cmdInfo.command === '/clear') {
-          log('Clearing session (resetting continuation)');
-          continuation = undefined;
-          clearStoredSessionId();
-          writeMessageOut({
-            id: generateId(),
-            kind: 'chat',
-            platform_id: routing.platformId,
-            channel_type: routing.channelType,
-            thread_id: routing.threadId,
-            content: JSON.stringify({ text: 'Session cleared.' }),
-          });
-          commandIds.push(msg.id);
-          continue;
-        }
-
-        // Other admin commands — pass through to agent
-        normalMessages.push(msg);
-        continue;
-      }
-
-      // passthrough or none
       normalMessages.push(msg);
     }
 
-    // Mark filtered/denied command messages as completed immediately
     if (commandIds.length > 0) {
       markCompleted(commandIds);
     }
 
-    // If all messages were filtered commands, skip processing
     if (normalMessages.length === 0) {
-      // Mark remaining processing IDs as completed
       const remainingIds = ids.filter((id) => !commandIds.includes(id));
       if (remainingIds.length > 0) markCompleted(remainingIds);
       log(`All ${messages.length} message(s) were commands, skipping query`);
@@ -192,10 +171,10 @@ export async function runPollLoop(config: PollLoopConfig): Promise<void> {
     const skippedSet = new Set(skipped);
     const processingIds = ids.filter((id) => !commandIds.includes(id) && !skippedSet.has(id));
     try {
-      const result = await processQuery(query, routing);
+      const result = await processQuery(query, routing, processingIds, config.providerName);
       if (result.continuation && result.continuation !== continuation) {
         continuation = result.continuation;
-        setStoredSessionId(continuation);
+        setContinuation(config.providerName, continuation);
       }
     } catch (err) {
       const errMsg = err instanceof Error ? err.message : String(err);
@@ -207,7 +186,7 @@ export async function runPollLoop(config: PollLoopConfig): Promise<void> {
       if (continuation && config.provider.isSessionInvalid(err)) {
         log(`Stale session detected (${continuation}) — clearing for next retry`);
         continuation = undefined;
-        clearStoredSessionId();
+        clearContinuation(config.providerName);
       }
 
       // Write error response so the user knows something went wrong
@@ -221,6 +200,8 @@ export async function runPollLoop(config: PollLoopConfig): Promise<void> {
       });
     }
 
+    // Ensure completed even if processQuery ended without a result event
+    // (e.g. stream closed unexpectedly).
     markCompleted(processingIds);
     log(`Completed ${ids.length} message(s)`);
   }
@@ -264,27 +245,34 @@ interface QueryResult {
   continuation?: string;
 }
 
-async function processQuery(query: AgentQuery, routing: RoutingContext): Promise<QueryResult> {
+async function processQuery(
+  query: AgentQuery,
+  routing: RoutingContext,
+  initialBatchIds: string[],
+  providerName: string,
+): Promise<QueryResult> {
   let queryContinuation: string | undefined;
   let done = false;
-  let lastEventTime = Date.now();
 
-  // Concurrent polling: push follow-ups, checkpoint WAL, detect idle
+  // Concurrent polling: push follow-ups into the active query as they arrive.
+  // We do NOT force-end the stream on silence — keeping the query open is
+  // strictly cheaper than close+reopen (no cold prompt cache, no reconnect).
+  // Stream liveness is decided host-side via the heartbeat file + processing
+  // claim age (see src/host-sweep.ts); if something is truly stuck, the host
+  // will kill the container and messages get reset to pending.
   const pollHandle = setInterval(() => {
     if (done) return;
 
-    // Skip system messages (MCP tool responses) and admin commands (need fresh query).
-    // Also defer messages whose thread_id differs from the active turn's routing
-    // — mixing threads into one streaming turn would send the reply to the wrong
-    // thread because `routing` is captured at turn start. The next turn will pick
-    // them up with fresh routing.
+    // Skip system messages (MCP tool responses) and /clear (needs fresh query).
+    // Thread routing is the router's concern — if a message landed in this
+    // session, the agent should see it. Per-thread sessions already isolate
+    // threads into separate containers; shared sessions intentionally merge
+    // everything. Filtering on thread_id here caused deadlocks when the
+    // initial batch and follow-ups had mismatched thread_ids (e.g. a
+    // host-generated welcome trigger with null thread vs a Discord DM reply).
     const newMessages = getPendingMessages().filter((m) => {
       if (m.kind === 'system') return false;
-      if (m.kind === 'chat' || m.kind === 'chat-sdk') {
-        const cmd = categorizeMessage(m);
-        if (cmd.category === 'admin') return false;
-      }
-      if ((m.thread_id ?? null) !== (routing.threadId ?? null)) return false;
+      if ((m.kind === 'chat' || m.kind === 'chat-sdk') && isClearCommand(m)) return false;
       return true;
     });
     if (newMessages.length > 0) {
@@ -296,26 +284,34 @@ async function processQuery(query: AgentQuery, routing: RoutingContext): Promise
       query.push(prompt);
 
       markCompleted(newIds);
-      lastEventTime = Date.now(); // new input counts as activity
-    }
-
-    // End stream when agent is idle: no SDK events and no pending messages
-    if (Date.now() - lastEventTime > IDLE_END_MS) {
-      log(`No SDK events for ${IDLE_END_MS / 1000}s, ending query`);
-      query.end();
     }
   }, ACTIVE_POLL_INTERVAL_MS);
 
   try {
     for await (const event of query.events) {
-      lastEventTime = Date.now();
       handleEvent(event, routing);
       touchHeartbeat();
 
       if (event.type === 'init') {
         queryContinuation = event.continuation;
-      } else if (event.type === 'result' && event.text) {
-        dispatchResultText(event.text, routing);
+        // Persist immediately so a mid-turn container crash still lets the
+        // next wake resume the conversation. Without this, the session id
+        // was only written after the full stream completed — if the
+        // container died between `init` and `result`, the SDK session was
+        // effectively orphaned and the next message started a blank
+        // Claude session with no prior context.
+        setContinuation(providerName, event.continuation);
+      } else if (event.type === 'result') {
+        // A result — with or without text — means the turn is done. Mark
+        // the initial batch completed now so the host sweep doesn't see
+        // stale 'processing' claims while the query stays open for
+        // follow-up pushes. The agent may have responded via MCP
+        // (send_message) mid-turn, or the message may not need a response
+        // at all — either way the turn is finished.
+        markCompleted(initialBatchIds);
+        if (event.text) {
+          dispatchResultText(event.text, routing);
+        }
       }
     }
   } finally {
@@ -384,10 +380,7 @@ function dispatchResultText(text: string, routing: RoutingContext): void {
     scratchpadParts.push(text.slice(lastIndex));
   }
 
-  const scratchpad = scratchpadParts
-    .join('')
-    .replace(/<internal>[\s\S]*?<\/internal>/g, '')
-    .trim();
+  const scratchpad = stripInternalTags(scratchpadParts.join(''));
 
   // Single-destination shortcut: the agent wrote plain text — send to
   // the session's originating channel (from session_routing) if available,
diff --git a/container/agent-runner/src/providers/claude.ts b/container/agent-runner/src/providers/claude.ts
index 97fe44a..fbb077c 100644
--- a/container/agent-runner/src/providers/claude.ts
+++ b/container/agent-runner/src/providers/claude.ts
@@ -3,6 +3,7 @@ import path from 'path';
 
 import { query as sdkQuery, type HookCallback, type PreCompactHookInput } from '@anthropic-ai/claude-agent-sdk';
 
+import { clearContainerToolInFlight, setContainerToolInFlight } from '../db/connection.js';
 import { registerProvider } from './provider-registry.js';
 import type { AgentProvider, AgentQuery, McpServerConfig, ProviderEvent, ProviderOptions, QueryInput } from './types.js';
 
@@ -10,10 +11,28 @@ function log(msg: string): void {
   console.error(`[claude-provider] ${msg}`);
 }
 
-// Deferred SDK builtins that would sidestep nanoclaw's own scheduling.
-// Scheduling goes through mcp__nanoclaw__schedule_task so that tasks are
-// durable across sessions/restarts and gated by our pre-task script hook.
-const SDK_DISALLOWED_TOOLS = ['CronCreate', 'CronDelete', 'CronList', 'ScheduleWakeup'];
+// Deferred SDK builtins that either sidestep nanoclaw's own scheduling or
+// don't fit our async message-passing model (they're designed for Claude
+// Code's interactive UI and would hang here).
+//
+// - CronCreate / CronDelete / CronList / ScheduleWakeup: we have durable
+//   scheduling via mcp__nanoclaw__schedule_task.
+// - AskUserQuestion: SDK returns a placeholder instead of blocking on a
+//   real answer — we have mcp__nanoclaw__ask_user_question that persists
+//   the question and blocks on the real reply.
+// - EnterPlanMode / ExitPlanMode / EnterWorktree / ExitWorktree: Claude
+//   Code UI affordances; in a headless container they'd appear stuck.
+const SDK_DISALLOWED_TOOLS = [
+  'CronCreate',
+  'CronDelete',
+  'CronList',
+  'ScheduleWakeup',
+  'AskUserQuestion',
+  'EnterPlanMode',
+  'ExitPlanMode',
+  'EnterWorktree',
+  'ExitWorktree',
+];
 
 // Tool allowlist for NanoClaw agent containers
 const TOOL_ALLOWLIST = [
@@ -122,6 +141,43 @@ function formatTranscriptMarkdown(messages: ParsedMessage[], title?: string | nu
   return lines.join('\n');
 }
 
+/**
+ * PreToolUse hook: record the current tool + its declared timeout so the host
+ * sweep can widen its stuck tolerance while Bash is running a long-declared
+ * script. Defense-in-depth: if SDK_DISALLOWED_TOOLS slips through somehow,
+ * block the call here instead of letting the agent hang.
+ */
+const preToolUseHook: HookCallback = async (input) => {
+  const i = input as { tool_name?: string; tool_input?: Record<string, unknown> };
+  const toolName = i.tool_name ?? '';
+  if (SDK_DISALLOWED_TOOLS.includes(toolName)) {
+    return {
+      decision: 'block',
+      stopReason: `Tool '${toolName}' is not available in this environment — use the nanoclaw equivalent.`,
+    } as unknown as ReturnType<HookCallback>;
+  }
+  // Bash exposes its timeout via the tool_input.timeout field (ms). Any other
+  // tool: no declared timeout.
+  const declaredTimeoutMs =
+    toolName === 'Bash' && typeof i.tool_input?.timeout === 'number' ? (i.tool_input.timeout as number) : null;
+  try {
+    setContainerToolInFlight(toolName, declaredTimeoutMs);
+  } catch (err) {
+    log(`PreToolUse: failed to record container_state: ${err instanceof Error ? err.message : String(err)}`);
+  }
+  return { continue: true };
+};
+
+/** Clear in-flight tool on PostToolUse / PostToolUseFailure. */
+const postToolUseHook: HookCallback = async () => {
+  try {
+    clearContainerToolInFlight();
+  } catch (err) {
+    log(`PostToolUse: failed to clear container_state: ${err instanceof Error ? err.message : String(err)}`);
+  }
+  return { continue: true };
+};
+
 function createPreCompactHook(assistantName?: string): HookCallback {
   return async (input) => {
     const preCompact = input as PreCompactHookInput;
@@ -215,6 +271,7 @@ export class ClaudeProvider implements AgentProvider {
         cwd: input.cwd,
         additionalDirectories: this.additionalDirectories,
         resume: input.continuation,
+        pathToClaudeCodeExecutable: '/pnpm/claude',
         systemPrompt: instructions ? { type: 'preset' as const, preset: 'claude_code' as const, append: instructions } : undefined,
         allowedTools: TOOL_ALLOWLIST,
         disallowedTools: SDK_DISALLOWED_TOOLS,
@@ -224,6 +281,9 @@ export class ClaudeProvider implements AgentProvider {
         settingSources: ['project', 'user'],
         mcpServers: this.mcpServers,
         hooks: {
+          PreToolUse: [{ hooks: [preToolUseHook] }],
+          PostToolUse: [{ hooks: [postToolUseHook] }],
+          PostToolUseFailure: [{ hooks: [postToolUseHook] }],
           PreCompact: [{ hooks: [createPreCompactHook(this.assistantName)] }],
         },
       },
diff --git a/container/agent-runner/src/timezone.test.ts b/container/agent-runner/src/timezone.test.ts
new file mode 100644
index 0000000..a4539e9
--- /dev/null
+++ b/container/agent-runner/src/timezone.test.ts
@@ -0,0 +1,93 @@
+import { describe, it, expect } from 'bun:test';
+
+import { formatLocalTime, isValidTimezone, parseZonedToUtc, resolveTimezone } from './timezone.js';
+
+// --- formatLocalTime ---
+
+describe('formatLocalTime', () => {
+  it('converts UTC to local time display', () => {
+    // 2026-02-04T18:30:00Z in America/New_York (EST, UTC-5) = 1:30 PM
+    const result = formatLocalTime('2026-02-04T18:30:00.000Z', 'America/New_York');
+    expect(result).toContain('1:30');
+    expect(result).toContain('PM');
+    expect(result).toContain('Feb');
+    expect(result).toContain('2026');
+  });
+
+  it('handles different timezones', () => {
+    // Same UTC time should produce different local times
+    const utc = '2026-06-15T12:00:00.000Z';
+    const ny = formatLocalTime(utc, 'America/New_York');
+    const tokyo = formatLocalTime(utc, 'Asia/Tokyo');
+    // NY is UTC-4 in summer (EDT), Tokyo is UTC+9
+    expect(ny).toContain('8:00');
+    expect(tokyo).toContain('9:00');
+  });
+
+  it('does not throw on invalid timezone, falls back to UTC', () => {
+    expect(() => formatLocalTime('2026-01-01T00:00:00.000Z', 'IST-2')).not.toThrow();
+    const result = formatLocalTime('2026-01-01T12:00:00.000Z', 'IST-2');
+    // Should format as UTC (noon UTC = 12:00 PM)
+    expect(result).toContain('12:00');
+    expect(result).toContain('PM');
+  });
+});
+
+describe('isValidTimezone', () => {
+  it('accepts valid IANA identifiers', () => {
+    expect(isValidTimezone('America/New_York')).toBe(true);
+    expect(isValidTimezone('UTC')).toBe(true);
+    expect(isValidTimezone('Asia/Tokyo')).toBe(true);
+    expect(isValidTimezone('Asia/Jerusalem')).toBe(true);
+  });
+
+  it('rejects invalid timezone strings', () => {
+    expect(isValidTimezone('IST-2')).toBe(false);
+    expect(isValidTimezone('XYZ+3')).toBe(false);
+  });
+
+  it('rejects empty and garbage strings', () => {
+    expect(isValidTimezone('')).toBe(false);
+    expect(isValidTimezone('NotATimezone')).toBe(false);
+  });
+});
+
+describe('resolveTimezone', () => {
+  it('returns the timezone if valid', () => {
+    expect(resolveTimezone('America/New_York')).toBe('America/New_York');
+  });
+
+  it('falls back to UTC for invalid timezone', () => {
+    expect(resolveTimezone('IST-2')).toBe('UTC');
+    expect(resolveTimezone('')).toBe('UTC');
+  });
+});
+
+describe('parseZonedToUtc', () => {
+  it('passes strings with Z suffix through unchanged', () => {
+    const d = parseZonedToUtc('2026-01-15T09:00:00Z', 'America/New_York');
+    expect(d.toISOString()).toBe('2026-01-15T09:00:00.000Z');
+  });
+
+  it('passes strings with numeric offset through unchanged', () => {
+    const d = parseZonedToUtc('2026-01-15T09:00:00+02:00', 'America/New_York');
+    expect(d.toISOString()).toBe('2026-01-15T07:00:00.000Z');
+  });
+
+  it('interprets naive ISO as wall-clock in the given timezone', () => {
+    // 09:00 naive in NY in January = 09:00 EST = 14:00 UTC
+    const d = parseZonedToUtc('2026-01-15T09:00:00', 'America/New_York');
+    expect(d.toISOString()).toBe('2026-01-15T14:00:00.000Z');
+  });
+
+  it('handles a different positive-offset zone', () => {
+    // 09:00 naive in Tokyo (UTC+9) = 00:00 UTC
+    const d = parseZonedToUtc('2026-06-15T09:00:00', 'Asia/Tokyo');
+    expect(d.toISOString()).toBe('2026-06-15T00:00:00.000Z');
+  });
+
+  it('treats invalid timezone as UTC', () => {
+    const d = parseZonedToUtc('2026-01-15T09:00:00', 'NotATimezone');
+    expect(d.toISOString()).toBe('2026-01-15T09:00:00.000Z');
+  });
+});
diff --git a/container/agent-runner/src/timezone.ts b/container/agent-runner/src/timezone.ts
new file mode 100644
index 0000000..d9a2e1b
--- /dev/null
+++ b/container/agent-runner/src/timezone.ts
@@ -0,0 +1,107 @@
+/**
+ * Timezone utilities — mirror of src/timezone.ts (host).
+ *
+ * The container can't import from src/ (separate tsconfig, different runtime).
+ * Kept deliberately byte-aligned with the host module so behaviour is the
+ * same on both sides of the session-DB boundary.
+ *
+ * TIMEZONE is resolved once at module load from process.env.TZ (which the host
+ * sets from its own TIMEZONE constant when spawning the container; see
+ * src/container-runner.ts). Invalid values fall back to UTC.
+ */
+
+/**
+ * Check whether a timezone string is a valid IANA identifier
+ * that Intl.DateTimeFormat can use.
+ */
+export function isValidTimezone(tz: string): boolean {
+  try {
+    Intl.DateTimeFormat(undefined, { timeZone: tz });
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Return the given timezone if valid IANA, otherwise fall back to UTC.
+ */
+export function resolveTimezone(tz: string): string {
+  return isValidTimezone(tz) ? tz : 'UTC';
+}
+
+/**
+ * Convert a UTC ISO timestamp to a localized display string.
+ * Uses the Intl API (no external dependencies).
+ * Falls back to UTC if the timezone is invalid.
+ */
+export function formatLocalTime(utcIso: string, timezone: string): string {
+  const date = new Date(utcIso);
+  return date.toLocaleString('en-US', {
+    timeZone: resolveTimezone(timezone),
+    year: 'numeric',
+    month: 'short',
+    day: 'numeric',
+    hour: 'numeric',
+    minute: '2-digit',
+    hour12: true,
+  });
+}
+
+function resolveContainerTimezone(): string {
+  const candidates = [process.env.TZ, Intl.DateTimeFormat().resolvedOptions().timeZone];
+  for (const tz of candidates) {
+    if (tz && isValidTimezone(tz)) return tz;
+  }
+  return 'UTC';
+}
+
+export const TIMEZONE = resolveContainerTimezone();
+
+/**
+ * Interpret a naive ISO-like timestamp (no trailing `Z`, no offset) as wall-clock
+ * time in `tz` and return the corresponding UTC Date. Strings that already carry
+ * offset info (`Z` or `±HH:MM`) are passed through to the Date constructor
+ * unchanged.
+ *
+ * Algorithm: treat the naive string as UTC, ask Intl.DateTimeFormat what that
+ * UTC instant is called in `tz`, then invert the offset. Near DST boundaries
+ * this can be off by an hour for ~1h of wall-clock time per year; acceptable
+ * for scheduling where the agent normally picks round-hour targets.
+ */
+export function parseZonedToUtc(input: string, tz: string): Date {
+  const hasOffset = /Z$|[+-]\d{2}:?\d{2}$/.test(input.trim());
+  if (hasOffset) return new Date(input);
+
+  const zone = resolveTimezone(tz);
+  const asIfUtc = new Date(input + 'Z');
+  if (Number.isNaN(asIfUtc.getTime())) return asIfUtc;
+
+  const fmt = new Intl.DateTimeFormat('en-US', {
+    timeZone: zone,
+    year: 'numeric',
+    month: '2-digit',
+    day: '2-digit',
+    hour: '2-digit',
+    minute: '2-digit',
+    second: '2-digit',
+    hour12: false,
+  });
+  const parts = Object.fromEntries(
+    fmt
+      .formatToParts(asIfUtc)
+      .filter((p) => p.type !== 'literal')
+      .map((p) => [p.type, p.value]),
+  );
+  const hour = parts.hour === '24' ? '00' : parts.hour;
+  const zonedAsUtcMs = Date.UTC(
+    Number(parts.year),
+    Number(parts.month) - 1,
+    Number(parts.day),
+    Number(hour),
+    Number(parts.minute),
+    Number(parts.second),
+  );
+  const offsetMs = zonedAsUtcMs - asIfUtc.getTime();
+  return new Date(asIfUtc.getTime() - offsetMs);
+}
diff --git a/container/build.sh b/container/build.sh
index fd5210d..ae0c3d9 100755
--- a/container/build.sh
+++ b/container/build.sh
@@ -9,9 +9,15 @@
 set -e
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
 cd "$SCRIPT_DIR"
 
-IMAGE_NAME="nanoclaw-agent"
+# Derive the image name from the project root so two NanoClaw installs on the
+# same host don't overwrite each other's `nanoclaw-agent:latest` tag. Matches
+# setup/lib/install-slug.sh + src/install-slug.ts.
+# shellcheck source=../setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+IMAGE_NAME="$(container_image_base)"
 TAG="${1:-latest}"
 CONTAINER_RUNTIME="${CONTAINER_RUNTIME:-docker}"
 
diff --git a/container/skills/self-customize/SKILL.md b/container/skills/self-customize/SKILL.md
index e1d5588..c8bad16 100644
--- a/container/skills/self-customize/SKILL.md
+++ b/container/skills/self-customize/SKILL.md
@@ -11,9 +11,9 @@ You can modify your own environment. Different kinds of changes have different w
 
 **What needs to change?**
 
-- **Your CLAUDE.md or files in your workspace** → Edit directly, no approval needed. Your workspace (`/workspace/agent/`) is persisted on the host.
-- **System package (apt) or global npm package** → `install_packages` → `request_rebuild`. Requires admin approval.
-- **MCP server** → `add_mcp_server` → `request_rebuild`. No approval needed, but rebuild required to apply.
+- **`CLAUDE.local.md` or files in your workspace** → Edit directly, no approval needed. Your workspace (`/workspace/agent/`) is persisted on the host. (Note: the composed `CLAUDE.md` itself is read-only and regenerated every spawn — write to `CLAUDE.local.md` instead.)
+- **System package (apt) or global npm package** → `install_packages`. Requires admin approval. On approval, image rebuild + container restart happen automatically.
+- **MCP server** → `add_mcp_server`. Requires admin approval. On approval, container restarts with the new server wired up (no rebuild — bun runs TS directly).
 - **Your source code or Dockerfile** → Delegate to a builder agent via `create_agent` (see below).
 - **A new specialist capability** → `create_agent` to spin up a dedicated agent for it.
 
@@ -25,7 +25,7 @@ For anything that requires editing source files (your own code, Dockerfile, etc.
 2. Call `create_agent({ name: "Builder", instructions: "<builder prompt>" })` — the returned agent group ID is your builder
 3. Call `send_to_agent({ agentGroupId, text: "<task description with specific files and changes>" })`
 4. The builder works in its own container, makes the changes, and reports back
-5. You review the builder's summary, confirm with the user, then call `request_rebuild` if the changes require it
+5. You review the builder's summary and confirm with the user. Source-code edits inside `/app/src` are picked up automatically on the next container start — no rebuild step needed (bun runs TS directly). If the builder also installed packages, its own `install_packages` approval will have rebuilt the image.
 
 ### Builder Agent Instructions (use as CLAUDE.md when creating)
 
@@ -64,12 +64,11 @@ The limits are **per builder task**, not per session. A 500-line feature is fine
 User: "Can you add a tool for reading RSS feeds?"
 
 1. Check [mcp.so](https://mcp.so) for an existing RSS MCP server
-2. If one exists → `add_mcp_server({ name: "rss", command: "npx", args: ["some-rss-mcp"] })` → `request_rebuild` → done
+2. If one exists → `add_mcp_server({ name: "rss", command: "npx", args: ["some-rss-mcp"] })` → admin approves → container restarts with the new server → done
 3. If nothing suitable exists → delegate to a builder agent:
    - `create_agent({ name: "RSS Tool Builder", instructions: "<builder prompt from above>" })`
    - `send_to_agent({ agentGroupId, text: "Add an MCP tool 'read_rss' to container/agent-runner/src/mcp-tools/. It should fetch an RSS URL and return the latest N items. Register it in mcp-tools/index.ts. Target: <200 new lines." })`
-   - Wait for builder's report
-   - `request_rebuild` if needed
+   - Wait for builder's report — new tool code is picked up on the next container start (bun runs TS directly)
 
 ## Example: Installing a System Tool
 
@@ -78,10 +77,8 @@ User: "Can you transcribe audio?"
 1. Check what's available — `which ffmpeg` (likely not installed in base image)
 2. Decide approach: `@xenova/transformers` (npm, workspace-local) or `whisper.cpp` (apt + compile)
 3. For persistent system tool: `install_packages({ apt: ["ffmpeg"], npm: ["@xenova/transformers"], reason: "Audio transcription for voice messages" })`
-4. Wait for admin approval
-5. `request_rebuild({ reason: "Apply audio transcription packages" })`
-6. Wait for admin approval
-7. Test the new capability once the container restarts
+4. Wait for admin approval — on approve, the image is rebuilt and your container is restarted automatically
+5. Test the new capability once the container restarts
 
 ## When NOT to Self-Customize
 
diff --git a/container/skills/welcome/SKILL.md b/container/skills/welcome/SKILL.md
index fe3245b..c540e86 100644
--- a/container/skills/welcome/SKILL.md
+++ b/container/skills/welcome/SKILL.md
@@ -3,26 +3,83 @@ name: welcome
 description: Introduce yourself to a newly connected channel. Triggered automatically when a channel is first wired. Send a friendly greeting and brief overview of what you can do.
 ---
 
-# /welcome — Channel Onboarding
+# /welcome — Channel Onboarding (Updated)
 
-You've just been connected to a new messaging channel. Introduce yourself to the user.
+You've just been connected to a new user. This your time to shine and make a strong first impression. Introduce yourself and guide the user through what you can do. you got this!
 
 ## What to do
 
 1. Send a short, warm greeting using `send_message`
-2. Mention your name (from your CLAUDE.md)
-3. Make it clear you can do a lot — but do NOT list your tools or skills upfront. Keep it open-ended and intriguing
-4. End by asking: would they like to explore what you can do, or jump straight into building/creating something?
+2. State your name (from your system prompt / CLAUDE.md)
+3. Signal that you're capable of a lot — but don't list everything upfront. Be intriguing, not encyclopedic
+4. Ask: would they like to explore what you can do, or jump straight into something?
 
-**If they want to explore:** show one skill or capability at a time. Briefly explain what it does, offer to demo it or let them try it, then ask if they want to see the next one or move on. Drip-feed — never dump a list.
+**If they want to explore:** drip-feed one capability at a time. Briefly explain it, offer to demo a compelling example or let them try it. Never dump a full list.
 
-**If they want to jump in:** just go. Help them with whatever they ask.
+**If they want to jump in:** just go.
+
+---
+
+## Capabilities to reveal (in order)
+
+Reveal these one at a time, in this sequence. Each should be 2–4 sentences max.
+
+### 1. Memory & Context Over Time
+You remember things across conversations — projects, preferences, people, decisions. Users don't have to re-explain context every session. The more they work with you, the more situationally aware you become.
+
+### 2. Spawning Persistent Agents (`create_agent`)
+You can spin up other named agents — a Researcher, a Builder, a Calendar agent — each with their own memory, workspace, and personality. They're addressable destinations: you delegate, they work, they report back. These aren't one-shot tasks; they accumulate context across sessions.
+
+### 3. Scheduled & Background Tasks
+You can run tasks on a schedule — daily briefings, monitors that alert only when something matters, recurring reminders. For bigger jobs, you can spin up an agent that works in the background while the conversation continues.
+
+### 4. Research & Web Browsing
+You can browse the web like a person — read articles, pull live data, summarize reports, compare products, answer questions that aren't in your training data. Ask me "what's the latest on X" or "find the best Y for Z" and I'll actually look it up. Very powerful when combined with scheduled tasks.
+
+### 5. Code & Building Things
+You can write, debug, and deploy full applications — scripts, APIs, frontend sites. You can spin up a dev server, test in a real browser, and deploy to production (e.g. Vercel). Concept to live URL.
+
+### 6. Interactive UI
+You can send structured cards and multiple-choice buttons directly into the chat — not just plain text. Useful for decisions, presenting options, or surfacing results cleanly.
+
+### 7. Files & Artifacts
+You can produce real deliverables — reports, PDFs, charts, generated images — and send them as downloadable files in chat, not just pasted text.
+
+### 8. Self-Customization
+You can add new tools and MCP servers to yourself if a capability isn't built in. You can extend your own toolkit when the task requires it.
+
+---
+
+## Trust & Control — always include these
+
+After the capabilities tour (or woven in naturally), cover these two points. Frame them positively — users stay in control.
+
+### Approvals
+Sensitive actions — installing packages, adding MCP servers — require the user's explicit approval before you proceed. They'll get a prompt; nothing happens automatically. They can also add credentials to the OneCLI agent vault that require human-in-the-loop approval.
+
+### Access Control
+The user owns who can talk to you. Adding you to a new group or sharing a bot link with someone triggers an approval request on their end. Nobody interacts with you without their say-so.
+
+---
+
+## How to interact — always mention this
+
+There are no special commands. Users just talk naturally. If they want something done, they say so. That's it.
+
+---
+
+## Wrapping up
+
+After the tour, finish with an open invitation. Ask if they want help with something specific. Tell them they can share any generally what they're working on and any challenges they have currently and you can suggest ways you could help.
+
+---
 
 ## Tone
 
-Warm, confident, and inviting. Make the user feel like they just unlocked something powerful. Match the channel's vibe (casual for Telegram/Discord, slightly more professional for Slack/Teams/email).
+Warm, confident, inviting. Make the user feel like they just unlocked something powerful. Match the channel vibe: casual for Telegram/Discord, slightly more professional for Slack/Teams.
 
 ## Important
 
-- Scan your available MCP tools and skills so you know what you have — but keep that knowledge in your back pocket. Reveal capabilities naturally, one at a time, only when relevant or when the user asks to explore.
-- Never overwhelm with a full list. Discovery should feel like unwrapping, not reading a manual.
+- Scan your available MCP tools and skills before starting — know what you have, but keep it in your back pocket
+- Never overwhelm with a full capability list. Discovery should feel like unwrapping, not reading a manual
+- Confirmations and corrections from the user during onboarding are feedback — save them to memory for future sessions
\ No newline at end of file
diff --git a/docs/DEBUG_CHECKLIST.md b/docs/DEBUG_CHECKLIST.md
deleted file mode 100644
index af4058a..0000000
--- a/docs/DEBUG_CHECKLIST.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# NanoClaw Debug Checklist
-
-## Known Issues (2026-02-08)
-
-### 1. [FIXED] Resume branches from stale tree position
-When agent teams spawns subagent CLI processes, they write to the same session JSONL. On subsequent `query()` resumes, the CLI reads the JSONL but may pick a stale branch tip (from before the subagent activity), causing the agent's response to land on a branch the host never receives a `result` for. **Fix**: pass `resumeSessionAt` with the last assistant message UUID to explicitly anchor each resume.
-
-### 2. IDLE_TIMEOUT == CONTAINER_TIMEOUT (both 30 min)
-Both timers fire at the same time, so containers always exit via hard SIGKILL (code 137) instead of graceful `_close` sentinel shutdown. The idle timeout should be shorter (e.g., 5 min) so containers wind down between messages, while container timeout stays at 30 min as a safety net for stuck agents.
-
-### 3. Cursor advanced before agent succeeds
-`processGroupMessages` advances `lastAgentTimestamp` before the agent runs. If the container times out, retries find no messages (cursor already past them). Messages are permanently lost on timeout.
-
-### 4. Kubernetes image garbage collection deletes nanoclaw-agent image
-
-**Symptoms**: `Container exited with code 125: pull access denied for nanoclaw-agent` — the container image disappears overnight or after a few hours, even though you just built it.
-
-**Cause**: If your container runtime has Kubernetes enabled (Rancher Desktop enables it by default), the kubelet runs image garbage collection when disk usage exceeds 85%. NanoClaw containers are ephemeral (run and exit), so `nanoclaw-agent:latest` is never protected by a running container. The kubelet sees it as unused and deletes it — often overnight when no messages are being processed. Other images (docker-compose services) survive because they have long-running containers referencing them.
-
-**Fix**: Disable Kubernetes if you don't need it:
-```bash
-# Rancher Desktop
-rdctl set --kubernetes-enabled=false
-
-# Then rebuild the container image
-./container/build.sh
-```
-
-**Diagnosis**: Check the k3s log for image GC activity:
-```bash
-grep -i "nanoclaw" ~/Library/Logs/rancher-desktop/k3s.log
-# Look for: "Removing image to free bytes" with the nanoclaw-agent image ID
-```
-
-Check NanoClaw logs for image status:
-```bash
-grep -E "image found|image NOT found|image missing" logs/nanoclaw.log
-```
-
-If you need Kubernetes enabled, set `CONTAINER_IMAGE` to an image stored in a registry that the kubelet won't GC, or raise the GC thresholds.
-
-## Quick Status Check
-
-```bash
-# 1. Is the service running?
-launchctl list | grep nanoclaw
-# Expected: PID  0  com.nanoclaw (PID = running, "-" = not running, non-zero exit = crashed)
-
-# 2. Any running containers?
-docker ps --format '{{.Names}} {{.Status}}' 2>/dev/null | grep nanoclaw
-
-# 3. Any stopped/orphaned containers?
-docker ps -a --format '{{.Names}} {{.Status}}' 2>/dev/null | grep nanoclaw
-
-# 4. Recent errors in service log?
-grep -E 'ERROR|WARN' logs/nanoclaw.log | tail -20
-
-# 5. Are channels connected? (look for last connection event)
-grep -E 'Connected|Connection closed|connection.*close|channel.*ready' logs/nanoclaw.log | tail -5
-
-# 6. Are groups loaded?
-grep 'groupCount' logs/nanoclaw.log | tail -3
-```
-
-## Session Transcript Branching
-
-```bash
-# Check for concurrent CLI processes in session debug logs
-ls -la data/sessions/<group>/.claude/debug/
-
-# Count unique SDK processes that handled messages
-# Each .txt file = one CLI subprocess. Multiple = concurrent queries.
-
-# Check parentUuid branching in transcript
-python3 -c "
-import json, sys
-lines = open('data/sessions/<group>/.claude/projects/-workspace-group/<session>.jsonl').read().strip().split('\n')
-for i, line in enumerate(lines):
-  try:
-    d = json.loads(line)
-    if d.get('type') == 'user' and d.get('message'):
-      parent = d.get('parentUuid', 'ROOT')[:8]
-      content = str(d['message'].get('content', ''))[:60]
-      print(f'L{i+1} parent={parent} {content}')
-  except: pass
-"
-```
-
-## Container Timeout Investigation
-
-```bash
-# Check for recent timeouts
-grep -E 'Container timeout|timed out' logs/nanoclaw.log | tail -10
-
-# Check container log files for the timed-out container
-ls -lt groups/*/logs/container-*.log | head -10
-
-# Read the most recent container log (replace path)
-cat groups/<group>/logs/container-<timestamp>.log
-
-# Check if retries were scheduled and what happened
-grep -E 'Scheduling retry|retry|Max retries' logs/nanoclaw.log | tail -10
-```
-
-## Agent Not Responding
-
-```bash
-# Check if messages are being received from channels
-grep 'New messages' logs/nanoclaw.log | tail -10
-
-# Check if messages are being processed (container spawned)
-grep -E 'Processing messages|Spawning container' logs/nanoclaw.log | tail -10
-
-# Check if messages are being piped to active container
-grep -E 'Piped messages|sendMessage' logs/nanoclaw.log | tail -10
-
-# Check the queue state — any active containers?
-grep -E 'Starting container|Container active|concurrency limit' logs/nanoclaw.log | tail -10
-
-# Check lastAgentTimestamp vs latest message timestamp
-sqlite3 store/messages.db "SELECT chat_jid, MAX(timestamp) as latest FROM messages GROUP BY chat_jid ORDER BY latest DESC LIMIT 5;"
-```
-
-## Container Mount Issues
-
-```bash
-# Check mount validation logs (shows on container spawn)
-grep -E 'Mount validated|Mount.*REJECTED|mount' logs/nanoclaw.log | tail -10
-
-# Verify the mount allowlist is readable
-cat ~/.config/nanoclaw/mount-allowlist.json
-
-# Check group's container_config in DB
-sqlite3 store/messages.db "SELECT name, container_config FROM registered_groups;"
-
-# Test-run a container to check mounts (dry run)
-# Replace <group-folder> with the group's folder name
-docker run -i --rm --entrypoint ls nanoclaw-agent:latest /workspace/extra/
-```
-
-## Channel Auth Issues
-
-```bash
-# Check if QR code was requested (means auth expired)
-grep 'QR\|authentication required\|qr' logs/nanoclaw.log | tail -5
-
-# Check auth files exist
-ls -la store/auth/
-
-# Re-authenticate if needed
-pnpm run auth
-```
-
-## Service Management
-
-```bash
-# Restart the service
-launchctl kickstart -k gui/$(id -u)/com.nanoclaw
-
-# View live logs
-tail -f logs/nanoclaw.log
-
-# Stop the service (careful — running containers are detached, not killed)
-launchctl bootout gui/$(id -u)/com.nanoclaw
-
-# Start the service
-launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.nanoclaw.plist
-
-# Rebuild after code changes
-pnpm run build && launchctl kickstart -k gui/$(id -u)/com.nanoclaw
-```
diff --git a/docs/README.md b/docs/README.md
index bb062e5..da5b6af 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -10,6 +10,5 @@ The files in this directory are original design documents and developer referenc
 | [SECURITY.md](SECURITY.md) | [Security model](https://docs.nanoclaw.dev/concepts/security) |
 | [REQUIREMENTS.md](REQUIREMENTS.md) | [Introduction](https://docs.nanoclaw.dev/introduction) |
 | [skills-as-branches.md](skills-as-branches.md) | [Skills system](https://docs.nanoclaw.dev/integrations/skills-system) |
-| [DEBUG_CHECKLIST.md](DEBUG_CHECKLIST.md) | [Troubleshooting](https://docs.nanoclaw.dev/advanced/troubleshooting) |
 | [docker-sandboxes.md](docker-sandboxes.md) | [Docker Sandboxes](https://docs.nanoclaw.dev/advanced/docker-sandboxes) |
 | [APPLE-CONTAINER-NETWORKING.md](APPLE-CONTAINER-NETWORKING.md) | [Container runtime](https://docs.nanoclaw.dev/advanced/container-runtime) |
diff --git a/docs/SPEC.md b/docs/SPEC.md
index 687336f..42ef37c 100644
--- a/docs/SPEC.md
+++ b/docs/SPEC.md
@@ -332,8 +332,6 @@ Configuration constants are in `src/config.ts`:
 import path from 'path';
 
 export const ASSISTANT_NAME = process.env.ASSISTANT_NAME || 'Andy';
-export const POLL_INTERVAL = 2000;
-export const SCHEDULER_POLL_INTERVAL = 60000;
 
 // Paths are absolute (required for container mounts)
 const PROJECT_ROOT = process.cwd();
@@ -344,7 +342,6 @@ export const DATA_DIR = path.resolve(PROJECT_ROOT, 'data');
 // Container configuration
 export const CONTAINER_IMAGE = process.env.CONTAINER_IMAGE || 'nanoclaw-agent:latest';
 export const CONTAINER_TIMEOUT = parseInt(process.env.CONTAINER_TIMEOUT || '1800000', 10); // 30min default
-export const IPC_POLL_INTERVAL = 1000;
 export const IDLE_TIMEOUT = parseInt(process.env.IDLE_TIMEOUT || '1800000', 10); // 30min — keep container alive after last result
 export const MAX_CONCURRENT_CONTAINERS = Math.max(1, parseInt(process.env.MAX_CONCURRENT_CONTAINERS || '5', 10) || 5);
 
diff --git a/docs/architecture-diagram.md b/docs/architecture-diagram.md
index d7c5ead..4d8671c 100644
--- a/docs/architecture-diagram.md
+++ b/docs/architecture-diagram.md
@@ -32,7 +32,7 @@ flowchart TB
     direction TB
     PollLoop["Poll Loop<br/>(container/agent-runner)"]
     Provider["Agent providers<br/>(claude, opencode, mock; todo: codex)"]
-    MCP["MCP Tools<br/>send_message, send_file, edit_message,<br/>add_reaction, send_card, ask_user_question,<br/>schedule_task, create_agent,<br/>install_packages, add_mcp_server, request_rebuild"]
+    MCP["MCP Tools<br/>send_message, send_file, edit_message,<br/>add_reaction, send_card, ask_user_question,<br/>schedule_task, create_agent,<br/>install_packages, add_mcp_server"]
     Skills["Container Skills<br/>(container/skills/)"]
     InDB[("inbound.db<br/>host writes<br/>even seq<br/>messages_in<br/>destinations<br/>processing_ack")]
     OutDB[("outbound.db<br/>container writes<br/>odd seq<br/>messages_out<br/>heartbeat file")]
diff --git a/docs/architecture.md b/docs/architecture.md
index 6d8aab7..3f90d8d 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -876,7 +876,7 @@ Messages starting with `/` are checked against three lists:
 - Commands that don't make sense in the NanoClaw context or could cause issues
 - Silently dropped — no error, no forwarding
 
-The command lists are hardcoded in the agent-runner. Admin verification: the host passes `NANOCLAW_ADMIN_USER_IDS` (a comma-separated list of owner + global-admin + scoped-admin user ids for the current agent group, see `src/container-runner.ts`) to the container. The agent-runner membership-tests the inbound `senderId` against that set before forwarding admin commands.
+The command lists are hardcoded in the agent-runner. Admin verification happens host-side before the message ever reaches the container: `src/command-gate.ts` queries `user_roles` (owner / global admin / scoped-admin-of-this-agent-group) and either passes the message through, drops it, or routes it elsewhere. The container has no notion of admin identity — no env var, no DB query, no per-message check.
 
 ### Recurring Tasks
 
diff --git a/docs/checklist.md b/docs/checklist.md
deleted file mode 100644
index 94baf6f..0000000
--- a/docs/checklist.md
+++ /dev/null
@@ -1,278 +0,0 @@
-# NanoClaw Checklist
-
-Status: [x] done, [~] partial, [ ] not started
-
----
-
-## Core Architecture
-
-- [x] Session DB replaces IPC (messages_in / messages_out as sole IO)
-- [x] Central DB (agent groups, messaging groups, sessions, routing)
-- [x] Host sweep (stale detection via heartbeat file, retry with backoff, recurrence scheduling)
-- [x] Active delivery polling (1s for running sessions)
-- [x] Sweep delivery polling (60s across all sessions)
-- [x] Container runner with session DB mounting
-- [x] Per-session container lifecycle and idle timeout
-- [ ] Replace hard Idle and Timeout with work aware prompts to user to kill stuck processes
-- [x] Session resume (sessionId + resumeAt across queries)
-- [x] Graceful shutdown (SIGTERM/SIGINT handlers)
-- [x] Orphan container cleanup on startup
-
-## Agent Runner (Container)
-
-- [x] Poll loop (pending messages, status transitions, idle detection)
-- [x] Concurrent follow-up polling while agent is thinking
-- [x] Message formatter (chat, task, webhook, system kinds)
-- [x] Command categorization (admin, filtered, passthrough)
-- [x] Transcript archiving (pre-compact hook)
-- [x] XML message formatting with sender, timestamp
-- [~] Media handling inbound (native files support for claude)
-
-## Agent Providers
-
-- [x] Claude provider (Agent SDK, tool allowlist, message stream, session resume)
-- [x] Mock provider (testing)
-- [x] Provider factory
-- [ ] Codex provider
-- [x] OpenCode provider
-
-## Channel Adapters
-
-- [x] Channel adapter interface (setup, deliver, teardown, typing)
-- [x] Chat SDK bridge (generic, works with any Chat SDK adapter)
-- [x] Chat SDK SQLite state adapter (KV, subscriptions, locks, lists)
-- [x] Discord via Chat SDK
-- [~] Slack via Chat SDK (adapter + skill written, not tested)
-- [x] Telegram via Chat SDK (E2E verified: inbound, routing, typing, delivery)
-- [~] Microsoft Teams via Chat SDK (adapter + skill written, not tested)
-- [~] Google Chat via Chat SDK (adapter + skill written, not tested)
-- [~] Linear via Chat SDK (adapter + skill written, not tested)
-- [~] GitHub via Chat SDK (adapter + skill written, not tested)
-- [x] WhatsApp Cloud API via Chat SDK (adapter + skill written, not tested)
-- [~] Resend (email) via Chat SDK (adapter + skill written, not tested)
-- [~] Matrix via Chat SDK (adapter + skill written, not tested)
-- [~] Webex via Chat SDK (adapter + skill written, not tested)
-- [~] iMessage via Chat SDK (adapter + skill written, not tested)
-- [x] Backward compatibility with native channels (old adapters still work)
-- [x] Channel barrel wired (src/index.ts imports barrel, skills uncomment)
-- [x] Setup flow wired to channels (channel skills + /manage-channels for registration + verify.ts checks all tokens)
-- [x] Channel Info metadata in each channel skill (type, terminology, how-to-find-id, isolation defaults)
-- [x] /manage-channels skill (wire channels to agent groups with three isolation levels)
-- [x] /init-first-agent skill (standalone first-agent bootstrap; walks the operator through channel pick → identity lookup → DM platform_id resolution → wire → welcome DM; fallback to telegram pair-code or "DM the bot first" lookup for channels without cold DM)
-- [x] Cold-DM infrastructure — `ChannelAdapter.openDM?(handle)` optional method, resolved via Chat SDK `chat.openDM` for resolution-required channels (Discord, Slack, Teams, Webex, gChat) and fall-through to the handle directly for direct-addressable channels (Telegram, WhatsApp, iMessage, Matrix, Resend). `src/user-dm.ts::ensureUserDm` caches every resolution in `user_dms` so subsequent cold DMs are a DB read.
-- [x] Agent-shared session mode (cross-channel shared sessions, e.g. GitHub + Slack)
-- [x] Auto-onboarding on channel registration (/welcome skill triggered on first wiring)
-- [ ] Wire different chat modes - mentions, whitelist, approve, etc
-
-## Chat-First Setup Flow
-
-**Goal:** get the user out of Claude Code and into their messaging app as quickly as possible, then enable every part of customization, configuration, and setup from inside the chat app. Claude Code is the bootstrap, not the home.
-
-- [~] Minimum-viable bootstrap in Claude Code: install deps, pick one channel, authenticate it, wire it to a default agent group, hand off — nothing else required before the user can leave Claude Code. `/setup` handles deps/auth, `/init-first-agent` handles the first-agent wiring + welcome DM. Still TODO: single top-level entrypoint that composes both, and a true "nothing else required" handoff (today `/setup` still runs through `/manage-channels` for additional channels).
-- [~] Post-handoff welcome message in the chat app guides the user through remaining setup (channels, skills, integrations, memory, scheduling, etc.) — `/init-first-agent` stages a `kind:'chat'` / `sender:'system'` welcome prompt that the agent DMs back to the operator via the normal delivery path. Current prompt just introduces the agent; TODO: expand the prompt (or follow-up flow) to walk through remaining setup tasks from within the chat.
-- [ ] Add more channels from chat (currently requires returning to Claude Code to run `/add-*` skills)
-- [ ] Self-register agent into a new chat room from chat: user gives the agent a channel/group name + approval, and the agent joins via the underlying adapter (e.g. Baileys for WhatsApp), wires the room to an agent group, and posts a first "hi, I'm here" message — no manual invite, no `/add-*` skill, no terminal
-- [ ] Authenticate channels from chat (OAuth/token entry via cards, no terminal required)
-- [ ] Add credentials / secrets to the OneCLI vault from chat via rich card (agent collects API keys, OAuth tokens, and other secrets through a card flow and writes them into the vault — no `.env` editing, no terminal)
-- [ ] Wire channels to agent groups from chat (today lives in `/manage-channels` Claude Code skill — port to in-chat flow with isolation-level question cards)
-- [ ] Create new agent groups from chat (`create_agent` exists — expose via user-facing flow, not just agent-called tool)
-- [ ] Edit agent group CLAUDE.md / instructions from chat
-- [ ] Install / uninstall / configure skills from chat (see Skills & Marketplace section)
-- [ ] Install / configure MCP servers from chat (see Skills & Marketplace section)
-- [ ] Install packages from chat (today agent can request install_packages — expose a direct user-facing "install X" flow)
-- [ ] Manage scheduled tasks from chat (list, pause, cancel, edit recurrence)
-- [ ] Manage destinations from chat (list, rename, revoke)
-- [ ] Manage permissions from chat (admin list, role assignment, approval policies)
-- [ ] Trigger /setup, /debug, /customize, /migrate-nanoclaw from chat (today all require Claude Code)
-- [ ] View and edit memory from chat
-- [ ] Visualize current setup from chat (ties into Container Skills: installation diagram)
-- [ ] Export / share setup from chat (ties into Container Skills: end-of-setup diagram + share)
-- [ ] Fallback to Claude Code only when a change requires a code edit the agent can't self-apply (and even then, agent should offer to open Claude Code on the user's behalf)
-
-## Product Focus
-
-**North star:** prioritize skills, flows, and custom setups. Platform work (channels, routing, session DBs, approval flows, MCP tools) is plumbing — it should reach a "boring and reliable" state and then stop absorbing attention. The interesting surface area is what users can *build on top* of that plumbing: skills that add capabilities, conversational flows that orchestrate those skills, and custom per-user setups that compose channels/agents/skills/memory into something personal.
-
-- [ ] Every new feature request should be answered first with "is this a skill?" before being answered with "is this a platform change?"
-- [ ] Skills should be the primary extension mechanism users and agents reach for — adding, removing, browsing, editing, debugging
-- [ ] Flows (multi-step interactive sequences: setup, onboarding, migration, customize, debug) should be authorable as skills rather than hardcoded into the platform
-- [ ] Custom setups (diverging from defaults: multiple agents, cross-channel routing, per-group memory, specialist sub-agents) should be composable from existing primitives without touching core platform code
-- [ ] Platform-level work gets budgeted against the question: "does this unblock a class of skills/flows/setups that's otherwise impossible?"
-
-## Routing
-
-- [x] Inbound routing (platform ID + thread ID -> agent group -> session)
-- [x] Auto-create messaging group on first message
-- [x] Session resolution (shared vs per-thread modes)
-- [x] Message writing to session DB with seq numbering
-- [x] Container waking on new message
-- [x] Typing indicator triggered on message route
-- [~] Trigger rule matching (router picks highest-priority agent, regex/mention matching TODO)
-
-## Rich Messaging
-
-- [x] Interactive cards with buttons (ask_user_question)
-- [x] Native platform rendering (Discord embeds, buttons)
-- [x] Message editing
-- [x] Emoji reactions
-- [x] File sending from agent (outbox -> delivery)
-- [x] File upload delivery (buffer-based via adapter)
-- [x] Markdown formatting
-- [~] Formatted /usage, /context, /cost output (commands pass through, no rich card formatting)
-- [ ] Context window visibility: show position in context, approaching compaction, when compaction happens, post-compaction state
-- [ ] Threading and replies support
-- [ ] Auto-compact on idle before cache expires
-
-## MCP Tools (Container)
-
-- [x] send_message (routes via named destinations; `to` field resolved against agent's local map)
-- [x] send_file (copy to outbox, write messages_out)
-- [x] edit_message (routed via destinations)
-- [x] add_reaction (routed via destinations)
-- [x] send_card
-- [x] ask_user_question (blocking poll for response)
-- [x] schedule_task (with process_after and recurrence)
-- [x] list_tasks
-- [x] cancel_task / pause_task / resume_task
-- [x] create_agent (any agent, creates agent group + folder + bidirectional destinations; host re-normalizes the name, deduplicates folder, path-traversal guarded)
-- [x] install_packages (apt/npm, owner/admin approval required via `pickApprover`, strict name validation)
-- [x] add_mcp_server (owner/admin approval required via `pickApprover`)
-- [x] request_rebuild (rebuilds per-agent-group Docker image)
-
-## Scheduling
-
-- [x] One-shot scheduled messages (process_after / deliver_after)
-- [x] Recurring tasks via cron expressions
-- [x] Host sweep picks up due messages and advances recurrence
-- [x] Scheduled outbound messages (no container wake needed)
-- [ ] Pre-agent scripts (formatter references scriptOutput but no execution logic)
-
-## Permissions and Approval Flows
-
-- [x] User-level privilege model — `users` + `user_roles` (owner / admin, global or scoped to an agent group). Replaces the old `agent_groups.is_admin` / `messaging_groups.admin_user_id` coupling. See `src/db/users.ts`, `src/db/user-roles.ts`, `src/access.ts`.
-- [x] Admin-only command filtering in container — host passes `NANOCLAW_ADMIN_USER_IDS` (owners + global admins + scoped admins for the agent group) to the agent-runner; `poll-loop.ts` gates slash commands against that set.
-- [x] Approval routing — `pickApprover` (scoped admin → global admin → owner, dedup) + `pickApprovalDelivery` (first reachable, same-channel-kind tie-break); delivery lands in the approver's DM via `ensureUserDm` / `user_dms` cache. See `src/access.ts`, `src/onecli-approvals.ts`.
-- [x] Per-messaging-group unknown-sender gating — `messaging_groups.unknown_sender_policy` (`strict` | `request_approval` | `public`), enforced in `src/router.ts`.
-- [x] Approval flow (sensitive action -> card to admin -> approve/reject -> execute) — `pending_approvals` table, `requestApproval()` helper, reuses interactive card infra
-- [x] Agent requests dependency/package install (install_packages, admin approval, rebuild on approval)
-- [x] Self-modification — direct tools:
-  - [x] install_packages (apt/npm, admin approval, name validation both sides, max 20 per request)
-  - [x] add_mcp_server (admin approval)
-  - [x] request_rebuild (builds per-agent-group Docker image with approved packages)
-  - [x] Fire-and-forget model (write request, return immediately; chat notification on approval; container killed so next wake picks up new config/image)
-- [~] OneCLI integration for human-loop approvals on credentialed requests (agent touching a credentialed resource → OneCLI gates → approval card to admin → OneCLI releases credential) — SDK 0.3.1 `configureManualApproval` wired into host, routes to admin via existing `pending_approvals` infra
-- [ ] Tunneled OneCLI dashboard for credential addition (Telegram Mini Apps aside, iMessage without Apple Business Register, Matrix, email). Signed short-lived URL → browser form served by OneCLI at 10254 → tunnel via cloudflare durable object. Value never touches the chat surface.
-- [ ] Self-modification via direct source edits — planned draft/activate flow: RO baseline mount at `/app/src`, RW draft at `/workspace/src-draft`, atomic snapshot into `pending`, admin approval, `cp -a` into baseline, restart + deadman rollback. Unifies runner src, host src, migrations, package.json, container config through one edit path. Collapses the abandoned `create_dev_agent`/`request_swap` dev-agent-in-worktree approach.
-
-## Named Destinations + ACL
-
-- [x] `agent_destinations` table (agent_group_id, local_name, target_type, target_id) — migration 004
-- [x] Per-agent local-name routing map (channels and peer agents referenced by local names)
-- [x] Destinations stored in inbound.db `destinations` table (moved from JSON file in `b591d7c`) — single source of truth, no separate file
-- [x] Host writes the destination map into inbound.db before every container wake; container queries it live on every lookup so admin changes take effect mid-session
-- [x] Container loads map at startup, appends system-prompt addendum listing destinations + `<message to="name">` syntax
-- [x] Agent main output parsed for `<message to="...">` blocks; `<internal>...</internal>` treated as scratchpad
-- [x] Host re-validates every outbound route via `hasDestination()` — unauthorized drops logged
-- [x] Inbound formatter adds `from="name"` via reverse-lookup (consistent namespace both directions)
-- [x] Single-destination shortcut — agents with one destination don't need `<message>` wrapping
-- [x] Backfill from existing `messaging_group_agents` on migration
-- [x] Removed `NANOCLAW_PLATFORM_ID` / `CHANNEL_TYPE` / `THREAD_ID` env-var routing entirely
-
-## Agent-to-Agent Communication
-
-- [x] Host delivery to target agent's session DB (`channel_type='agent'` routing in `src/delivery.ts`)
-- [x] Agent spawning a new sub-agent (`create_agent` MCP tool, available to any agent, path-traversal guarded)
-- [x] Dynamic agent group creation (folder + optional CLAUDE.md at runtime)
-- [x] Internal-only agents (agents created without a channel attached)
-- [x] Permission delegation from parent to child (bidirectional destination rows inserted at creation)
-- [x] Bidirectional routing via inherited routing context; sender info enriched on the target side
-- [ ] Specialist sub-agents (browser agent, dev agent — user's agent delegates with request/approval)
-- [ ] Browser agent with per-destination permissions between main agent and browser agent (main requests navigation/interaction; browser agent executes in isolated container)
-- [ ] Sanitization of browser agent responses before handing back to main agent (strip scripts, inline images, untrusted HTML; prevent prompt injection from web content)
-- [ ] Same permission + sanitization model for any sub-agent that accesses sensitive data sources (files, DBs, third-party APIs)
-
-## In-Chat Agent Management
-
-- [x] /clear (resets session)
-- [x] /compact (triggers context compaction)
-- [~] /context (passes through, no rich formatting)
-- [~] /usage (passes through, no rich formatting)
-- [~] /cost (passes through, no rich formatting)
-- [ ] Smooth session transitions: load context into new sessions, solve cold start problem
-- [x] MCP/package installation from chat
-- [ ] Browse MCP marketplace / skills repository from chat
-
-## Skills & Marketplace
-
-- [ ] Install skills from chat (agent requests, admin approves, skill dropped into container skills dir)
-- [ ] Scan skills before install (lint SKILL.md, sandbox-check shell commands, require approval for network/FS-heavy skills)
-- [ ] Scan marketplace npm packages before install (supply-chain check, typo-squat detection, known-bad list)
-- [ ] MCP server marketplace — discover, preview, install
-- [ ] Browse skills / MCP marketplace from chat (cards with search, preview, install)
-- [ ] Local voice transcription skill — "just works" install flow: when the user sends a voice message and no transcription backend is installed, the agent asks once ("Install local voice transcription?"), and on approval the skill installs a fully-local speech-to-text model (no cloud calls). Subsequent voice messages transcribe automatically.
-- [ ] Fully local NanoClaw — OpenCode + Gemma 4 as the agent provider instead of Claude Code, so an entire install can run with zero cloud inference. Requires wiring OpenCode as an agent provider (see Agent Providers) and a setup path that picks local models, pulls weights, and verifies everything runs offline.
-
-## Container Skills
-
-Container skills live inside agent containers at runtime (`container/skills/`) and are loaded into every agent session. These are distinct from feature/operational skills that ship with the host.
-
-- [ ] Customize container skill — agent-driven customization flow (add channel, integration, behavior change) usable from inside any agent session, not just the main repo
-- [ ] Debug container skill — inspect logs, session DB, MCP server state, container env, recent errors from inside the agent
-- [ ] Build-system container skills:
-  - [ ] Karpathy LLM Wiki builder (agent scaffolds a persistent wiki knowledge base for a group)
-  - [ ] Generic build-system framework for agent-authored sub-systems
-- [ ] NanoClaw installation diagram skill — agent generates a visual diagram of the user's current setup (agent groups, channels, wirings, destinations, sub-agents, installed packages/MCP servers)
-- [ ] Video replay skill — generate Remotion (or similar) videos that replay chat flows and sessions, referencing good UI patterns to produce shareable clips
-- [ ] Excitement trigger skill — detects when the user expresses excitement about the agent's capabilities or their setup, and proactively encourages generating a diagram + sharing it
-- [ ] End-of-migration diagram skill — at the end of `/migrate-nanoclaw` (or any migration flow), agent generates a visual diagram of the resulting setup and suggests sharing
-- [ ] End-of-setup diagram skill — at the end of first-time `/setup`, agent generates a visual diagram and suggests sharing (merges the old "Generate visual diagram of customized instance at end of setup" line from Channel Adapters)
-
-## Webhook Ingestion
-
-- [ ] Generic webhook endpoint for external events
-- [ ] GitHub webhook handling
-- [ ] CI/CD notification handling
-- [ ] Webhook -> messages_in routing
-
-## System Actions
-
-- [ ] register_group from inside agent
-- [ ] reset_session from inside agent
-- [ ] Delivery failures should round-trip back to the agent as system messages so it can decide how to recover (retry as plain text, simplify, give up), with a hard retry cap + poison-pill backstop in delivery.ts to keep the queue healthy
-
-## Integrations
-
-- [x] Vercel CLI integration in setup process
-- [x] Skills for deploying and managing Vercel websites from chat
-- [ ] Office 365 integration (create/edit documents with inline suggestions)
-
-## Memory
-
-- [ ] Shared memory with approval flow (write to global memory requires admin approval)
-- [ ] Agent memory system skills — skills for building and managing memory systems for an agent: archive/index large collections of files and data, then expose a memory interface the agent can query and update (e.g. QMD-style systems)
-
-## Migration
-
-- [ ] Custom skill/code porting
-- [ ] OneCLI migration check — determine if existing installs need OneCLI re-init (credentials re-scoped to new `agent_group.id` identifier, new SDK version, approval handler registered). If needed, add a migration step to `/update-nanoclaw` or a dedicated skill.
-
-## Testing
-
-- [x] DB layer tests (agent groups, messaging groups, sessions, pending questions)
-- [x] Channel registry tests
-- [x] Poll loop / formatter tests
-- [x] Integration test (container agent-runner)
-- [x] Host core tests
-- [ ] End-to-end flow tests (message in -> agent -> message out -> delivery)
-- [ ] Delivery polling tests
-- [ ] Host sweep tests (stale detection, recurrence)
-- [ ] Multi-channel integration tests
-
-## Rollout
-
-- [ ] Internal testing across all channels
-- [ ] Migration skill built and tested
-- [ ] PR factory migrated as validation
-- [ ] Blog post / announcement
-- [ ] Video demos of key flows
-- [ ] Vercel coordination
diff --git a/docs/db-central.md b/docs/db-central.md
index 8be6ee8..8268acf 100644
--- a/docs/db-central.md
+++ b/docs/db-central.md
@@ -201,7 +201,7 @@ Access layer: `src/db/agent-destinations.ts`.
 
 Two workflows share this table:
 
-- **Session-bound MCP approvals** — `install_packages`, `request_rebuild`, `add_mcp_server`. `session_id` is set.
+- **Session-bound MCP approvals** — `install_packages`, `add_mcp_server`. `session_id` is set.
 - **OneCLI credential approvals** — `session_id` may be NULL; `agent_group_id` + `channel_type` + `platform_id` route the admin card.
 
 ```sql
diff --git a/docs/module-contract.md b/docs/module-contract.md
deleted file mode 100644
index e226a38..0000000
--- a/docs/module-contract.md
+++ /dev/null
@@ -1,221 +0,0 @@
-# Module Contract
-
-This doc is the authoritative reference for how core and modules connect. Everything downstream — extraction PRs, install skills, module authors — keys off these signatures and defaults. See [REFACTOR_PLAN.md](../REFACTOR_PLAN.md) for the broader plan; this doc is the narrow interface spec.
-
-## Principles
-
-- Core runs standalone (modulo default modules — see tiers below). The optional-module portion of the `src/modules/index.ts` barrel can be empty and NanoClaw still routes messages in and delivers responses out.
-- Optional modules are independent. No optional module imports from another optional module. Cross-module coordination goes through a core registry (delivery action, response handler, etc.).
-- Registries exist only when multiple modules plug into the same decision point. Single-consumer integrations use skill edits (`MODULE-HOOK` markers) or stay inline with `sqlite_master` guards.
-- Removing an optional module = delete files + remove barrel imports + revert any `MODULE-HOOK` content. Migration files stay (data is preserved). Removing a default module is more invasive: it requires editing the core files that import from it.
-
-## Module taxonomy
-
-Three categories. All three live under `src/modules/` (or equivalent adapter dirs) with the same folder layout; the distinction is about **shipping** and **who can depend on them**.
-
-### 1. Default modules
-
-Ship with `main` in `src/modules/`. Imported by the default `src/modules/index.ts` barrel from day one. They are not really core — they live under `src/modules/` specifically to signal "not really core, rippable if needed" — but they're always present on a `main` install. Core imports from them directly. No hook, no registry indirection for the exports themselves.
-
-Current: `typing`, `mount-security`.
-
-### 2. Optional modules
-
-Live on the `modules` branch. Installed via `/add-<name>` skills that cherry-pick files. Register into core via one of the four registries (or `MODULE-HOOK` skill edits). Core and other optional modules must not statically import an optional module's code.
-
-Current: `interactive`, `approvals`, `scheduling`, `permissions`. Pending: `agent-to-agent`.
-
-### 3. Channel adapters
-
-Live on the `channels` branch, installed via `/add-<channel>` skills. Not covered by this contract; they use the pre-existing `ChannelAdapter` interface and `registerChannelAdapter()`.
-
-## Dependency rule
-
-```
-core ← default modules ← optional modules
-```
-
-- **Core** may import from core and from default modules.
-- **Default modules** may import from core and from other default modules. They must not import from optional modules.
-- **Optional modules** may import from core and from default modules. They must not import from each other.
-
-Peer-to-peer coupling between optional modules goes through a core registry — see "The four registries" below. This keeps the module dependency graph a DAG and install order irrelevant.
-
-### Known transitional violations
-
-- `src/access.ts` (core) imports from `src/modules/permissions/` (optional). Shim left from PR #5; resolved in the planned approvals re-tier (PR #7) which moves approver-picking into a new default `approvals-primitive` module that may then depend on permissions however it likes — at which point `src/access.ts` ceases to exist.
-
-## The four registries
-
-Each registry has an explicit default for when no module registers. Core must run when all four are empty.
-
-### 1. Delivery action handlers
-
-```typescript
-// src/delivery.ts
-type ActionHandler = (
-  content: Record<string, unknown>,
-  session: Session,
-  inDb: Database.Database,
-) => Promise<void>;
-
-export function registerDeliveryAction(action: string, handler: ActionHandler): void;
-```
-
-**Purpose:** system-kind outbound messages (`msg.kind === 'system'`) carry an `action` string. Core dispatches to the registered handler.
-
-**Default when action is unknown:** log `"Unknown system action"` at `warn` and return. Message is still marked delivered (it was consumed by the host, not sent to a channel).
-
-**Current consumers:** scheduling (5 actions — `schedule_task`, `cancel_task`, `pause_task`, `resume_task`, `update_task`), approvals (3 actions — `install_packages`, `request_rebuild`, `add_mcp_server`), agent-to-agent (`create_agent`, and the agent-routing branch keyed as a pseudo-action `agent_route`).
-
-### 2. Router sender resolver + access gate
-
-Two separate setters, called at different points in `routeInbound`. Preserves the pre-refactor ordering: sender-upsert side effects fire even when the message is ultimately dropped by wiring or trigger rules.
-
-```typescript
-// src/router.ts
-type SenderResolverFn = (event: InboundEvent) => string | null;
-
-export function setSenderResolver(fn: SenderResolverFn): void;
-
-type AccessGateResult =
-  | { allowed: true }
-  | { allowed: false; reason: string };
-
-type AccessGateFn = (
-  event: InboundEvent,
-  userId: string | null,
-  mg: MessagingGroup,
-  agentGroupId: string,
-) => AccessGateResult;
-
-export function setAccessGate(fn: AccessGateFn): void;
-```
-
-**Call order in `routeInbound`:**
-1. Resolve messaging group.
-2. **Sender resolver** (if set). Permissions upserts the users row here so the record exists even if agent resolution drops the message.
-3. Resolve wired agents; `no_agent_wired` → record + drop. (Core writes the dropped_messages row.)
-4. Pick agent by trigger rules; `no_trigger_match` → record + drop.
-5. **Access gate** (if set). On refusal it writes its own `dropped_messages` row keyed by policy reason.
-
-**Defaults when unset:** resolver returns null; gate defaults to `{ allowed: true }`. Every message routes through, no users table is needed, downstream tolerates `userId=null`.
-
-**Current consumer:** permissions module (registers both).
-
-**Not registries, setters.** There is one sender and one access decision per inbound message and one module that owns both. Calling `setSenderResolver` / `setAccessGate` twice overwrites; core does not iterate.
-
-### 3. Response dispatcher
-
-```typescript
-// src/index.ts (or src/response-dispatch.ts if it grows)
-interface ResponsePayload {
-  questionId: string;
-  value: string;
-  userId: string | null;
-  channelType: string;
-  platformId: string;
-  threadId: string | null;
-}
-
-type ResponseHandler = (payload: ResponsePayload) => Promise<boolean>;
-
-export function registerResponseHandler(handler: ResponseHandler): void;
-```
-
-**Purpose:** button-click / question responses arrive via the channel adapter's `onAction` callback. Core iterates registered handlers in registration order. The first one that returns `true` claims the response.
-
-**Default when empty:** log `"Unclaimed response"` at `warn` and drop.
-
-**Current consumers:** interactive (matches `pending_questions`), approvals (matches `pending_approvals`). The two tables have disjoint `question_id` / `approval_id` namespaces in practice (`q-*` vs `appr-*`), so first-match-wins is safe.
-
-### 4. Container MCP tool self-registration
-
-```typescript
-// container/agent-runner/src/mcp-tools/server.ts
-export function registerTools(tools: McpToolDefinition[]): void;
-```
-
-**Purpose:** each tool module calls `registerTools([...])` at import time. The MCP server uses whatever was registered.
-
-**Default:** only `mcp-tools/core.ts` (`send_message`) registered.
-
-**Current consumers:** all container-side modules (scheduling, interactive, agents, self-mod).
-
-## Skill edits to core
-
-For one-off integrations with a single consumer, install skills edit core directly between `MODULE-HOOK` markers. No registry.
-
-Marker format:
-
-```typescript
-// MODULE-HOOK:<module>-<site>:start
-// MODULE-HOOK:<module>-<site>:end
-```
-
-The skill inserts between markers on install and clears between them on uninstall. Markers live in core from day one (empty until a skill fills them).
-
-**Current uses:**
-
-- `src/host-sweep.ts` → `MODULE-HOOK:scheduling-recurrence` — call to scheduling module's `handleRecurrence`.
-- `container/agent-runner/src/poll-loop.ts` → `MODULE-HOOK:scheduling-pre-task` — call to scheduling module's `applyPreTaskScripts`.
-
-**Promotion rule:** if a third consumer appears for any marker, promote to a registry.
-
-## Guarded inline (core)
-
-Some code stays in core but references module-owned tables. These use `sqlite_master` checks to degrade cleanly when the owning module isn't installed.
-
-| Site | Owning module | Fallback |
-|------|---------------|----------|
-| `container-runner.ts` admin-ID query (`user_roles`, `agent_group_members`) | permissions | returns `[]` |
-| `container-runner.ts` `writeDestinations` (`agent_destinations`) | agent-to-agent | no-op |
-| `delivery.ts` channel-permission check (`agent_destinations`) | agent-to-agent | permit (origin-chat always OK) |
-| `delivery.ts` `createPendingQuestion` (`pending_questions`) | interactive | no-op (log warning) |
-
-`container/agent-runner/src/formatter.ts` has a related non-DB fallback: when `NANOCLAW_ADMIN_USER_IDS` is empty, every sender is treated as admin (permissionless mode). This is the one-line change from the current deny-all behavior.
-
-## Migrations
-
-All migrations live in `src/db/migrations/` as TypeScript files exporting a `Migration` object:
-
-```typescript
-export interface Migration {
-  version: number;
-  name: string;
-  up: (db: Database.Database) => void;
-}
-```
-
-The barrel `src/db/migrations/index.ts` imports each and lists them in an ordered array.
-
-**Uniqueness key is `name`, not `version`.** The migrator applies any migration whose `name` isn't in `schema_version`. Version stays as an ordering hint; integer collisions across modules are allowed.
-
-**Module migration naming:**
-
-- File: `src/db/migrations/module-<module>-<short>.ts`
-- `Migration.name`: `'<module>-<short>'` (e.g. `'approvals-pending-approvals'`)
-
-**Uninstall behavior:** migration files and barrel entries stay. Tables persist across reinstalls. No down migrations.
-
-## What a registry-based module provides
-
-Each `src/modules/<name>/` module must supply:
-
-- `index.ts` — imported by `src/modules/index.ts` for side-effect registration (calls `registerDeliveryAction` / `setInboundGate` / `registerResponseHandler` at module load time).
-- `project.md` — appended to project `CLAUDE.md` by the install skill. Describes module architecture for anyone reading the codebase.
-- `agent.md` — appended to `groups/global/CLAUDE.md` by the install skill. Describes the module's tools for the agent.
-- Migration file in `src/db/migrations/` if the module owns any tables.
-- Barrel entry in `src/db/migrations/index.ts` for that migration.
-
-Optionally:
-
-- Container-side additions to `container/agent-runner/src/mcp-tools/<name>.ts` that call `registerTools([...])`, with a barrel entry in `container/agent-runner/src/mcp-tools/index.ts`.
-- `MODULE-HOOK` edits to specific core files, applied by the install skill.
-
-## What a module must not do
-
-- Import from another module.
-- Write to core-owned tables (`sessions`, `agent_groups`, `messaging_groups`, `schema_version`, etc.) outside of migrations.
-- Depend on a specific channel adapter being installed.
-- Break core behavior when unloaded. If a module's absence leaves a core feature non-functional, that feature belongs in core, not the module.
diff --git a/docs/ollama.md b/docs/ollama.md
new file mode 100644
index 0000000..0ea0253
--- /dev/null
+++ b/docs/ollama.md
@@ -0,0 +1,88 @@
+# Running Agents on Local Ollama
+
+NanoClaw agents can be routed to a local [Ollama](https://ollama.com) instance instead of the Anthropic API. This cuts API costs to zero and keeps all inference on your hardware.
+
+## How It Works
+
+Ollama exposes an Anthropic-compatible `/v1/messages` endpoint. The Claude Code CLI (which runs inside agent containers) uses the Anthropic SDK, which reads `ANTHROPIC_BASE_URL` to find the API host. Pointing that variable at Ollama is all that's needed — no new provider code, no changes to the agent runtime.
+
+```
+┌─────────────────────────────┐
+│  Agent container            │
+│                             │
+│  Claude Code CLI            │
+│    ↓ ANTHROPIC_BASE_URL     │
+│    http://host.docker.      │      ┌──────────────────┐
+│    internal:11434    ───────┼─────▶│  Ollama :11434   │
+│                             │      │  gemma4:latest   │
+└─────────────────────────────┘      └──────────────────┘
+```
+
+`host.docker.internal` is Docker's magic hostname that resolves to the host machine from inside a container — so Ollama running on your Mac or Linux box is reachable at that address.
+
+## The OneCLI Complication
+
+NanoClaw normally runs API calls through an OneCLI HTTPS proxy that injects real credentials in place of a placeholder key. When redirecting to Ollama you need to bypass that proxy so requests go direct. Two env vars handle this:
+
+- `NO_PROXY=host.docker.internal` — tells the Anthropic SDK's HTTP client to skip the proxy for that hostname
+- `no_proxy=host.docker.internal` — lowercase variant for tools that check the lowercase form
+
+Both are set in the agent group's `container.json` alongside `ANTHROPIC_BASE_URL`.
+
+## Network Isolation
+
+Setting `ANTHROPIC_BASE_URL` redirects requests but doesn't prevent a misconfigured agent from accidentally reaching `api.anthropic.com` directly. The `blockedHosts` field in `container.json` adds a Docker `--add-host` flag that resolves the domain to `0.0.0.0`, making it physically unreachable from inside the container:
+
+```json
+"blockedHosts": ["api.anthropic.com"]
+```
+
+With this in place, even if the model setting drifts back to a Claude model name, the API call will fail immediately rather than silently billing your account.
+
+## Model Selection
+
+The Claude Code CLI reads its model from `~/.claude/settings.json` inside the container, which NanoClaw bind-mounts from `data/v2-sessions/<agent-group-id>/.claude-shared/settings.json`. Set `"model": "gemma4:latest"` (or whatever Ollama model you've pulled) there. Use the exact name from `ollama list`.
+
+Model selection considerations for Apple Silicon:
+
+| Model | Size | Quality | Speed (M4 Pro) |
+|-------|------|---------|----------------|
+| `gemma4:latest` | 12B | Good general-purpose | Fast |
+| `qwen3-coder:latest` | 32B | Excellent for coding tasks | Moderate |
+| `llama3.2:latest` | 3B | Basic | Very fast |
+
+The agent uses tool calls extensively (read/write files, shell commands). Models that support tool use reliably work best. Gemma 4 and Qwen 3 Coder both handle structured tool calls well.
+
+## What Changes at the Code Level
+
+Three files need to support this feature. See `/add-ollama-provider` for the exact changes.
+
+**`src/container-config.ts`** — `ContainerConfig` interface needs `env` and `blockedHosts` fields so the per-group JSON can carry them.
+
+**`src/container-runner.ts`** — At container spawn time, `env` entries become `-e KEY=VAL` Docker flags (applied after OneCLI's injected vars so they win), and `blockedHosts` entries become `--add-host HOST:0.0.0.0` flags.
+
+**`container/Dockerfile`** — The container runs as the host user's uid (e.g. 501 on macOS), not as the `node` user (uid 1000). The home directory must be `chmod 777` so any uid can write `~/.claude.json` and `~/.claude/settings.json`.
+
+## Tradeoffs
+
+| | Ollama (local) | Anthropic API |
+|---|---|---|
+| Cost | Free | Pay-per-token |
+| Privacy | Fully local | Data sent to Anthropic |
+| Model quality | Good (open-weight) | Excellent (Claude) |
+| Cold start | 5–30s (model load) | ~1s |
+| Context window | Varies by model | 200k tokens (Sonnet) |
+| Tool use reliability | Good (large models) | Excellent |
+| Hardware req. | 16GB+ RAM | None |
+
+For personal automation on capable hardware, the tradeoff favors local. For complex multi-step tasks requiring large context or high reliability, Claude is still ahead.
+
+## Reverting to Claude
+
+Remove the `env` and `blockedHosts` keys from `groups/<folder>/container.json`, remove `"model"` from the shared settings file, and restart the service. No rebuild needed.
+
+## See Also
+
+- `/add-ollama-provider` — step-by-step skill to configure any agent group for Ollama
+- [Ollama Anthropic compatibility docs](https://ollama.com/blog/openai-compatibility) — upstream docs on the API bridge
+- `docs/architecture.md` — how the container spawn and env injection pipeline works
diff --git a/docs/setup-flow.md b/docs/setup-flow.md
new file mode 100644
index 0000000..800411c
--- /dev/null
+++ b/docs/setup-flow.md
@@ -0,0 +1,226 @@
+# Setup flow
+
+This document is the contract for NanoClaw's end-to-end scripted setup
+(`bash nanoclaw.sh` → `pnpm run setup:auto`). Read it before adding a new
+step, fixing a regression, or changing how output is rendered.
+
+## The three output levels
+
+Every setup step produces output at **three distinct levels**. They have
+different audiences, go to different places, and are formatted differently.
+Don't conflate them.
+
+| Level | Audience | Destination | Format |
+|---|---|---|---|
+| 1. User-facing | The operator running setup | Terminal (via clack) | Branded, concise, informational — "product content" |
+| 2. Progression | Future debuggers, AI agents reviewing a failed run, release support | `logs/setup.log` (one file, append-only) | Structured per-step blocks, linear chronology, human + machine readable |
+| 3. Raw | Whoever is deep-debugging a specific step | `logs/setup-steps/NN-step-name.log` (one file per step) | Full raw child stdout + stderr, verbatim |
+
+Think of it as: the user sees a **summary**, the progression log is an
+**index with key facts**, the raw logs are the **evidence**.
+
+### Level 1: user-facing (clack)
+
+Rendered by `setup/auto.ts` via `@clack/prompts`. This is our *product
+surface* for setup — every line should read as if we designed it for a
+stranger on day one.
+
+- Clack spinners for in-progress work. Show elapsed time.
+- `p.log.success` / `p.log.step` / `p.log.warn` for permanent status
+  markers.
+- `p.note` for multi-line information (pairing code, next steps).
+- `p.text` / `p.select` / `p.password` for prompts.
+- Brand palette: `brand()` / `brandBold()` / `brandChip()` helpers in
+  `setup/auto.ts`. Truecolor when the terminal supports it, 16-color
+  cyan fallback otherwise, plain text when piped / `NO_COLOR`.
+
+Rules:
+- **No discontinuity.** Every sub-step belongs to the same visual flow.
+  The only exception is Anthropic credential registration (see below).
+- **No raw child output.** Never `stdio: 'inherit'` a child whose output
+  wasn't written by us. Capture it and show it on failure only.
+- **No debug-style prefixes** (`[add-telegram] …`, `INFO …`, timestamps).
+  Those belong in levels 2 and 3.
+- **No emoji** unless the clack glyph requires it.
+
+### Level 2: progression log
+
+`logs/setup.log` — one file per setup run, append-only, cumulative across
+a multi-run install (if a run fails midway and is re-attempted, the new
+entries append). It's the thing you'd ask an operator to paste when they
+report a setup bug, and the thing an AI agent would read to understand
+what happened.
+
+Entry format:
+
+```
+=== [2026-04-22T22:14:12Z] bootstrap [45.1s] → success ===
+  platform: linux
+  is_wsl: false
+  node_version: 22.22.2
+  deps_ok: true
+  native_ok: true
+  raw: logs/setup-steps/01-bootstrap.log
+
+=== [2026-04-22T22:14:57Z] environment [2.3s] → success ===
+  docker: running
+  apple_container: not_found
+  raw: logs/setup-steps/02-environment.log
+
+=== [2026-04-22T22:15:00Z] container [92.4s] → success ===
+  runtime: docker
+  image: nanoclaw-agent:latest
+  build_ok: true
+  raw: logs/setup-steps/03-container.log
+```
+
+Design constraints:
+- Start-time timestamp (UTC, ISO-8601) on the opening line so a `grep`
+  gives you the sequence.
+- Duration in seconds with one decimal — fast steps read as "0.5s", not
+  "0ms".
+- Status is one of: `success`, `skipped`, `failed`, `aborted`.
+- Fields are step-specific but **must** be short scalar values. No JSON,
+  no multi-line. If a value is long, put it in the raw log and reference
+  it.
+- Always emit a `raw:` pointer, even on success — makes debugging the
+  second failure easier.
+- **User choices** are their own entries, not nested inside a step:
+
+  ```
+  === [2026-04-22T22:17:44Z] user-input → display_name ===
+    value: gav
+
+  === [2026-04-22T22:17:51Z] user-input → channel_choice ===
+    value: telegram
+  ```
+
+  These matter because the path through the setup flow depends on them.
+
+The log opens with a header block identifying the run, and closes with
+a completion block:
+
+```
+## 2026-04-22T22:14:12Z · setup:auto started
+  user: exedev
+  cwd: /home/exedev/nanoclaw
+  branch: branded-setup
+  commit: 6e0d742
+
+… (step entries) …
+
+## 2026-04-22T22:18:54Z · completed (total 4m42s)
+```
+
+On failure the completion block names the failing step and its error:
+
+```
+## 2026-04-22T22:16:40Z · aborted at container (err=cache_miss)
+```
+
+### Level 3: raw per-step logs
+
+`logs/setup-steps/NN-step-name.log` — one file per step, numbered in
+execution order (zero-padded 2-digit prefix for natural sorting). Full
+verbatim stdout + stderr from the child process. Truncated and rewritten
+on each run (not appended).
+
+Contents are whatever the step emits: apt output, docker build layers,
+pnpm install spam, `curl` bodies, etc. This is the evidence plane —
+"what did the shell actually see?" Nothing is filtered.
+
+## Contract for a new step
+
+When you add a step (either a TS step in `setup/<name>.ts` or a bash
+installer invoked from `auto.ts`), it must:
+
+1. **Receive a raw-log path** from the caller. Write all stdout + stderr
+   there. Don't write to the terminal directly.
+2. **Emit a single terminal status block** at the end, containing
+   `STATUS: success|skipped|failed` and any step-specific fields:
+
+   ```
+   === NANOCLAW SETUP: STEP_NAME ===
+   STATUS: success
+   KEY: value
+   KEY: value
+   === END ===
+   ```
+
+   Field names are `UPPER_SNAKE_CASE`. Values are short scalars.
+
+3. If it's a long-running step, optionally emit **sub-status blocks**
+   mid-stream. `auto.ts` parses them live and can render intermediate
+   UI (as `pair-telegram` does with `PAIR_TELEGRAM_CODE` /
+   `PAIR_TELEGRAM_ATTEMPT`).
+
+4. **Exit non-zero** on hard failure so `auto.ts` can distinguish
+   "step ran to completion and reported failed" from "step crashed".
+
+The driver handles the rest: spinner in level 1, structured append to
+level 2, raw capture to level 3.
+
+## The Anthropic exception
+
+Anthropic credential registration (`setup/register-claude-token.sh`) is
+the **one** permitted break in the visual flow. Why:
+
+- `claude setup-token` opens a browser, runs its own OAuth prompt, and
+  prints the token. It owns the TTY via `script(1)`.
+- We don't want to re-implement the OAuth device flow ourselves.
+- We don't want to intercept / mirror the token (it appears in the
+  user's terminal already — mirroring it adds attack surface).
+
+So during this step:
+- The clack flow explicitly pauses (a `p.log.step` marker says "this
+  part is interactive, you're handing off to Anthropic").
+- The child inherits stdio fully.
+- When control returns, clack resumes on the next line with a success
+  marker.
+
+The level-2 log still gets an entry (`auth [interactive] → success`
+with the method — subscription / oauth-token / api-key). Level-3 captures
+are optional here; mirroring `script -q` output is tricky and the risk of
+leaking the token to disk outweighs the debugging value.
+
+## File reference
+
+| File | Role |
+|---|---|
+| `nanoclaw.sh` | Top-level wrapper. Phase 1 (bootstrap) and phase 2 (setup:auto) orchestration. Writes bootstrap's raw log + progression entry. |
+| `setup.sh` | Phase 1 bootstrap: Node, pnpm, native-module verify. Emits its own `BOOTSTRAP` status block (historically printed to stdout; now goes to the bootstrap raw log). |
+| `setup/auto.ts` | Phase 2 driver. Orchestrates the clack UI, step execution, user prompts, and writes to all three log levels for every step it spawns. |
+| `setup/logs.ts` | The logging primitives (`logStep`, `logUserInput`, `logComplete`, `stepRawLog`, `initSetupLog`). Single source of truth for level 2/3 formatting and file paths. |
+| `setup/<step>.ts` | Individual step implementations. Must emit one terminal status block; must not write directly to the terminal. |
+| `setup/register-claude-token.sh` | The Anthropic exception. Inherits stdio, prints its own UI, returns a status to the driver. |
+| `setup/add-telegram.sh` | Non-interactive adapter installer. Reads `TELEGRAM_BOT_TOKEN` from env; never prompts. User-facing bits live in `auto.ts`. |
+| `setup/pair-telegram.ts` | Emits `PAIR_TELEGRAM_CODE` / `PAIR_TELEGRAM_ATTEMPT` / `PAIR_TELEGRAM` status blocks. Never prints UI. The driver renders it via clack notes. |
+
+## Common pitfalls
+
+- **Printing debug output from inside a step.** Tempting during
+  development; forbidden in checked-in code. All runtime messaging goes
+  through status blocks (level 2) or raw log writes (level 3).
+- **Adding a `console.log` that "just this once" goes to the terminal.**
+  It breaks the clack flow — the spinner line gets torn. Use
+  `log.info` / `log.error` from `src/log.ts` (writes to the raw log)
+  instead.
+- **`stdio: 'inherit'` for a non-exception child.** See Anthropic above.
+  Anything else needs `pipe` + explicit capture.
+- **Tee-ing to stderr.** Clack's spinner owns the terminal during a step.
+  Even stderr writes tear the frame. Pipe everything, then choose what
+  to surface.
+- **UTF-8 in bash `$VAR…` positions.** Bash's lexer can pull the first
+  byte of a multi-byte character into the variable name and trip
+  `set -u`. Always brace: `${VAR}…`.
+
+## Future work (not yet implemented)
+
+- **Progression log rotation.** Today's implementation truncates on each
+  run. Future: roll prior runs to `logs/setup.log.1`, `.2`, etc.
+- **Raw log rotation for multi-run installs.** Currently each run
+  overwrites. Fine for now; revisit if support needs to compare
+  successive attempts.
+- **Structured output from `register-claude-token.sh`.** The interactive
+  step emits no machine-readable status today. Future could add a
+  post-interaction status block with the method used.
diff --git a/nanoclaw.sh b/nanoclaw.sh
new file mode 100755
index 0000000..f8b58e7
--- /dev/null
+++ b/nanoclaw.sh
@@ -0,0 +1,262 @@
+#!/usr/bin/env bash
+#
+# NanoClaw — end-to-end setup entry point.
+#
+# Runs two parts from the user's perspective as one continuous flow:
+#   - bash-side: install the basics (Node + pnpm + native modules) under a
+#     bash-rendered clack-alike spinner. Can't use setup/auto.ts here since
+#     tsx isn't available until pnpm install completes.
+#   - hand off to `pnpm run setup:auto`, which renders the rest with
+#     @clack/prompts. The wordmark is printed once here so setup:auto can
+#     skip it and the flow reads as a single sequence.
+#
+# Obeys the three-level output contract (see docs/setup-flow.md):
+#   1. User-facing       — concise status line with elapsed time
+#   2. Progression log   — logs/setup.log (header + one entry per step)
+#   3. Raw per-step log  — logs/setup-steps/NN-name.log (full verbatim output)
+#
+# Config via env — passed through unchanged:
+#   NANOCLAW_SKIP  comma-separated setup:auto step names to skip
+#   SECRET_NAME    OneCLI secret name (default: Anthropic)
+#   HOST_PATTERN   OneCLI host pattern (default: api.anthropic.com)
+
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$PROJECT_ROOT"
+
+LOGS_DIR="$PROJECT_ROOT/logs"
+STEPS_DIR="$LOGS_DIR/setup-steps"
+PROGRESS_LOG="$LOGS_DIR/setup.log"
+
+# Diagnostics: persisted install-id + fire-and-forget emit. Sourced early
+# so `setup_launched` covers dropoff before bootstrap even starts.
+# shellcheck source=setup/lib/diagnostics.sh
+source "$PROJECT_ROOT/setup/lib/diagnostics.sh"
+ph_event setup_launched \
+  platform="$(uname -s | tr 'A-Z' 'a-z')" \
+  is_wsl="$([ -f /proc/version ] && grep -qi 'microsoft\|wsl' /proc/version 2>/dev/null && echo true || echo false)"
+
+# ─── log helpers ────────────────────────────────────────────────────────
+
+ts_utc() { date -u +%Y-%m-%dT%H:%M:%SZ; }
+
+write_header() {
+  local ts
+  ts=$(ts_utc)
+  local branch commit
+  branch=$(git branch --show-current 2>/dev/null || echo unknown)
+  commit=$(git rev-parse --short HEAD 2>/dev/null || echo unknown)
+  {
+    echo "## ${ts} · setup:auto started"
+    echo "  invocation: nanoclaw.sh"
+    echo "  user: $(whoami)"
+    echo "  cwd: ${PROJECT_ROOT}"
+    echo "  branch: ${branch}"
+    echo "  commit: ${commit}"
+    echo ""
+  } > "$PROGRESS_LOG"
+}
+
+# grep_field FIELD FILE — first value of FIELD: from a status block.
+grep_field() {
+  grep "^$1:" "$2" 2>/dev/null | head -1 | sed "s/^$1: *//" || true
+}
+
+write_bootstrap_entry() {
+  local status=$1 dur=$2 raw=$3
+  local ts
+  ts=$(ts_utc)
+  local platform is_wsl node_version deps_ok native_ok has_build_tools
+  platform=$(grep_field PLATFORM "$raw")
+  is_wsl=$(grep_field IS_WSL "$raw")
+  node_version=$(grep_field NODE_VERSION "$raw" | head -1)
+  deps_ok=$(grep_field DEPS_OK "$raw")
+  native_ok=$(grep_field NATIVE_OK "$raw")
+  has_build_tools=$(grep_field HAS_BUILD_TOOLS "$raw")
+  {
+    echo "=== [${ts}] bootstrap [${dur}s] → ${status} ==="
+    [ -n "$platform" ]        && echo "  platform: ${platform}"
+    [ -n "$is_wsl" ]          && echo "  is_wsl: ${is_wsl}"
+    [ -n "$node_version" ]    && echo "  node_version: ${node_version}"
+    [ -n "$deps_ok" ]         && echo "  deps_ok: ${deps_ok}"
+    [ -n "$native_ok" ]       && echo "  native_ok: ${native_ok}"
+    [ -n "$has_build_tools" ] && echo "  has_build_tools: ${has_build_tools}"
+    # Emit the raw path relative to PROJECT_ROOT so the progression log
+     # is portable and matches the TS-side format (logs/setup-steps/NN-…).
+    echo "  raw: ${raw#${PROJECT_ROOT}/}"
+    echo ""
+  } >> "$PROGRESS_LOG"
+}
+
+write_abort_entry() {
+  local step=$1 error=$2
+  local ts
+  ts=$(ts_utc)
+  echo "## ${ts} · aborted at ${step} (${error})" >> "$PROGRESS_LOG"
+}
+
+# ─── bash-side "clack-alike" status line ────────────────────────────────
+
+use_ansi() { [ -t 1 ] && [ -z "${NO_COLOR:-}" ]; }
+dim()     { use_ansi && printf '\033[2m%s\033[0m' "$1" || printf '%s' "$1"; }
+gray()    { use_ansi && printf '\033[90m%s\033[0m' "$1" || printf '%s' "$1"; }
+red()     { use_ansi && printf '\033[31m%s\033[0m' "$1" || printf '%s' "$1"; }
+bold()    { use_ansi && printf '\033[1m%s\033[0m' "$1" || printf '%s' "$1"; }
+# brand cyan (≈ #2BB7CE) — truecolor when supported, 16-color cyan fallback.
+brand_bold() {
+  if use_ansi; then
+    if [ "${COLORTERM:-}" = "truecolor" ] || [ "${COLORTERM:-}" = "24bit" ]; then
+      printf '\033[1;38;2;43;183;206m%s\033[0m' "$1"
+    else
+      printf '\033[1;36m%s\033[0m' "$1"
+    fi
+  else
+    printf '%s' "$1"
+  fi
+}
+clear_line() { use_ansi && printf '\r\033[2K' || printf '\n'; }
+
+spinner_start()   { printf '%s  %s…' "$(gray '◒')" "$1"; }
+spinner_update()  { clear_line; printf '%s  %s… %s' "$(gray '◒')" "$1" "$(dim "(${2}s)")"; }
+spinner_success() { clear_line; printf '%s  %s %s\n' "$(gray '◇')" "$1" "$(dim "(${2}s)")"; }
+spinner_failure() { clear_line; printf '%s  %s %s\n' "$(red '✗')"  "$1" "$(dim "(${2}s)")"; }
+
+# ─── fresh-run setup ────────────────────────────────────────────────────
+
+rm -rf "$STEPS_DIR"
+rm -f  "$PROGRESS_LOG"
+mkdir -p "$STEPS_DIR" "$LOGS_DIR"
+write_header
+
+# NanoClaw wordmark + subtitle — setup:auto will see NANOCLAW_BOOTSTRAPPED=1
+# and skip printing these again, so the flow stays visually continuous.
+printf '\n  %s%s\n' "$(bold 'Nano')" "$(brand_bold 'Claw')"
+printf '  %s\n\n' "$(dim 'Setting up your personal AI assistant')"
+
+# ─── pre-flight: Homebrew on macOS ─────────────────────────────────────
+# setup/install-node.sh and setup/install-docker.sh both require `brew` on
+# macOS. On a factory Mac there's no brew, and those helpers would fail
+# later inside the bootstrap spinner with a cryptic error. Prompt here,
+# before the spinner starts, so the user knows what's about to happen and
+# brew's own interactive sudo/CLT prompts stay readable.
+if [ "$(uname -s)" = "Darwin" ] && ! command -v brew >/dev/null 2>&1; then
+  printf '  %s\n' \
+    "$(dim "Homebrew isn't installed. NanoClaw uses it to install Node and Docker on your Mac.")"
+  printf '  %s\n\n' \
+    "$(dim "This also installs Apple's Command Line Tools, which can take 5-10 minutes.")"
+  read -r -p "  $(bold 'Install Homebrew now?') [Y/n] " BREW_ANS </dev/tty
+
+  case "${BREW_ANS:-Y}" in
+    [Yy]*|'')
+      printf '\n'
+      # Official installer. Runs interactively, triggers xcode-select --install
+      # for Command Line Tools, and prompts for the user's password for sudo.
+      # `|| true` so a user-cancelled install doesn't kill us via `set -e`;
+      # the PATH check below is the real gate.
+      /bin/bash -c \
+        "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" \
+        || true
+
+      # Put brew on PATH for this session (the installer writes to
+      # .zprofile/.bash_profile for future shells, but not this one).
+      if [ -x /opt/homebrew/bin/brew ]; then
+        eval "$(/opt/homebrew/bin/brew shellenv)"
+      elif [ -x /usr/local/bin/brew ]; then
+        eval "$(/usr/local/bin/brew shellenv)"
+      fi
+
+      if ! command -v brew >/dev/null 2>&1; then
+        printf '\n  %s %s\n' "$(red '✗')" "Homebrew install didn't complete."
+        printf '  %s\n\n' \
+          "$(dim 'Install manually from https://brew.sh and re-run: bash nanoclaw.sh')"
+        exit 1
+      fi
+      printf '\n'
+      ;;
+    *)
+      printf '\n  %s\n\n' \
+        "$(dim 'NanoClaw needs Homebrew. Install it from https://brew.sh and re-run.')"
+      exit 1
+      ;;
+  esac
+fi
+
+# ─── first step: install the basics (Node + pnpm + native modules) ─────
+
+BOOTSTRAP_RAW="${STEPS_DIR}/01-bootstrap.log"
+BOOTSTRAP_LABEL="Installing the basics"
+BOOTSTRAP_START=$(date +%s)
+
+# One-line "why" that teaches a differentiator while the user waits.
+printf '%s  %s\n' "$(gray '│')" \
+  "$(dim "NanoClaw is small and runs entirely on your machine. Yours to modify.")"
+spinner_start "$BOOTSTRAP_LABEL"
+
+# Run in the background so we can tick elapsed time. Capture exit code via
+# a tmpfile (subshell $? is lost after the while loop finishes).
+BOOTSTRAP_EXIT_FILE=$(mktemp -t nanoclaw-bootstrap-exit.XXXXXX)
+(
+  # setup.sh's legacy `log()` writes to a file; point it at the raw log
+  # so its verbose entries land alongside the stdout we're capturing.
+  export NANOCLAW_BOOTSTRAP_LOG="$BOOTSTRAP_RAW"
+  if bash setup.sh > "$BOOTSTRAP_RAW" 2>&1; then
+    echo 0 > "$BOOTSTRAP_EXIT_FILE"
+  else
+    echo $? > "$BOOTSTRAP_EXIT_FILE"
+  fi
+) &
+BOOTSTRAP_PID=$!
+
+while kill -0 "$BOOTSTRAP_PID" 2>/dev/null; do
+  sleep 1
+  if kill -0 "$BOOTSTRAP_PID" 2>/dev/null; then
+    spinner_update "$BOOTSTRAP_LABEL" "$(( $(date +%s) - BOOTSTRAP_START ))"
+  fi
+done
+# `wait` surfaces the child's exit code; we've already captured it.
+wait "$BOOTSTRAP_PID" 2>/dev/null || true
+
+BOOTSTRAP_RC=$(cat "$BOOTSTRAP_EXIT_FILE")
+rm -f "$BOOTSTRAP_EXIT_FILE"
+BOOTSTRAP_DUR=$(( $(date +%s) - BOOTSTRAP_START ))
+
+if [ "$BOOTSTRAP_RC" -eq 0 ]; then
+  spinner_success "Basics installed" "$BOOTSTRAP_DUR"
+  write_bootstrap_entry success "$BOOTSTRAP_DUR" "$BOOTSTRAP_RAW"
+else
+  spinner_failure "Couldn't install the basics" "$BOOTSTRAP_DUR"
+  write_bootstrap_entry failed "$BOOTSTRAP_DUR" "$BOOTSTRAP_RAW"
+  write_abort_entry bootstrap "exit-${BOOTSTRAP_RC}"
+
+  echo
+  echo "$(dim '── last 40 lines of ')$(dim "$BOOTSTRAP_RAW")$(dim ' ──')"
+  tail -40 "$BOOTSTRAP_RAW"
+  echo
+  echo "$(dim "Full raw log: $BOOTSTRAP_RAW")"
+  echo "$(dim "Progression:  $PROGRESS_LOG")"
+  exit 1
+fi
+
+# ─── hand off to setup:auto ────────────────────────────────────────────
+
+# NANOCLAW_BOOTSTRAPPED=1 tells setup/auto.ts to skip the wordmark (we
+# already printed it) and to append to the progression log rather than
+# wipe it.
+export NANOCLAW_BOOTSTRAPPED=1
+
+# setup.sh may have just installed pnpm via npm into a prefix that's not on
+# our PATH (custom `npm config set prefix`, or the default prefix missing
+# from the shell's login PATH). Its PATH mutation doesn't propagate back
+# to us — so replay the same lookup here before the exec.
+if ! command -v pnpm >/dev/null 2>&1 && command -v npm >/dev/null 2>&1; then
+  NPM_PREFIX="$(npm config get prefix 2>/dev/null)"
+  if [ -n "$NPM_PREFIX" ] && [ -x "$NPM_PREFIX/bin/pnpm" ]; then
+    export PATH="$NPM_PREFIX/bin:$PATH"
+  fi
+fi
+
+# --silent suppresses pnpm's `> nanoclaw@2.0.0 setup:auto / > tsx setup/auto.ts`
+# preamble so the flow continues visually from "Basics installed" straight
+# into setup:auto's spinner. exec so signals (Ctrl-C) propagate directly.
+exec pnpm --silent run setup:auto
diff --git a/package.json b/package.json
index e2af027..6029e0b 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "nanoclaw",
-  "version": "1.2.52",
+  "version": "2.0.13",
   "description": "Personal Claude assistant. Lightweight, secure, customizable.",
   "type": "module",
   "packageManager": "pnpm@10.33.0",
@@ -15,6 +15,7 @@
     "format:check": "prettier --check \"src/**/*.ts\"",
     "prepare": "husky",
     "setup": "tsx setup/index.ts",
+    "setup:auto": "tsx setup/auto.ts",
     "chat": "tsx scripts/chat.ts",
     "auth": "tsx src/whatsapp-auth.ts",
     "lint": "eslint src/",
@@ -23,10 +24,13 @@
     "test:watch": "vitest"
   },
   "dependencies": {
+    "@clack/core": "^1.2.0",
+    "@clack/prompts": "^1.2.0",
     "@onecli-sh/sdk": "^0.3.1",
     "better-sqlite3": "11.10.0",
     "chat": "^4.24.0",
-    "cron-parser": "5.5.0"
+    "cron-parser": "5.5.0",
+    "kleur": "^4.1.5"
   },
   "devDependencies": {
     "@eslint/js": "^9.35.0",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index c1aa197..3f74033 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -8,6 +8,12 @@ importers:
 
   .:
     dependencies:
+      '@clack/core':
+        specifier: ^1.2.0
+        version: 1.2.0
+      '@clack/prompts':
+        specifier: ^1.2.0
+        version: 1.2.0
       '@onecli-sh/sdk':
         specifier: ^0.3.1
         version: 0.3.1
@@ -20,6 +26,9 @@ importers:
       cron-parser:
         specifier: 5.5.0
         version: 5.5.0
+      kleur:
+        specifier: ^4.1.5
+        version: 4.1.5
     devDependencies:
       '@eslint/js':
         specifier: ^9.35.0
@@ -60,6 +69,12 @@ importers:
 
 packages:
 
+  '@clack/core@1.2.0':
+    resolution: {integrity: sha512-qfxof/3T3t9DPU/Rj3OmcFyZInceqj/NVtO9rwIuJqCUgh32gwPjpFQQp/ben07qKlhpwq7GzfWpST4qdJ5Drg==}
+
+  '@clack/prompts@1.2.0':
+    resolution: {integrity: sha512-4jmztR9fMqPMjz6H/UZXj0zEmE43ha1euENwkckKKel4XpSfokExPo5AiVStdHSAlHekz4d0CA/r45Ok1E4D3w==}
+
   '@emnapi/core@1.9.2':
     resolution: {integrity: sha512-UC+ZhH3XtczQYfOlu3lNEkdW/p4dsJ1r/bP7H8+rhao3TTTMO1ATq/4DdIi23XuGoFY+Cz0JmCbdVl0hz9jZcA==}
 
@@ -748,6 +763,15 @@ packages:
   fast-levenshtein@2.0.6:
     resolution: {integrity: sha512-DCXu6Ifhqcks7TZKY3Hxp3y6qphY5SJZmrWMDrKcERSOXWQdMhU9Ig/PYrzyw/ul9jOIyh0N4M0tbC5hodg8dw==}
 
+  fast-string-truncated-width@1.2.1:
+    resolution: {integrity: sha512-Q9acT/+Uu3GwGj+5w/zsGuQjh9O1TyywhIwAxHudtWrgF09nHOPrvTLhQevPbttcxjr/SNN7mJmfOw/B1bXgow==}
+
+  fast-string-width@1.1.0:
+    resolution: {integrity: sha512-O3fwIVIH5gKB38QNbdg+3760ZmGz0SZMgvwJbA1b2TGXceKE6A2cOlfogh1iw8lr049zPyd7YADHy+B7U4W9bQ==}
+
+  fast-wrap-ansi@0.1.6:
+    resolution: {integrity: sha512-HlUwET7a5gqjURj70D5jl7aC3Zmy4weA1SHUfM0JFI0Ptq987NH2TwbBFLoERhfwk+E+eaq4EK3jXoT+R3yp3w==}
+
   fdir@6.5.0:
     resolution: {integrity: sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg==}
     engines: {node: '>=12.0.0'}
@@ -866,6 +890,10 @@ packages:
   keyv@4.5.4:
     resolution: {integrity: sha512-oxVHkHR/EJf2CNXnWxRLW6mg7JyCCUcG0DtEGmL2ctUo1PNTin1PUil+r/+4r5MpVgC/fn1kjsx7mjSujKqIpw==}
 
+  kleur@4.1.5:
+    resolution: {integrity: sha512-o+NO+8WrRiQEE4/7nwRJhN1HWpVmJm511pBHUxPLtp0BUISzlBplORYSmTclCnJvQq2tKu/sgl3xVpkc7ZWuQQ==}
+    engines: {node: '>=6'}
+
   levn@0.4.1:
     resolution: {integrity: sha512-+bT2uH4E5LGE7h/n3evcS/sQlJXCpIp6ym8OWJ5eV6+67Dsql/LaaT7qJBAt2rzfoa/5QBGBhxDix1dMt2kQKQ==}
     engines: {node: '>= 0.8.0'}
@@ -1239,6 +1267,9 @@ packages:
   simple-get@4.0.1:
     resolution: {integrity: sha512-brv7p5WgH0jmQJr1ZDDfKDOSeWWg+OVypG99A/5vYGPqJ6pxiaHLy8nxtFjBA7oMa01ebA9gfh1uMCFqOuXxvA==}
 
+  sisteransi@1.0.5:
+    resolution: {integrity: sha512-bLGGlR1QxBcynn2d5YmDX4MGjlZvy2MRBDRNHLJ8VI6l6+9FUiyTFNJ0IveOSP0bcXgVDPRcfGqA0pjaqUpfVg==}
+
   source-map-js@1.2.1:
     resolution: {integrity: sha512-UXWMKhLOwVKb728IUtQPXxfYU+usdybtUrK/8uGE8CQMvrhOpwvzDBwj0QhSL7MQc7vIsISBG8VQ8+IDQxpfQA==}
     engines: {node: '>=0.10.0'}
@@ -1462,6 +1493,18 @@ packages:
 
 snapshots:
 
+  '@clack/core@1.2.0':
+    dependencies:
+      fast-wrap-ansi: 0.1.6
+      sisteransi: 1.0.5
+
+  '@clack/prompts@1.2.0':
+    dependencies:
+      '@clack/core': 1.2.0
+      fast-string-width: 1.1.0
+      fast-wrap-ansi: 0.1.6
+      sisteransi: 1.0.5
+
   '@emnapi/core@1.9.2':
     dependencies:
       '@emnapi/wasi-threads': 1.2.1
@@ -2105,6 +2148,16 @@ snapshots:
 
   fast-levenshtein@2.0.6: {}
 
+  fast-string-truncated-width@1.2.1: {}
+
+  fast-string-width@1.1.0:
+    dependencies:
+      fast-string-truncated-width: 1.2.1
+
+  fast-wrap-ansi@0.1.6:
+    dependencies:
+      fast-string-width: 1.1.0
+
   fdir@6.5.0(picomatch@4.0.4):
     optionalDependencies:
       picomatch: 4.0.4
@@ -2191,6 +2244,8 @@ snapshots:
     dependencies:
       json-buffer: 3.0.1
 
+  kleur@4.1.5: {}
+
   levn@0.4.1:
     dependencies:
       prelude-ls: 1.2.1
@@ -2735,6 +2790,8 @@ snapshots:
       once: 1.4.0
       simple-concat: 1.0.1
 
+  sisteransi@1.0.5: {}
+
   source-map-js@1.2.1: {}
 
   stackback@0.0.2: {}
diff --git a/repo-tokens/badge.svg b/repo-tokens/badge.svg
index 52e70b4..0dfb9a2 100644
--- a/repo-tokens/badge.svg
+++ b/repo-tokens/badge.svg
@@ -1,22 +1,22 @@
-<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="97" height="20" role="img" aria-label="43.7k tokens, 22% of context window">
-  <title>43.7k tokens, 22% of context window</title>
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="90" height="20" role="img" aria-label="132k tokens, 66% of context window">
+  <title>132k tokens, 66% of context window</title>
   <linearGradient id="s" x2="0" y2="100%">
     <stop offset="0" stop-color="#bbb" stop-opacity=".1"/>
     <stop offset="1" stop-opacity=".1"/>
   </linearGradient>
   <clipPath id="r">
-    <rect width="97" height="20" rx="3" fill="#fff"/>
+    <rect width="90" height="20" rx="3" fill="#fff"/>
   </clipPath>
   <a xlink:href="https://github.com/qwibitai/nanoclaw/tree/main/repo-tokens">
     <g clip-path="url(#r)">
       <rect width="52" height="20" fill="#555"/>
-      <rect x="52" width="45" height="20" fill="#4c1"/>
-      <rect width="97" height="20" fill="url(#s)"/>
+      <rect x="52" width="38" height="20" fill="#dfb317"/>
+      <rect width="90" height="20" fill="url(#s)"/>
       <g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" font-size="11">
         <text aria-hidden="true" x="26" y="15" fill="#010101" fill-opacity=".3">tokens</text>
         <text x="26" y="14">tokens</text>
-        <text aria-hidden="true" x="74" y="15" fill="#010101" fill-opacity=".3">43.7k</text>
-        <text x="74" y="14">43.7k</text>
+        <text aria-hidden="true" x="71" y="15" fill="#010101" fill-opacity=".3">132k</text>
+        <text x="71" y="14">132k</text>
       </g>
     </g>
   </a>
diff --git a/scripts/init-cli-agent.ts b/scripts/init-cli-agent.ts
new file mode 100644
index 0000000..4a56827
--- /dev/null
+++ b/scripts/init-cli-agent.ts
@@ -0,0 +1,170 @@
+/**
+ * Initialize the scratch CLI agent used during `/new-setup`.
+ *
+ * Creates the synthetic `cli:local` user, grants owner role if no owner
+ * exists yet, builds an agent group with a minimal CLAUDE.md, and wires it
+ * to the CLI messaging group so `pnpm run chat` works immediately.
+ *
+ * No welcome is staged — the operator's first `pnpm run chat` is the
+ * natural wake, and the agent introduces itself on first contact per its
+ * CLAUDE.md.
+ *
+ * Runs alongside the service (WAL-mode sqlite) — does NOT initialize
+ * channel adapters, so there's no Gateway conflict.
+ *
+ * Usage:
+ *   pnpm exec tsx scripts/init-cli-agent.ts \
+ *     --display-name "Gavriel" \
+ *     [--agent-name "Andy"]
+ */
+import path from 'path';
+
+import { DATA_DIR } from '../src/config.js';
+import { createAgentGroup, getAgentGroupByFolder } from '../src/db/agent-groups.js';
+import { initDb } from '../src/db/connection.js';
+import {
+  createMessagingGroup,
+  createMessagingGroupAgent,
+  getMessagingGroupAgentByPair,
+  getMessagingGroupByPlatform,
+} from '../src/db/messaging-groups.js';
+import { runMigrations } from '../src/db/migrations/index.js';
+import { normalizeName } from '../src/modules/agent-to-agent/db/agent-destinations.js';
+import { upsertUser } from '../src/modules/permissions/db/users.js';
+import { initGroupFilesystem } from '../src/group-init.js';
+import type { AgentGroup, MessagingGroup } from '../src/types.js';
+
+const CLI_CHANNEL = 'cli';
+const CLI_PLATFORM_ID = 'local';
+const CLI_SYNTHETIC_USER_ID = `${CLI_CHANNEL}:${CLI_PLATFORM_ID}`;
+
+interface Args {
+  displayName: string;
+  agentName: string;
+}
+
+function parseArgs(argv: string[]): Args {
+  let displayName: string | undefined;
+  let agentName: string | undefined;
+  for (let i = 0; i < argv.length; i++) {
+    const key = argv[i];
+    const val = argv[i + 1];
+    if (key === '--display-name') {
+      displayName = val;
+      i++;
+    } else if (key === '--agent-name') {
+      agentName = val;
+      i++;
+    }
+  }
+
+  if (!displayName) {
+    console.error('Missing required arg: --display-name');
+    console.error('See scripts/init-cli-agent.ts header for usage.');
+    process.exit(2);
+  }
+
+  return {
+    displayName,
+    agentName: agentName?.trim() || displayName,
+  };
+}
+
+function generateId(prefix: string): string {
+  return `${prefix}-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+}
+
+async function main(): Promise<void> {
+  const args = parseArgs(process.argv.slice(2));
+
+  const db = initDb(path.join(DATA_DIR, 'v2.db'));
+  runMigrations(db);
+
+  const now = new Date().toISOString();
+
+  // 1. Synthetic CLI user + owner grant if none exists.
+  upsertUser({
+    id: CLI_SYNTHETIC_USER_ID,
+    kind: CLI_CHANNEL,
+    display_name: args.displayName,
+    created_at: now,
+  });
+
+  // Owner grant deferred to init-first-agent when the real channel user is
+  // wired — cli:local is a scratch identity, not the operator.
+  const promotedToOwner = false;
+
+  // 2. Agent group + filesystem.
+  const folder = `cli-with-${normalizeName(args.displayName)}`;
+  let ag: AgentGroup | undefined = getAgentGroupByFolder(folder);
+  if (!ag) {
+    const agId = generateId('ag');
+    createAgentGroup({
+      id: agId,
+      name: args.agentName,
+      folder,
+      agent_provider: null,
+      created_at: now,
+    });
+    ag = getAgentGroupByFolder(folder)!;
+    console.log(`Created agent group: ${ag.id} (${folder})`);
+  } else {
+    console.log(`Reusing agent group: ${ag.id} (${folder})`);
+  }
+  initGroupFilesystem(ag, {
+    instructions:
+      `# ${args.agentName}\n\n` +
+      `You are ${args.agentName}, a personal NanoClaw agent for ${args.displayName}. ` +
+      'When the user first reaches out, introduce yourself briefly and invite them to chat. Keep replies concise.',
+  });
+
+  // 3. CLI messaging group + wiring.
+  let cliMg: MessagingGroup | undefined = getMessagingGroupByPlatform(CLI_CHANNEL, CLI_PLATFORM_ID);
+  if (!cliMg) {
+    cliMg = {
+      id: generateId('mg'),
+      channel_type: CLI_CHANNEL,
+      platform_id: CLI_PLATFORM_ID,
+      name: 'Local CLI',
+      is_group: 0,
+      unknown_sender_policy: 'public',
+      created_at: now,
+    };
+    createMessagingGroup(cliMg);
+    console.log(`Created CLI messaging group: ${cliMg.id}`);
+  }
+
+  const existing = getMessagingGroupAgentByPair(cliMg.id, ag.id);
+  if (!existing) {
+    createMessagingGroupAgent({
+      id: generateId('mga'),
+      messaging_group_id: cliMg.id,
+      agent_group_id: ag.id,
+      engage_mode: 'pattern',
+      engage_pattern: '.',
+      sender_scope: 'all',
+      ignored_message_policy: 'drop',
+      session_mode: 'shared',
+      priority: 0,
+      created_at: now,
+    });
+    console.log(`Wired cli: ${cliMg.id} -> ${ag.id}`);
+  } else {
+    console.log(`Wiring already exists: ${existing.id}`);
+  }
+
+  console.log('');
+  console.log('Init complete.');
+  console.log(
+    `  owner:   ${CLI_SYNTHETIC_USER_ID}${promotedToOwner ? ' (promoted on first owner)' : ''}`,
+  );
+  console.log(`  agent:   ${ag.name} [${ag.id}] @ groups/${folder}`);
+  console.log(`  channel: cli/${CLI_PLATFORM_ID}`);
+  console.log('');
+  console.log('Run `pnpm run chat hi` to talk to your agent.');
+}
+
+main().catch((err) => {
+  console.error(err);
+  process.exit(1);
+});
diff --git a/scripts/init-first-agent.ts b/scripts/init-first-agent.ts
index c40f07f..61a17d6 100644
--- a/scripts/init-first-agent.ts
+++ b/scripts/init-first-agent.ts
@@ -1,43 +1,36 @@
 /**
- * Init the first (or Nth) NanoClaw v2 agent.
+ * Init the first (or Nth) NanoClaw v2 agent for a DM channel.
  *
- * Two modes:
+ * Wires a real DM channel (discord, telegram, etc.) to a new agent group,
+ * then hands a welcome message to the running service via the CLI socket
+ * (admin transport). The service routes that message into the DM session,
+ * which wakes the container synchronously — the agent processes the welcome
+ * and DMs the operator through the normal delivery path.
  *
- * 1. **DM channel mode** (default): wires a real DM channel (discord, telegram,
- *    etc.) + the CLI channel to the same agent, stages a welcome into the DM
- *    session so the agent greets the operator over that channel.
- *
- * 2. **CLI-only mode** (`--cli-only`): wires only the CLI channel. Used by
- *    `/new-setup` to get to a working 2-way CLI chat with the bare minimum.
- *    Owner grant uses a synthetic `cli:local` user so admin-gated flows work.
+ * CLI channel wiring is handled separately by `scripts/init-cli-agent.ts`.
  *
  * Creates/reuses: user, owner grant (if none), agent group + filesystem,
- * messaging group(s), wiring, session. Stages a system welcome message so
- * the host sweep wakes the container and the agent sends the greeting via
- * the normal delivery path.
+ * messaging group(s), wiring.
  *
- * Runs alongside the service (WAL-mode sqlite) — does NOT initialize
- * channel adapters, so there's no Gateway conflict.
+ * Runs alongside the service (WAL-mode sqlite + CLI socket IPC) — does NOT
+ * initialize channel adapters, so there's no Gateway conflict. Requires
+ * the service to be running: the welcome hand-off goes over the CLI socket
+ * and fails loudly if the service isn't up.
  *
  * Usage:
- *   # DM mode
  *   pnpm exec tsx scripts/init-first-agent.ts \
  *     --channel discord \
  *     --user-id discord:1470183333427675709 \
  *     --platform-id discord:@me:1491573333382523708 \
  *     --display-name "Gavriel" \
  *     [--agent-name "Andy"] \
- *     [--welcome "System instruction: ..."]
- *
- *   # CLI-only mode
- *   pnpm exec tsx scripts/init-first-agent.ts --cli-only \
- *     --display-name "Gavriel" \
- *     [--agent-name "Andy"] \
- *     [--welcome "System instruction: ..."]
+ *     [--welcome "System instruction: ..."] \
+ *     [--role owner|admin|member]    # default: owner
  *
  * For direct-addressable channels (telegram, whatsapp, etc.), --platform-id
  * is typically the same as the handle in --user-id, with the channel prefix.
  */
+import net from 'net';
 import path from 'path';
 
 import { DATA_DIR } from '../src/config.js';
@@ -51,38 +44,36 @@ import {
 } from '../src/db/messaging-groups.js';
 import { runMigrations } from '../src/db/migrations/index.js';
 import { normalizeName } from '../src/modules/agent-to-agent/db/agent-destinations.js';
-import { grantRole, hasAnyOwner } from '../src/modules/permissions/db/user-roles.js';
+import { addMember } from '../src/modules/permissions/db/agent-group-members.js';
+import { getUserRoles, grantRole } from '../src/modules/permissions/db/user-roles.js';
 import { upsertUser } from '../src/modules/permissions/db/users.js';
 import { initGroupFilesystem } from '../src/group-init.js';
-import { resolveSession, writeSessionMessage } from '../src/session-manager.js';
+import { namespacedPlatformId } from '../src/platform-id.js';
 import type { AgentGroup, MessagingGroup } from '../src/types.js';
 
+type Role = 'owner' | 'admin' | 'member';
+
 interface Args {
-  cliOnly: boolean;
   channel: string;
   userId: string;
   platformId: string;
   displayName: string;
   agentName: string;
   welcome: string;
+  role: Role;
 }
 
 const DEFAULT_WELCOME =
   'System instruction: run /welcome to introduce yourself to the user on this new channel.';
 
-const CLI_CHANNEL = 'cli';
-const CLI_PLATFORM_ID = 'local';
-const CLI_SYNTHETIC_USER_ID = `${CLI_CHANNEL}:${CLI_PLATFORM_ID}`;
+const DEFAULT_ROLE: Role = 'owner';
 
 function parseArgs(argv: string[]): Args {
-  const out: Partial<Args> = { cliOnly: false };
+  const out: Partial<Args> = {};
   for (let i = 0; i < argv.length; i++) {
     const key = argv[i];
     const val = argv[i + 1];
     switch (key) {
-      case '--cli-only':
-        out.cliOnly = true;
-        break;
       case '--channel':
         out.channel = (val ?? '').toLowerCase();
         i++;
@@ -107,44 +98,39 @@ function parseArgs(argv: string[]): Args {
         out.welcome = val;
         i++;
         break;
+      case '--role': {
+        const raw = (val ?? '').toLowerCase();
+        if (raw !== 'owner' && raw !== 'admin' && raw !== 'member') {
+          console.error(
+            `Invalid --role: ${raw} (expected 'owner', 'admin', or 'member')`,
+          );
+          process.exit(2);
+        }
+        out.role = raw;
+        i++;
+        break;
+      }
     }
   }
 
-  if (!out.displayName) {
-    console.error('Missing required arg: --display-name');
-    console.error('See scripts/init-first-agent.ts header for usage.');
-    process.exit(2);
-  }
-
-  if (out.cliOnly) {
-    // CLI-only: channel/user/platform default to the synthetic local CLI identity.
-    return {
-      cliOnly: true,
-      channel: CLI_CHANNEL,
-      userId: CLI_SYNTHETIC_USER_ID,
-      platformId: CLI_PLATFORM_ID,
-      displayName: out.displayName,
-      agentName: out.agentName?.trim() || out.displayName,
-      welcome: out.welcome?.trim() || DEFAULT_WELCOME,
-    };
-  }
-
-  const required: (keyof Args)[] = ['channel', 'userId', 'platformId'];
+  const required: (keyof Args)[] = ['channel', 'userId', 'platformId', 'displayName'];
   const missing = required.filter((k) => !out[k]);
   if (missing.length) {
-    console.error(`Missing required args: ${missing.map((k) => `--${k.replace(/([A-Z])/g, '-$1').toLowerCase()}`).join(', ')}`);
+    console.error(
+      `Missing required args: ${missing.map((k) => `--${k.replace(/([A-Z])/g, '-$1').toLowerCase()}`).join(', ')}`,
+    );
     console.error('See scripts/init-first-agent.ts header for usage.');
     process.exit(2);
   }
 
   return {
-    cliOnly: false,
     channel: out.channel!,
     userId: out.userId!,
     platformId: out.platformId!,
-    displayName: out.displayName,
-    agentName: out.agentName?.trim() || out.displayName,
+    displayName: out.displayName!,
+    agentName: out.agentName?.trim() || out.displayName!,
     welcome: out.welcome?.trim() || DEFAULT_WELCOME,
+    role: out.role ?? DEFAULT_ROLE,
   };
 }
 
@@ -152,38 +138,11 @@ function namespacedUserId(channel: string, raw: string): string {
   return raw.includes(':') ? raw : `${channel}:${raw}`;
 }
 
-function namespacedPlatformId(channel: string, raw: string): string {
-  return raw.startsWith(`${channel}:`) ? raw : `${channel}:${raw}`;
-}
-
 function generateId(prefix: string): string {
   return `${prefix}-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
 }
 
-function ensureCliMessagingGroup(now: string): MessagingGroup {
-  let cliMg = getMessagingGroupByPlatform(CLI_CHANNEL, CLI_PLATFORM_ID);
-  if (cliMg) return cliMg;
-
-  cliMg = {
-    id: generateId('mg'),
-    channel_type: CLI_CHANNEL,
-    platform_id: CLI_PLATFORM_ID,
-    name: 'Local CLI',
-    is_group: 0,
-    unknown_sender_policy: 'public',
-    created_at: now,
-  };
-  createMessagingGroup(cliMg);
-  console.log(`Created CLI messaging group: ${cliMg.id}`);
-  return cliMg;
-}
-
-function wireIfMissing(
-  mg: MessagingGroup,
-  ag: AgentGroup,
-  now: string,
-  label: string,
-): void {
+function wireIfMissing(mg: MessagingGroup, ag: AgentGroup, now: string, label: string): void {
   const existing = getMessagingGroupAgentByPair(mg.id, ag.id);
   if (existing) {
     console.log(`Wiring already exists: ${existing.id} (${label})`);
@@ -193,8 +152,13 @@ function wireIfMissing(
     id: generateId('mga'),
     messaging_group_id: mg.id,
     agent_group_id: ag.id,
-    trigger_rules: null,
-    response_scope: 'all',
+    // DM / CLI (is_group=0) default to "respond to everything" via a '.' regex.
+    // Group chats default to mention-only; admins can upgrade to mention-sticky
+    // via /manage-channels once the agent is in use.
+    engage_mode: mg.is_group === 0 ? 'pattern' : 'mention',
+    engage_pattern: mg.is_group === 0 ? '.' : null,
+    sender_scope: 'all',
+    ignored_message_policy: 'drop',
     session_mode: 'shared',
     priority: 0,
     created_at: now,
@@ -211,7 +175,6 @@ async function main(): Promise<void> {
   const now = new Date().toISOString();
 
   // 1. User + (conditional) owner grant.
-  // In cli-only mode, the synthetic `cli:local` user becomes the first owner.
   const userId = namespacedUserId(args.channel, args.userId);
   upsertUser({
     id: userId,
@@ -220,22 +183,11 @@ async function main(): Promise<void> {
     created_at: now,
   });
 
-  let promotedToOwner = false;
-  if (!hasAnyOwner()) {
-    grantRole({
-      user_id: userId,
-      role: 'owner',
-      agent_group_id: null,
-      granted_by: null,
-      granted_at: now,
-    });
-    promotedToOwner = true;
-  }
+  // Owner grant is deferred until after the agent group is resolved, since
+  // an admin grant is scoped to that group. See step 2b.
 
-  // 2. Agent group + filesystem
-  const folder = args.cliOnly
-    ? `cli-with-${normalizeName(args.displayName)}`
-    : `dm-with-${normalizeName(args.displayName)}`;
+  // 2. Agent group + filesystem.
+  const folder = `dm-with-${normalizeName(args.displayName)}`;
   let ag: AgentGroup | undefined = getAgentGroupByFolder(folder);
   if (!ag) {
     const agId = generateId('ag');
@@ -255,85 +207,170 @@ async function main(): Promise<void> {
     instructions:
       `# ${args.agentName}\n\n` +
       `You are ${args.agentName}, a personal NanoClaw agent for ${args.displayName}. ` +
-      'When you receive a system welcome prompt, introduce yourself briefly and invite them to chat. Keep replies concise.',
+      'When the user first reaches out (or you receive a system welcome prompt), introduce yourself briefly and invite them to chat. Keep replies concise.',
   });
 
-  // 3. Primary messaging group + wiring + welcome session.
-  // In DM mode: the DM messaging group is primary, CLI is wired as a bonus.
-  // In cli-only mode: the CLI messaging group is primary; no DM group.
-  const cliMg = ensureCliMessagingGroup(now);
-
-  let primaryMg: MessagingGroup;
-  if (args.cliOnly) {
-    primaryMg = cliMg;
-  } else {
-    const platformId = namespacedPlatformId(args.channel, args.platformId);
-    let dmMg = getMessagingGroupByPlatform(args.channel, platformId);
-    if (!dmMg) {
-      const mgId = generateId('mg');
-      createMessagingGroup({
-        id: mgId,
-        channel_type: args.channel,
-        platform_id: platformId,
-        name: args.displayName,
-        is_group: 0,
-        unknown_sender_policy: 'strict',
-        created_at: now,
+  // 2b. Assign the user a role for this agent group. The caller picks via
+  // --role; the channel drivers default to 'owner' for the self-host case.
+  //  - owner:  global owner (agent_group_id=null). Cross-channel access.
+  //  - admin:  scoped admin for this agent group only.
+  //  - member: no role grant, just the membership row below.
+  // grantRole inserts a new row per call — idempotence check against
+  // getUserRoles prevents duplicates on re-runs.
+  const existingRoles = getUserRoles(userId);
+  if (args.role === 'owner') {
+    const alreadyOwner = existingRoles.some(
+      (r) => r.role === 'owner' && r.agent_group_id === null,
+    );
+    if (!alreadyOwner) {
+      grantRole({
+        user_id: userId,
+        role: 'owner',
+        agent_group_id: null,
+        granted_by: null,
+        granted_at: now,
+      });
+    }
+  } else if (args.role === 'admin') {
+    const alreadyAdmin = existingRoles.some(
+      (r) => r.role === 'admin' && r.agent_group_id === ag.id,
+    );
+    if (!alreadyAdmin) {
+      grantRole({
+        user_id: userId,
+        role: 'admin',
+        agent_group_id: ag.id,
+        granted_by: null,
+        granted_at: now,
       });
-      dmMg = getMessagingGroupByPlatform(args.channel, platformId)!;
-      console.log(`Created messaging group: ${dmMg.id} (${platformId})`);
-    } else {
-      console.log(`Reusing messaging group: ${dmMg.id} (${platformId})`);
     }
-    primaryMg = dmMg;
   }
 
-  // Wire primary (DM or CLI), auto-creates companion agent_destinations row.
-  wireIfMissing(primaryMg, ag, now, args.cliOnly ? 'cli' : 'dm');
-
-  // In DM mode also wire CLI so `pnpm run chat` works immediately.
-  if (!args.cliOnly) {
-    wireIfMissing(cliMg, ag, now, 'cli-bonus');
-  }
-
-  // 4. Session + staged welcome (on the primary messaging group)
-  const { session, created } = resolveSession(ag.id, primaryMg.id, null, 'shared');
-  console.log(`${created ? 'Created' : 'Reusing'} session: ${session.id}`);
-
-  writeSessionMessage(ag.id, session.id, {
-    id: generateId('sys-welcome'),
-    kind: 'chat',
-    timestamp: now,
-    platformId: primaryMg.platform_id,
-    channelType: primaryMg.channel_type,
-    threadId: null,
-    content: JSON.stringify({
-      text: args.welcome,
-      sender: 'system',
-      senderId: 'system',
-    }),
+  // Always add a membership row so the access gate has a straightforward
+  // yes/no even for users without a role grant. INSERT OR IGNORE, so this
+  // is a no-op when the row already exists (e.g. re-runs, owners whose
+  // access already passes via role).
+  addMember({
+    user_id: userId,
+    agent_group_id: ag.id,
+    added_by: null,
+    added_at: now,
   });
 
+  // 3. DM messaging group.
+  const platformId = namespacedPlatformId(args.channel, args.platformId);
+  let dmMg = getMessagingGroupByPlatform(args.channel, platformId);
+  if (!dmMg) {
+    const mgId = generateId('mg');
+    createMessagingGroup({
+      id: mgId,
+      channel_type: args.channel,
+      platform_id: platformId,
+      name: args.displayName,
+      is_group: 0,
+      unknown_sender_policy: 'strict',
+      created_at: now,
+    });
+    dmMg = getMessagingGroupByPlatform(args.channel, platformId)!;
+    console.log(`Created messaging group: ${dmMg.id} (${platformId})`);
+  } else {
+    console.log(`Reusing messaging group: ${dmMg.id} (${platformId})`);
+  }
+
+  // 4. Wire DM messaging group to the agent.
+  wireIfMissing(dmMg, ag, now, 'dm');
+
+  // 5. Welcome delivery over the CLI socket. Router picks up the line,
+  // writes the message into the DM session's inbound.db, and wakes the
+  // container synchronously — no sweep wait. The paired user's identity is
+  // passed so the sender resolver sees the real owner, not cli:local.
+  await sendWelcomeViaCliSocket(dmMg, args.welcome, {
+    senderId: userId,
+    sender: args.displayName,
+  });
+
+  const roleLabel =
+    args.role === 'owner'
+      ? 'owner (global)'
+      : args.role === 'admin'
+        ? `admin (scoped to ${ag.id})`
+        : 'member';
+
   console.log('');
   console.log('Init complete.');
-  console.log(`  owner:   ${userId}${promotedToOwner ? ' (promoted on first owner)' : ''}`);
+  console.log(`  user:    ${userId}`);
+  console.log(`  role:    ${roleLabel}`);
   console.log(`  agent:   ${ag.name} [${ag.id}] @ groups/${folder}`);
-  if (args.cliOnly) {
-    console.log(`  channel: cli/${CLI_PLATFORM_ID}`);
-  } else {
-    console.log(`  channel: ${args.channel} ${primaryMg.platform_id}`);
-    console.log(`  cli:     cli/${CLI_PLATFORM_ID} wired — try \`pnpm run chat hi\``);
-  }
-  console.log(`  session: ${session.id}`);
+  console.log(`  channel: ${args.channel} ${dmMg.platform_id}`);
   console.log('');
-  console.log(
-    args.cliOnly
-      ? 'Host sweep (<=60s) will wake the container. Try `pnpm run chat hi`.'
-      : 'Host sweep (<=60s) will wake the container and the agent will send the welcome DM.',
-  );
+  console.log('Welcome DM queued — the agent will greet you shortly.');
+}
+
+/**
+ * Hand the welcome to the running service via its CLI Unix socket. The
+ * service's CLI adapter receives `{text, to}`, builds an InboundEvent
+ * targeting the DM messaging group, and calls routeInbound(). Router writes
+ * the message into inbound.db and wakes the container synchronously.
+ *
+ * Throws if the socket isn't reachable — this script requires the service
+ * to be running.
+ */
+async function sendWelcomeViaCliSocket(
+  dmMg: MessagingGroup,
+  welcome: string,
+  identity: { senderId: string; sender: string },
+): Promise<void> {
+  const sockPath = path.join(DATA_DIR, 'cli.sock');
+
+  await new Promise<void>((resolve, reject) => {
+    const socket = net.connect(sockPath);
+    let settled = false;
+
+    const settle = (err: Error | null) => {
+      if (settled) return;
+      settled = true;
+      try {
+        socket.end();
+      } catch {
+        /* noop */
+      }
+      if (err) reject(err);
+      else resolve();
+    };
+
+    socket.once('error', (err) =>
+      settle(
+        new Error(
+          `CLI socket at ${sockPath} not reachable: ${err.message}. Is the NanoClaw service running?`,
+        ),
+      ),
+    );
+    socket.once('connect', () => {
+      const payload =
+        JSON.stringify({
+          text: welcome,
+          senderId: identity.senderId,
+          sender: identity.sender,
+          to: {
+            channelType: dmMg.channel_type,
+            platformId: dmMg.platform_id,
+            threadId: dmMg.platform_id,
+          },
+        }) + '\n';
+      socket.write(payload, (err) => {
+        if (err) {
+          settle(err);
+          return;
+        }
+        // Brief flush delay so the router picks up the line before we close.
+        // Router handles it synchronously once read, so 50ms is plenty.
+        setTimeout(() => settle(null), 50);
+      });
+    });
+  });
 }
 
 main().catch((err) => {
-  console.error(err);
+  console.error(err instanceof Error ? err.message : err);
   process.exit(1);
 });
diff --git a/scripts/migrate-group-claude-md.ts b/scripts/migrate-group-claude-md.ts
deleted file mode 100644
index dd16faf..0000000
--- a/scripts/migrate-group-claude-md.ts
+++ /dev/null
@@ -1,113 +0,0 @@
-/**
- * One-shot migration: wire each existing group up to global memory via
- * an in-tree symlink + @-import.
- *
- * Claude Code's @-import only follows paths inside cwd, so a direct
- * `@/workspace/global/CLAUDE.md` or `@../global/CLAUDE.md` silently does
- * nothing (the import line is parsed but the target file is never
- * loaded into context). The working approach:
- *
- *   1. Symlink `groups/<folder>/.claude-global.md` →
- *      `/workspace/global/CLAUDE.md` (container path; dangling on host,
- *      valid inside the container via the /workspace/global mount).
- *   2. Have the group's CLAUDE.md import the symlink:
- *      `@./.claude-global.md`.
- *
- * This script:
- *   - Creates the symlink if missing.
- *   - Replaces any existing broken `@/workspace/global/CLAUDE.md` or
- *     `@../global/CLAUDE.md` import line with the symlink form.
- *   - Prepends the symlink import if neither form is present.
- *   - Skips entirely if `groups/global/CLAUDE.md` doesn't exist.
- *
- * Idempotent — safe to re-run.
- *
- * Usage: pnpm exec tsx scripts/migrate-group-claude-md.ts
- */
-import fs from 'fs';
-import path from 'path';
-
-import { GROUPS_DIR } from '../src/config.js';
-
-const GLOBAL_CLAUDE_MD = path.join(GROUPS_DIR, 'global', 'CLAUDE.md');
-const GLOBAL_MEMORY_CONTAINER_PATH = '/workspace/global/CLAUDE.md';
-const GLOBAL_MEMORY_LINK_NAME = '.claude-global.md';
-const IMPORT_LINE = `@./${GLOBAL_MEMORY_LINK_NAME}`;
-
-// Match any existing @-import that points at global/CLAUDE.md, whether
-// via absolute path, relative path, or the new symlink form.
-const EXISTING_IMPORT_REGEX =
-  /^@(?:\/workspace\/global\/CLAUDE\.md|\.\.\/global\/CLAUDE\.md|\.\/\.claude-global\.md)\s*$/m;
-
-if (!fs.existsSync(GLOBAL_CLAUDE_MD)) {
-  console.error(`No global CLAUDE.md at ${GLOBAL_CLAUDE_MD} — nothing to migrate.`);
-  process.exit(1);
-}
-
-if (!fs.existsSync(GROUPS_DIR)) {
-  console.error(`No groups dir at ${GROUPS_DIR} — nothing to migrate.`);
-  process.exit(1);
-}
-
-const entries = fs.readdirSync(GROUPS_DIR, { withFileTypes: true });
-let updated = 0;
-let alreadyWired = 0;
-let missingClaudeMd = 0;
-let symlinksCreated = 0;
-
-for (const entry of entries) {
-  if (!entry.isDirectory()) continue;
-  if (entry.name === 'global') continue;
-
-  const groupDir = path.join(GROUPS_DIR, entry.name);
-
-  // Symlink (idempotent — skip if already present)
-  const linkPath = path.join(groupDir, GLOBAL_MEMORY_LINK_NAME);
-  let linkExists = false;
-  try {
-    fs.lstatSync(linkPath);
-    linkExists = true;
-  } catch {
-    /* missing */
-  }
-  if (!linkExists) {
-    fs.symlinkSync(GLOBAL_MEMORY_CONTAINER_PATH, linkPath);
-    console.log(`[link]  ${entry.name}: created ${GLOBAL_MEMORY_LINK_NAME}`);
-    symlinksCreated++;
-  }
-
-  // CLAUDE.md import wiring
-  const claudeMd = path.join(groupDir, 'CLAUDE.md');
-  if (!fs.existsSync(claudeMd)) {
-    console.log(`[skip]  ${entry.name}: no CLAUDE.md`);
-    missingClaudeMd++;
-    continue;
-  }
-
-  const body = fs.readFileSync(claudeMd, 'utf-8');
-  const match = body.match(EXISTING_IMPORT_REGEX);
-
-  if (match && match[0] === IMPORT_LINE) {
-    console.log(`[wired] ${entry.name}: already imports ${IMPORT_LINE}`);
-    alreadyWired++;
-    continue;
-  }
-
-  let newBody: string;
-  if (match) {
-    // Replace the broken import with the working form
-    newBody = body.replace(EXISTING_IMPORT_REGEX, IMPORT_LINE);
-    console.log(`[fix]   ${entry.name}: rewrote ${match[0]} → ${IMPORT_LINE}`);
-  } else {
-    // Prepend fresh
-    newBody = `${IMPORT_LINE}\n\n${body}`;
-    console.log(`[ok]    ${entry.name}: prepended ${IMPORT_LINE}`);
-  }
-
-  fs.writeFileSync(claudeMd, newBody);
-  updated++;
-}
-
-console.log(
-  `\nDone. updated=${updated} alreadyWired=${alreadyWired} missingClaudeMd=${missingClaudeMd} symlinksCreated=${symlinksCreated}`,
-);
diff --git a/scripts/seed-discord.ts b/scripts/seed-discord.ts
index 9aed1c5..3ea24e8 100644
--- a/scripts/seed-discord.ts
+++ b/scripts/seed-discord.ts
@@ -58,8 +58,12 @@ try {
     id: 'mga-discord',
     messaging_group_id: MESSAGING_GROUP_ID,
     agent_group_id: AGENT_GROUP_ID,
-    trigger_rules: null,
-    response_scope: 'all',
+    // Discord group channel → mention-sticky default. Mention once, stay
+    // subscribed to the thread. Admins can tune via /manage-channels.
+    engage_mode: 'mention-sticky',
+    engage_pattern: null,
+    sender_scope: 'all',
+    ignored_message_policy: 'drop',
     session_mode: 'shared',
     priority: 0,
     created_at: new Date().toISOString(),
diff --git a/scripts/test-v2-channel-e2e.ts b/scripts/test-v2-channel-e2e.ts
index fc0a570..6721ff0 100644
--- a/scripts/test-v2-channel-e2e.ts
+++ b/scripts/test-v2-channel-e2e.ts
@@ -53,8 +53,10 @@ createMessagingGroupAgent({
   id: 'mga-chan',
   messaging_group_id: 'mg-chan',
   agent_group_id: 'ag-chan',
-  trigger_rules: null,
-  response_scope: 'all',
+  engage_mode: 'pattern',
+  engage_pattern: '.',
+  sender_scope: 'all',
+  ignored_message_policy: 'drop',
   session_mode: 'shared',
   priority: 0,
   created_at: new Date().toISOString(),
@@ -105,7 +107,15 @@ registerChannelAdapter('mock', { factory: () => mockAdapter });
 
 // Init channel adapters — this calls setup() with conversation configs from central DB
 await initChannelAdapters((adapter) => ({
-  conversations: [{ platformId: 'mock-channel-1', agentGroupId: 'ag-chan', requiresTrigger: false, sessionMode: 'shared' }],
+  conversations: [
+    {
+      platformId: 'mock-channel-1',
+      agentGroupId: 'ag-chan',
+      engageMode: 'pattern',
+      engagePattern: '.',
+      sessionMode: 'shared',
+    },
+  ],
   onInbound(platformId, threadId, message) {
     routeInbound({
       channelType: adapter.channelType,
diff --git a/scripts/test-v2-host.ts b/scripts/test-v2-host.ts
index b82bc99..2e49a3b 100644
--- a/scripts/test-v2-host.ts
+++ b/scripts/test-v2-host.ts
@@ -55,8 +55,10 @@ createMessagingGroupAgent({
   id: 'mga-e2e',
   messaging_group_id: 'mg-e2e',
   agent_group_id: 'ag-e2e',
-  trigger_rules: null,
-  response_scope: 'all',
+  engage_mode: 'pattern',
+  engage_pattern: '.',
+  sender_scope: 'all',
+  ignored_message_policy: 'drop',
   session_mode: 'shared',
   priority: 0,
   created_at: new Date().toISOString(),
diff --git a/setup.sh b/setup.sh
index af2c5e5..9ca73c1 100755
--- a/setup.sh
+++ b/setup.sh
@@ -6,9 +6,17 @@ set -euo pipefail
 # This is the only bash script in the setup flow.
 
 PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-LOG_FILE="$PROJECT_ROOT/logs/setup.log"
 
-mkdir -p "$PROJECT_ROOT/logs"
+# Where verbose bootstrap logs go. nanoclaw.sh captures setup.sh's stdout to
+# the per-step raw log, but legacy code in this script + install-node.sh
+# also calls `log` which writes to a file. Route those to the raw log so
+# they don't contaminate the progression log (logs/setup.log).
+# Default: write to the raw bootstrap log if nanoclaw.sh pointed us there,
+# else fall back to a dedicated bootstrap log (keeps standalone `bash
+# setup.sh` invocations working).
+LOG_FILE="${NANOCLAW_BOOTSTRAP_LOG:-${PROJECT_ROOT}/logs/bootstrap.log}"
+
+mkdir -p "$(dirname "$LOG_FILE")"
 
 log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] [bootstrap] $*" >> "$LOG_FILE"; }
 
@@ -72,19 +80,63 @@ install_deps() {
 
   cd "$PROJECT_ROOT"
 
-  # Enable corepack so `pnpm` shim lands on PATH.
-  log "Enabling corepack"
-  corepack enable >> "$LOG_FILE" 2>&1 || true
+  # Corepack's first-use "Do you want to continue? [Y/n]" prompt would hang
+  # the script since we redirect stdout/stderr to the log file — the prompt
+  # is invisible but corepack still blocks on stdin. Auto-accept.
+  export COREPACK_ENABLE_DOWNLOAD_PROMPT=0
 
-  # On Linux/WSL with system-wide Node (e.g. apt-installed to /usr/bin),
-  # corepack needs root to symlink /usr/bin/pnpm. Retry with sudo when pnpm
-  # isn't on PATH. macOS Homebrew installs land in a user-writable prefix,
-  # and a sudo retry there would create root-owned shims inside /opt/homebrew
-  # that later break brew — so the retry is Linux-only.
-  if ! command -v pnpm >/dev/null 2>&1 && [ "$PLATFORM" = "linux" ] \
-      && command -v sudo >/dev/null 2>&1; then
-    log "pnpm not on PATH after corepack enable — retrying with sudo"
-    sudo corepack enable >> "$LOG_FILE" 2>&1 || true
+  # Preferred path: enable corepack so `pnpm` shim lands on PATH.
+  if command -v corepack >/dev/null 2>&1; then
+    log "Enabling corepack"
+    corepack enable >> "$LOG_FILE" 2>&1 || true
+
+    # On Linux/WSL with system-wide Node (e.g. apt-installed to /usr/bin),
+    # corepack needs root to symlink /usr/bin/pnpm. macOS Homebrew installs
+    # land in a user-writable prefix, and a sudo retry there would create
+    # root-owned shims inside /opt/homebrew that later break brew — so the
+    # retry is Linux-only.
+    if ! command -v pnpm >/dev/null 2>&1 && [ "$PLATFORM" = "linux" ] \
+        && command -v sudo >/dev/null 2>&1; then
+      log "pnpm not on PATH after corepack enable — retrying with sudo"
+      sudo corepack enable >> "$LOG_FILE" 2>&1 || true
+    fi
+  else
+    log "corepack not available — will fall back to npm-install pnpm"
+  fi
+
+  # Fallback: some Node installs (older nvm, node@22 keg-only, minimal
+  # distro packages) don't include corepack. Install pnpm directly at the
+  # version pinned via package.json's `packageManager` field.
+  if ! command -v pnpm >/dev/null 2>&1 && command -v npm >/dev/null 2>&1; then
+    local pinned
+    pinned=$(grep -E '"packageManager"' "$PROJECT_ROOT/package.json" 2>/dev/null \
+      | head -1 \
+      | sed -E 's/.*"pnpm@([^"]+)".*/\1/')
+    [ -z "$pinned" ] && pinned="latest"
+    log "Installing pnpm@${pinned} via npm"
+    npm install -g "pnpm@${pinned}" >> "$LOG_FILE" 2>&1 \
+      || ([ "$PLATFORM" = "linux" ] && command -v sudo >/dev/null 2>&1 \
+            && sudo npm install -g "pnpm@${pinned}" >> "$LOG_FILE" 2>&1) \
+      || true
+  fi
+
+  # `npm install -g` writes to npm's global prefix, which isn't always on the
+  # shell PATH — common on macOS where the user has `npm config set prefix
+  # ~/.npm-global` to avoid sudo, or on Linux where /usr/local/bin isn't in
+  # PATH. Discover the prefix and prepend its bin dir so `command -v pnpm`
+  # sees the new install.
+  if ! command -v pnpm >/dev/null 2>&1 && command -v npm >/dev/null 2>&1; then
+    local npm_prefix
+    npm_prefix=$(npm config get prefix 2>/dev/null)
+    if [ -n "$npm_prefix" ] && [ -x "$npm_prefix/bin/pnpm" ]; then
+      export PATH="$npm_prefix/bin:$PATH"
+      log "Prepended npm prefix bin to PATH: $npm_prefix/bin"
+    fi
+  fi
+
+  if ! command -v pnpm >/dev/null 2>&1; then
+    log "pnpm not on PATH after corepack + npm fallback"
+    return
   fi
 
   log "Running pnpm install --frozen-lockfile"
@@ -131,6 +183,16 @@ log "=== Bootstrap started ==="
 detect_platform
 
 check_node
+if [ "$NODE_OK" = "false" ]; then
+  log "Node missing or too old — running setup/install-node.sh"
+  echo "Node not found — installing via setup/install-node.sh"
+  if bash "$PROJECT_ROOT/setup/install-node.sh" 2>&1 | tee -a "$LOG_FILE"; then
+    hash -r 2>/dev/null || true
+    check_node
+  else
+    log "install-node.sh failed"
+  fi
+fi
 install_deps
 check_build_tools
 
@@ -144,11 +206,20 @@ elif [ "$NATIVE_OK" = "false" ]; then
   STATUS="native_failed"
 fi
 
-# Anonymous setup start event (non-blocking, best-effort)
-curl -sS --max-time 3 -X POST https://us.i.posthog.com/capture/ \
-  -H 'Content-Type: application/json' \
-  -d "{\"api_key\":\"phc_fx1Hhx9ucz8GuaJC8LVZWO8u03yXZZJJ6ObS4yplnaP\",\"event\":\"setup_start\",\"distinct_id\":\"$(uuidgen 2>/dev/null || cat /proc/sys/kernel/random/uuid 2>/dev/null || echo unknown)\",\"properties\":{\"platform\":\"$PLATFORM\",\"is_wsl\":\"$IS_WSL\",\"is_root\":\"$IS_ROOT\",\"node_version\":\"$NODE_VERSION\",\"deps_ok\":\"$DEPS_OK\",\"native_ok\":\"$NATIVE_OK\",\"has_build_tools\":\"$HAS_BUILD_TOOLS\"}}" \
-  >/dev/null 2>&1 &
+# Anonymous setup start event (non-blocking, best-effort). Uses the
+# persisted distinct_id from data/install-id so bash-side events and the
+# node-side funnel share one id.
+# shellcheck source=setup/lib/diagnostics.sh
+source "$PROJECT_ROOT/setup/lib/diagnostics.sh"
+ph_event setup_start \
+  platform="$PLATFORM" \
+  is_wsl="$IS_WSL" \
+  is_root="$IS_ROOT" \
+  node_version="$NODE_VERSION" \
+  deps_ok="$DEPS_OK" \
+  native_ok="$NATIVE_OK" \
+  has_build_tools="$HAS_BUILD_TOOLS" \
+  status="$STATUS"
 
 cat <<EOF
 === NANOCLAW SETUP: BOOTSTRAP ===
diff --git a/setup/add-discord.sh b/setup/add-discord.sh
new file mode 100755
index 0000000..fa614fd
--- /dev/null
+++ b/setup/add-discord.sh
@@ -0,0 +1,130 @@
+#!/usr/bin/env bash
+#
+# Install the Discord adapter, persist DISCORD_BOT_TOKEN / APPLICATION_ID /
+# PUBLIC_KEY to .env + data/env/env, and restart the service. Non-interactive —
+# the operator-facing "Create a bot" walkthrough, owner confirmation, and
+# server-invite step live in setup/channels/discord.ts. Credentials come in via
+# env vars: DISCORD_BOT_TOKEN, DISCORD_APPLICATION_ID, DISCORD_PUBLIC_KEY.
+#
+# Emits exactly one status block on stdout (ADD_DISCORD) at the end. All chatty
+# progress messages go to stderr so setup:auto's raw-log capture sees the full
+# story without cluttering the final block for the parser.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-discord/SKILL.md.
+ADAPTER_VERSION="@chat-adapter/discord@4.26.0"
+
+# Resolve which remote carries the channels branch — handles forks where
+# upstream lives on a different remote than `origin`.
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  echo "=== NANOCLAW SETUP: ADD_DISCORD ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_VERSION: ${ADAPTER_VERSION}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-discord] $*" >&2; }
+
+if [ -z "${DISCORD_BOT_TOKEN:-}" ]; then
+  emit_status failed "DISCORD_BOT_TOKEN env var not set"
+  exit 1
+fi
+if [ -z "${DISCORD_APPLICATION_ID:-}" ]; then
+  emit_status failed "DISCORD_APPLICATION_ID env var not set"
+  exit 1
+fi
+if [ -z "${DISCORD_PUBLIC_KEY:-}" ]; then
+  emit_status failed "DISCORD_PUBLIC_KEY env var not set"
+  exit 1
+fi
+
+need_install() {
+  [ ! -f src/channels/discord.ts ] && return 0
+  ! grep -q "^import './discord.js';" src/channels/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  log "Copying adapter from ${CHANNELS_BRANCH}…"
+  git show "${CHANNELS_BRANCH}:src/channels/discord.ts" > src/channels/discord.ts
+
+  # Append self-registration import if missing.
+  if ! grep -q "^import './discord.js';" src/channels/index.ts; then
+    echo "import './discord.js';" >> src/channels/index.ts
+  fi
+
+  log "Installing ${ADAPTER_VERSION}…"
+  pnpm install "${ADAPTER_VERSION}" >&2 2>/dev/null || {
+    emit_status failed "pnpm install ${ADAPTER_VERSION} failed"
+    exit 1
+  }
+
+  log "Building…"
+  pnpm run build >&2 2>/dev/null || {
+    emit_status failed "pnpm run build failed"
+    exit 1
+  }
+else
+  log "Adapter files already installed — skipping install phase."
+fi
+
+# Persist credentials. auto.ts validates before this point, so bad values here
+# would be an internal bug rather than operator input.
+touch .env
+upsert_env() {
+  local key=$1 value=$2
+  if grep -q "^${key}=" .env; then
+    awk -v k="$key" -v v="$value" \
+        'BEGIN{FS=OFS="="} $1==k {print k "=" v; next} {print}' \
+      .env > .env.tmp && mv .env.tmp .env
+  else
+    echo "${key}=${value}" >> .env
+  fi
+}
+upsert_env DISCORD_BOT_TOKEN "$DISCORD_BOT_TOKEN"
+upsert_env DISCORD_APPLICATION_ID "$DISCORD_APPLICATION_ID"
+upsert_env DISCORD_PUBLIC_KEY "$DISCORD_PUBLIC_KEY"
+
+# Container reads from data/env/env (the host mounts it).
+mkdir -p data/env
+cp .env data/env/env
+
+log "Restarting service so the new adapter picks up the credentials…"
+# shellcheck source=setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+case "$(uname -s)" in
+  Darwin)
+    launchctl kickstart -k "gui/$(id -u)/$(launchd_label)" >&2 2>/dev/null || true
+    ;;
+  Linux)
+    systemctl --user restart "$(systemd_unit)" >&2 2>/dev/null \
+      || sudo systemctl restart "$(systemd_unit)" >&2 2>/dev/null \
+      || true
+    ;;
+esac
+
+# Give the Discord adapter a moment to finish gateway handshake before
+# init-first-agent attempts delivery.
+sleep 5
+
+emit_status success
diff --git a/setup/add-imessage.sh b/setup/add-imessage.sh
new file mode 100755
index 0000000..ea19862
--- /dev/null
+++ b/setup/add-imessage.sh
@@ -0,0 +1,160 @@
+#!/usr/bin/env bash
+#
+# Install the iMessage adapter, persist mode/creds to .env + data/env/env,
+# and restart the service. Non-interactive — the Full Disk Access walkthrough
+# (local mode) and Photon URL/key prompts (remote mode) live in
+# setup/channels/imessage.ts. Creds come in via env vars:
+#   IMESSAGE_LOCAL   'true' | 'false'  (required)
+#   IMESSAGE_ENABLED 'true'            (required when IMESSAGE_LOCAL=true)
+#   IMESSAGE_SERVER_URL                (required when IMESSAGE_LOCAL=false)
+#   IMESSAGE_API_KEY                   (required when IMESSAGE_LOCAL=false)
+#
+# Emits exactly one status block on stdout (ADD_IMESSAGE) at the end.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-imessage/SKILL.md.
+ADAPTER_VERSION="chat-adapter-imessage@0.1.1"
+
+# Resolve which remote carries the channels branch — handles forks where
+# upstream lives on a different remote than `origin`.
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  local mode=${IMESSAGE_LOCAL:-}
+  echo "=== NANOCLAW SETUP: ADD_IMESSAGE ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_VERSION: ${ADAPTER_VERSION}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$mode" ] && echo "MODE: $([ "$mode" = "true" ] && echo local || echo remote)"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-imessage] $*" >&2; }
+
+# Validate creds based on mode.
+if [ -z "${IMESSAGE_LOCAL:-}" ]; then
+  emit_status failed "IMESSAGE_LOCAL env var not set (expected true|false)"
+  exit 1
+fi
+if [ "${IMESSAGE_LOCAL}" = "true" ]; then
+  if [ -z "${IMESSAGE_ENABLED:-}" ]; then
+    emit_status failed "IMESSAGE_ENABLED env var not set for local mode"
+    exit 1
+  fi
+  if [ "$(uname -s)" != "Darwin" ]; then
+    emit_status failed "local mode requires macOS"
+    exit 1
+  fi
+else
+  if [ -z "${IMESSAGE_SERVER_URL:-}" ]; then
+    emit_status failed "IMESSAGE_SERVER_URL env var not set for remote mode"
+    exit 1
+  fi
+  if [ -z "${IMESSAGE_API_KEY:-}" ]; then
+    emit_status failed "IMESSAGE_API_KEY env var not set for remote mode"
+    exit 1
+  fi
+fi
+
+need_install() {
+  [ ! -f src/channels/imessage.ts ] && return 0
+  ! grep -q "^import './imessage.js';" src/channels/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  log "Copying adapter from ${CHANNELS_BRANCH}…"
+  git show "${CHANNELS_BRANCH}:src/channels/imessage.ts" > src/channels/imessage.ts
+
+  # Append self-registration import if missing.
+  if ! grep -q "^import './imessage.js';" src/channels/index.ts; then
+    echo "import './imessage.js';" >> src/channels/index.ts
+  fi
+
+  log "Installing ${ADAPTER_VERSION}…"
+  pnpm install "${ADAPTER_VERSION}" >&2 2>/dev/null || {
+    emit_status failed "pnpm install ${ADAPTER_VERSION} failed"
+    exit 1
+  }
+
+  log "Building…"
+  pnpm run build >&2 2>/dev/null || {
+    emit_status failed "pnpm run build failed"
+    exit 1
+  }
+else
+  log "Adapter files already installed — skipping install phase."
+fi
+
+touch .env
+upsert_env() {
+  local key=$1 value=$2
+  if grep -q "^${key}=" .env; then
+    awk -v k="$key" -v v="$value" \
+        'BEGIN{FS=OFS="="} $1==k {print k "=" v; next} {print}' \
+      .env > .env.tmp && mv .env.tmp .env
+  else
+    echo "${key}=${value}" >> .env
+  fi
+}
+
+remove_env() {
+  local key=$1
+  if grep -q "^${key}=" .env 2>/dev/null; then
+    grep -v "^${key}=" .env > .env.tmp && mv .env.tmp .env
+  fi
+}
+
+# Write the canonical keys for the chosen mode, strip the opposite mode's
+# keys so stale values can't confuse the adapter's factory.
+upsert_env IMESSAGE_LOCAL "$IMESSAGE_LOCAL"
+if [ "$IMESSAGE_LOCAL" = "true" ]; then
+  upsert_env IMESSAGE_ENABLED "$IMESSAGE_ENABLED"
+  remove_env IMESSAGE_SERVER_URL
+  remove_env IMESSAGE_API_KEY
+else
+  upsert_env IMESSAGE_SERVER_URL "$IMESSAGE_SERVER_URL"
+  upsert_env IMESSAGE_API_KEY "$IMESSAGE_API_KEY"
+  remove_env IMESSAGE_ENABLED
+fi
+
+# Container reads from data/env/env (the host mounts it).
+mkdir -p data/env
+cp .env data/env/env
+
+log "Restarting service so the new adapter picks up the creds…"
+# shellcheck source=setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+case "$(uname -s)" in
+  Darwin)
+    launchctl kickstart -k "gui/$(id -u)/$(launchd_label)" >&2 2>/dev/null || true
+    ;;
+  Linux)
+    systemctl --user restart "$(systemd_unit)" >&2 2>/dev/null \
+      || sudo systemctl restart "$(systemd_unit)" >&2 2>/dev/null \
+      || true
+    ;;
+esac
+
+# Give the adapter a moment to open chat.db (local) or handshake with
+# Photon (remote) before emitting success.
+sleep 3
+
+emit_status success
diff --git a/setup/add-signal.sh b/setup/add-signal.sh
new file mode 100755
index 0000000..8ebf2b9
--- /dev/null
+++ b/setup/add-signal.sh
@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+#
+# Install the Signal adapter in an already-running NanoClaw checkout.
+# Non-interactive — the operator-facing "install signal-cli" + QR scan
+# live in setup/channels/signal.ts. This script only:
+#
+#   1. Fetches src/channels/signal.ts + signal.test.ts from the channels
+#      branch.
+#   2. Appends the self-registration import to src/channels/index.ts.
+#   3. Installs qrcode (for setup-flow QR rendering — adapter itself has
+#      no npm deps).
+#   4. Builds.
+#
+# SIGNAL_ACCOUNT is persisted separately by the driver once signal-cli
+# link has produced a number; that keeps this script idempotent and
+# re-runnable without re-auth.
+#
+# Emits exactly one status block on stdout (ADD_SIGNAL) at the end. All
+# chatty progress goes to stderr so setup:auto's raw-log capture sees
+# the full story without cluttering the final block for the parser.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-signal/SKILL.md.
+QRCODE_VERSION="qrcode@1.5.4"
+QRCODE_TYPES_VERSION="@types/qrcode@1.5.6"
+
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  echo "=== NANOCLAW SETUP: ADD_SIGNAL ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-signal] $*" >&2; }
+
+need_install() {
+  [ ! -f src/channels/signal.ts ] && return 0
+  ! grep -q "^import './signal.js';" src/channels/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  log "Copying adapter files from ${CHANNELS_BRANCH}…"
+  for f in \
+    src/channels/signal.ts \
+    src/channels/signal.test.ts
+  do
+    git show "${CHANNELS_BRANCH}:$f" > "$f" || {
+      emit_status failed "git show ${CHANNELS_BRANCH}:$f failed"
+      exit 1
+    }
+  done
+
+  if ! grep -q "^import './signal.js';" src/channels/index.ts; then
+    echo "import './signal.js';" >> src/channels/index.ts
+  fi
+fi
+
+# qrcode is needed by setup/signal-auth.ts to render the linking URL as a
+# terminal QR. Install idempotently — if it's already present (e.g. from a
+# prior WhatsApp install) pnpm is a no-op.
+if ! node -e "require.resolve('qrcode')" >/dev/null 2>&1; then
+  log "Installing ${QRCODE_VERSION}…"
+  pnpm install "${QRCODE_VERSION}" "${QRCODE_TYPES_VERSION}" >&2 2>/dev/null || {
+    emit_status failed "pnpm install ${QRCODE_VERSION} failed"
+    exit 1
+  }
+fi
+
+log "Building…"
+pnpm run build >&2 2>/dev/null || {
+  emit_status failed "pnpm run build failed"
+  exit 1
+}
+
+emit_status success
diff --git a/setup/add-slack.sh b/setup/add-slack.sh
new file mode 100755
index 0000000..3eea3e5
--- /dev/null
+++ b/setup/add-slack.sh
@@ -0,0 +1,125 @@
+#!/usr/bin/env bash
+#
+# Install the Slack adapter, persist SLACK_BOT_TOKEN + SLACK_SIGNING_SECRET to
+# .env + data/env/env, and restart the service. Non-interactive — the
+# operator-facing app creation walkthrough + credential paste live in
+# setup/channels/slack.ts. Credentials come in via env vars:
+# SLACK_BOT_TOKEN, SLACK_SIGNING_SECRET.
+#
+# Emits exactly one status block on stdout (ADD_SLACK) at the end. All chatty
+# progress messages go to stderr so setup:auto's raw-log capture sees the full
+# story without cluttering the final block for the parser.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-slack/SKILL.md.
+ADAPTER_VERSION="@chat-adapter/slack@4.26.0"
+
+# Resolve which remote carries the channels branch — handles forks where
+# upstream lives on a different remote than `origin`.
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  echo "=== NANOCLAW SETUP: ADD_SLACK ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_VERSION: ${ADAPTER_VERSION}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-slack] $*" >&2; }
+
+if [ -z "${SLACK_BOT_TOKEN:-}" ]; then
+  emit_status failed "SLACK_BOT_TOKEN env var not set"
+  exit 1
+fi
+if [ -z "${SLACK_SIGNING_SECRET:-}" ]; then
+  emit_status failed "SLACK_SIGNING_SECRET env var not set"
+  exit 1
+fi
+
+need_install() {
+  [ ! -f src/channels/slack.ts ] && return 0
+  ! grep -q "^import './slack.js';" src/channels/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  log "Copying adapter from ${CHANNELS_BRANCH}…"
+  git show "${CHANNELS_BRANCH}:src/channels/slack.ts" > src/channels/slack.ts
+
+  # Append self-registration import if missing.
+  if ! grep -q "^import './slack.js';" src/channels/index.ts; then
+    echo "import './slack.js';" >> src/channels/index.ts
+  fi
+
+  log "Installing ${ADAPTER_VERSION}…"
+  pnpm install "${ADAPTER_VERSION}" >&2 2>/dev/null || {
+    emit_status failed "pnpm install ${ADAPTER_VERSION} failed"
+    exit 1
+  }
+
+  log "Building…"
+  pnpm run build >&2 2>/dev/null || {
+    emit_status failed "pnpm run build failed"
+    exit 1
+  }
+else
+  log "Adapter files already installed — skipping install phase."
+fi
+
+# Persist credentials. auto.ts validates via auth.test before this point, so
+# bad values here would be an internal bug rather than operator input.
+touch .env
+upsert_env() {
+  local key=$1 value=$2
+  if grep -q "^${key}=" .env; then
+    awk -v k="$key" -v v="$value" \
+        'BEGIN{FS=OFS="="} $1==k {print k "=" v; next} {print}' \
+      .env > .env.tmp && mv .env.tmp .env
+  else
+    echo "${key}=${value}" >> .env
+  fi
+}
+upsert_env SLACK_BOT_TOKEN "$SLACK_BOT_TOKEN"
+upsert_env SLACK_SIGNING_SECRET "$SLACK_SIGNING_SECRET"
+
+# Container reads from data/env/env (the host mounts it).
+mkdir -p data/env
+cp .env data/env/env
+
+log "Restarting service so the new adapter picks up the credentials…"
+# shellcheck source=setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+case "$(uname -s)" in
+  Darwin)
+    launchctl kickstart -k "gui/$(id -u)/$(launchd_label)" >&2 2>/dev/null || true
+    ;;
+  Linux)
+    systemctl --user restart "$(systemd_unit)" >&2 2>/dev/null \
+      || sudo systemctl restart "$(systemd_unit)" >&2 2>/dev/null \
+      || true
+    ;;
+esac
+
+# Give the Slack adapter a moment to finish starting the webhook listener
+# before emitting success.
+sleep 3
+
+emit_status success
diff --git a/setup/add-teams.sh b/setup/add-teams.sh
new file mode 100755
index 0000000..273cad6
--- /dev/null
+++ b/setup/add-teams.sh
@@ -0,0 +1,139 @@
+#!/usr/bin/env bash
+#
+# Install the Teams adapter, persist TEAMS_APP_ID / _PASSWORD / _TENANT_ID /
+# _TYPE to .env + data/env/env, and restart the service. Non-interactive —
+# the operator-facing Azure portal walkthroughs live in
+# setup/channels/teams.ts. Credentials come in via env vars:
+#   TEAMS_APP_ID            (required)
+#   TEAMS_APP_PASSWORD      (required — client secret value from Azure)
+#   TEAMS_APP_TYPE          (required — SingleTenant | MultiTenant)
+#   TEAMS_APP_TENANT_ID     (required when type=SingleTenant)
+#
+# Emits exactly one status block on stdout (ADD_TEAMS) at the end. All chatty
+# progress messages go to stderr so setup:auto's raw-log capture sees the
+# full story without cluttering the final block for the parser.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-teams/SKILL.md.
+ADAPTER_VERSION="@chat-adapter/teams@4.26.0"
+
+# Resolve which remote carries the channels branch — handles forks where
+# upstream lives on a different remote than `origin`.
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  echo "=== NANOCLAW SETUP: ADD_TEAMS ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_VERSION: ${ADAPTER_VERSION}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-teams] $*" >&2; }
+
+if [ -z "${TEAMS_APP_ID:-}" ]; then
+  emit_status failed "TEAMS_APP_ID env var not set"
+  exit 1
+fi
+if [ -z "${TEAMS_APP_PASSWORD:-}" ]; then
+  emit_status failed "TEAMS_APP_PASSWORD env var not set"
+  exit 1
+fi
+if [ -z "${TEAMS_APP_TYPE:-}" ]; then
+  emit_status failed "TEAMS_APP_TYPE env var not set (SingleTenant|MultiTenant)"
+  exit 1
+fi
+if [ "${TEAMS_APP_TYPE}" = "SingleTenant" ] && [ -z "${TEAMS_APP_TENANT_ID:-}" ]; then
+  emit_status failed "TEAMS_APP_TENANT_ID required when TEAMS_APP_TYPE=SingleTenant"
+  exit 1
+fi
+
+need_install() {
+  [ ! -f src/channels/teams.ts ] && return 0
+  ! grep -q "^import './teams.js';" src/channels/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  log "Copying adapter from ${CHANNELS_BRANCH}…"
+  git show "${CHANNELS_BRANCH}:src/channels/teams.ts" > src/channels/teams.ts
+
+  # Append self-registration import if missing.
+  if ! grep -q "^import './teams.js';" src/channels/index.ts; then
+    echo "import './teams.js';" >> src/channels/index.ts
+  fi
+
+  log "Installing ${ADAPTER_VERSION}…"
+  pnpm install "${ADAPTER_VERSION}" >&2 2>/dev/null || {
+    emit_status failed "pnpm install ${ADAPTER_VERSION} failed"
+    exit 1
+  }
+
+  log "Building…"
+  pnpm run build >&2 2>/dev/null || {
+    emit_status failed "pnpm run build failed"
+    exit 1
+  }
+else
+  log "Adapter files already installed — skipping install phase."
+fi
+
+# Persist credentials.
+touch .env
+upsert_env() {
+  local key=$1 value=$2
+  if grep -q "^${key}=" .env; then
+    awk -v k="$key" -v v="$value" \
+        'BEGIN{FS=OFS="="} $1==k {print k "=" v; next} {print}' \
+      .env > .env.tmp && mv .env.tmp .env
+  else
+    echo "${key}=${value}" >> .env
+  fi
+}
+upsert_env TEAMS_APP_ID "$TEAMS_APP_ID"
+upsert_env TEAMS_APP_PASSWORD "$TEAMS_APP_PASSWORD"
+upsert_env TEAMS_APP_TYPE "$TEAMS_APP_TYPE"
+if [ -n "${TEAMS_APP_TENANT_ID:-}" ]; then
+  upsert_env TEAMS_APP_TENANT_ID "$TEAMS_APP_TENANT_ID"
+fi
+
+# Container reads from data/env/env (the host mounts it).
+mkdir -p data/env
+cp .env data/env/env
+
+log "Restarting service so the new adapter picks up the credentials…"
+# shellcheck source=setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+case "$(uname -s)" in
+  Darwin)
+    launchctl kickstart -k "gui/$(id -u)/$(launchd_label)" >&2 2>/dev/null || true
+    ;;
+  Linux)
+    systemctl --user restart "$(systemd_unit)" >&2 2>/dev/null \
+      || sudo systemctl restart "$(systemd_unit)" >&2 2>/dev/null \
+      || true
+    ;;
+esac
+
+# Give the Teams adapter a moment to register its webhook before the driver
+# continues.
+sleep 5
+
+emit_status success
diff --git a/setup/add-telegram.sh b/setup/add-telegram.sh
new file mode 100755
index 0000000..c81fc6d
--- /dev/null
+++ b/setup/add-telegram.sh
@@ -0,0 +1,164 @@
+#!/usr/bin/env bash
+#
+# Install the Telegram adapter, persist the bot token to .env + data/env/env,
+# restart the service, and open the bot's chat page in the local Telegram
+# client. Non-interactive — the operator-facing "Create a bot" instructions
+# and token paste live in setup/auto.ts. The token comes in via the
+# TELEGRAM_BOT_TOKEN env var.
+#
+# Emits exactly one status block on stdout (ADD_TELEGRAM) at the end. All
+# chatty progress messages go to stderr so setup:auto's raw-log capture
+# sees the full story without cluttering the final block for the parser.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-telegram/SKILL.md.
+ADAPTER_VERSION="@chat-adapter/telegram@4.26.0"
+
+# Resolve which remote carries the channels branch — handles forks where
+# upstream lives on a different remote than `origin`.
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  local username=${BOT_USERNAME:-}
+  echo "=== NANOCLAW SETUP: ADD_TELEGRAM ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_VERSION: ${ADAPTER_VERSION}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$username" ] && echo "BOT_USERNAME: ${username}"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-telegram] $*" >&2; }
+
+if [ -z "${TELEGRAM_BOT_TOKEN:-}" ]; then
+  emit_status failed "TELEGRAM_BOT_TOKEN env var not set"
+  exit 1
+fi
+
+if ! [[ "$TELEGRAM_BOT_TOKEN" =~ ^[0-9]+:[A-Za-z0-9_-]{35,}$ ]]; then
+  emit_status failed "token format invalid (expected <digits>:<chars>)"
+  exit 1
+fi
+
+need_install() {
+  [ ! -f src/channels/telegram.ts ] && return 0
+  ! grep -q "^import './telegram.js';" src/channels/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  # pair-telegram.ts is maintained in this branch (setup-auto), so it's NOT
+  # in this list — do not overwrite the local version with the channels copy.
+  log "Copying adapter files from ${CHANNELS_BRANCH}…"
+  for f in \
+    src/channels/telegram.ts \
+    src/channels/telegram-pairing.ts \
+    src/channels/telegram-pairing.test.ts \
+    src/channels/telegram-markdown-sanitize.ts \
+    src/channels/telegram-markdown-sanitize.test.ts
+  do
+    git show "${CHANNELS_BRANCH}:$f" > "$f"
+  done
+
+  # Append self-registration import if missing.
+  if ! grep -q "^import './telegram.js';" src/channels/index.ts; then
+    echo "import './telegram.js';" >> src/channels/index.ts
+  fi
+
+  # Register pair-telegram step if not already in the STEPS map.
+  # Uses node (not sed) since sed's in-place + escape semantics differ
+  # between BSD (macOS) and GNU.
+  node -e '
+    const fs = require("fs");
+    const p = "setup/index.ts";
+    let s = fs.readFileSync(p, "utf-8");
+    if (!s.includes("\047pair-telegram\047")) {
+      s = s.replace(
+        /(register: \(\) => import\(\x27\.\/register\.js\x27\),)/,
+        "$1\n  \x27pair-telegram\x27: () => import(\x27./pair-telegram.js\x27),"
+      );
+      fs.writeFileSync(p, s);
+    }
+  '
+
+  log "Installing ${ADAPTER_VERSION}…"
+  pnpm install "${ADAPTER_VERSION}" >&2 2>/dev/null || {
+    emit_status failed "pnpm install ${ADAPTER_VERSION} failed"
+    exit 1
+  }
+
+  log "Building…"
+  pnpm run build >&2 2>/dev/null || {
+    emit_status failed "pnpm run build failed"
+    exit 1
+  }
+else
+  log "Adapter files already installed — skipping install phase."
+fi
+
+# Persist token. auto.ts validates before this point, so a bad token here
+# would be an internal bug rather than operator input.
+touch .env
+if grep -q '^TELEGRAM_BOT_TOKEN=' .env; then
+  awk -v tok="$TELEGRAM_BOT_TOKEN" \
+      '/^TELEGRAM_BOT_TOKEN=/{print "TELEGRAM_BOT_TOKEN=" tok; next} {print}' \
+    .env > .env.tmp && mv .env.tmp .env
+else
+  echo "TELEGRAM_BOT_TOKEN=${TELEGRAM_BOT_TOKEN}" >> .env
+fi
+
+# Look up the bot username (auto.ts already validated; we re-query here so
+# standalone invocations still work — BOT_USERNAME is emitted in the status
+# block for parent drivers to display).
+INFO=$(curl -fsS --max-time 8 \
+  "https://api.telegram.org/bot${TELEGRAM_BOT_TOKEN}/getMe" 2>/dev/null || true)
+BOT_USERNAME=""
+if echo "$INFO" | grep -q '"ok":true'; then
+  BOT_USERNAME=$(echo "$INFO" | sed -nE 's/.*"username":"([^"]+)".*/\1/p')
+fi
+
+# Container reads from data/env/env (the host mounts it).
+mkdir -p data/env
+cp .env data/env/env
+
+# Browser/app deep-link is done by the parent driver (setup/channels/telegram.ts)
+# BEFORE this script runs — gated on a clack confirm so focus-stealing doesn't
+# surprise the user. Keeping it out of here means this script stays pure
+# non-interactive install.
+
+log "Restarting service so the new adapter picks up the token…"
+# shellcheck source=setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+case "$(uname -s)" in
+  Darwin)
+    launchctl kickstart -k "gui/$(id -u)/$(launchd_label)" >&2 2>/dev/null || true
+    ;;
+  Linux)
+    systemctl --user restart "$(systemd_unit)" >&2 2>/dev/null \
+      || sudo systemctl restart "$(systemd_unit)" >&2 2>/dev/null \
+      || true
+    ;;
+esac
+
+# Give the Telegram adapter a moment to finish starting before pair-telegram
+# begins polling for the user's code message.
+sleep 5
+
+emit_status success
diff --git a/setup/add-whatsapp.sh b/setup/add-whatsapp.sh
new file mode 100755
index 0000000..c7356af
--- /dev/null
+++ b/setup/add-whatsapp.sh
@@ -0,0 +1,120 @@
+#!/usr/bin/env bash
+#
+# Install the native WhatsApp (Baileys) adapter and its whatsapp-auth + groups
+# setup steps. No credentials in env — WhatsApp uses linked-device auth, run
+# by the whatsapp-auth step as a separate process. The adapter's factory
+# returns null until store/auth/creds.json exists, so it's safe to install
+# this before auth runs; the driver restarts the service *after* auth
+# succeeds.
+#
+# Emits exactly one status block on stdout (ADD_WHATSAPP) at the end. All
+# chatty progress messages go to stderr so setup:auto's raw-log capture sees
+# the full story without cluttering the final block for the parser.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+# Keep in sync with .claude/skills/add-whatsapp/SKILL.md.
+BAILEYS_VERSION="@whiskeysockets/baileys@6.17.16"
+QRCODE_VERSION="qrcode@1.5.4"
+QRCODE_TYPES_VERSION="@types/qrcode@1.5.6"
+PINO_VERSION="pino@9.6.0"
+
+# Resolve which remote carries the channels branch — handles forks where
+# upstream lives on a different remote than `origin`.
+# shellcheck source=setup/lib/channels-remote.sh
+source "$PROJECT_ROOT/setup/lib/channels-remote.sh"
+CHANNELS_REMOTE=$(resolve_channels_remote)
+CHANNELS_BRANCH="${CHANNELS_REMOTE}/channels"
+
+emit_status() {
+  local status=$1 error=${2:-}
+  local already=${ADAPTER_ALREADY_INSTALLED:-false}
+  echo "=== NANOCLAW SETUP: ADD_WHATSAPP ==="
+  echo "STATUS: ${status}"
+  echo "ADAPTER_ALREADY_INSTALLED: ${already}"
+  [ -n "$error" ] && echo "ERROR: ${error}"
+  echo "=== END ==="
+}
+
+log() { echo "[add-whatsapp] $*" >&2; }
+
+need_install() {
+  [ ! -f src/channels/whatsapp.ts ] && return 0
+  [ ! -f setup/groups.ts ] && return 0
+  ! grep -q "^import './whatsapp.js';" src/channels/index.ts 2>/dev/null && return 0
+  ! grep -q "'whatsapp-auth':" setup/index.ts 2>/dev/null && return 0
+  ! grep -q "^  groups:" setup/index.ts 2>/dev/null && return 0
+  return 1
+}
+
+ADAPTER_ALREADY_INSTALLED=true
+if need_install; then
+  ADAPTER_ALREADY_INSTALLED=false
+  log "Fetching channels branch…"
+  git fetch "$CHANNELS_REMOTE" channels >&2 2>/dev/null || {
+    emit_status failed "git fetch ${CHANNELS_REMOTE} channels failed"
+    exit 1
+  }
+
+  # whatsapp-auth.ts is maintained in this branch (setup-auto) — do not copy
+  # from channels. Matches the pair-telegram.ts pattern.
+  log "Copying adapter + group step from ${CHANNELS_BRANCH}…"
+  git show "${CHANNELS_BRANCH}:src/channels/whatsapp.ts" > src/channels/whatsapp.ts
+  git show "${CHANNELS_BRANCH}:setup/groups.ts"          > setup/groups.ts
+
+  # Append self-registration import if missing.
+  if ! grep -q "^import './whatsapp.js';" src/channels/index.ts; then
+    echo "import './whatsapp.js';" >> src/channels/index.ts
+  fi
+
+  # Register the setup steps in setup/index.ts's STEPS map. node (not sed) —
+  # sed's in-place + escape semantics differ between BSD (macOS) and GNU.
+  node -e '
+    const fs = require("fs");
+    const p = "setup/index.ts";
+    let s = fs.readFileSync(p, "utf-8");
+    let changed = false;
+    if (!s.includes("\047whatsapp-auth\047:")) {
+      s = s.replace(
+        /(register: \(\) => import\(\x27\.\/register\.js\x27\),)/,
+        "$1\n  \x27whatsapp-auth\x27: () => import(\x27./whatsapp-auth.js\x27),"
+      );
+      changed = true;
+    }
+    if (!/^\s*groups:\s/m.test(s)) {
+      s = s.replace(
+        /(register: \(\) => import\(\x27\.\/register\.js\x27\),)/,
+        "$1\n  groups: () => import(\x27./groups.js\x27),"
+      );
+      changed = true;
+    }
+    if (changed) fs.writeFileSync(p, s);
+  '
+
+  log "Installing Baileys + QR + pino (pinned)…"
+  pnpm install \
+    "${BAILEYS_VERSION}" \
+    "${QRCODE_VERSION}" \
+    "${QRCODE_TYPES_VERSION}" \
+    "${PINO_VERSION}" \
+    >&2 2>/dev/null || {
+    emit_status failed "pnpm install failed"
+    exit 1
+  }
+
+  log "Building…"
+  pnpm run build >&2 2>/dev/null || {
+    emit_status failed "pnpm run build failed"
+    exit 1
+  }
+else
+  log "Adapter + setup steps already installed — skipping install phase."
+fi
+
+# No service restart here — the adapter factory returns null without
+# store/auth/creds.json, so restarting now would no-op. The driver restarts
+# the service AFTER whatsapp-auth completes so the adapter picks up creds.
+
+emit_status success
diff --git a/setup/auto.ts b/setup/auto.ts
new file mode 100644
index 0000000..cff2f63
--- /dev/null
+++ b/setup/auto.ts
@@ -0,0 +1,1012 @@
+/**
+ * Non-interactive setup driver — the step sequencer for `pnpm run setup:auto`.
+ *
+ * Responsibility: orchestrate the sequence of steps end-to-end and route
+ * between them. The runner, spawning, status parsing, spinner, abort, and
+ * prompt primitives live in `setup/lib/runner.ts`; theming in
+ * `setup/lib/theme.ts`; Telegram's full flow in `setup/channels/telegram.ts`.
+ *
+ * Config via env:
+ *   NANOCLAW_DISPLAY_NAME  how the agents address the operator — skips the
+ *                          prompt. Defaults to $USER.
+ *   NANOCLAW_AGENT_NAME    messaging-channel agent name (consumed by the
+ *                          channel flow). The CLI scratch agent is always
+ *                          "Terminal Agent".
+ *   NANOCLAW_SKIP          comma-separated step names to skip
+ *                          (environment|container|onecli|auth|mounts|
+ *                           service|cli-agent|timezone|channel|verify|
+ *                           first-chat)
+ *
+ * Timezone is auto-detected after the CLI agent step. UTC resolves are
+ * confirmed with the user, and free-text replies fall through to a
+ * headless `claude -p` call for IANA-zone resolution.
+ */
+import { spawn, spawnSync } from 'child_process';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import { runDiscordChannel } from './channels/discord.js';
+import { runIMessageChannel } from './channels/imessage.js';
+import { runSignalChannel } from './channels/signal.js';
+import { runSlackChannel } from './channels/slack.js';
+import { runTeamsChannel } from './channels/teams.js';
+import { runTelegramChannel } from './channels/telegram.js';
+import { runWhatsAppChannel } from './channels/whatsapp.js';
+import { pingCliAgent, type PingResult } from './lib/agent-ping.js';
+import { brightSelect } from './lib/bright-select.js';
+import { offerClaudeAssist } from './lib/claude-assist.js';
+import { runWindowedStep } from './lib/windowed-runner.js';
+import { getLaunchdLabel, getSystemdUnit } from '../src/install-slug.js';
+import {
+  claudeCliAvailable,
+  resolveTimezoneViaClaude,
+} from './lib/tz-from-claude.js';
+import * as setupLog from './logs.js';
+import { ensureAnswer, fail, runQuietChild, runQuietStep } from './lib/runner.js';
+import { emit as phEmit } from './lib/diagnostics.js';
+import { brandBold, brandChip, dimWrap, fitToWidth, wrapForGutter } from './lib/theme.js';
+import { isValidTimezone } from '../src/timezone.js';
+
+const CLI_AGENT_NAME = 'Terminal Agent';
+const RUN_START = Date.now();
+
+type ChannelChoice =
+  | 'telegram'
+  | 'discord'
+  | 'whatsapp'
+  | 'signal'
+  | 'teams'
+  | 'slack'
+  | 'imessage'
+  | 'skip';
+
+async function main(): Promise<void> {
+  printIntro();
+  initProgressionLog();
+  phEmit('auto_started');
+
+  const skip = new Set(
+    (process.env.NANOCLAW_SKIP ?? '')
+      .split(',')
+      .map((s) => s.trim())
+      .filter(Boolean),
+  );
+
+  if (!skip.has('environment')) {
+    const res = await runQuietStep('environment', {
+      running: 'Checking your system…',
+      done: 'Your system looks good.',
+    });
+    if (!res.ok) {
+      await fail(
+        'environment',
+        "Your system doesn't look quite right.",
+        'See logs/setup-steps/ for details, then retry.',
+      );
+    }
+  }
+
+  if (!skip.has('container')) {
+    p.log.message(
+      dimWrap(
+        'Your assistant lives in its own sandbox. It can only see what you explicitly share.',
+        4,
+      ),
+    );
+    p.log.message(
+      dimWrap(
+        'The first build pulls a base image and installs a few tools. On a fresh machine this usually takes 3–10 minutes.',
+        4,
+      ),
+    );
+    const res = await runWindowedStep('container', {
+      running: "Preparing your assistant's sandbox…",
+      done: 'Sandbox ready.',
+      failed: "Couldn't prepare the sandbox.",
+    });
+    if (!res.ok) {
+      const err = res.terminal?.fields.ERROR;
+      if (err === 'runtime_not_available') {
+        await fail(
+          'container',
+          "Docker isn't available.",
+          'Install Docker Desktop (or start it if already installed), then retry.',
+        );
+      }
+      if (err === 'docker_group_not_active') {
+        await fail(
+          'container',
+          "Docker was just installed but your shell doesn't know yet.",
+          'Log out and back in (or run `newgrp docker` in a new shell), then retry.',
+        );
+      }
+      await fail(
+        'container',
+        "Couldn't build the sandbox.",
+        'If Docker has a stale cache, try: `docker builder prune -f`, then retry.',
+      );
+    }
+    maybeReexecUnderSg();
+  }
+
+  if (!skip.has('onecli')) {
+    p.log.message(
+      dimWrap(
+        'Your assistant never gets your API keys directly. The vault adds them to approved requests as they leave the sandbox.',
+        4,
+      ),
+    );
+
+    // Respect an existing OneCLI install. Re-running the installer would
+    // rebind the listener and knock any other app using that gateway
+    // offline — confirm with the user before doing that.
+    const existing = detectExistingOnecli();
+    let reuse = false;
+    if (existing) {
+      const choice = ensureAnswer(
+        await brightSelect({
+          message: `Found an existing OneCLI at ${existing.apiHost}. What would you like to do?`,
+          options: [
+            {
+              value: 'reuse',
+              label: 'Use the existing instance',
+              hint: 'recommended — keeps other apps bound to this vault working',
+            },
+            {
+              value: 'fresh',
+              label: 'Install a fresh instance for NanoClaw',
+              hint: 'reinstalls onecli; other apps may need to reconnect',
+            },
+          ],
+        }),
+      ) as 'reuse' | 'fresh';
+      setupLog.userInput('onecli_choice', choice);
+      reuse = choice === 'reuse';
+    }
+
+    const res = await runQuietStep(
+      'onecli',
+      {
+        running: reuse
+          ? 'Hooking up to your existing OneCLI…'
+          : "Setting up OneCLI, your agent's vault…",
+        done: 'OneCLI vault ready.',
+      },
+      reuse ? ['--reuse'] : [],
+    );
+    if (!res.ok) {
+      const err = res.terminal?.fields.ERROR;
+      if (err === 'onecli_not_on_path_after_install') {
+        await fail(
+          'onecli',
+          'OneCLI was installed but your shell needs to refresh to see it.',
+          'Open a new shell or run `export PATH="$HOME/.local/bin:$PATH"`, then retry.',
+        );
+      }
+      await fail(
+        'onecli',
+        `Couldn't set up OneCLI (${err ?? 'unknown error'}).`,
+        'Make sure curl is installed and ~/.local/bin is writable, then retry.',
+      );
+    }
+  }
+
+  if (!skip.has('auth')) {
+    await runAuthStep();
+  }
+
+  if (!skip.has('mounts')) {
+    const res = await runQuietStep(
+      'mounts',
+      {
+        running: "Setting your assistant's access rules…",
+        done: 'Access rules set.',
+        skipped: 'Access rules already set.',
+      },
+      ['--empty'],
+    );
+    if (!res.ok) {
+      await fail('mounts', "Couldn't write access rules.");
+    }
+  }
+
+  if (!skip.has('service')) {
+    const res = await runQuietStep('service', {
+      running: 'Starting NanoClaw in the background…',
+      done: 'NanoClaw is running.',
+    });
+    if (!res.ok) {
+      await fail(
+        'service',
+        "Couldn't start NanoClaw.",
+        'See logs/nanoclaw.error.log for details.',
+      );
+    }
+    if (res.terminal?.fields.DOCKER_GROUP_STALE === 'true') {
+      p.log.warn(
+        "NanoClaw's permissions need a tweak before it can reach Docker.",
+      );
+      p.log.message(
+        '  sudo setfacl -m u:$(whoami):rw /var/run/docker.sock\n' +
+          `  systemctl --user restart ${getSystemdUnit()}`,
+      );
+    }
+  }
+
+  let displayName: string | undefined;
+  const needsDisplayName = !skip.has('cli-agent') || !skip.has('channel');
+  if (needsDisplayName) {
+    const fallback = process.env.USER?.trim() || 'Operator';
+    const preset = process.env.NANOCLAW_DISPLAY_NAME?.trim();
+    displayName = preset || (await askDisplayName(fallback));
+  }
+
+  if (!skip.has('cli-agent')) {
+    const res = await runQuietStep(
+      'cli-agent',
+      {
+        running: 'Bringing your assistant online…',
+        done: 'Assistant wired up.',
+      },
+      ['--display-name', displayName!, '--agent-name', CLI_AGENT_NAME],
+    );
+    if (!res.ok) {
+      await fail(
+        'cli-agent',
+        "Couldn't bring your assistant online.",
+        `You can retry later with \`pnpm exec tsx scripts/init-cli-agent.ts --display-name "${displayName!}" --agent-name "${CLI_AGENT_NAME}"\`.`,
+      );
+    }
+    if (!skip.has('first-chat')) {
+      p.log.message(
+        dimWrap(
+          "Your assistant runs in an isolated sandbox. I'm going to send it a quick test message (ping) and wait for a reply (pong) to confirm it's responding. First startup typically takes 30–60 seconds while the sandbox warms up.",
+          4,
+        ),
+      );
+      const ping = await confirmAssistantResponds();
+      if (ping === 'ok') {
+        phEmit('first_chat_ready');
+        const next = ensureAnswer(
+          await p.select({
+            message: 'What next?',
+            options: [
+              {
+                value: 'continue',
+                label: 'Continue with setup',
+                hint: 'recommended',
+              },
+              {
+                value: 'chat',
+                label: 'Pause here and chat with your agent from the terminal',
+              },
+            ],
+          }),
+        ) as 'continue' | 'chat';
+        setupLog.userInput('first_chat_choice', next);
+        if (next === 'chat') await runFirstChat();
+      } else {
+        phEmit('first_chat_failed', { reason: ping });
+        renderPingFailureNote(ping);
+        await offerClaudeAssist({
+          stepName: 'cli-agent',
+          msg:
+            ping === 'socket_error'
+              ? "NanoClaw service isn't listening on its CLI socket."
+              : "No reply from the assistant within 30 seconds.",
+          hint:
+            ping === 'socket_error'
+              ? 'Socket at data/cli.sock did not accept a connection.'
+              : 'Agent container may be failing to start or authenticate.',
+        });
+      }
+    }
+  }
+
+  if (!skip.has('timezone')) {
+    await runTimezoneStep();
+  }
+
+  let channelChoice: ChannelChoice = 'skip';
+  if (!skip.has('channel')) {
+    channelChoice = await askChannelChoice();
+    if (channelChoice === 'telegram') {
+      await runTelegramChannel(displayName!);
+    } else if (channelChoice === 'discord') {
+      await runDiscordChannel(displayName!);
+    } else if (channelChoice === 'whatsapp') {
+      await runWhatsAppChannel(displayName!);
+    } else if (channelChoice === 'signal') {
+      await runSignalChannel(displayName!);
+    } else if (channelChoice === 'teams') {
+      await runTeamsChannel(displayName!);
+    } else if (channelChoice === 'slack') {
+      await runSlackChannel(displayName!);
+    } else if (channelChoice === 'imessage') {
+      await runIMessageChannel(displayName!);
+    } else {
+      p.log.info(
+        wrapForGutter(
+          'No messaging app for now. You can add one later (like Telegram, Discord, WhatsApp, Teams, Slack, or iMessage).',
+          4,
+        ),
+      );
+    }
+  }
+
+  if (!skip.has('verify')) {
+    const res = await runQuietStep('verify', {
+      running: 'Making sure everything works together…',
+      done: "Everything's connected.",
+      failed: 'A few things still need your attention.',
+    });
+    if (!res.ok) {
+      const notes: string[] = [];
+      if (res.terminal?.fields.CREDENTIALS !== 'configured') {
+        notes.push('• Your Claude account isn\'t connected. Re-run setup and try again.');
+      }
+      const service = res.terminal?.fields.SERVICE;
+      if (service === 'running_other_checkout') {
+        const label = getLaunchdLabel();
+        notes.push(
+          wrapForGutter(
+            [
+              '• Your NanoClaw service is running from a different folder on this machine.',
+              '  Point it at this checkout with:',
+              `    launchctl bootout gui/$(id -u)/${label}`,
+              `    launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/${label}.plist`,
+            ].join('\n'),
+            6,
+          ),
+        );
+      } else {
+        const agentPing = res.terminal?.fields.AGENT_PING;
+        if (agentPing && agentPing !== 'ok' && agentPing !== 'skipped') {
+          notes.push(
+            "• Your assistant didn't reply to a test message. " +
+              'Check `logs/nanoclaw.log` for clues, then try `pnpm run chat hi`.',
+          );
+        }
+      }
+      if (!res.terminal?.fields.CONFIGURED_CHANNELS) {
+        notes.push('• Want to chat from your phone? Add a messaging app with `/add-telegram`, `/add-slack`, or `/add-discord`.');
+      }
+      if (notes.length > 0) {
+        p.note(notes.join('\n'), "What's left");
+      }
+      // "What's left" is a soft failure — we don't abort like fail(), but the
+      // user is still stuck and a fix is exactly what claude-assist is for.
+      const summary = notes
+        .map((n) => n.replace(/^•\s*/, '').split('\n')[0].trim())
+        .filter(Boolean)
+        .join(' · ');
+      phEmit('setup_incomplete', {
+        unresolved_count: notes.length,
+        service_running: res.terminal?.fields.SERVICE === 'running',
+        has_credentials: res.terminal?.fields.CREDENTIALS === 'configured',
+        agent_responds: res.terminal?.fields.AGENT_PING === 'ok',
+      });
+      await offerClaudeAssist({
+        stepName: 'verify',
+        msg: summary || 'Verification completed with unresolved issues.',
+        hint: `Terminal block: ${JSON.stringify(res.terminal?.fields ?? {})}`,
+        rawLogPath: res.rawLog,
+      });
+      p.outro(k.yellow('Almost there. A few things still need your attention.'));
+      return;
+    }
+  }
+
+  const rows: [string, string][] = [
+    ['Chat in the terminal:', 'pnpm run chat hi'],
+    ["See what's happening:", 'tail -f logs/nanoclaw.log'],
+    ['Open Claude Code:', 'claude'],
+  ];
+  const labelWidth = Math.max(...rows.map(([l]) => l.length));
+  const nextSteps = rows
+    .map(([l, c]) => `${k.cyan(l.padEnd(labelWidth))}  ${c}`)
+    .join('\n');
+  p.note(nextSteps, 'Try these');
+
+  // Always-on warning goes before the "check your DMs" directive so the
+  // caveat doesn't land after the user's already looked away at their phone.
+  p.note(
+    wrapForGutter(
+      "NanoClaw runs on this machine. It's only reachable while this computer is on and connected to the internet. For always-on availability, run it on a cloud VM — or keep this machine awake.",
+      6,
+    ),
+    'Heads up',
+  );
+
+  setupLog.complete(Date.now() - RUN_START);
+  phEmit('setup_completed', { duration_ms: Date.now() - RUN_START });
+
+  const dmTarget = channelDmLabel(channelChoice);
+  if (dmTarget) {
+    // Bright framed banner (not dim) — the whole point of the feedback was
+    // that the welcome-message signal was too easy to miss. Use p.note so it
+    // renders with a visible box, cyan-bold the directive line, and put it
+    // as the last thing before outro.
+    p.note(
+      `${brandBold('→')} ${k.bold(`Check your ${dmTarget} — your assistant is saying hi.`)}`,
+      'Go say hi',
+    );
+    p.outro(k.green("You're set."));
+  } else {
+    p.outro(k.green("You're ready! Chat with `pnpm run chat hi`."));
+  }
+}
+
+function channelDmLabel(choice: ChannelChoice): string | null {
+  switch (choice) {
+    case 'telegram':
+      return 'Telegram';
+    case 'discord':
+      return 'Discord DMs';
+    case 'whatsapp':
+      return 'WhatsApp';
+    case 'signal':
+      return 'Signal';
+    case 'teams':
+      return 'Teams';
+    case 'imessage':
+      return 'iMessage';
+    case 'slack':
+      // Slack install doesn't wire an agent or send a welcome DM — the
+      // driver prints its own "finish in your Slack app" note. Falling
+      // through to null avoids a misleading "check your Slack DMs" banner.
+      return null;
+    default:
+      return null;
+  }
+}
+
+// ─── first-chat step ───────────────────────────────────────────────────
+
+/**
+ * Round-trip ping against the CLI socket before we ask the user to chat.
+ * Renders its own spinner with elapsed time because a cold-start container
+ * boot can take 30–60s — the elapsed counter is the difference between
+ * "patient" and "is this hung?". Returns the raw result so the caller can
+ * branch between the chat loop (ok) and a diagnostic note (anything else).
+ */
+async function confirmAssistantResponds(): Promise<PingResult> {
+  const s = p.spinner();
+  const start = Date.now();
+  const label = 'Waking your assistant…';
+  s.start(fitToWidth(label, ' (999s)'));
+  const tick = setInterval(() => {
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    const suffix = ` (${elapsed}s)`;
+    s.message(`${fitToWidth(label, suffix)}${k.dim(suffix)}`);
+  }, 1000);
+
+  const result = await pingCliAgent();
+
+  clearInterval(tick);
+  const elapsed = Math.round((Date.now() - start) / 1000);
+  const suffix = ` (${elapsed}s)`;
+  if (result === 'ok') {
+    s.stop(`${k.bold(fitToWidth('Your assistant is ready.', suffix))}${k.dim(suffix)}`);
+  } else {
+    const msg =
+      result === 'socket_error'
+        ? "Couldn't reach the NanoClaw service."
+        : "Your assistant didn't reply in time.";
+    s.stop(`${k.bold(fitToWidth(msg, suffix))}${k.dim(suffix)}`, 1);
+  }
+  return result;
+}
+
+function renderPingFailureNote(result: PingResult): void {
+  const body =
+    result === 'socket_error'
+      ? [
+          wrapForGutter(
+            "The NanoClaw service isn't listening on its local socket. Try restarting it, then chat with `pnpm run chat hi`:",
+            6,
+          ),
+          '',
+          `  macOS:  launchctl kickstart -k gui/$(id -u)/${getLaunchdLabel()}`,
+          `  Linux:  systemctl --user restart ${getSystemdUnit()}`,
+        ].join('\n')
+      : wrapForGutter(
+          'No reply from your assistant within 30 seconds. Check `logs/nanoclaw.log` for clues, then try `pnpm run chat hi`.',
+          6,
+        );
+  p.note(body, 'Skipping the first chat');
+}
+
+/**
+ * Chat loop. Each message is piped through `pnpm run chat`, which uses
+ * the same Unix-socket path the ping just exercised, so output streams
+ * back inline as the agent replies. An empty input ends the loop.
+ *
+ * The intro note teaches the sandbox mental model — users reported being
+ * confused about what the terminal chat *is* (vs the phone channel they'd
+ * set up next) and what happens to the agent when they walk away. We
+ * explain once, then offer "message or Enter to continue" so the chat is
+ * clearly optional.
+ */
+async function runFirstChat(): Promise<void> {
+  p.note(
+    wrapForGutter(
+      [
+        'Your assistant runs in a sandbox on this machine.',
+        'It wakes up when you send a message and goes back to sleep when',
+        "you're not talking — so it isn't burning resources in the background.",
+        'Its memory and environment persist between conversations.',
+      ].join(' '),
+      6,
+    ),
+    'How this works',
+  );
+  let first = true;
+  while (true) {
+    const answer = ensureAnswer(
+      await p.text({
+        message: first
+          ? 'Try a quick hello — or press Enter to continue setup'
+          : 'Another message? Press Enter to continue setup',
+        placeholder: first
+          ? 'e.g. "hi, what can you do?"'
+          : 'press Enter to continue',
+      }),
+    );
+    first = false;
+    const text = ((answer as string | undefined) ?? '').trim();
+    if (!text) return;
+    await sendChatMessage(text);
+  }
+}
+
+function sendChatMessage(message: string): Promise<void> {
+  return new Promise((resolve) => {
+    // `pnpm --silent` suppresses the `> nanoclaw@… chat` preamble so the
+    // agent's reply reads as a clean block under the prompt. Splitting on
+    // whitespace mirrors `pnpm run chat hello world` — chat.ts joins argv
+    // with spaces on the far side.
+    const child = spawn(
+      'pnpm',
+      ['--silent', 'run', 'chat', ...message.split(/\s+/)],
+      { stdio: ['ignore', 'inherit', 'inherit'] },
+    );
+    child.on('close', () => resolve());
+    child.on('error', () => resolve());
+  });
+}
+
+// ─── auth step (select → branch) ────────────────────────────────────────
+
+async function runAuthStep(): Promise<void> {
+  if (anthropicSecretExists()) {
+    p.log.success('Your Claude account is already connected.');
+    setupLog.step('auth', 'skipped', 0, { REASON: 'secret-already-present' });
+    return;
+  }
+
+  const method = ensureAnswer(
+    await brightSelect({
+      message: 'How would you like to connect to Claude?',
+      options: [
+        {
+          value: 'subscription',
+          label: 'Sign in with my Claude subscription',
+          hint: 'recommended if you have Pro or Max',
+        },
+        {
+          value: 'oauth',
+          label: 'Paste an OAuth token I already have',
+          hint: 'sk-ant-oat…',
+        },
+        {
+          value: 'api',
+          label: 'Paste an Anthropic API key',
+          hint: 'pay-per-use via console.anthropic.com',
+        },
+      ],
+    }),
+  ) as 'subscription' | 'oauth' | 'api';
+  setupLog.userInput('auth_method', method);
+  phEmit('auth_method_chosen', { method });
+
+  if (method === 'subscription') {
+    await runSubscriptionAuth();
+  } else {
+    await runPasteAuth(method);
+  }
+}
+
+async function runSubscriptionAuth(): Promise<void> {
+  p.log.step("Opening the Claude sign-in flow…");
+  console.log(
+    k.dim('   (a browser will open for sign-in; this part is interactive)'),
+  );
+  console.log();
+  const start = Date.now();
+  const code = await runInheritScript('bash', [
+    'setup/register-claude-token.sh',
+  ]);
+  const durationMs = Date.now() - start;
+  console.log();
+  if (code !== 0) {
+    setupLog.step('auth', 'failed', durationMs, {
+      EXIT_CODE: code,
+      METHOD: 'subscription',
+    });
+    await fail(
+      'auth',
+      "Couldn't complete the Claude sign-in.",
+      'Re-run setup and try again, or choose a paste option instead.',
+    );
+  }
+  setupLog.step('auth', 'interactive', durationMs, { METHOD: 'subscription' });
+  p.log.success('Claude account connected.');
+}
+
+async function runPasteAuth(method: 'oauth' | 'api'): Promise<void> {
+  const label = method === 'oauth' ? 'OAuth token' : 'API key';
+  const prefix = method === 'oauth' ? 'sk-ant-oat' : 'sk-ant-api';
+
+  const answer = ensureAnswer(
+    await p.password({
+      message: `Paste your ${label}`,
+      validate: (v) => {
+        if (!v || !v.trim()) return 'Required';
+        if (!v.trim().startsWith(prefix)) {
+          return `Should start with ${prefix}…`;
+        }
+        return undefined;
+      },
+    }),
+  );
+  const token = (answer as string).trim();
+
+  const res = await runQuietChild(
+    'auth',
+    'onecli',
+    [
+      'secrets', 'create',
+      '--name', 'Anthropic',
+      '--type', 'anthropic',
+      '--value', token,
+      '--host-pattern', 'api.anthropic.com',
+    ],
+    {
+      running: `Saving your ${label} to your OneCLI vault…`,
+      done: 'Claude account connected.',
+    },
+    {
+      extraFields: { METHOD: method },
+    },
+  );
+  if (!res.ok) {
+    await fail(
+      'auth',
+      `Couldn't save your ${label} to the vault.`,
+      'Make sure OneCLI is running (`onecli version`), then retry.',
+    );
+  }
+}
+
+// ─── timezone step ─────────────────────────────────────────────────────
+
+/**
+ * Auto-detect TZ, confirm with the user when it comes back as UTC (a
+ * common sign we're on a VPS that wasn't localised), and persist through
+ * the usual `--step timezone -- --tz <zone>` path. Free-text answers get
+ * a headless `claude -p` pass to resolve them to a real IANA zone.
+ */
+async function runTimezoneStep(): Promise<void> {
+  const res = await runQuietStep('timezone', {
+    running: 'Checking your timezone…',
+    done: 'Timezone set.',
+  });
+  if (!res.ok && res.terminal?.fields.NEEDS_USER_INPUT !== 'true') {
+    await fail('timezone', "Couldn't determine your timezone.");
+  }
+
+  const fields = res.terminal?.fields ?? {};
+  const resolvedTz = fields.RESOLVED_TZ;
+  const needsInput = fields.NEEDS_USER_INPUT === 'true';
+  const isUtc =
+    resolvedTz === 'UTC' ||
+    resolvedTz === 'Etc/UTC' ||
+    resolvedTz === 'Universal';
+
+  // Three branches:
+  //   - no TZ detected: ask where they are (or leave as UTC)
+  //   - detected UTC: confirm (likely VPS, but worth checking)
+  //   - detected specific zone: confirm explicitly rather than silently
+  //     persisting — users shouldn't be surprised the agent "already knew"
+  //     their timezone from system settings they didn't think about.
+  if (!needsInput && !isUtc && resolvedTz && resolvedTz !== 'none') {
+    const confirmed = ensureAnswer(
+      await p.confirm({
+        message: `I detected ${resolvedTz} from your computer settings. Is that right?`,
+        initialValue: true,
+      }),
+    );
+    setupLog.userInput('timezone_confirm_detected', String(confirmed));
+    if (confirmed) return;
+  }
+
+  const message = needsInput
+    ? "Your system didn't expose a timezone. Which one are you in?"
+    : !isUtc
+      ? "Where are you, then?"
+      : "Your system reports UTC as the timezone. Is that right, or are you somewhere else?";
+
+  // For the non-UTC "detected-but-wrong" branch we skip the select and jump
+  // straight to the free-text prompt — the user already said "not that".
+  let choice: 'keep' | 'answer' = 'answer';
+  if (needsInput || isUtc) {
+    choice = ensureAnswer(
+      await brightSelect({
+        message,
+        options: needsInput
+          ? [
+              { value: 'answer', label: "I'll tell you where I am" },
+              { value: 'keep', label: 'Leave it as UTC' },
+            ]
+          : [
+              { value: 'keep', label: 'Keep UTC', hint: 'remote server / happy with UTC' },
+              { value: 'answer', label: "I'm somewhere else" },
+            ],
+      }),
+    ) as 'keep' | 'answer';
+    setupLog.userInput('timezone_choice', choice);
+  }
+
+  if (choice === 'keep') return;
+
+  const answer = ensureAnswer(
+    await p.text({
+      message: "Where are you? (city, region, or IANA zone)",
+      placeholder: 'e.g. New York, London, Asia/Tokyo',
+      validate: (v) => (v && v.trim() ? undefined : 'Required'),
+    }),
+  );
+  const raw = (answer as string).trim();
+  setupLog.userInput('timezone_input', raw);
+
+  let tz: string | null = isValidTimezone(raw) ? raw : null;
+  if (!tz) {
+    if (claudeCliAvailable()) {
+      tz = await resolveTimezoneViaClaude(raw);
+    } else {
+      p.log.warn(
+        wrapForGutter(
+          "That's not a standard IANA zone and I can't call Claude to interpret it here — try again with a zone like `America/New_York` or `Europe/London`.",
+          4,
+        ),
+      );
+    }
+  }
+
+  if (!tz) {
+    // One retry with a direct-IANA ask; if that fails too, leave the
+    // previously-detected value in .env and move on rather than looping.
+    const retryAnswer = ensureAnswer(
+      await p.text({
+        message: 'Enter an IANA timezone string',
+        placeholder: 'e.g. America/New_York',
+        validate: (v) => {
+          const s = (v ?? '').trim();
+          if (!s) return 'Required';
+          if (!isValidTimezone(s)) return 'Not a valid IANA zone';
+          return undefined;
+        },
+      }),
+    );
+    tz = (retryAnswer as string).trim();
+    setupLog.userInput('timezone_retry', tz);
+  }
+
+  const persist = await runQuietStep(
+    'timezone',
+    {
+      running: `Saving timezone ${tz}…`,
+      done: `Timezone set to ${tz}.`,
+    },
+    ['--tz', tz],
+  );
+  if (!persist.ok) {
+    await fail('timezone', `Couldn't save timezone ${tz}.`);
+  }
+}
+
+// ─── prompts owned by the sequencer ────────────────────────────────────
+
+async function askDisplayName(fallback: string): Promise<string> {
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'What should your assistant call you?',
+      placeholder: fallback,
+      defaultValue: fallback,
+    }),
+  );
+  const value = (answer as string).trim() || fallback;
+  setupLog.userInput('display_name', value);
+  return value;
+}
+
+async function askChannelChoice(): Promise<ChannelChoice> {
+  const isMac = process.platform === 'darwin';
+  const choice = ensureAnswer(
+    await brightSelect<ChannelChoice>({
+      message: 'Want to chat with your assistant from your phone?',
+      options: [
+        { value: 'telegram', label: 'Yes, connect Telegram', hint: 'recommended' },
+        { value: 'discord', label: 'Yes, connect Discord' },
+        { value: 'whatsapp', label: 'Yes, connect WhatsApp' },
+        {
+          value: 'signal',
+          label: 'Yes, connect Signal',
+          hint: 'needs signal-cli installed',
+        },
+        {
+          value: 'imessage',
+          label: 'Yes, connect iMessage (experimental)',
+          hint: isMac ? 'local macOS mode' : 'remote Photon only',
+        },
+        {
+          value: 'slack',
+          label: 'Yes, connect Slack (experimental)',
+          hint: 'needs public URL',
+        },
+        { value: 'teams', label: 'Yes, connect Microsoft Teams', hint: 'complex setup' },
+        { value: 'skip', label: 'Skip for now', hint: "I'll just use the terminal" },
+      ],
+    }),
+  );
+  setupLog.userInput('channel_choice', String(choice));
+  phEmit('channel_chosen', { channel: String(choice) });
+  return choice;
+}
+
+// ─── interactive / env helpers ─────────────────────────────────────────
+
+function anthropicSecretExists(): boolean {
+  try {
+    const res = spawnSync('onecli', ['secrets', 'list'], {
+      encoding: 'utf-8',
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    if (res.status !== 0) return false;
+    return /anthropic/i.test(res.stdout ?? '');
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Probe the host for a working OneCLI install so we can offer to reuse it
+ * instead of re-running the installer (which rebinds the listener and breaks
+ * any other app already using that gateway).
+ */
+function detectExistingOnecli(): { version: string; apiHost: string } | null {
+  try {
+    const ver = spawnSync('onecli', ['version'], {
+      encoding: 'utf-8',
+      stdio: ['ignore', 'pipe', 'ignore'],
+    });
+    if (ver.status !== 0) return null;
+    const version = (ver.stdout ?? '').trim();
+    if (!version) return null;
+
+    const host = spawnSync('onecli', ['config', 'get', 'api-host'], {
+      encoding: 'utf-8',
+      stdio: ['ignore', 'pipe', 'ignore'],
+    });
+    if (host.status !== 0) return null;
+    const raw = (host.stdout ?? '').trim();
+    if (!raw) return null;
+
+    // onecli 1.3+ emits JSON by default. Older versions would print raw text.
+    try {
+      const parsed = JSON.parse(raw) as { data?: unknown; value?: unknown };
+      const val = parsed.data ?? parsed.value;
+      if (typeof val === 'string' && val.trim()) {
+        return { version, apiHost: val.trim() };
+      }
+    } catch {
+      // not JSON — try to extract a URL directly
+    }
+    const m = raw.match(/https?:\/\/[\w.\-]+(?::\d+)?/);
+    return m ? { version, apiHost: m[0] } : null;
+  } catch {
+    return null;
+  }
+}
+
+function runInheritScript(cmd: string, args: string[]): Promise<number> {
+  return new Promise((resolve) => {
+    const child = spawn(cmd, args, { stdio: 'inherit' });
+    child.on('close', (code) => resolve(code ?? 1));
+  });
+}
+
+/**
+ * After installing Docker, this process's supplementary groups are still
+ * frozen from login — subsequent steps that talk to /var/run/docker.sock
+ * (onecli install, service start, …) fail with EACCES even though the
+ * daemon is up. Detect that and re-exec the whole driver under `sg docker`
+ * so the rest of the run inherits the docker group without a re-login.
+ */
+function maybeReexecUnderSg(): void {
+  if (process.env.NANOCLAW_REEXEC_SG === '1') return;
+  if (process.platform !== 'linux') return;
+  const info = spawnSync('docker', ['info'], { encoding: 'utf-8' });
+  if (info.status === 0) return;
+  const err = `${info.stderr ?? ''}\n${info.stdout ?? ''}`;
+  if (!/permission denied/i.test(err)) return;
+  if (spawnSync('which', ['sg'], { stdio: 'ignore' }).status !== 0) return;
+
+  p.log.warn('Docker socket not accessible in current group. Re-executing under `sg docker`.');
+  const res = spawnSync('sg', ['docker', '-c', 'pnpm run setup:auto'], {
+    stdio: 'inherit',
+    env: { ...process.env, NANOCLAW_REEXEC_SG: '1' },
+  });
+  process.exit(res.status ?? 1);
+}
+
+// ─── intro + progression-log init ──────────────────────────────────────
+
+function printIntro(): void {
+  const isReexec = process.env.NANOCLAW_REEXEC_SG === '1';
+  const wordmark = `${k.bold('Nano')}${brandBold('Claw')}`;
+
+  if (isReexec) {
+    p.intro(
+      `${brandChip(' Welcome ')}  ${wordmark}  ${k.dim('· picking up where we left off')}`,
+    );
+    return;
+  }
+
+  // Always include the wordmark inside the clack intro line. When bash ran
+  // first (NANOCLAW_BOOTSTRAPPED=1) it already printed its own wordmark
+  // above us; the small repeat is worth it to keep the brand anchored at
+  // the visible top of the clack session once the bash output scrolls away.
+  p.intro(`${wordmark}  ${k.dim("Let's get you set up.")}`);
+}
+
+/**
+ * Bootstrap (nanoclaw.sh) normally initializes logs/setup.log and writes
+ * the bootstrap entry before we even boot. If someone runs `pnpm run
+ * setup:auto` directly, start a fresh progression log here so we don't
+ * append to a stale one from a previous run.
+ */
+function initProgressionLog(): void {
+  if (process.env.NANOCLAW_BOOTSTRAPPED === '1') return;
+  let commit = '';
+  try {
+    commit = spawnSync('git', ['rev-parse', '--short', 'HEAD'], {
+      encoding: 'utf-8',
+    }).stdout.trim();
+  } catch {
+    // git not available or not a repo — skip
+  }
+  let branch = '';
+  try {
+    branch = spawnSync('git', ['branch', '--show-current'], {
+      encoding: 'utf-8',
+    }).stdout.trim();
+  } catch {
+    // skip
+  }
+  setupLog.reset({
+    invocation: 'setup:auto (standalone)',
+    user: process.env.USER ?? 'unknown',
+    cwd: process.cwd(),
+    branch: branch || 'unknown',
+    commit: commit || 'unknown',
+  });
+}
+
+main().catch((err) => {
+  p.log.error(err instanceof Error ? err.message : String(err));
+  p.cancel('Setup aborted.');
+  process.exit(1);
+});
diff --git a/setup/channels/discord.ts b/setup/channels/discord.ts
new file mode 100644
index 0000000..3668686
--- /dev/null
+++ b/setup/channels/discord.ts
@@ -0,0 +1,518 @@
+/**
+ * Discord channel flow for setup:auto.
+ *
+ * `runDiscordChannel(displayName)` owns the full branch from "do you have a
+ * bot?" through the welcome DM:
+ *
+ *   1. Ask if they have a bot already; walk them through Dev Portal creation
+ *      if not
+ *   2. Paste the bot token (clack password) — format-validated
+ *   3. GET /users/@me to confirm the token and resolve bot username
+ *   4. GET /oauth2/applications/@me to derive application_id, verify_key
+ *      (public key), and owner — no separate paste needed in the common case
+ *   5. Confirm owner identity (falls back to a manual user-id prompt with
+ *      Developer Mode instructions if declined or if the app is team-owned)
+ *   6. Print the OAuth invite URL, open it, wait for "I've added the bot"
+ *   7. Install the adapter via setup/add-discord.sh (non-interactive)
+ *   8. POST /users/@me/channels to open the DM channel (yields dm channel id)
+ *   9. Ask for the messaging-agent name (defaulting to "Nano")
+ *  10. Wire the agent via scripts/init-first-agent.ts, which sends the welcome
+ *      DM through the normal delivery path
+ *
+ * All output obeys the three-level contract: clack UI for the user, structured
+ * entries in logs/setup.log, full raw output in per-step files under
+ * logs/setup-steps/. See docs/setup-flow.md.
+ */
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { brightSelect } from '../lib/bright-select.js';
+import { confirmThenOpen } from '../lib/browser.js';
+import { askOperatorRole } from '../lib/role-prompt.js';
+import { ensureAnswer, fail, runQuietChild } from '../lib/runner.js';
+
+const DEFAULT_AGENT_NAME = 'Nano';
+const DISCORD_API = 'https://discord.com/api/v10';
+
+// Send Messages (0x800) + Add Reactions (0x40) + Attach Files (0x8000)
+//   + Read Message History (0x10000) = 100416.
+// Matches the permissions set documented in .claude/skills/add-discord/SKILL.md.
+const INVITE_PERMISSIONS = '100416';
+
+interface AppInfo {
+  applicationId: string;
+  publicKey: string;
+  owner: { id: string; username: string } | null;
+}
+
+export async function runDiscordChannel(displayName: string): Promise<void> {
+  const hasBot = await askHasBotToken();
+  if (!hasBot) {
+    await walkThroughBotCreation();
+  }
+  // Even users who said "yes" often can't find the token on demand — the
+  // Dev Portal resets it if you don't store it, and people forget which
+  // app it belongs to. A quick reminder before the paste prompt is cheap.
+  showTokenLocationReminder(hasBot);
+
+  const token = await collectDiscordToken();
+  const botUsername = await validateDiscordToken(token);
+  const app = await fetchApplicationInfo(token);
+
+  const ownerUserId = await resolveOwnerUserId(app.owner);
+
+  // Before inviting: do they have a server to invite into? Walkthrough if
+  // not — a fresh Discord account without a server makes the invite page a
+  // dead end.
+  if (!(await askHasDiscordServer())) {
+    await walkThroughServerCreation();
+  }
+
+  await promptInviteBot(app.applicationId, botUsername);
+
+  const install = await runQuietChild(
+    'discord-install',
+    'bash',
+    ['setup/add-discord.sh'],
+    {
+      running: `Connecting Discord to @${botUsername}…`,
+      done: 'Discord connected.',
+    },
+    {
+      env: {
+        DISCORD_BOT_TOKEN: token,
+        DISCORD_APPLICATION_ID: app.applicationId,
+        DISCORD_PUBLIC_KEY: app.publicKey,
+      },
+      extraFields: {
+        BOT_USERNAME: botUsername,
+        APPLICATION_ID: app.applicationId,
+      },
+    },
+  );
+  if (!install.ok) {
+    await fail(
+      'discord-install',
+      "Couldn't connect Discord.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+
+  const dmChannelId = await openDmChannel(token, ownerUserId);
+  const platformId = `discord:@me:${dmChannelId}`;
+
+  const role = await askOperatorRole('Discord');
+  setupLog.userInput('discord_role', role);
+
+  const agentName = await resolveAgentName();
+
+  const init = await runQuietChild(
+    'init-first-agent',
+    'pnpm',
+    [
+      'exec', 'tsx', 'scripts/init-first-agent.ts',
+      '--channel', 'discord',
+      '--user-id', `discord:${ownerUserId}`,
+      '--platform-id', platformId,
+      '--display-name', displayName,
+      '--agent-name', agentName,
+      '--role', role,
+    ],
+    {
+      running: `Connecting ${agentName} to your Discord DMs…`,
+      done: `${agentName} is ready. Check Discord for a welcome message.`,
+    },
+    {
+      extraFields: {
+        CHANNEL: 'discord',
+        AGENT_NAME: agentName,
+        PLATFORM_ID: platformId,
+      },
+    },
+  );
+  if (!init.ok) {
+    await fail(
+      'init-first-agent',
+      `Couldn't finish connecting ${agentName}.`,
+      'Most likely the bot and you don\'t share a server yet — invite the bot, then retry later with `/manage-channels`.',
+    );
+  }
+}
+
+async function askHasBotToken(): Promise<boolean> {
+  const answer = ensureAnswer(
+    await brightSelect({
+      message: 'Do you already have a Discord bot?',
+      options: [
+        { value: 'yes', label: 'Yes, I have a bot token ready' },
+        { value: 'no', label: "No, walk me through creating one" },
+      ],
+    }),
+  );
+  return answer === 'yes';
+}
+
+async function walkThroughBotCreation(): Promise<void> {
+  const url = 'https://discord.com/developers/applications';
+  p.note(
+    [
+      "You'll create a Discord bot in the Developer Portal. It's free and takes about a minute.",
+      '',
+      '  1. Click "New Application", give it a name (e.g. "NanoClaw")',
+      '  2. In the "Bot" tab, click "Reset Token" and copy the token',
+      '  3. On the same tab, enable "Message Content Intent"',
+      '     (under Privileged Gateway Intents)',
+      '',
+      k.dim(url),
+    ].join('\n'),
+    'Create a Discord bot',
+  );
+  await confirmThenOpen(url, 'Press Enter to open the Developer Portal');
+
+  ensureAnswer(
+    await p.confirm({
+      message: "Got your bot token?",
+      initialValue: true,
+    }),
+  );
+}
+
+function showTokenLocationReminder(hasExistingBot: boolean): void {
+  // If we just walked them through creating a bot, they're staring at the
+  // token. If they came in with an existing one, they may still need a nudge
+  // to find it — tokens in the Dev Portal aren't visible after first reveal,
+  // and "Reset Token" issues a new one.
+  if (hasExistingBot) {
+    p.note(
+      [
+        "Where to find your bot token:",
+        '',
+        '  1. discord.com/developers/applications → pick your app',
+        '  2. "Bot" tab → "Reset Token" (the old one stops working)',
+        '  3. Copy the new token',
+      ].join('\n'),
+      'Reminder',
+    );
+  }
+}
+
+async function askHasDiscordServer(): Promise<boolean> {
+  const answer = ensureAnswer(
+    await brightSelect({
+      message: 'Do you have a Discord server you can add the bot to?',
+      options: [
+        { value: 'yes', label: 'Yes, I have a server' },
+        { value: 'no', label: "No, walk me through creating one" },
+      ],
+    }),
+  );
+  setupLog.userInput('discord_has_server', String(answer));
+  return answer === 'yes';
+}
+
+async function walkThroughServerCreation(): Promise<void> {
+  // Discord doesn't have a stable deep-link for "create server" so we open
+  // the web client and rely on the + button being visible. The steps below
+  // are the same whether they're in the desktop app or the browser.
+  const url = 'https://discord.com/channels/@me';
+  p.note(
+    [
+      "A Discord server is just a private space for you and the bot. Free and takes 30 seconds.",
+      '',
+      '  1. In Discord, click the "+" at the bottom of the server list',
+      '  2. Choose "Create My Own" → "For me and my friends"',
+      '  3. Give it any name (e.g. "NanoClaw")',
+      '',
+      k.dim(url),
+    ].join('\n'),
+    'Create a Discord server',
+  );
+  await confirmThenOpen(url, 'Press Enter to open Discord');
+
+  ensureAnswer(
+    await p.confirm({
+      message: "Server created?",
+      initialValue: true,
+    }),
+  );
+}
+
+async function collectDiscordToken(): Promise<string> {
+  const answer = ensureAnswer(
+    await p.password({
+      message: 'Paste your bot token',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'Token is required';
+        // Discord bot tokens are base64url segments separated by dots.
+        // Be lenient on length; the real check is /users/@me.
+        if (!/^[A-Za-z0-9._-]{50,}$/.test(t)) {
+          return "That doesn't look like a Discord bot token";
+        }
+        return undefined;
+      },
+    }),
+  );
+  const token = (answer as string).trim();
+  setupLog.userInput(
+    'discord_token',
+    `${token.slice(0, 10)}…${token.slice(-4)}`,
+  );
+  return token;
+}
+
+async function validateDiscordToken(token: string): Promise<string> {
+  const s = p.spinner();
+  const start = Date.now();
+  s.start('Checking your bot token…');
+  try {
+    const res = await fetch(`${DISCORD_API}/users/@me`, {
+      headers: { Authorization: `Bot ${token}` },
+    });
+    const data = (await res.json()) as {
+      id?: string;
+      username?: string;
+      message?: string;
+    };
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    if (res.ok && data.username) {
+      s.stop(`Found your bot: @${data.username}. ${k.dim(`(${elapsedS}s)`)}`);
+      setupLog.step('discord-validate', 'success', Date.now() - start, {
+        BOT_USERNAME: data.username,
+        BOT_ID: data.id ?? '',
+      });
+      return data.username;
+    }
+    const reason = data.message ?? `HTTP ${res.status}`;
+    s.stop(`Discord didn't accept that token: ${reason}`, 1);
+    setupLog.step('discord-validate', 'failed', Date.now() - start, {
+      ERROR: reason,
+    });
+    await fail(
+      'discord-validate',
+      "Discord didn't accept that token.",
+      'Copy the token again from the Developer Portal and retry setup.',
+    );
+  } catch (err) {
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    s.stop(`Couldn't reach Discord. ${k.dim(`(${elapsedS}s)`)}`, 1);
+    const message = err instanceof Error ? err.message : String(err);
+    setupLog.step('discord-validate', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    await fail(
+      'discord-validate',
+      "Couldn't reach Discord.",
+      'Check your internet connection and retry setup.',
+    );
+  }
+}
+
+async function fetchApplicationInfo(token: string): Promise<AppInfo> {
+  const s = p.spinner();
+  const start = Date.now();
+  s.start('Looking up your bot application…');
+  try {
+    const res = await fetch(`${DISCORD_API}/oauth2/applications/@me`, {
+      headers: { Authorization: `Bot ${token}` },
+    });
+    const data = (await res.json()) as {
+      id?: string;
+      verify_key?: string;
+      owner?: { id: string; username: string } | null;
+      team?: unknown;
+      message?: string;
+    };
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    if (!res.ok || !data.id || !data.verify_key) {
+      const reason = data.message ?? `HTTP ${res.status}`;
+      s.stop(`Couldn't read application info: ${reason}`, 1);
+      setupLog.step('discord-app-info', 'failed', Date.now() - start, {
+        ERROR: reason,
+      });
+      await fail(
+        'discord-app-info',
+        "Couldn't read your Discord application details.",
+        'Re-run setup. If it keeps failing, check the bot token has the right scopes.',
+      );
+    }
+    s.stop(`Got your application details. ${k.dim(`(${elapsedS}s)`)}`);
+    // owner is populated for solo applications; team-owned apps return a
+    // team object instead and we'll fall back to a manual user-id prompt.
+    const owner =
+      data.owner && data.owner.id && data.owner.username
+        ? { id: data.owner.id, username: data.owner.username }
+        : null;
+    setupLog.step('discord-app-info', 'success', Date.now() - start, {
+      APPLICATION_ID: data.id,
+      OWNER_USERNAME: owner?.username ?? '',
+      TEAM_OWNED: data.team ? 'true' : 'false',
+    });
+    return {
+      applicationId: data.id,
+      publicKey: data.verify_key,
+      owner,
+    };
+  } catch (err) {
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    s.stop(`Couldn't reach Discord. ${k.dim(`(${elapsedS}s)`)}`, 1);
+    const message = err instanceof Error ? err.message : String(err);
+    setupLog.step('discord-app-info', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    await fail(
+      'discord-app-info',
+      "Couldn't reach Discord.",
+      'Check your internet connection and retry setup.',
+    );
+  }
+}
+
+async function resolveOwnerUserId(
+  owner: { id: string; username: string } | null,
+): Promise<string> {
+  if (owner) {
+    const confirmed = ensureAnswer(
+      await p.confirm({
+        message: `Is @${owner.username} your Discord account?`,
+        initialValue: true,
+      }),
+    );
+    if (confirmed === true) {
+      setupLog.userInput('discord_owner_confirmed', owner.username);
+      return owner.id;
+    }
+  } else {
+    p.log.info(
+      "Your bot is owned by a Developer Team, so we need your Discord user ID directly.",
+    );
+  }
+  return await promptForUserIdWithDevMode();
+}
+
+async function promptForUserIdWithDevMode(): Promise<string> {
+  p.note(
+    [
+      "To get your Discord user ID:",
+      '',
+      '  1. Open Discord → Settings (⚙️) → Advanced',
+      '  2. Turn on "Developer Mode"',
+      '  3. Right-click your own name/avatar → "Copy User ID"',
+    ].join('\n'),
+    'Find your Discord user ID',
+  );
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'Paste your Discord user ID',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'User ID is required';
+        if (!/^\d{17,20}$/.test(t)) {
+          return "That doesn't look like a Discord user ID (17-20 digits)";
+        }
+        return undefined;
+      },
+    }),
+  );
+  const id = (answer as string).trim();
+  setupLog.userInput('discord_user_id', id);
+  return id;
+}
+
+async function promptInviteBot(
+  applicationId: string,
+  botUsername: string,
+): Promise<void> {
+  const url =
+    `https://discord.com/api/oauth2/authorize` +
+    `?client_id=${applicationId}` +
+    `&scope=bot` +
+    `&permissions=${INVITE_PERMISSIONS}`;
+
+  p.note(
+    [
+      `@${botUsername} needs to share a server with you before it can DM you.`,
+      '',
+      '  1. Pick any server you\'re in (a personal one is fine)',
+      '  2. Click "Authorize"',
+      '',
+      k.dim(url),
+    ].join('\n'),
+    'Add bot to a server',
+  );
+  await confirmThenOpen(url, 'Press Enter to open the invite page');
+
+  ensureAnswer(
+    await p.confirm({
+      message: "I've added the bot to a server",
+      initialValue: true,
+    }),
+  );
+}
+
+async function openDmChannel(token: string, userId: string): Promise<string> {
+  const s = p.spinner();
+  const start = Date.now();
+  s.start('Opening a DM channel…');
+  try {
+    const res = await fetch(`${DISCORD_API}/users/@me/channels`, {
+      method: 'POST',
+      headers: {
+        Authorization: `Bot ${token}`,
+        'Content-Type': 'application/json',
+      },
+      body: JSON.stringify({ recipient_id: userId }),
+    });
+    const data = (await res.json()) as { id?: string; message?: string };
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    if (!res.ok || !data.id) {
+      const reason = data.message ?? `HTTP ${res.status}`;
+      s.stop(`Couldn't open a DM channel: ${reason}`, 1);
+      setupLog.step('discord-open-dm', 'failed', Date.now() - start, {
+        ERROR: reason,
+      });
+      await fail(
+        'discord-open-dm',
+        "Couldn't open a DM channel with you.",
+        'Make sure the bot is in a server you\'re also in, then retry setup.',
+      );
+    }
+    s.stop(`DM channel ready. ${k.dim(`(${elapsedS}s)`)}`);
+    setupLog.step('discord-open-dm', 'success', Date.now() - start, {
+      DM_CHANNEL_ID: data.id,
+    });
+    return data.id;
+  } catch (err) {
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    s.stop(`Couldn't reach Discord. ${k.dim(`(${elapsedS}s)`)}`, 1);
+    const message = err instanceof Error ? err.message : String(err);
+    setupLog.step('discord-open-dm', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    await fail(
+      'discord-open-dm',
+      "Couldn't reach Discord.",
+      'Check your internet connection and retry setup.',
+    );
+  }
+}
+
+async function resolveAgentName(): Promise<string> {
+  const preset = process.env.NANOCLAW_AGENT_NAME?.trim();
+  if (preset) {
+    setupLog.userInput('agent_name', preset);
+    return preset;
+  }
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'What should your assistant be called?',
+      placeholder: DEFAULT_AGENT_NAME,
+      defaultValue: DEFAULT_AGENT_NAME,
+    }),
+  );
+  const value = (answer as string).trim() || DEFAULT_AGENT_NAME;
+  setupLog.userInput('agent_name', value);
+  return value;
+}
+
diff --git a/setup/channels/imessage.ts b/setup/channels/imessage.ts
new file mode 100644
index 0000000..d8b129f
--- /dev/null
+++ b/setup/channels/imessage.ts
@@ -0,0 +1,314 @@
+/**
+ * iMessage channel flow for setup:auto.
+ *
+ * `runIMessageChannel(displayName)` covers both deployment modes:
+ *
+ *   Local (macOS): the bot runs on this Mac and talks via the signed-in
+ *   iMessage account. Reading chat.db needs Full Disk Access granted to
+ *   the Node binary — we open the directory for them so they can drag
+ *   the `node` file into System Settings.
+ *
+ *   Remote (Photon API): the bot talks to a separate server (Photon)
+ *   that owns an iMessage account on another Mac. Used when this host
+ *   is Linux, or when the operator wants to keep their daily-driver
+ *   Mac's chat history out of the loop.
+ *
+ * Flow:
+ *   1. Pick mode (auto-defaults to local on macOS, remote elsewhere)
+ *   2. Local: FDA walkthrough (open node bin directory, wait for ack)
+ *      Remote: prompt for Photon server URL + API key
+ *   3. Ask for the phone or email the operator messages from — this is
+ *      the platform-id for first-agent wiring
+ *   4. Install the adapter (setup/add-imessage.sh, non-interactive)
+ *   5. Wire the agent via scripts/init-first-agent.ts — the welcome
+ *      iMessage goes out through the normal delivery path
+ *
+ * All output obeys the three-level contract. See docs/setup-flow.md.
+ */
+import { execSync } from 'child_process';
+import os from 'os';
+import path from 'path';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { brightSelect } from '../lib/bright-select.js';
+import { askOperatorRole } from '../lib/role-prompt.js';
+import { ensureAnswer, fail, runQuietChild } from '../lib/runner.js';
+import { wrapForGutter } from '../lib/theme.js';
+
+const DEFAULT_AGENT_NAME = 'Nano';
+
+type Mode = 'local' | 'remote';
+
+interface RemoteCreds {
+  serverUrl: string;
+  apiKey: string;
+}
+
+export async function runIMessageChannel(displayName: string): Promise<void> {
+  const isMac = os.platform() === 'darwin';
+
+  const mode = await askMode(isMac);
+  let remoteCreds: RemoteCreds | null = null;
+
+  if (mode === 'local') {
+    if (!isMac) {
+      await fail(
+        'imessage',
+        "Local iMessage mode only works on macOS.",
+        'Choose remote mode (Photon API) on Linux/WSL, or run setup from your Mac.',
+      );
+    }
+    await walkThroughFullDiskAccess();
+  } else {
+    remoteCreds = await collectRemoteCreds();
+  }
+
+  const handle = await askOperatorHandle();
+
+  const install = await runQuietChild(
+    'imessage-install',
+    'bash',
+    ['setup/add-imessage.sh'],
+    {
+      running:
+        mode === 'local'
+          ? "Connecting the iMessage adapter to this Mac…"
+          : `Connecting the iMessage adapter to ${remoteCreds!.serverUrl}…`,
+      done: 'iMessage adapter installed.',
+    },
+    {
+      env:
+        mode === 'local'
+          ? { IMESSAGE_LOCAL: 'true', IMESSAGE_ENABLED: 'true' }
+          : {
+              IMESSAGE_LOCAL: 'false',
+              IMESSAGE_SERVER_URL: remoteCreds!.serverUrl,
+              IMESSAGE_API_KEY: remoteCreds!.apiKey,
+            },
+      extraFields: { MODE: mode },
+    },
+  );
+  if (!install.ok) {
+    await fail(
+      'imessage-install',
+      "Couldn't install the iMessage adapter.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+
+  const role = await askOperatorRole('iMessage');
+  setupLog.userInput('imessage_role', role);
+
+  const agentName = await resolveAgentName();
+
+  const init = await runQuietChild(
+    'init-first-agent',
+    'pnpm',
+    [
+      'exec', 'tsx', 'scripts/init-first-agent.ts',
+      '--channel', 'imessage',
+      '--user-id', handle,
+      '--platform-id', handle,
+      '--display-name', displayName,
+      '--agent-name', agentName,
+      '--role', role,
+    ],
+    {
+      running: `Connecting ${agentName} to iMessage…`,
+      done: `${agentName} is ready. Check iMessage for a welcome message.`,
+    },
+    {
+      extraFields: {
+        CHANNEL: 'imessage',
+        AGENT_NAME: agentName,
+        PLATFORM_ID: handle,
+        MODE: mode,
+      },
+    },
+  );
+  if (!init.ok) {
+    await fail(
+      'init-first-agent',
+      `Couldn't finish connecting ${agentName}.`,
+      'Double-check Full Disk Access (local mode) or Photon credentials (remote), then retry.',
+    );
+  }
+}
+
+async function askMode(isMac: boolean): Promise<Mode> {
+  const choice = ensureAnswer(
+    await brightSelect<Mode>({
+      message: 'How should iMessage run?',
+      initialValue: isMac ? 'local' : 'remote',
+      options: isMac
+        ? [
+            {
+              value: 'local',
+              label: 'Local (this Mac)',
+              hint: "uses this machine's iMessage account",
+            },
+            {
+              value: 'remote',
+              label: 'Remote (Photon API)',
+              hint: 'the bot lives on another server',
+            },
+          ]
+        : [
+            {
+              value: 'remote',
+              label: 'Remote (Photon API)',
+              hint: 'only option off macOS',
+            },
+          ],
+    }),
+  );
+  setupLog.userInput('imessage_mode', String(choice));
+  return choice;
+}
+
+/**
+ * Grant Full Disk Access to the Node binary the host runs under — without
+ * it, the adapter can't read chat.db and inbound messages never arrive.
+ * Opening the containing directory in Finder makes the drag-and-drop
+ * target obvious; falling back to printing the path keeps us working in
+ * SSH/headless contexts where `open` is a no-op.
+ */
+async function walkThroughFullDiskAccess(): Promise<void> {
+  let nodePath = process.execPath;
+  try {
+    // `which node` picks up the user's shell-resolved node, which may differ
+    // from process.execPath (e.g. they launched setup under a different
+    // Node via `nvm`). If it succeeds and is resolvable, prefer it.
+    const which = execSync('which node', { encoding: 'utf-8' }).trim();
+    if (which) nodePath = which;
+  } catch {
+    // fall back to process.execPath
+  }
+  const nodeDir = path.dirname(nodePath);
+
+  p.note(
+    wrapForGutter(
+      [
+        `iMessage needs Full Disk Access granted to the Node binary:`,
+        '',
+        `  ${nodePath}`,
+        '',
+        '  1. System Settings → Privacy & Security → Full Disk Access',
+        `  2. Click +, then drag the "node" file from the Finder window`,
+        '     we just opened for you',
+        '  3. Toggle it on, then come back here',
+      ].join('\n'),
+      6,
+    ),
+    'Grant Full Disk Access',
+  );
+
+  try {
+    execSync(`open "${nodeDir}"`, { stdio: 'ignore' });
+  } catch {
+    // No Finder (SSH/headless) — user sees the path in the note above.
+  }
+
+  ensureAnswer(
+    await p.confirm({
+      message: "Granted Full Disk Access?",
+      initialValue: true,
+    }),
+  );
+  setupLog.userInput('imessage_fda_confirmed', 'true');
+}
+
+async function collectRemoteCreds(): Promise<RemoteCreds> {
+  p.note(
+    [
+      "Photon is a separate service that owns an iMessage account and",
+      "exposes it over HTTP. NanoClaw will talk to it via its API.",
+      '',
+      '  1. Set up a Photon server: https://photon.im',
+      '  2. Copy the server URL and API key from your Photon dashboard',
+    ].join('\n'),
+    'Remote iMessage via Photon',
+  );
+
+  const urlAnswer = ensureAnswer(
+    await p.text({
+      message: 'Photon server URL',
+      placeholder: 'https://photon.example.com',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'URL is required';
+        if (!/^https?:\/\//i.test(t)) return 'Must start with http:// or https://';
+        return undefined;
+      },
+    }),
+  );
+  const serverUrl = (urlAnswer as string).trim();
+
+  const keyAnswer = ensureAnswer(
+    await p.password({
+      message: 'Photon API key',
+      validate: (v) => ((v ?? '').trim() ? undefined : 'API key is required'),
+    }),
+  );
+  const apiKey = (keyAnswer as string).trim();
+
+  setupLog.userInput('imessage_server_url', serverUrl);
+  setupLog.userInput(
+    'imessage_api_key',
+    `${apiKey.slice(0, 4)}…${apiKey.slice(-4)}`,
+  );
+  return { serverUrl, apiKey };
+}
+
+async function askOperatorHandle(): Promise<string> {
+  p.note(
+    [
+      "What phone number or email do you iMessage with?",
+      "That's where your assistant will send its welcome message.",
+      '',
+      k.dim('  • Phone: full E.164, e.g. +15551234567'),
+      k.dim('  • Email: whatever iMessage recognises (Apple ID, iCloud alias, …)'),
+    ].join('\n'),
+    'Your iMessage handle',
+  );
+
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'Phone number or email',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'Required';
+        const isPhone = /^\+\d{8,15}$/.test(t);
+        const isEmail = /^[^\s@]+@[^\s@]+\.[^\s@]+$/.test(t);
+        if (!isPhone && !isEmail) {
+          return "Use a +E.164 phone number or an email address";
+        }
+        return undefined;
+      },
+    }),
+  );
+  const handle = (answer as string).trim();
+  setupLog.userInput('imessage_handle', handle);
+  return handle;
+}
+
+async function resolveAgentName(): Promise<string> {
+  const preset = process.env.NANOCLAW_AGENT_NAME?.trim();
+  if (preset) {
+    setupLog.userInput('agent_name', preset);
+    return preset;
+  }
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'What should your assistant be called?',
+      placeholder: DEFAULT_AGENT_NAME,
+      defaultValue: DEFAULT_AGENT_NAME,
+    }),
+  );
+  const value = (answer as string).trim() || DEFAULT_AGENT_NAME;
+  setupLog.userInput('agent_name', value);
+  return value;
+}
diff --git a/setup/channels/signal.ts b/setup/channels/signal.ts
new file mode 100644
index 0000000..9e54cb9
--- /dev/null
+++ b/setup/channels/signal.ts
@@ -0,0 +1,357 @@
+/**
+ * Signal channel flow for setup:auto.
+ *
+ * `runSignalChannel(displayName)` owns the full branch from signal-cli
+ * presence check through the welcome DM:
+ *
+ *   1. Probe signal-cli on PATH (or SIGNAL_CLI_PATH). On macOS without it,
+ *      offer `brew install signal-cli` inline. On Linux, surface the
+ *      GitHub releases URL and bail with an actionable error.
+ *   2. Install the adapter + qrcode via setup/add-signal.sh (idempotent).
+ *   3. Run the signal-auth step, rendering each SIGNAL_AUTH_QR block as
+ *      a terminal QR the operator scans from Signal → Linked Devices.
+ *   4. Persist SIGNAL_ACCOUNT to .env (+ data/env/env).
+ *   5. Kick the service so the adapter picks up the new credentials.
+ *   6. Ask operator role + agent name.
+ *   7. Wire the agent via scripts/init-first-agent.ts; the existing welcome
+ *      DM path delivers the greeting through the adapter.
+ *
+ * Signal's `link` flow creates a *secondary* device. The phone number
+ * comes from the primary (the phone that scanned the QR); this host then
+ * sends/receives as that primary number. No registration of new numbers.
+ *
+ * Output obeys the three-level contract: clack UI for the user, structured
+ * entries in logs/setup.log, full raw output in per-step files under
+ * logs/setup-steps/. See docs/setup-flow.md.
+ */
+import { spawnSync } from 'child_process';
+import fs from 'fs';
+import path from 'path';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { getLaunchdLabel, getSystemdUnit } from '../../src/install-slug.js';
+import {
+  type Block,
+  type StepResult,
+  dumpTranscriptOnFailure,
+  ensureAnswer,
+  fail,
+  runQuietChild,
+  spawnStep,
+  writeStepEntry,
+} from '../lib/runner.js';
+import { askOperatorRole } from '../lib/role-prompt.js';
+
+const DEFAULT_AGENT_NAME = 'Nano';
+
+export async function runSignalChannel(displayName: string): Promise<void> {
+  await ensureSignalCli();
+
+  const install = await runQuietChild(
+    'signal-install',
+    'bash',
+    ['setup/add-signal.sh'],
+    {
+      running: 'Installing the Signal adapter…',
+      done: 'Signal adapter installed.',
+      skipped: 'Signal adapter already installed.',
+    },
+  );
+  if (!install.ok) {
+    await fail(
+      'signal-install',
+      "Couldn't install the Signal adapter.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+
+  const auth = await runSignalAuth();
+  if (!auth.ok) {
+    const reason = auth.terminal?.fields.ERROR ?? 'unknown';
+    await fail(
+      'signal-auth',
+      `Signal link failed (${reason}).`,
+      reason === 'qr_timeout'
+        ? 'The code expired. Re-run setup to get a fresh one.'
+        : 'Re-run setup to try again.',
+    );
+  }
+
+  const account = auth.terminal?.fields.ACCOUNT;
+  if (!account) {
+    await fail(
+      'signal-auth',
+      'Linked with Signal but couldn\'t read the phone number back.',
+      'Run `signal-cli listAccounts` to confirm, then re-run setup.',
+    );
+  }
+
+  writeSignalAccount(account!);
+  await restartService();
+
+  const role = await askOperatorRole('Signal');
+  setupLog.userInput('signal_role', role);
+
+  const agentName = await resolveAgentName();
+
+  const init = await runQuietChild(
+    'init-first-agent',
+    'pnpm',
+    [
+      'exec', 'tsx', 'scripts/init-first-agent.ts',
+      '--channel', 'signal',
+      '--user-id', account!,
+      '--platform-id', account!,
+      '--display-name', displayName,
+      '--agent-name', agentName,
+      '--role', role,
+    ],
+    {
+      running: `Connecting ${agentName} to Signal…`,
+      done: `${agentName} is ready. Check Signal for a welcome message.`,
+    },
+    {
+      extraFields: {
+        CHANNEL: 'signal',
+        AGENT_NAME: agentName,
+        PLATFORM_ID: account!,
+        ROLE: role,
+      },
+    },
+  );
+  if (!init.ok) {
+    await fail(
+      'init-first-agent',
+      `Couldn't finish connecting ${agentName}.`,
+      'You can retry later with `/manage-channels`.',
+    );
+  }
+}
+
+async function ensureSignalCli(): Promise<void> {
+  const cli = process.env.SIGNAL_CLI_PATH || 'signal-cli';
+  const probe = spawnSync(cli, ['--version'], {
+    stdio: ['ignore', 'pipe', 'pipe'],
+  });
+  if (!probe.error && probe.status === 0) return;
+
+  if (process.platform === 'darwin') {
+    p.note(
+      [
+        "NanoClaw talks to Signal through signal-cli, which isn't installed yet.",
+        '',
+        'The quickest way on macOS is Homebrew:',
+        '',
+        k.cyan('  brew install signal-cli'),
+        '',
+        "Install it in another terminal, then re-run setup.",
+      ].join('\n'),
+      'signal-cli not found',
+    );
+  } else {
+    p.note(
+      [
+        "NanoClaw talks to Signal through signal-cli, which isn't installed yet.",
+        '',
+        'Grab the latest release from GitHub:',
+        '',
+        k.cyan('  https://github.com/AsamK/signal-cli/releases'),
+        '',
+        "Install it, make sure `signal-cli --version` works, then re-run setup.",
+      ].join('\n'),
+      'signal-cli not found',
+    );
+  }
+  await fail(
+    'signal-install',
+    'signal-cli is required but not installed.',
+    'Install it and re-run setup.',
+  );
+}
+
+async function runSignalAuth(): Promise<
+  StepResult & { rawLog: string; durationMs: number }
+> {
+  const rawLog = setupLog.stepRawLog('signal-auth');
+  const start = Date.now();
+  const s = p.spinner();
+  s.start('Starting Signal link…');
+  let spinnerActive = true;
+
+  const stopSpinner = (msg: string, code?: number): void => {
+    if (spinnerActive) {
+      s.stop(msg, code);
+      spinnerActive = false;
+    }
+  };
+
+  // Tracks how many lines the QR block occupies so we can wipe it in-place
+  // once linking succeeds (Signal's link URL doesn't rotate like WhatsApp's,
+  // but we still want to erase the QR from screen once it's served).
+  let qrLinesPrinted = 0;
+
+  const result = await spawnStep(
+    'signal-auth',
+    [],
+    (block: Block) => {
+      if (block.type === 'SIGNAL_AUTH_QR') {
+        const qr = block.fields.QR ?? '';
+        if (!qr) return;
+        void renderQr(qr).then((lines) => {
+          stopSpinner('Scan this QR from Signal → Settings → Linked Devices.');
+          process.stdout.write(lines.join('\n') + '\n');
+          qrLinesPrinted = lines.length;
+          s.start('Waiting for you to scan…');
+          spinnerActive = true;
+        });
+      } else if (block.type === 'SIGNAL_AUTH') {
+        const status = block.fields.STATUS;
+        // Wipe the QR block regardless of outcome — it's either scanned
+        // and useless, or expired and misleading.
+        if (qrLinesPrinted > 0) {
+          process.stdout.write(`\x1b[${qrLinesPrinted}A\x1b[0J`);
+          qrLinesPrinted = 0;
+        }
+        const account = block.fields.ACCOUNT;
+        if (status === 'skipped') {
+          stopSpinner(
+            account
+              ? `Signal already linked as ${k.cyan(account)}.`
+              : 'Signal already linked.',
+          );
+        } else if (status === 'success') {
+          stopSpinner(`Signal linked as ${k.cyan(String(account ?? ''))}.`);
+        } else if (status === 'failed') {
+          const err = block.fields.ERROR ?? 'unknown';
+          stopSpinner(`Signal link failed: ${err}`, 1);
+        }
+      }
+    },
+    rawLog,
+  );
+  const durationMs = Date.now() - start;
+
+  if (spinnerActive) {
+    stopSpinner(
+      result.ok ? 'Done.' : 'Signal link ended unexpectedly.',
+      result.ok ? 0 : 1,
+    );
+    if (!result.ok) dumpTranscriptOnFailure(result.transcript);
+  }
+
+  writeStepEntry('signal-auth', result, durationMs, rawLog);
+  return { ...result, rawLog, durationMs };
+}
+
+/**
+ * Render the raw linking URL as a block-art QR, returned line-by-line so
+ * the caller can count lines for in-place cleanup. Uses small-mode so the
+ * code stays scannable on 24-row terminals. If qrcode isn't installed
+ * (add-signal.sh should have handled it, but we're defensive), fall back
+ * to the raw URL and ask the user to paste it into an external renderer.
+ */
+async function renderQr(url: string): Promise<string[]> {
+  try {
+    const QRCode = await import('qrcode');
+    const qrText = await QRCode.toString(url, { type: 'terminal', small: true });
+    const caption = k.dim(
+      '   Signal → Settings → Linked Devices → Link New Device → scan.',
+    );
+    return [...qrText.trimEnd().split('\n'), '', caption];
+  } catch {
+    return [
+      'Linking URL (render at https://qr.io or similar):',
+      '',
+      url,
+      '',
+      k.dim('Signal → Settings → Linked Devices → Link New Device → scan.'),
+    ];
+  }
+}
+
+/** Persist SIGNAL_ACCOUNT to .env and mirror to data/env/env for the container. */
+function writeSignalAccount(account: string): void {
+  const envPath = path.join(process.cwd(), '.env');
+  let contents = '';
+  try {
+    contents = fs.readFileSync(envPath, 'utf-8');
+  } catch {
+    contents = '';
+  }
+  if (/^SIGNAL_ACCOUNT=/m.test(contents)) {
+    contents = contents.replace(
+      /^SIGNAL_ACCOUNT=.*$/m,
+      `SIGNAL_ACCOUNT=${account}`,
+    );
+  } else {
+    if (contents.length > 0 && !contents.endsWith('\n')) contents += '\n';
+    contents += `SIGNAL_ACCOUNT=${account}\n`;
+  }
+  fs.writeFileSync(envPath, contents);
+
+  const containerEnvDir = path.join(process.cwd(), 'data', 'env');
+  fs.mkdirSync(containerEnvDir, { recursive: true });
+  fs.copyFileSync(envPath, path.join(containerEnvDir, 'env'));
+
+  setupLog.userInput('signal_account', account);
+}
+
+async function restartService(): Promise<void> {
+  const s = p.spinner();
+  s.start('Restarting NanoClaw so it sees your Signal account…');
+  const start = Date.now();
+  const platform = process.platform;
+  try {
+    if (platform === 'darwin') {
+      spawnSync(
+        'launchctl',
+        ['kickstart', '-k', `gui/${process.getuid?.() ?? 501}/${getLaunchdLabel()}`],
+        { stdio: 'ignore' },
+      );
+    } else if (platform === 'linux') {
+      const unit = getSystemdUnit();
+      const user = spawnSync('systemctl', ['--user', 'restart', unit], {
+        stdio: 'ignore',
+      });
+      if (user.status !== 0) {
+        spawnSync('sudo', ['systemctl', 'restart', unit], { stdio: 'ignore' });
+      }
+    }
+    // Give the adapter a moment to connect to signal-cli before
+    // init-first-agent's welcome DM hits the delivery path.
+    await new Promise((r) => setTimeout(r, 5000));
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    s.stop(`NanoClaw restarted. ${k.dim(`(${elapsed}s)`)}`);
+    setupLog.step('signal-restart', 'success', Date.now() - start, {
+      PLATFORM: platform,
+    });
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err);
+    s.stop(`Restart may have failed: ${message}`, 1);
+    setupLog.step('signal-restart', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    // Non-fatal — the user can restart manually if init-first-agent fails.
+  }
+}
+
+async function resolveAgentName(): Promise<string> {
+  const preset = process.env.NANOCLAW_AGENT_NAME?.trim();
+  if (preset) {
+    setupLog.userInput('agent_name', preset);
+    return preset;
+  }
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'What should your assistant be called?',
+      placeholder: DEFAULT_AGENT_NAME,
+      defaultValue: DEFAULT_AGENT_NAME,
+    }),
+  );
+  const value = (answer as string).trim() || DEFAULT_AGENT_NAME;
+  setupLog.userInput('agent_name', value);
+  return value;
+}
diff --git a/setup/channels/slack.ts b/setup/channels/slack.ts
new file mode 100644
index 0000000..f66c29a
--- /dev/null
+++ b/setup/channels/slack.ts
@@ -0,0 +1,249 @@
+/**
+ * Slack channel flow for setup:auto.
+ *
+ * `runSlackChannel(displayName)` walks the operator from a bare Slack
+ * workspace through a running bot, then stops before wiring an agent:
+ *
+ *   1. Walk through creating a Slack app (api.slack.com/apps) — scopes,
+ *      event subscriptions, and signing secret
+ *   2. Paste the bot token + signing secret (clack password prompts)
+ *   3. Validate via auth.test → resolves workspace + bot identity
+ *   4. Install the adapter (setup/add-slack.sh, non-interactive)
+ *   5. Print the post-install checklist: set the public webhook URL in
+ *      Slack's Event Subscriptions, DM the bot to bootstrap the channel,
+ *      then `/manage-channels` to wire an agent.
+ *
+ * Why no welcome DM here: unlike Discord/Telegram (gateway / long-poll),
+ * Slack needs a public Event Subscriptions URL for inbound events, and
+ * opening an unsolicited DM would need `im:write` scope we don't force
+ * the SKILL.md to require. Shipping a honest "here's what's left" note
+ * is better than a welcome DM the user won't receive until they
+ * configure the webhook anyway.
+ *
+ * All output obeys the three-level contract. See docs/setup-flow.md.
+ */
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { confirmThenOpen } from '../lib/browser.js';
+import { ensureAnswer, fail, runQuietChild } from '../lib/runner.js';
+import { wrapForGutter } from '../lib/theme.js';
+
+const SLACK_API = 'https://slack.com/api';
+const SLACK_APPS_URL = 'https://api.slack.com/apps';
+
+interface WorkspaceInfo {
+  teamName: string;
+  teamId: string;
+  botName: string;
+  botUserId: string;
+}
+
+// displayName is reserved for when we start wiring the first agent here.
+// Kept to match the `run<X>Channel(displayName)` signature every other
+// channel driver uses, so auto.ts can dispatch without a branch.
+export async function runSlackChannel(_displayName: string): Promise<void> {
+  await walkThroughAppCreation();
+
+  const token = await collectBotToken();
+  const signingSecret = await collectSigningSecret();
+  const info = await validateSlackToken(token);
+
+  const install = await runQuietChild(
+    'slack-install',
+    'bash',
+    ['setup/add-slack.sh'],
+    {
+      running: `Connecting Slack to @${info.botName} (${info.teamName})…`,
+      done: 'Slack adapter installed.',
+    },
+    {
+      env: {
+        SLACK_BOT_TOKEN: token,
+        SLACK_SIGNING_SECRET: signingSecret,
+      },
+      extraFields: {
+        BOT_NAME: info.botName,
+        TEAM_NAME: info.teamName,
+        TEAM_ID: info.teamId,
+      },
+    },
+  );
+  if (!install.ok) {
+    await fail(
+      'slack-install',
+      "Couldn't connect Slack.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+
+  showPostInstallChecklist(info);
+}
+
+async function walkThroughAppCreation(): Promise<void> {
+  p.note(
+    [
+      "You'll create a Slack app that the assistant talks through.",
+      "Free and stays inside the workspaces you pick.",
+      '',
+      '  1. Create a new app "From scratch", name it, pick a workspace',
+      '  2. OAuth & Permissions → add Bot Token Scopes:',
+      '     chat:write, channels:history, groups:history, im:history,',
+      '     channels:read, groups:read, users:read, reactions:write',
+      '  3. App Home → enable "Messages Tab" and "Allow users to send',
+      '     slash commands and messages from the messages tab"',
+      '  4. Basic Information → copy the "Signing Secret"',
+      '  5. Install to Workspace → copy the "Bot User OAuth Token" (xoxb-…)',
+      '',
+      k.dim(SLACK_APPS_URL),
+    ].join('\n'),
+    'Create a Slack app',
+  );
+  await confirmThenOpen(SLACK_APPS_URL, 'Press Enter to open Slack app settings');
+
+  ensureAnswer(
+    await p.confirm({
+      message: 'Got your bot token and signing secret?',
+      initialValue: true,
+    }),
+  );
+}
+
+async function collectBotToken(): Promise<string> {
+  const answer = ensureAnswer(
+    await p.password({
+      message: 'Paste your Slack bot token',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'Token is required';
+        if (!t.startsWith('xoxb-')) return 'Bot tokens start with xoxb-';
+        if (t.length < 24) return "That's shorter than a real Slack bot token";
+        return undefined;
+      },
+    }),
+  );
+  const token = (answer as string).trim();
+  setupLog.userInput(
+    'slack_bot_token',
+    `${token.slice(0, 10)}…${token.slice(-4)}`,
+  );
+  return token;
+}
+
+async function collectSigningSecret(): Promise<string> {
+  const answer = ensureAnswer(
+    await p.password({
+      message: 'Paste your Slack signing secret',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'Signing secret is required';
+        // Slack signing secrets are 32-char hex strings, but newer apps
+        // sometimes emit longer variants — leniently require hex only.
+        if (!/^[a-f0-9]{16,}$/i.test(t)) {
+          return 'Signing secrets are a string of hex characters';
+        }
+        return undefined;
+      },
+    }),
+  );
+  const secret = (answer as string).trim();
+  setupLog.userInput(
+    'slack_signing_secret',
+    `${secret.slice(0, 4)}…${secret.slice(-4)}`,
+  );
+  return secret;
+}
+
+async function validateSlackToken(token: string): Promise<WorkspaceInfo> {
+  const s = p.spinner();
+  const start = Date.now();
+  s.start('Checking your bot token…');
+  try {
+    const res = await fetch(`${SLACK_API}/auth.test`, {
+      method: 'POST',
+      headers: {
+        Authorization: `Bearer ${token}`,
+        'Content-Type': 'application/x-www-form-urlencoded',
+      },
+    });
+    const data = (await res.json()) as {
+      ok?: boolean;
+      team?: string;
+      team_id?: string;
+      user?: string;
+      user_id?: string;
+      error?: string;
+    };
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    if (data.ok && data.team && data.user) {
+      s.stop(
+        `Connected to ${data.team} as @${data.user}. ${k.dim(`(${elapsedS}s)`)}`,
+      );
+      const info: WorkspaceInfo = {
+        teamName: data.team,
+        teamId: data.team_id ?? '',
+        botName: data.user,
+        botUserId: data.user_id ?? '',
+      };
+      setupLog.step('slack-validate', 'success', Date.now() - start, {
+        BOT_NAME: info.botName,
+        BOT_USER_ID: info.botUserId,
+        TEAM_NAME: info.teamName,
+        TEAM_ID: info.teamId,
+      });
+      return info;
+    }
+    const reason = data.error ?? `HTTP ${res.status}`;
+    s.stop(`Slack didn't accept that token: ${reason}`, 1);
+    setupLog.step('slack-validate', 'failed', Date.now() - start, {
+      ERROR: reason,
+    });
+    await fail(
+      'slack-validate',
+      "Slack didn't accept that token.",
+      reason === 'invalid_auth' || reason === 'token_revoked'
+        ? 'Copy the token again from OAuth & Permissions and retry setup.'
+        : `Slack said "${reason}". Check the token scopes and workspace install, then retry.`,
+    );
+  } catch (err) {
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    s.stop(`Couldn't reach Slack. ${k.dim(`(${elapsedS}s)`)}`, 1);
+    const message = err instanceof Error ? err.message : String(err);
+    setupLog.step('slack-validate', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    await fail(
+      'slack-validate',
+      "Couldn't reach Slack.",
+      'Check your internet connection and retry setup.',
+    );
+  }
+}
+
+function showPostInstallChecklist(info: WorkspaceInfo): void {
+  p.note(
+    wrapForGutter(
+      [
+        `The Slack adapter is installed and your creds are saved. ${info.teamName} still needs two things before it can talk to you:`,
+        '',
+        '  1. A public URL so Slack can deliver events.',
+        '     NanoClaw serves a webhook on port 3000 by default — expose it',
+        '     via ngrok, Cloudflare Tunnel, or a reverse proxy on a VPS.',
+        '',
+        '  2. In your Slack app → Event Subscriptions:',
+        '     • Toggle "Enable Events" on',
+        `     • Request URL: https://<your-public-host>/webhook/slack`,
+        '     • Subscribe to bot events: message.channels, message.groups,',
+        '       message.im, app_mention',
+        '     • Save, then reinstall the app when Slack prompts',
+        '',
+        `  3. DM @${info.botName} from Slack once — that bootstraps the`,
+        '     messaging group. Then run `/manage-channels` in `claude` to',
+        '     wire an agent to it.',
+      ].join('\n'),
+      6,
+    ),
+    'Finish setting up Slack',
+  );
+}
diff --git a/setup/channels/teams.ts b/setup/channels/teams.ts
new file mode 100644
index 0000000..fb4d878
--- /dev/null
+++ b/setup/channels/teams.ts
@@ -0,0 +1,673 @@
+/**
+ * Microsoft Teams channel flow for setup:auto.
+ *
+ * Teams is the most complex channel NanoClaw supports — the Slack/Discord
+ * "paste a token" shortcut doesn't exist. The operator has to walk through
+ * ~7 Azure portal steps (app registration, client secret, Azure Bot
+ * resource, messaging endpoint, Teams channel enable, manifest, sideload).
+ *
+ * This driver's job is to make each of those steps as guided as possible
+ * inside the terminal:
+ *   1. Print a clack note with the exact sub-steps and the portal URL.
+ *   2. Ask for the value(s) that step yields (App ID, secret, tenant, etc.).
+ *   3. At every step boundary, offer `stepGate` — a Done / Stuck / Show-again
+ *      select. "Stuck" hands off to interactive Claude with full context.
+ *
+ * Text/password prompts also accept `?` as an answer to trigger the handoff,
+ * so the operator can escape at any paste point without scrolling back to a
+ * step boundary.
+ *
+ * What's deferred (known limitation, instruct user how to finish manually):
+ *   - Wait-for-first-DM to capture the auto-generated Teams platformId.
+ *     Unlike Discord/Telegram, the Teams platform_id is only discoverable
+ *     after the first inbound activity. The driver installs the adapter and
+ *     stops there; the operator DMs the bot, NanoClaw auto-creates the
+ *     messaging group, and they wire an agent via `/manage-channels`.
+ */
+import os from 'os';
+import path from 'path';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import { brightSelect } from '../lib/bright-select.js';
+import { confirmThenOpen } from '../lib/browser.js';
+import {
+  isHelpEscape,
+  offerClaudeHandoff,
+  validateWithHelpEscape,
+  type HandoffContext,
+} from '../lib/claude-handoff.js';
+import { ensureAnswer, fail, runQuietChild } from '../lib/runner.js';
+import { buildTeamsAppPackage } from '../lib/teams-manifest.js';
+import * as setupLog from '../logs.js';
+
+const CHANNEL = 'teams';
+const MANIFEST_DIR = path.join(process.cwd(), 'data', 'teams');
+const AZURE_PORTAL_URL = 'https://portal.azure.com';
+
+interface Collected {
+  publicUrl?: string;
+  appId?: string;
+  tenantId?: string;
+  appType?: 'SingleTenant' | 'MultiTenant';
+  appPassword?: string;
+  agentName?: string;
+}
+
+export async function runTeamsChannel(_displayName: string): Promise<void> {
+  const collected: Collected = {};
+  const completed: string[] = [];
+
+  printIntro();
+
+  await confirmPrereqs({ collected, completed });
+  await stepPublicUrl({ collected, completed });
+  await stepAppRegistration({ collected, completed });
+  await stepClientSecret({ collected, completed });
+  await stepAzureBot({ collected, completed });
+  await stepEnableTeamsChannel({ collected, completed });
+  const manifestResult = await stepGenerateManifest({ collected, completed });
+  await stepSideload({ collected, completed, zipPath: manifestResult.zipPath });
+
+  await installAdapter(collected);
+  completed.push('Adapter installed and service restarted.');
+
+  await finishWithHandoff(collected, completed);
+}
+
+// ─── step: intro / prereqs ──────────────────────────────────────────────
+
+function printIntro(): void {
+  p.note(
+    [
+      'Setting up Teams is more involved than the other channels — about',
+      '7 steps across the Azure portal and Teams admin.',
+      '',
+      k.dim("At any prompt you can type '?' and press Enter to hand off"),
+      k.dim("to Claude interactive mode with your current progress."),
+      k.dim("You can also pick 'Stuck' at any Done/Stuck/Show-again prompt."),
+    ].join('\n'),
+    'Microsoft Teams setup',
+  );
+}
+
+async function confirmPrereqs(args: { collected: Collected; completed: string[] }): Promise<void> {
+  p.note(
+    [
+      'Before we start, confirm you have:',
+      '',
+      '  • A Microsoft 365 tenant where you can sideload custom apps',
+      '    (free personal Teams does NOT support this — you need a',
+      '     Microsoft 365 Business / EDU / developer tenant)',
+      '  • Teams admin or developer tenant rights',
+      '  • A way to expose an HTTPS endpoint from this machine',
+      '    (ngrok, Cloudflare Tunnel, or a reverse-proxied VPS)',
+    ].join('\n'),
+    'Prereqs',
+  );
+
+  await stepGate({
+    stepName: 'teams-prereqs',
+    stepDescription: 'confirming they have the right Microsoft 365 tenant and tunnel',
+    reshow: () => confirmPrereqs(args),
+    args,
+  });
+  args.completed.push('Prereqs confirmed.');
+}
+
+// ─── step: public URL ──────────────────────────────────────────────────
+
+async function stepPublicUrl(args: { collected: Collected; completed: string[] }): Promise<void> {
+  p.note(
+    [
+      "Azure Bot Service delivers messages to an HTTPS endpoint you",
+      "control. The endpoint needs to reach this machine's webhook",
+      "server at /api/webhooks/teams.",
+      '',
+      k.dim('Examples:'),
+      k.dim('  ngrok http 3000      → https://abcd1234.ngrok.io'),
+      k.dim('  cloudflared tunnel …  → https://<tunnel>.trycloudflare.com'),
+      k.dim('  or a reverse proxy on your own domain'),
+      '',
+      "If you don't have a tunnel running yet, start one in another",
+      "terminal, then come back here.",
+    ].join('\n'),
+    'Public HTTPS URL',
+  );
+
+  while (true) {
+    const answer = ensureAnswer(
+      await p.text({
+        message: 'Paste your public base URL (e.g. https://abcd1234.ngrok.io)',
+        placeholder: 'https://…',
+        validate: validateWithHelpEscape((v) => {
+          const t = (v ?? '').trim();
+          if (!t) return 'Required';
+          if (!/^https:\/\/[^\s/]+/.test(t)) {
+            return 'Must be an https:// URL (Azure rejects http)';
+          }
+          return undefined;
+        }),
+      }),
+    );
+    if (isHelpEscape(answer)) {
+      await offerHandoff({
+        step: 'teams-public-url',
+        stepDescription:
+          'setting up a public HTTPS tunnel to reach this machine on port 3000',
+        args,
+      });
+      continue;
+    }
+    const url = (answer as string).trim().replace(/\/$/, '');
+    args.collected.publicUrl = url;
+    setupLog.userInput('teams_public_url', url);
+    break;
+  }
+
+  args.completed.push(`Public URL: ${args.collected.publicUrl}`);
+}
+
+// ─── step: Azure App Registration ──────────────────────────────────────
+
+async function stepAppRegistration(args: {
+  collected: Collected;
+  completed: string[];
+}): Promise<void> {
+  p.note(
+    [
+      `1. In ${AZURE_PORTAL_URL}, search "App registrations" → "New registration"`,
+      '2. Name it (e.g. "NanoClaw")',
+      '3. Supported account types: Single tenant (your org only) OR',
+      '   Multi tenant (any Microsoft 365 tenant can add the bot)',
+      '4. Click Register',
+      '5. On the Overview page, copy:',
+      '     • Application (client) ID',
+      '     • Directory (tenant) ID',
+    ].join('\n'),
+    'Step 1 of 6 — Create Azure App Registration',
+  );
+  await confirmThenOpen(
+    AZURE_PORTAL_URL,
+    'Press Enter to open the Azure portal',
+  );
+
+  args.collected.appType = await askAppType(args);
+  args.collected.appId = await askUuid(
+    'Paste the Application (client) ID',
+    'teams-app-id',
+    args,
+  );
+  if (args.collected.appType === 'SingleTenant') {
+    args.collected.tenantId = await askUuid(
+      'Paste the Directory (tenant) ID',
+      'teams-tenant-id',
+      args,
+    );
+  }
+
+  await stepGate({
+    stepName: 'teams-app-registration',
+    stepDescription: 'registering an app in Azure and collecting App ID + tenant type',
+    reshow: () => stepAppRegistration(args),
+    args,
+  });
+  args.completed.push(
+    `App registered: ${args.collected.appId} (${args.collected.appType})`,
+  );
+}
+
+async function askAppType(args: {
+  collected: Collected;
+  completed: string[];
+}): Promise<'SingleTenant' | 'MultiTenant'> {
+  while (true) {
+    const choice = ensureAnswer(
+      await brightSelect({
+        message: 'Which account type did you pick?',
+        options: [
+          {
+            value: 'SingleTenant',
+            label: 'Single tenant',
+            hint: 'your org only — most common for self-host',
+          },
+          {
+            value: 'MultiTenant',
+            label: 'Multi tenant',
+            hint: 'any Microsoft 365 tenant can install the bot',
+          },
+          { value: 'help', label: 'Stuck — hand me off to Claude' },
+        ],
+      }),
+    );
+    if (choice === 'help') {
+      await offerHandoff({
+        step: 'teams-app-type',
+        stepDescription: "deciding between Single tenant and Multi tenant for their Azure app",
+        args,
+      });
+      continue;
+    }
+    return choice as 'SingleTenant' | 'MultiTenant';
+  }
+}
+
+// ─── step: client secret ───────────────────────────────────────────────
+
+async function stepClientSecret(args: {
+  collected: Collected;
+  completed: string[];
+}): Promise<void> {
+  p.note(
+    [
+      `1. In your app registration, open "Certificates & secrets"`,
+      '2. Click "New client secret"',
+      '     Description: nanoclaw',
+      '     Expires: 180 days (recommended) or longer',
+      '3. Click Add',
+      '4. ' + k.yellow('COPY THE VALUE NOW — Azure only shows it once'),
+      '   (the Value column, not the Secret ID)',
+    ].join('\n'),
+    'Step 2 of 6 — Create a client secret',
+  );
+
+  while (true) {
+    const answer = ensureAnswer(
+      await p.password({
+        message: 'Paste the client secret Value',
+        validate: validateWithHelpEscape((v) => {
+          const t = (v ?? '').trim();
+          if (!t) return 'Required';
+          if (t.length < 20) return "That looks too short — make sure you copied the Value, not the Secret ID";
+          return undefined;
+        }),
+      }),
+    );
+    if (isHelpEscape(answer)) {
+      await offerHandoff({
+        step: 'teams-client-secret',
+        stepDescription: 'creating and copying the client secret value from Azure',
+        args,
+      });
+      continue;
+    }
+    args.collected.appPassword = (answer as string).trim();
+    setupLog.userInput(
+      'teams_client_secret',
+      `${args.collected.appPassword.slice(0, 4)}…${args.collected.appPassword.slice(-4)}`,
+    );
+    break;
+  }
+
+  await stepGate({
+    stepName: 'teams-client-secret',
+    stepDescription: 'creating and copying the client secret',
+    reshow: () => stepClientSecret(args),
+    args,
+  });
+  args.completed.push('Client secret captured.');
+}
+
+// ─── step: Azure Bot resource ──────────────────────────────────────────
+
+async function stepAzureBot(args: {
+  collected: Collected;
+  completed: string[];
+}): Promise<void> {
+  const endpoint = `${args.collected.publicUrl}/api/webhooks/teams`;
+  const tenantFlag =
+    args.collected.appType === 'SingleTenant'
+      ? `--tenant-id ${args.collected.tenantId} `
+      : '';
+  const cliCommand =
+    `az bot create \\\n` +
+    `  --resource-group nanoclaw-rg \\\n` +
+    `  --name nanoclaw-bot \\\n` +
+    `  --app-type ${args.collected.appType} \\\n` +
+    `  --appid ${args.collected.appId} \\\n` +
+    `  ${tenantFlag}--endpoint "${endpoint}"`;
+
+  p.note(
+    [
+      `In ${AZURE_PORTAL_URL}, search "Azure Bot" → Create.`,
+      '',
+      '  • Bot handle: unique name, e.g. nanoclaw-bot',
+      `  • Type of App: ${args.collected.appType}`,
+      '  • Creation type: Use existing app registration',
+      `  • App ID: ${args.collected.appId ?? '<pending>'}`,
+      ...(args.collected.appType === 'SingleTenant'
+        ? [`  • App tenant ID: ${args.collected.tenantId ?? '<pending>'}`]
+        : []),
+      '',
+      'After creating, open the bot → Configuration and set:',
+      `  Messaging endpoint: ${k.cyan(endpoint)}`,
+      '',
+      k.dim('Or via Azure CLI (if you have az installed):'),
+      k.dim(cliCommand),
+    ].join('\n'),
+    'Step 3 of 6 — Create Azure Bot resource',
+  );
+
+  await stepGate({
+    stepName: 'teams-azure-bot',
+    stepDescription:
+      'creating an Azure Bot resource linked to the app registration and setting the messaging endpoint',
+    reshow: () => stepAzureBot(args),
+    args,
+  });
+  args.completed.push('Azure Bot created; messaging endpoint configured.');
+}
+
+// ─── step: enable Teams channel ────────────────────────────────────────
+
+async function stepEnableTeamsChannel(args: {
+  collected: Collected;
+  completed: string[];
+}): Promise<void> {
+  p.note(
+    [
+      '1. Open your Azure Bot resource → Channels',
+      '2. Click Microsoft Teams → Accept terms → Apply',
+      '',
+      k.dim('CLI alternative:'),
+      k.dim('  az bot msteams create --resource-group nanoclaw-rg --name nanoclaw-bot'),
+    ].join('\n'),
+    'Step 4 of 6 — Enable Teams channel on the bot',
+  );
+  await stepGate({
+    stepName: 'teams-enable-channel',
+    stepDescription: 'enabling the Microsoft Teams channel on the Azure Bot resource',
+    reshow: () => stepEnableTeamsChannel(args),
+    args,
+  });
+  args.completed.push('Teams channel enabled on the bot.');
+}
+
+// ─── step: manifest zip ────────────────────────────────────────────────
+
+async function stepGenerateManifest(args: {
+  collected: Collected;
+  completed: string[];
+}): Promise<{ zipPath: string }> {
+  if (!args.collected.appId) {
+    fail(
+      'teams-manifest',
+      'Missing Azure App ID.',
+      "That's an internal bug — open an issue or retry setup.",
+    );
+  }
+  const shortName =
+    process.env.NANOCLAW_AGENT_NAME?.trim() || 'NanoClaw';
+
+  const s = p.spinner();
+  s.start('Generating your Teams app package…');
+  try {
+    const result = buildTeamsAppPackage({
+      appId: args.collected.appId!,
+      shortName,
+      longDescription: `${shortName} personal assistant powered by NanoClaw.`,
+      websiteUrl: args.collected.publicUrl!,
+      outDir: MANIFEST_DIR,
+    });
+    s.stop(`Package ready: ${k.cyan(shortPath(result.zipPath))}`);
+    setupLog.step('teams-manifest', 'success', 0, {
+      ZIP: result.zipPath,
+    });
+    args.completed.push(`Generated manifest zip at ${shortPath(result.zipPath)}.`);
+    return { zipPath: result.zipPath };
+  } catch (err) {
+    s.stop("Couldn't build the manifest zip.", 1);
+    const message = err instanceof Error ? err.message : String(err);
+    setupLog.step('teams-manifest', 'failed', 0, { ERROR: message });
+    fail(
+      'teams-manifest',
+      "Couldn't generate the Teams app package.",
+      'Make sure `zip` is available on your PATH, then retry.',
+    );
+  }
+}
+
+// ─── step: sideload ────────────────────────────────────────────────────
+
+async function stepSideload(args: {
+  collected: Collected;
+  completed: string[];
+  zipPath: string;
+}): Promise<void> {
+  p.note(
+    [
+      '1. Open Microsoft Teams',
+      '2. Go to Apps → Manage your apps → Upload an app',
+      '3. Click "Upload a custom app" (or "Upload for me or my teams")',
+      `4. Select: ${k.cyan(args.zipPath)}`,
+      '5. Click Add',
+      '',
+      k.dim('If "Upload a custom app" is missing, your tenant admin has'),
+      k.dim('disabled sideloading. Enable it in Teams Admin Center →'),
+      k.dim('Teams apps → Setup policies → Global → Upload custom apps = On'),
+    ].join('\n'),
+    'Step 5 of 6 — Sideload the app into Teams',
+  );
+  await stepGate({
+    stepName: 'teams-sideload',
+    stepDescription: 'uploading the generated zip into Teams as a custom app',
+    reshow: () => stepSideload(args),
+    args,
+  });
+  args.completed.push('App sideloaded into Teams.');
+}
+
+// ─── step: install adapter ─────────────────────────────────────────────
+
+async function installAdapter(collected: Collected): Promise<void> {
+  const env: Record<string, string> = {
+    TEAMS_APP_ID: collected.appId!,
+    TEAMS_APP_PASSWORD: collected.appPassword!,
+    TEAMS_APP_TYPE: collected.appType!,
+  };
+  if (collected.appType === 'SingleTenant') {
+    env.TEAMS_APP_TENANT_ID = collected.tenantId!;
+  }
+
+  const install = await runQuietChild(
+    'teams-install',
+    'bash',
+    ['setup/add-teams.sh'],
+    {
+      running: 'Installing the Teams adapter and restarting the service…',
+      done: 'Teams adapter installed.',
+    },
+    {
+      env,
+      extraFields: {
+        APP_ID: collected.appId!,
+        APP_TYPE: collected.appType!,
+      },
+    },
+  );
+  if (!install.ok) {
+    fail(
+      'teams-install',
+      "Couldn't install the Teams adapter.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+}
+
+// ─── post-install: hand off to Claude for the final wiring ────────────
+
+async function finishWithHandoff(
+  collected: Collected,
+  completed: string[],
+): Promise<void> {
+  p.note(
+    [
+      'The Teams adapter is live and the service is running.',
+      '',
+      "One thing left: your Teams bot's platform ID (which NanoClaw needs",
+      'to wire to an agent group) only becomes known after you DM the bot',
+      'for the first time. Claude can walk you through that interactively —',
+      'watch the logs for your first inbound, find the auto-created',
+      'messaging group in the DB, run scripts/init-first-agent.ts with',
+      'the right flags, and verify end-to-end.',
+    ].join('\n'),
+    'Step 6 of 6 — Finish wiring',
+  );
+
+  const choice = ensureAnswer(
+    await brightSelect({
+      message: 'Ready to finish?',
+      options: [
+        {
+          value: 'handoff',
+          label: 'Hand me off to Claude to walk me through it',
+          hint: 'recommended',
+        },
+        { value: 'self', label: "I'll do it myself" },
+      ],
+    }),
+  );
+
+  if (choice === 'self') {
+    p.note(
+      [
+        '  1. Find your bot in Teams (search by name, or via the sideloaded',
+        '     app) and send it a message ("hi" is fine)',
+        '  2. Tail ' + k.cyan('logs/nanoclaw.log') + ' for the inbound; the router',
+        '     auto-creates a row in ' + k.cyan('messaging_groups') + ' in data/v2.db',
+        '  3. Run ' + k.cyan('scripts/init-first-agent.ts') + ' with --channel teams,',
+        '     the discovered platform_id, and your AAD user id, OR use',
+        '     ' + k.cyan('/manage-channels') + ' to wire interactively',
+      ].join('\n'),
+      'Manual finish',
+    );
+    return;
+  }
+
+  await offerClaudeHandoff({
+    channel: CHANNEL,
+    step: 'teams-finish-wiring',
+    stepDescription:
+      'finishing the Teams wiring: watch for the first inbound, discover the auto-created messaging group in data/v2.db, and run scripts/init-first-agent.ts to wire it to an agent group',
+    completedSteps: completed,
+    collectedValues: redactCollected(collected),
+    files: [
+      'scripts/init-first-agent.ts',
+      'src/router.ts',
+      'src/db/messaging-groups.ts',
+      'logs/nanoclaw.log',
+      '.claude/skills/manage-channels/SKILL.md',
+    ],
+  });
+}
+
+// ─── shared step gate ──────────────────────────────────────────────────
+
+async function stepGate(args: {
+  stepName: string;
+  stepDescription: string;
+  reshow: () => Promise<void> | Promise<unknown>;
+  args: { collected: Collected; completed: string[] };
+}): Promise<void> {
+  while (true) {
+    const choice = ensureAnswer(
+      await brightSelect({
+        message: 'How did that go?',
+        options: [
+          { value: 'done', label: "Done — let's continue" },
+          { value: 'help', label: 'Stuck — hand me off to Claude' },
+          { value: 'reshow', label: 'Show me the steps again' },
+        ],
+      }),
+    );
+    if (choice === 'done') return;
+    if (choice === 'help') {
+      await offerHandoff({
+        step: args.stepName,
+        stepDescription: args.stepDescription,
+        args: args.args,
+      });
+      continue;
+    }
+    if (choice === 'reshow') {
+      await args.reshow();
+      return;
+    }
+  }
+}
+
+async function offerHandoff(args: {
+  step: string;
+  stepDescription: string;
+  args: { collected: Collected; completed: string[] };
+}): Promise<void> {
+  const ctx: HandoffContext = {
+    channel: CHANNEL,
+    step: args.step,
+    stepDescription: args.stepDescription,
+    completedSteps: args.args.completed.slice(),
+    collectedValues: redactCollected(args.args.collected),
+    files: ['setup/channels/teams.ts', 'setup/add-teams.sh'],
+  };
+  await offerClaudeHandoff(ctx);
+}
+
+function redactCollected(c: Collected): Record<string, string> {
+  const out: Record<string, string> = {};
+  if (c.publicUrl) out.publicUrl = c.publicUrl;
+  if (c.appId) out.appId = c.appId;
+  if (c.tenantId) out.tenantId = c.tenantId;
+  if (c.appType) out.appType = c.appType;
+  if (c.appPassword) {
+    out.appPassword = `${c.appPassword.slice(0, 4)}…${c.appPassword.slice(-4)}`;
+  }
+  return out;
+}
+
+// ─── shared: UUID paste with help escape ───────────────────────────────
+
+async function askUuid(
+  message: string,
+  logKey: string,
+  args: { collected: Collected; completed: string[] },
+): Promise<string> {
+  while (true) {
+    const answer = ensureAnswer(
+      await p.text({
+        message,
+        placeholder: '00000000-0000-0000-0000-000000000000',
+        validate: validateWithHelpEscape((v) => {
+          const t = (v ?? '').trim();
+          if (!t) return 'Required';
+          if (!/^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$/.test(t)) {
+            return 'Expected a UUID like 00000000-0000-0000-0000-000000000000';
+          }
+          return undefined;
+        }),
+      }),
+    );
+    if (isHelpEscape(answer)) {
+      await offerHandoff({
+        step: logKey,
+        stepDescription: `entering a UUID for ${logKey}`,
+        args,
+      });
+      continue;
+    }
+    const value = (answer as string).trim().toLowerCase();
+    setupLog.userInput(logKey, value);
+    return value;
+  }
+}
+
+// ─── path helpers ──────────────────────────────────────────────────────
+
+function shortPath(abs: string): string {
+  const home = os.homedir();
+  const cwd = process.cwd();
+  if (abs.startsWith(`${cwd}/`)) return abs.slice(cwd.length + 1);
+  if (abs.startsWith(`${home}/`)) return `~/${abs.slice(home.length + 1)}`;
+  return abs;
+}
+
diff --git a/setup/channels/telegram.ts b/setup/channels/telegram.ts
new file mode 100644
index 0000000..df97fcf
--- /dev/null
+++ b/setup/channels/telegram.ts
@@ -0,0 +1,302 @@
+/**
+ * Telegram channel flow for setup:auto.
+ *
+ * `runTelegramChannel(displayName)` owns the full branch from the
+ * BotFather instructions through the welcome DM:
+ *
+ *   1. BotFather instructions (clack note)
+ *   2. Paste the bot token (clack password) — format-validated
+ *   3. getMe via the Bot API to resolve the bot's username
+ *   4. Confirm + deep-link into the bot's Telegram chat (tg://resolve)
+ *   5. Install the adapter (setup/add-telegram.sh, non-interactive)
+ *   6. Run the pair-telegram step, rendering code events as clack notes
+ *   7. Ask for the messaging-agent name (defaulting to "Nano")
+ *   8. Wire the agent via scripts/init-first-agent.ts
+ *
+ * All output obeys the three-level contract: clack UI for the user,
+ * structured entries in logs/setup.log, full raw output in per-step files
+ * under logs/setup-steps/. See docs/setup-flow.md.
+ */
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { confirmThenOpen } from '../lib/browser.js';
+import { askOperatorRole } from '../lib/role-prompt.js';
+import {
+  type Block,
+  type StepResult,
+  dumpTranscriptOnFailure,
+  ensureAnswer,
+  fail,
+  runQuietChild,
+  spawnStep,
+  writeStepEntry,
+} from '../lib/runner.js';
+import { brandBold } from '../lib/theme.js';
+
+const DEFAULT_AGENT_NAME = 'Nano';
+
+export async function runTelegramChannel(displayName: string): Promise<void> {
+  const token = await collectTelegramToken();
+  const botUsername = await validateTelegramToken(token);
+
+  // Deep-link the user into the bot's chat so they're on the right screen
+  // by the time pair-telegram prints the code. https://t.me/<bot> works
+  // everywhere: browsers show an "Open in Telegram" button when the app is
+  // installed, or the bot's web profile if not. tg://resolve?domain= is
+  // more direct but silently fails when the scheme isn't registered.
+  const botUrl = `https://t.me/${botUsername}`;
+  p.note(
+    [
+      `Opening @${botUsername} in Telegram so it's ready when the pairing code shows up.`,
+      '',
+      k.dim(botUrl),
+    ].join('\n'),
+    'Open Telegram',
+  );
+  await confirmThenOpen(botUrl, 'Press Enter to open Telegram');
+
+  const install = await runQuietChild(
+    'telegram-install',
+    'bash',
+    ['setup/add-telegram.sh'],
+    {
+      running: `Connecting Telegram to @${botUsername}…`,
+      done: 'Telegram connected.',
+    },
+    {
+      env: { TELEGRAM_BOT_TOKEN: token },
+      extraFields: { BOT_USERNAME: botUsername },
+    },
+  );
+  if (!install.ok) {
+    await fail(
+      'telegram-install',
+      "Couldn't connect Telegram.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+
+  const pair = await runPairTelegram();
+  if (!pair.ok) {
+    await fail(
+      'pair-telegram',
+      "Couldn't pair with Telegram.",
+      'Re-run setup to try again.',
+    );
+  }
+
+  const platformId = pair.terminal?.fields.PLATFORM_ID;
+  const pairedUserId = pair.terminal?.fields.PAIRED_USER_ID;
+  if (!platformId || !pairedUserId) {
+    await fail(
+      'pair-telegram',
+      'Pairing completed but came back incomplete.',
+      'Re-run setup to try again.',
+    );
+  }
+
+  const role = await askOperatorRole('Telegram');
+  setupLog.userInput('telegram_role', role);
+
+  const agentName = await resolveAgentName();
+
+  const init = await runQuietChild(
+    'init-first-agent',
+    'pnpm',
+    [
+      'exec', 'tsx', 'scripts/init-first-agent.ts',
+      '--channel', 'telegram',
+      '--user-id', pairedUserId,
+      '--platform-id', platformId,
+      '--display-name', displayName,
+      '--agent-name', agentName,
+      '--role', role,
+    ],
+    {
+      running: `Connecting ${agentName} to your Telegram chat…`,
+      done: `${agentName} is ready. Check Telegram for a welcome message.`,
+    },
+    {
+      extraFields: { CHANNEL: 'telegram', AGENT_NAME: agentName, PLATFORM_ID: platformId },
+    },
+  );
+  if (!init.ok) {
+    await fail(
+      'init-first-agent',
+      `Couldn't finish connecting ${agentName}.`,
+      'You can retry later with `/manage-channels`.',
+    );
+  }
+}
+
+async function collectTelegramToken(): Promise<string> {
+  p.note(
+    [
+      "Your assistant talks to you through a Telegram bot you create.",
+      "Here's how:",
+      '',
+      '  1. Open Telegram and message @BotFather',
+      '  2. Send /newbot and follow the prompts',
+      '  3. Copy the token it gives you (it looks like <digits>:<chars>)',
+      '',
+      k.dim('Planning to add your assistant to group chats? In @BotFather:'),
+      k.dim('    /mybots → your bot → Bot Settings → Group Privacy → OFF'),
+    ].join('\n'),
+    'Set up your Telegram bot',
+  );
+
+  const answer = ensureAnswer(
+    await p.password({
+      message: 'Paste your bot token',
+      validate: (v) => {
+        if (!v || !v.trim()) return "Token is required";
+        if (!/^[0-9]+:[A-Za-z0-9_-]{35,}$/.test(v.trim())) {
+          return "That doesn't look right. It should be <digits>:<chars>";
+        }
+        return undefined;
+      },
+    }),
+  );
+  const token = (answer as string).trim();
+  setupLog.userInput(
+    'telegram_token',
+    `${token.slice(0, 12)}…${token.slice(-4)}`,
+  );
+  return token;
+}
+
+async function validateTelegramToken(token: string): Promise<string> {
+  const s = p.spinner();
+  const start = Date.now();
+  s.start('Checking your bot token…');
+  try {
+    const res = await fetch(`https://api.telegram.org/bot${token}/getMe`);
+    const data = (await res.json()) as {
+      ok?: boolean;
+      result?: { username?: string; id?: number };
+      description?: string;
+    };
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    if (data.ok && data.result?.username) {
+      const username = data.result.username;
+      s.stop(`Found your bot: @${username}. ${k.dim(`(${elapsedS}s)`)}`);
+      setupLog.step('telegram-validate', 'success', Date.now() - start, {
+        BOT_USERNAME: username,
+        BOT_ID: data.result.id ?? '',
+      });
+      return username;
+    }
+    const reason = data.description ?? 'token rejected by Telegram';
+    s.stop(`Telegram didn't accept that token: ${reason}`, 1);
+    setupLog.step('telegram-validate', 'failed', Date.now() - start, {
+      ERROR: reason,
+    });
+    await fail(
+      'telegram-validate',
+      "Telegram didn't accept that token.",
+      'Copy the token again from @BotFather and try setup once more.',
+    );
+  } catch (err) {
+    const elapsedS = Math.round((Date.now() - start) / 1000);
+    s.stop(`Couldn't reach Telegram. ${k.dim(`(${elapsedS}s)`)}`, 1);
+    const message = err instanceof Error ? err.message : String(err);
+    setupLog.step('telegram-validate', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    await fail(
+      'telegram-validate',
+      "Couldn't reach Telegram.",
+      'Check your internet connection and retry setup.',
+    );
+  }
+}
+
+async function runPairTelegram(): Promise<
+  StepResult & { rawLog: string; durationMs: number }
+> {
+  const rawLog = setupLog.stepRawLog('pair-telegram');
+  const start = Date.now();
+  const s = p.spinner();
+  s.start('Generating a secret code for your bot…');
+  let spinnerActive = true;
+
+  const stopSpinner = (msg: string, code?: number) => {
+    if (spinnerActive) {
+      s.stop(msg, code);
+      spinnerActive = false;
+    }
+  };
+
+  const result = await spawnStep(
+    'pair-telegram',
+    ['--intent', 'main'],
+    (block: Block) => {
+      if (block.type === 'PAIR_TELEGRAM_CODE') {
+        const reason = block.fields.REASON ?? 'initial';
+        if (reason === 'initial') {
+          stopSpinner('Your secret code is ready.');
+        } else {
+          stopSpinner("Old code expired. Here's a fresh one.");
+        }
+        p.note(formatCodeCard(block.fields.CODE ?? '????'), 'Secret code');
+        s.start('Waiting for you to send the code from Telegram…');
+        spinnerActive = true;
+      } else if (block.type === 'PAIR_TELEGRAM_ATTEMPT') {
+        stopSpinner(`Got "${block.fields.CANDIDATE ?? '?'}", not a match.`);
+        s.start('Waiting for the correct code…');
+        spinnerActive = true;
+      } else if (block.type === 'PAIR_TELEGRAM') {
+        if (block.fields.STATUS === 'success') {
+          stopSpinner('Telegram paired.');
+        } else {
+          stopSpinner(`Pairing failed: ${block.fields.ERROR ?? 'unknown'}`, 1);
+        }
+      }
+    },
+    rawLog,
+  );
+  const durationMs = Date.now() - start;
+
+  // Safety net: if the child died without emitting a terminal block, make
+  // sure we don't leave the spinner running.
+  if (spinnerActive) {
+    stopSpinner(
+      result.ok ? 'Done.' : 'Pairing ended unexpectedly.',
+      result.ok ? 0 : 1,
+    );
+    if (!result.ok) dumpTranscriptOnFailure(result.transcript);
+  }
+
+  writeStepEntry('pair-telegram', result, durationMs, rawLog);
+  return { ...result, rawLog, durationMs };
+}
+
+function formatCodeCard(code: string): string {
+  const spaced = code.split('').join('   ');
+  return [
+    '',
+    `   ${brandBold(spaced)}`,
+    '',
+    k.dim('   Send this code to your bot from Telegram.'),
+  ].join('\n');
+}
+
+async function resolveAgentName(): Promise<string> {
+  const preset = process.env.NANOCLAW_AGENT_NAME?.trim();
+  if (preset) {
+    setupLog.userInput('agent_name', preset);
+    return preset;
+  }
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'What should your assistant be called?',
+      placeholder: DEFAULT_AGENT_NAME,
+      defaultValue: DEFAULT_AGENT_NAME,
+    }),
+  );
+  const value = (answer as string).trim() || DEFAULT_AGENT_NAME;
+  setupLog.userInput('agent_name', value);
+  return value;
+}
diff --git a/setup/channels/whatsapp.ts b/setup/channels/whatsapp.ts
new file mode 100644
index 0000000..85c9866
--- /dev/null
+++ b/setup/channels/whatsapp.ts
@@ -0,0 +1,473 @@
+/**
+ * WhatsApp (community/Baileys) channel flow for setup:auto.
+ *
+ * `runWhatsAppChannel(displayName)` owns the full branch from auth-method
+ * picker through the welcome DM:
+ *
+ *   1. Ask how to authenticate (QR code in terminal, default, or pairing code)
+ *   2. If pairing-code: collect the phone number
+ *   3. Install the adapter + Baileys + QR + pino via setup/add-whatsapp.sh
+ *   4. Run the whatsapp-auth step, rendering status blocks as clack UI:
+ *      - WHATSAPP_AUTH_QR (repeating): render the QR as terminal block art
+ *        inside a clack note. On rotation we clear the previous QR in-place
+ *        via ANSI escapes so the terminal doesn't fill up with stale codes.
+ *      - WHATSAPP_AUTH_PAIRING_CODE (one-shot): centred code card.
+ *   5. Read store/auth/creds.json → extract the authenticated (bot) phone
+ *   6. Kick the service so the adapter picks up the new credentials
+ *   7. Ask the operator for the phone they'll chat from (defaults to the
+ *      authed number). Different number ⇒ dedicated mode ⇒ also writes
+ *      ASSISTANT_HAS_OWN_NUMBER=true so outbound replies aren't prefixed
+ *   8. Ask for the messaging-agent name (defaulting to "Nano")
+ *   9. Wire the agent via scripts/init-first-agent.ts; the existing welcome
+ *      DM path delivers the greeting through the adapter
+ *
+ * All output obeys the three-level contract: clack UI for the user, structured
+ * entries in logs/setup.log, full raw output in per-step files under
+ * logs/setup-steps/. See docs/setup-flow.md.
+ */
+import { spawnSync } from 'child_process';
+import fs from 'fs';
+import path from 'path';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { brightSelect } from '../lib/bright-select.js';
+import { getLaunchdLabel, getSystemdUnit } from '../../src/install-slug.js';
+import {
+  type Block,
+  type StepResult,
+  dumpTranscriptOnFailure,
+  ensureAnswer,
+  fail,
+  runQuietChild,
+  spawnStep,
+  writeStepEntry,
+} from '../lib/runner.js';
+import { askOperatorRole } from '../lib/role-prompt.js';
+import { brandBold } from '../lib/theme.js';
+
+const DEFAULT_AGENT_NAME = 'Nano';
+const AUTH_CREDS_PATH = path.join(process.cwd(), 'store', 'auth', 'creds.json');
+
+type AuthMethod = 'qr' | 'pairing-code';
+
+export async function runWhatsAppChannel(displayName: string): Promise<void> {
+  const method = await askAuthMethod();
+  const phone = method === 'pairing-code' ? await askPhoneNumber() : undefined;
+
+  const install = await runQuietChild(
+    'whatsapp-install',
+    'bash',
+    ['setup/add-whatsapp.sh'],
+    {
+      running: 'Installing the WhatsApp adapter…',
+      done: 'WhatsApp adapter installed.',
+      skipped: 'WhatsApp adapter already installed.',
+    },
+  );
+  if (!install.ok) {
+    fail(
+      'whatsapp-install',
+      "Couldn't install the WhatsApp adapter.",
+      'See logs/setup-steps/ for details, then retry setup.',
+    );
+  }
+
+  const auth = await runWhatsAppAuth(method, phone);
+  if (!auth.ok) {
+    const reason = auth.terminal?.fields.ERROR ?? 'unknown';
+    fail(
+      'whatsapp-auth',
+      `WhatsApp authentication failed (${reason}).`,
+      reason === 'qr_timeout' || reason === 'timeout'
+        ? 'The code expired. Re-run setup to get a fresh one.'
+        : 'Re-run setup to try again.',
+    );
+  }
+
+  const botPhone = readAuthedPhone();
+  if (!botPhone) {
+    fail(
+      'whatsapp-auth',
+      "Authenticated but couldn't read your WhatsApp number from the saved credentials.",
+      'Re-run setup to try again.',
+    );
+  }
+
+  await restartService();
+
+  const chatPhone = await askChatPhone(botPhone);
+  const isDedicated = chatPhone !== botPhone;
+  if (isDedicated) {
+    writeAssistantHasOwnNumber();
+  }
+
+  const role = await askOperatorRole('WhatsApp');
+  setupLog.userInput('whatsapp_role', role);
+
+  const agentName = await resolveAgentName();
+
+  const platformId = `${chatPhone}@s.whatsapp.net`;
+
+  const init = await runQuietChild(
+    'init-first-agent',
+    'pnpm',
+    [
+      'exec', 'tsx', 'scripts/init-first-agent.ts',
+      '--channel', 'whatsapp',
+      '--user-id', platformId,
+      '--platform-id', platformId,
+      '--display-name', displayName,
+      '--agent-name', agentName,
+      '--role', role,
+    ],
+    {
+      running: `Connecting ${agentName} to WhatsApp…`,
+      done: isDedicated
+        ? `${agentName} is ready. Check WhatsApp for a welcome message.`
+        : `${agentName} is ready. Look in your "You" chat on WhatsApp for the welcome.`,
+    },
+    {
+      extraFields: {
+        CHANNEL: 'whatsapp',
+        AGENT_NAME: agentName,
+        PLATFORM_ID: platformId,
+        MODE: isDedicated ? 'dedicated' : 'shared',
+        ROLE: role,
+      },
+    },
+  );
+  if (!init.ok) {
+    fail(
+      'init-first-agent',
+      `Couldn't finish connecting ${agentName}.`,
+      'You can retry later with `/manage-channels`.',
+    );
+  }
+}
+
+async function askAuthMethod(): Promise<AuthMethod> {
+  const choice = ensureAnswer(
+    await brightSelect({
+      message: 'How would you like to authenticate with WhatsApp?',
+      options: [
+        {
+          value: 'qr',
+          label: 'Scan a QR code in this terminal',
+          hint: 'recommended',
+        },
+        {
+          value: 'pairing-code',
+          label: 'Enter a pairing code on your phone',
+          hint: 'no camera needed',
+        },
+      ],
+    }),
+  ) as AuthMethod;
+  setupLog.userInput('whatsapp_auth_method', choice);
+  return choice;
+}
+
+async function askPhoneNumber(): Promise<string> {
+  p.note(
+    [
+      "Enter your phone number the way WhatsApp expects it:",
+      '',
+      '  • Digits only — no +, spaces, or dashes',
+      '  • Country code first, then the rest of the number',
+      '',
+      k.dim('Example: 14155551234 (country code 1, then 4155551234)'),
+    ].join('\n'),
+    'Your phone number',
+  );
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'Phone number',
+      validate: (v) => {
+        const t = (v ?? '').trim();
+        if (!t) return 'Phone number is required';
+        if (!/^\d{8,15}$/.test(t)) {
+          return "That doesn't look right. Digits only, country code included.";
+        }
+        return undefined;
+      },
+    }),
+  );
+  const phone = (answer as string).trim();
+  setupLog.userInput('whatsapp_phone', phone);
+  return phone;
+}
+
+async function runWhatsAppAuth(
+  method: AuthMethod,
+  phone: string | undefined,
+): Promise<StepResult & { rawLog: string; durationMs: number }> {
+  const rawLog = setupLog.stepRawLog('whatsapp-auth');
+  const start = Date.now();
+  const s = p.spinner();
+  s.start('Starting WhatsApp authentication…');
+  let spinnerActive = true;
+
+  const stopSpinner = (msg: string, code?: number) => {
+    if (spinnerActive) {
+      s.stop(msg, code);
+      spinnerActive = false;
+    }
+  };
+
+  // Tracks the QR render so we can overwrite it in-place on rotation. null
+  // before the first QR is printed.
+  let qrLinesPrinted = 0;
+
+  const extra =
+    method === 'pairing-code' && phone
+      ? ['--method', 'pairing-code', '--phone', phone]
+      : ['--method', 'qr'];
+
+  const result = await spawnStep(
+    'whatsapp-auth',
+    extra,
+    (block: Block) => {
+      if (block.type === 'WHATSAPP_AUTH_QR') {
+        const qr = block.fields.QR ?? '';
+        if (!qr) return;
+        // Fire-and-forget — await inside spawnStep's sync onBlock is fine
+        // since spawnStep's own logic keeps running in parallel.
+        void renderQr(qr).then((lines) => {
+          if (qrLinesPrinted === 0) {
+            stopSpinner('QR code ready — scan with WhatsApp.');
+          } else {
+            // Cursor up N lines + clear from there to end of screen. Wipes
+            // the previous QR + caption so the new one renders in place.
+            process.stdout.write(`\x1b[${qrLinesPrinted}A\x1b[0J`);
+          }
+          process.stdout.write(lines.join('\n') + '\n');
+          qrLinesPrinted = lines.length;
+        });
+      } else if (block.type === 'WHATSAPP_AUTH_PAIRING_CODE') {
+        const code = block.fields.CODE ?? '????';
+        stopSpinner('Your pairing code is ready.');
+        p.note(formatPairingCard(code), 'Pairing code');
+        s.start('Waiting for you to enter the code…');
+        spinnerActive = true;
+      } else if (block.type === 'WHATSAPP_AUTH') {
+        const status = block.fields.STATUS;
+        if (status === 'skipped') {
+          stopSpinner('WhatsApp is already authenticated.');
+        } else if (status === 'success') {
+          // Erase the QR block if one was on screen — it's served its purpose.
+          if (qrLinesPrinted > 0) {
+            process.stdout.write(`\x1b[${qrLinesPrinted}A\x1b[0J`);
+            qrLinesPrinted = 0;
+          }
+          // In QR flow the spinner was stopped when the first QR landed.
+          // Fall back to a plain success line so the user sees confirmation.
+          if (spinnerActive) {
+            stopSpinner('WhatsApp linked.');
+          } else {
+            p.log.success('WhatsApp linked.');
+          }
+        } else if (status === 'failed') {
+          if (qrLinesPrinted > 0) {
+            process.stdout.write(`\x1b[${qrLinesPrinted}A\x1b[0J`);
+            qrLinesPrinted = 0;
+          }
+          const err = block.fields.ERROR ?? 'unknown';
+          if (spinnerActive) {
+            stopSpinner(`Authentication failed: ${err}`, 1);
+          } else {
+            p.log.error(`Authentication failed: ${err}`);
+          }
+        }
+      }
+    },
+    rawLog,
+  );
+  const durationMs = Date.now() - start;
+
+  // Safety net — if the step died without emitting a terminal block, don't
+  // leave the spinner running.
+  if (spinnerActive) {
+    stopSpinner(
+      result.ok ? 'Done.' : 'Authentication ended unexpectedly.',
+      result.ok ? 0 : 1,
+    );
+    if (!result.ok) dumpTranscriptOnFailure(result.transcript);
+  }
+
+  writeStepEntry('whatsapp-auth', result, durationMs, rawLog);
+  return { ...result, rawLog, durationMs };
+}
+
+/**
+ * Render the raw QR string to an array of terminal lines (block-art QR +
+ * a caption). Returned as an array so the caller can count lines for the
+ * in-place rewrite on rotation. Uses the small-mode QR to keep the height
+ * manageable on 24-row terminals.
+ */
+async function renderQr(qr: string): Promise<string[]> {
+  try {
+    const QRCode = await import('qrcode');
+    const qrText = await QRCode.toString(qr, { type: 'terminal', small: true });
+    const caption = k.dim(
+      '   Open WhatsApp → Settings → Linked Devices → Link a Device → scan.',
+    );
+    return [...qrText.trimEnd().split('\n'), '', caption];
+  } catch {
+    return ['QR code (raw): ' + qr];
+  }
+}
+
+function formatPairingCard(code: string): string {
+  // WhatsApp pairing codes are 8 characters; render with two-wide gap so the
+  // digits read clearly in the terminal.
+  const spaced = code.split('').join('  ');
+  return [
+    '',
+    `   ${brandBold(spaced)}`,
+    '',
+    k.dim('   Open WhatsApp → Settings → Linked Devices → Link a Device'),
+    k.dim('   → "Link with phone number instead" → enter this code.'),
+    k.dim('   It expires in ~60 seconds.'),
+  ].join('\n');
+}
+
+/**
+ * Pull the authenticated WhatsApp phone out of store/auth/creds.json.
+ * `creds.me.id` looks like `14155551234:<device>@s.whatsapp.net` — we want
+ * just the leading digit run.
+ */
+function readAuthedPhone(): string {
+  try {
+    const raw = fs.readFileSync(AUTH_CREDS_PATH, 'utf-8');
+    const creds = JSON.parse(raw) as { me?: { id?: string } };
+    const id = creds.me?.id;
+    if (!id) return '';
+    return id.split(':')[0].split('@')[0];
+  } catch {
+    return '';
+  }
+}
+
+async function restartService(): Promise<void> {
+  const s = p.spinner();
+  s.start('Restarting NanoClaw so it sees your WhatsApp credentials…');
+  const start = Date.now();
+  const platform = process.platform;
+  try {
+    if (platform === 'darwin') {
+      spawnSync(
+        'launchctl',
+        ['kickstart', '-k', `gui/${process.getuid?.() ?? 501}/${getLaunchdLabel()}`],
+        { stdio: 'ignore' },
+      );
+    } else if (platform === 'linux') {
+      const unit = getSystemdUnit();
+      const user = spawnSync(
+        'systemctl',
+        ['--user', 'restart', unit],
+        { stdio: 'ignore' },
+      );
+      if (user.status !== 0) {
+        spawnSync('sudo', ['systemctl', 'restart', unit], {
+          stdio: 'ignore',
+        });
+      }
+    }
+    // Give the adapter a moment to reconnect before init-first-agent's
+    // welcome DM hits the delivery path.
+    await new Promise((r) => setTimeout(r, 5000));
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    s.stop(`NanoClaw restarted. ${k.dim(`(${elapsed}s)`)}`);
+    setupLog.step('whatsapp-restart', 'success', Date.now() - start, {
+      PLATFORM: platform,
+    });
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err);
+    s.stop(`Restart may have failed: ${message}`, 1);
+    setupLog.step('whatsapp-restart', 'failed', Date.now() - start, {
+      ERROR: message,
+    });
+    // Non-fatal — the user can restart manually if init-first-agent fails.
+  }
+}
+
+async function askChatPhone(authedPhone: string): Promise<string> {
+  p.note(
+    [
+      `Authenticated with ${k.cyan('+' + authedPhone)}.`,
+      '',
+      "What's the phone number you'll chat with your agent from?",
+      '',
+      k.dim(
+        'Same number = messages will land in your "You" / self-chat on WhatsApp\n' +
+          "(you won't be able to reply to yourself — use a different number for a\n" +
+          'two-way chat).',
+      ),
+    ].join('\n'),
+    'Your chat number',
+  );
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'Your personal phone number',
+      placeholder: authedPhone,
+      defaultValue: authedPhone,
+      validate: (v) => {
+        const t = (v ?? authedPhone).trim();
+        if (!/^\d{8,15}$/.test(t)) {
+          return 'Digits only, country code included.';
+        }
+        return undefined;
+      },
+    }),
+  );
+  const phone = ((answer as string) || authedPhone).trim();
+  setupLog.userInput('whatsapp_chat_phone', phone);
+  return phone;
+}
+
+/** Persist ASSISTANT_HAS_OWN_NUMBER=true to .env and data/env/env. */
+function writeAssistantHasOwnNumber(): void {
+  const envPath = path.join(process.cwd(), '.env');
+  let contents = '';
+  try {
+    contents = fs.readFileSync(envPath, 'utf-8');
+  } catch {
+    contents = '';
+  }
+  if (/^ASSISTANT_HAS_OWN_NUMBER=/m.test(contents)) {
+    contents = contents.replace(
+      /^ASSISTANT_HAS_OWN_NUMBER=.*$/m,
+      'ASSISTANT_HAS_OWN_NUMBER=true',
+    );
+  } else {
+    if (contents.length > 0 && !contents.endsWith('\n')) contents += '\n';
+    contents += 'ASSISTANT_HAS_OWN_NUMBER=true\n';
+  }
+  fs.writeFileSync(envPath, contents);
+
+  // Container reads from data/env/env.
+  const containerEnvDir = path.join(process.cwd(), 'data', 'env');
+  fs.mkdirSync(containerEnvDir, { recursive: true });
+  fs.copyFileSync(envPath, path.join(containerEnvDir, 'env'));
+}
+
+async function resolveAgentName(): Promise<string> {
+  const preset = process.env.NANOCLAW_AGENT_NAME?.trim();
+  if (preset) {
+    setupLog.userInput('agent_name', preset);
+    return preset;
+  }
+  const answer = ensureAnswer(
+    await p.text({
+      message: 'What should your assistant be called?',
+      placeholder: DEFAULT_AGENT_NAME,
+      defaultValue: DEFAULT_AGENT_NAME,
+    }),
+  );
+  const value = (answer as string).trim() || DEFAULT_AGENT_NAME;
+  setupLog.userInput('agent_name', value);
+  return value;
+}
diff --git a/setup/cli-agent.ts b/setup/cli-agent.ts
index e5a901d..d9a90c5 100644
--- a/setup/cli-agent.ts
+++ b/setup/cli-agent.ts
@@ -1,14 +1,13 @@
 /**
- * Step: cli-agent — Create the first agent wired to the CLI channel.
+ * Step: cli-agent — Create the scratch CLI agent for `/new-setup`.
  *
- * Thin wrapper around `scripts/init-first-agent.ts --cli-only`. Emits a
- * status block so /new-setup SKILL.md can parse the result without having
- * to read the script's plain stdout.
+ * Thin wrapper around `scripts/init-cli-agent.ts`. Emits a status block so
+ * /new-setup SKILL.md can parse the result without having to read the
+ * script's plain stdout.
  *
  * Args:
  *   --display-name <name>   (required) operator's display name
  *   --agent-name   <name>   (optional) agent persona name, defaults to display-name
- *   --welcome      <text>   (optional) system welcome instruction
  */
 import { execFileSync } from 'child_process';
 import path from 'path';
@@ -19,11 +18,9 @@ import { emitStatus } from './status.js';
 function parseArgs(args: string[]): {
   displayName: string;
   agentName?: string;
-  welcome?: string;
 } {
   let displayName: string | undefined;
   let agentName: string | undefined;
-  let welcome: string | undefined;
 
   for (let i = 0; i < args.length; i++) {
     const key = args[i];
@@ -37,10 +34,6 @@ function parseArgs(args: string[]): {
         agentName = val;
         i++;
         break;
-      case '--welcome':
-        welcome = val;
-        i++;
-        break;
     }
   }
 
@@ -53,20 +46,19 @@ function parseArgs(args: string[]): {
     process.exit(2);
   }
 
-  return { displayName, agentName, welcome };
+  return { displayName, agentName };
 }
 
 export async function run(args: string[]): Promise<void> {
-  const { displayName, agentName, welcome } = parseArgs(args);
+  const { displayName, agentName } = parseArgs(args);
 
   const projectRoot = process.cwd();
-  const script = path.join(projectRoot, 'scripts', 'init-first-agent.ts');
+  const script = path.join(projectRoot, 'scripts', 'init-cli-agent.ts');
 
-  const scriptArgs = ['exec', 'tsx', script, '--cli-only', '--display-name', displayName];
+  const scriptArgs = ['exec', 'tsx', script, '--display-name', displayName];
   if (agentName) scriptArgs.push('--agent-name', agentName);
-  if (welcome) scriptArgs.push('--welcome', welcome);
 
-  log.info('Invoking init-first-agent in cli-only mode', { displayName, agentName });
+  log.info('Invoking init-cli-agent', { displayName, agentName });
 
   try {
     execFileSync('pnpm', scriptArgs, {
@@ -76,7 +68,7 @@ export async function run(args: string[]): Promise<void> {
     });
   } catch (err) {
     const e = err as { stdout?: string; stderr?: string; status?: number };
-    log.error('init-first-agent failed', {
+    log.error('init-cli-agent failed', {
       status: e.status,
       stdout: e.stdout,
       stderr: e.stderr,
diff --git a/setup/container.ts b/setup/container.ts
index d810539..6ecd032 100644
--- a/setup/container.ts
+++ b/setup/container.ts
@@ -2,15 +2,74 @@
  * Step: container — Build container image and verify with test run.
  * Replaces 03-setup-container.sh
  */
-import { execSync } from 'child_process';
+import { execSync, spawnSync } from 'child_process';
 import path from 'path';
+import { setTimeout as sleep } from 'timers/promises';
 
 import { log } from '../src/log.js';
-import { commandExists } from './platform.js';
+import { getDefaultContainerImage } from '../src/install-slug.js';
+import { commandExists, getPlatform } from './platform.js';
 import { emitStatus } from './status.js';
 
+type DockerStatus = 'ok' | 'no-permission' | 'no-daemon' | 'other';
+
+function dockerStatus(): DockerStatus {
+  const res = spawnSync('docker', ['info'], { encoding: 'utf-8' });
+  if (res.status === 0) return 'ok';
+  const err = `${res.stderr ?? ''}\n${res.stdout ?? ''}`;
+  if (/permission denied/i.test(err)) return 'no-permission';
+  if (/cannot connect|is the docker daemon running|no such file/i.test(err)) return 'no-daemon';
+  return 'other';
+}
+
+function dockerRunning(): boolean {
+  return dockerStatus() === 'ok';
+}
+
+/**
+ * Try to start Docker if it's installed but idle. Poll up to 60s for the
+ * daemon to come up — but bail immediately if the socket is reachable and
+ * only blocked by a group-permission error, since that won't resolve by
+ * waiting (the caller handles the sg re-exec for that case).
+ */
+async function tryStartDocker(): Promise<DockerStatus> {
+  const platform = getPlatform();
+  log.info('Docker not running — attempting to start', { platform });
+
+  try {
+    if (platform === 'macos') {
+      execSync('open -a Docker', { stdio: 'ignore' });
+    } else if (platform === 'linux') {
+      // Inherit stdio so sudo can prompt for a password if needed.
+      execSync('sudo systemctl start docker', { stdio: 'inherit' });
+    } else {
+      return 'other';
+    }
+  } catch (err) {
+    log.warn('Start command failed', { err });
+    return 'other';
+  }
+
+  for (let i = 0; i < 30; i++) {
+    await sleep(2000);
+    const s = dockerStatus();
+    if (s === 'ok') {
+      log.info('Docker is up');
+      return 'ok';
+    }
+    if (s === 'no-permission') {
+      log.info('Docker daemon is up but socket is not accessible (group membership)');
+      return 'no-permission';
+    }
+  }
+  log.warn('Docker did not become ready within 60s');
+  return 'no-daemon';
+}
+
 function parseArgs(args: string[]): { runtime: string } {
-  let runtime = '';
+  // `--runtime` is still accepted for backwards compatibility with the /setup
+  // skill, but `docker` is the only supported value.
+  let runtime = 'docker';
   for (let i = 0; i < args.length; i++) {
     if (args[i] === '--runtime' && args[i + 1]) {
       runtime = args[i + 1];
@@ -23,66 +82,10 @@ function parseArgs(args: string[]): { runtime: string } {
 export async function run(args: string[]): Promise<void> {
   const projectRoot = process.cwd();
   const { runtime } = parseArgs(args);
-  const image = 'nanoclaw-agent:latest';
+  const image = getDefaultContainerImage(projectRoot);
   const logFile = path.join(projectRoot, 'logs', 'setup.log');
 
-  if (!runtime) {
-    emitStatus('SETUP_CONTAINER', {
-      RUNTIME: 'unknown',
-      IMAGE: image,
-      BUILD_OK: false,
-      TEST_OK: false,
-      STATUS: 'failed',
-      ERROR: 'missing_runtime_flag',
-      LOG: 'logs/setup.log',
-    });
-    process.exit(4);
-  }
-
-  // Validate runtime availability
-  if (runtime === 'apple-container' && !commandExists('container')) {
-    emitStatus('SETUP_CONTAINER', {
-      RUNTIME: runtime,
-      IMAGE: image,
-      BUILD_OK: false,
-      TEST_OK: false,
-      STATUS: 'failed',
-      ERROR: 'runtime_not_available',
-      LOG: 'logs/setup.log',
-    });
-    process.exit(2);
-  }
-
-  if (runtime === 'docker') {
-    if (!commandExists('docker')) {
-      emitStatus('SETUP_CONTAINER', {
-        RUNTIME: runtime,
-        IMAGE: image,
-        BUILD_OK: false,
-        TEST_OK: false,
-        STATUS: 'failed',
-        ERROR: 'runtime_not_available',
-        LOG: 'logs/setup.log',
-      });
-      process.exit(2);
-    }
-    try {
-      execSync('docker info', { stdio: 'ignore' });
-    } catch {
-      emitStatus('SETUP_CONTAINER', {
-        RUNTIME: runtime,
-        IMAGE: image,
-        BUILD_OK: false,
-        TEST_OK: false,
-        STATUS: 'failed',
-        ERROR: 'runtime_not_available',
-        LOG: 'logs/setup.log',
-      });
-      process.exit(2);
-    }
-  }
-
-  if (!['apple-container', 'docker'].includes(runtime)) {
+  if (runtime !== 'docker') {
     emitStatus('SETUP_CONTAINER', {
       RUNTIME: runtime,
       IMAGE: image,
@@ -95,9 +98,67 @@ export async function run(args: string[]): Promise<void> {
     process.exit(4);
   }
 
-  const buildCmd =
-    runtime === 'apple-container' ? 'container build' : 'docker build';
-  const runCmd = runtime === 'apple-container' ? 'container' : 'docker';
+  if (!commandExists('docker')) {
+    log.info('Docker not found — running setup/install-docker.sh');
+    try {
+      execSync('bash setup/install-docker.sh', { cwd: projectRoot, stdio: 'inherit' });
+    } catch (err) {
+      log.warn('install-docker.sh failed', { err });
+    }
+  }
+
+  if (!commandExists('docker')) {
+    emitStatus('SETUP_CONTAINER', {
+      RUNTIME: runtime,
+      IMAGE: image,
+      BUILD_OK: false,
+      TEST_OK: false,
+      STATUS: 'failed',
+      ERROR: 'runtime_not_available',
+      LOG: 'logs/setup.log',
+    });
+    process.exit(2);
+  }
+
+  {
+    let status = dockerStatus();
+    if (status !== 'ok') {
+      status = await tryStartDocker();
+    }
+
+    // Socket is unreachable due to group perms — current shell's supplementary
+    // groups are fixed at login, so `usermod -aG docker` (via install-docker.sh
+    // or a prior install) doesn't affect us until next login. Re-exec this
+    // step under `sg docker` so the child picks up docker as its primary
+    // group and can talk to /var/run/docker.sock without a logout.
+    if (status === 'no-permission' && getPlatform() === 'linux' && commandExists('sg')) {
+      log.info('Re-executing container step under `sg docker`');
+      const res = spawnSync(
+        'sg',
+        ['docker', '-c', 'pnpm exec tsx setup/index.ts --step container'],
+        { cwd: projectRoot, stdio: 'inherit' },
+      );
+      process.exit(res.status ?? 1);
+    }
+
+    if (status !== 'ok') {
+      const error =
+        status === 'no-permission' ? 'docker_group_not_active' : 'runtime_not_available';
+      emitStatus('SETUP_CONTAINER', {
+        RUNTIME: runtime,
+        IMAGE: image,
+        BUILD_OK: false,
+        TEST_OK: false,
+        STATUS: 'failed',
+        ERROR: error,
+        LOG: 'logs/setup.log',
+      });
+      process.exit(2);
+    }
+  }
+
+  const buildCmd = 'docker build';
+  const runCmd = 'docker';
 
   // Build-args from .env. Only INSTALL_CJK_FONTS is passed through today.
   // Keeps /setup and ./container/build.sh in sync — both read the same source.
@@ -114,19 +175,31 @@ export async function run(args: string[]): Promise<void> {
     // .env is optional; absence is normal on a fresh checkout
   }
 
-  // Build
+  // Build — stdio inherit so the parent setup runner can tail docker's
+  // per-step output and render it in a rolling window. Previously we used
+  // execSync which buffered everything; users couldn't tell whether a
+  // 3–10 minute build was making progress or hung.
   let buildOk = false;
   log.info('Building container', { runtime, buildArgs });
-  try {
-    const argsStr = buildArgs.length > 0 ? ' ' + buildArgs.join(' ') : '';
-    execSync(`${buildCmd}${argsStr} -t ${image} .`, {
+  const buildRes = spawnSync(
+    buildCmd.split(' ')[0],
+    [
+      ...buildCmd.split(' ').slice(1),
+      ...buildArgs.flatMap((a) => a.split(' ')),
+      '-t',
+      image,
+      '.',
+    ],
+    {
       cwd: path.join(projectRoot, 'container'),
-      stdio: ['ignore', 'pipe', 'pipe'],
-    });
+      stdio: 'inherit',
+    },
+  );
+  if (buildRes.status === 0) {
     buildOk = true;
     log.info('Container build succeeded');
-  } catch (err) {
-    log.error('Container build failed', { err });
+  } else {
+    log.error('Container build failed', { exitCode: buildRes.status });
   }
 
   // Test
diff --git a/setup/environment.test.ts b/setup/environment.test.ts
index deda62f..7765693 100644
--- a/setup/environment.test.ts
+++ b/setup/environment.test.ts
@@ -1,5 +1,7 @@
-import { describe, it, expect, beforeEach } from 'vitest';
+import { describe, it, expect, beforeEach, afterEach } from 'vitest';
 import fs from 'fs';
+import os from 'os';
+import path from 'path';
 
 import Database from 'better-sqlite3';
 
@@ -17,58 +19,63 @@ describe('environment detection', () => {
   });
 });
 
-describe('registered groups DB query', () => {
-  let db: Database.Database;
+describe('detectRegisteredGroups', () => {
+  let tempDir: string;
 
   beforeEach(() => {
-    db = new Database(':memory:');
-    db.exec(`CREATE TABLE IF NOT EXISTS registered_groups (
-      jid TEXT PRIMARY KEY,
-      name TEXT NOT NULL,
-      folder TEXT NOT NULL UNIQUE,
-      trigger_pattern TEXT NOT NULL,
-      added_at TEXT NOT NULL,
-      container_config TEXT,
-      requires_trigger INTEGER DEFAULT 1
-    )`);
+    tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'nanoclaw-env-test-'));
+    fs.mkdirSync(path.join(tempDir, 'data'), { recursive: true });
   });
 
-  it('returns 0 for empty table', () => {
-    const row = db
-      .prepare('SELECT COUNT(*) as count FROM registered_groups')
-      .get() as { count: number };
-    expect(row.count).toBe(0);
+  afterEach(() => {
+    fs.rmSync(tempDir, { recursive: true, force: true });
   });
 
-  it('returns correct count after inserts', () => {
-    db.prepare(
-      `INSERT INTO registered_groups (jid, name, folder, trigger_pattern, added_at, requires_trigger)
-       VALUES (?, ?, ?, ?, ?, ?)`,
-    ).run(
-      '123@g.us',
-      'Group 1',
-      'group-1',
-      '@Andy',
-      '2024-01-01T00:00:00.000Z',
-      1,
-    );
+  it('returns false when no registration state exists', async () => {
+    const { detectRegisteredGroups } = await import('./environment.js');
+    expect(detectRegisteredGroups(tempDir)).toBe(false);
+  });
 
-    db.prepare(
-      `INSERT INTO registered_groups (jid, name, folder, trigger_pattern, added_at, requires_trigger)
-       VALUES (?, ?, ?, ?, ?, ?)`,
-    ).run(
-      '456@g.us',
-      'Group 2',
-      'group-2',
-      '@Andy',
-      '2024-01-01T00:00:00.000Z',
-      1,
-    );
+  it('detects pre-migration registered_groups.json', async () => {
+    const { detectRegisteredGroups } = await import('./environment.js');
+    fs.writeFileSync(path.join(tempDir, 'data', 'registered_groups.json'), '[]');
+    expect(detectRegisteredGroups(tempDir)).toBe(true);
+  });
 
-    const row = db
-      .prepare('SELECT COUNT(*) as count FROM registered_groups')
-      .get() as { count: number };
-    expect(row.count).toBe(2);
+  it('returns false for an empty v2 central DB', async () => {
+    const { detectRegisteredGroups } = await import('./environment.js');
+    const db = new Database(path.join(tempDir, 'data', 'v2.db'));
+    db.exec(`
+      CREATE TABLE agent_groups (id TEXT PRIMARY KEY);
+      CREATE TABLE messaging_group_agents (
+        id TEXT PRIMARY KEY,
+        messaging_group_id TEXT NOT NULL,
+        agent_group_id TEXT NOT NULL
+      );
+    `);
+    db.close();
+
+    expect(detectRegisteredGroups(tempDir)).toBe(false);
+  });
+
+  it('detects wired agent groups in the v2 central DB', async () => {
+    const { detectRegisteredGroups } = await import('./environment.js');
+    const db = new Database(path.join(tempDir, 'data', 'v2.db'));
+    db.exec(`
+      CREATE TABLE agent_groups (id TEXT PRIMARY KEY);
+      CREATE TABLE messaging_group_agents (
+        id TEXT PRIMARY KEY,
+        messaging_group_id TEXT NOT NULL,
+        agent_group_id TEXT NOT NULL
+      );
+    `);
+    db.prepare('INSERT INTO agent_groups (id) VALUES (?)').run('ag-1');
+    db.prepare(
+      'INSERT INTO messaging_group_agents (id, messaging_group_id, agent_group_id) VALUES (?, ?, ?)',
+    ).run('mga-1', 'mg-1', 'ag-1');
+    db.close();
+
+    expect(detectRegisteredGroups(tempDir)).toBe(true);
   });
 });
 
diff --git a/setup/environment.ts b/setup/environment.ts
index 27de9f4..6986396 100644
--- a/setup/environment.ts
+++ b/setup/environment.ts
@@ -7,11 +7,35 @@ import path from 'path';
 
 import Database from 'better-sqlite3';
 
-import { STORE_DIR } from '../src/config.js';
 import { log } from '../src/log.js';
 import { commandExists, getPlatform, isHeadless, isWSL } from './platform.js';
 import { emitStatus } from './status.js';
 
+export function detectRegisteredGroups(projectRoot: string): boolean {
+  if (fs.existsSync(path.join(projectRoot, 'data', 'registered_groups.json'))) {
+    return true;
+  }
+
+  const dbPath = path.join(projectRoot, 'data', 'v2.db');
+  if (!fs.existsSync(dbPath)) return false;
+
+  let db: Database.Database | null = null;
+  try {
+    db = new Database(dbPath, { readonly: true });
+    const row = db
+      .prepare(
+        `SELECT COUNT(DISTINCT ag.id) as count FROM agent_groups ag
+         JOIN messaging_group_agents mga ON mga.agent_group_id = ag.id`,
+      )
+      .get() as { count: number };
+    return row.count > 0;
+  } catch {
+    return false;
+  } finally {
+    db?.close();
+  }
+}
+
 export async function run(_args: string[]): Promise<void> {
   const projectRoot = process.cwd();
 
@@ -21,12 +45,6 @@ export async function run(_args: string[]): Promise<void> {
   const wsl = isWSL();
   const headless = isHeadless();
 
-  // Check Apple Container
-  let appleContainer: 'installed' | 'not_found' = 'not_found';
-  if (commandExists('container')) {
-    appleContainer = 'installed';
-  }
-
   // Check Docker
   let docker: 'running' | 'installed_not_running' | 'not_found' = 'not_found';
   if (commandExists('docker')) {
@@ -45,26 +63,7 @@ export async function run(_args: string[]): Promise<void> {
   const authDir = path.join(projectRoot, 'store', 'auth');
   const hasAuth = fs.existsSync(authDir) && fs.readdirSync(authDir).length > 0;
 
-  let hasRegisteredGroups = false;
-  // Check JSON file first (pre-migration)
-  if (fs.existsSync(path.join(projectRoot, 'data', 'registered_groups.json'))) {
-    hasRegisteredGroups = true;
-  } else {
-    // Check SQLite directly using better-sqlite3 (no sqlite3 CLI needed)
-    const dbPath = path.join(STORE_DIR, 'messages.db');
-    if (fs.existsSync(dbPath)) {
-      try {
-        const db = new Database(dbPath, { readonly: true });
-        const row = db
-          .prepare('SELECT COUNT(*) as count FROM registered_groups')
-          .get() as { count: number };
-        if (row.count > 0) hasRegisteredGroups = true;
-        db.close();
-      } catch {
-        // Table might not exist yet
-      }
-    }
-  }
+  const hasRegisteredGroups = detectRegisteredGroups(projectRoot);
 
   // Check for existing OpenClaw installation
   const homedir = (await import('os')).homedir();
@@ -78,7 +77,6 @@ export async function run(_args: string[]): Promise<void> {
     {
       platform,
       wsl,
-      appleContainer,
       docker,
       hasEnv,
       hasAuth,
@@ -91,7 +89,6 @@ export async function run(_args: string[]): Promise<void> {
     PLATFORM: platform,
     IS_WSL: wsl,
     IS_HEADLESS: headless,
-    APPLE_CONTAINER: appleContainer,
     DOCKER: docker,
     HAS_ENV: hasEnv,
     HAS_AUTH: hasAuth,
diff --git a/setup/index.ts b/setup/index.ts
index 526ea7d..200b9e2 100644
--- a/setup/index.ts
+++ b/setup/index.ts
@@ -10,9 +10,13 @@ const STEPS: Record<
   () => Promise<{ run: (args: string[]) => Promise<void> }>
 > = {
   timezone: () => import('./timezone.js'),
+  'set-env': () => import('./set-env.js'),
   environment: () => import('./environment.js'),
   container: () => import('./container.js'),
   register: () => import('./register.js'),
+  groups: () => import('./groups.js'),
+  'whatsapp-auth': () => import('./whatsapp-auth.js'),
+  'signal-auth': () => import('./signal-auth.js'),
   mounts: () => import('./mounts.js'),
   service: () => import('./service.js'),
   verify: () => import('./verify.js'),
diff --git a/setup/install-claude.sh b/setup/install-claude.sh
new file mode 100755
index 0000000..485f0b4
--- /dev/null
+++ b/setup/install-claude.sh
@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Install the Claude Code CLI on the host via the official native installer.
+# Invoked from setup/register-claude-token.sh when the user picks the
+# subscription auth path and `claude` is missing. The other two auth paths
+# (paste OAuth token, paste API key) don't need the CLI, so this runs on
+# demand rather than up front.
+#
+# The native installer is Node-independent (downloads a prebuilt binary to
+# ~/.local/bin) and is the path Anthropic documents. This matches the
+# pattern used by install-docker.sh / install-node.sh: the script itself is
+# the allowlisted unit; the curl | bash pipe lives inside it.
+
+set -euo pipefail
+
+echo "=== NANOCLAW SETUP: INSTALL_CLAUDE ==="
+
+if command -v claude >/dev/null 2>&1; then
+  echo "STATUS: already-installed"
+  echo "CLAUDE_VERSION: $(claude --version 2>/dev/null || echo unknown)"
+  echo "=== END ==="
+  exit 0
+fi
+
+if ! command -v curl >/dev/null 2>&1; then
+  echo "STATUS: failed"
+  echo "ERROR: curl not available."
+  echo "=== END ==="
+  exit 1
+fi
+
+echo "STEP: claude-native-install"
+curl -fsSL https://claude.ai/install.sh | bash
+
+# Native installer writes to ~/.local/bin and appends a PATH line to the
+# user's rc file; that doesn't help this session, so put it on PATH now.
+if [ -d "$HOME/.local/bin" ] && [[ ":$PATH:" != *":$HOME/.local/bin:"* ]]; then
+  export PATH="$HOME/.local/bin:$PATH"
+fi
+hash -r 2>/dev/null || true
+
+if ! command -v claude >/dev/null 2>&1; then
+  echo "STATUS: failed"
+  echo "ERROR: claude not found on PATH after install."
+  echo "=== END ==="
+  exit 1
+fi
+
+echo "STATUS: installed"
+echo "CLAUDE_VERSION: $(claude --version 2>/dev/null || echo unknown)"
+echo "=== END ==="
diff --git a/setup/install-discord.sh b/setup/install-discord.sh
new file mode 100755
index 0000000..6f5a9c8
--- /dev/null
+++ b/setup/install-discord.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-discord — bundles the preflight + install commands
+# from the /add-discord skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Discord adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @chat-adapter/discord package;
+# builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_DISCORD ==="
+
+needs_install=false
+[[ -f src/channels/discord.ts ]] || needs_install=true
+grep -q "import './discord.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/discord"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/discord ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/discord.ts > src/channels/discord.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './discord.js';" src/channels/index.ts; then
+  printf "import './discord.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/discord@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-docker.sh b/setup/install-docker.sh
new file mode 100755
index 0000000..4aaadce
--- /dev/null
+++ b/setup/install-docker.sh
@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Setup helper: install-docker — bundles Docker install into one idempotent
+# script so /new-setup can run it without needing `curl | sh` in the allowlist
+# (pipelines split at matching time, and `sh` receiving stdin can't be
+# pre-approved safely).
+#
+# The script itself is the allowlisted unit; the pipes and sudo live inside
+# it. Starting the daemon (after install) stays separate — `open -a Docker`
+# and `sudo systemctl start docker` are already in the allowlist.
+set -euo pipefail
+
+echo "=== NANOCLAW SETUP: INSTALL_DOCKER ==="
+
+if command -v docker >/dev/null 2>&1; then
+  echo "STATUS: already-installed"
+  echo "DOCKER_VERSION: $(docker --version 2>/dev/null || echo unknown)"
+  echo "=== END ==="
+  exit 0
+fi
+
+case "$(uname -s)" in
+  Darwin)
+    echo "STEP: brew-install-docker"
+    if ! command -v brew >/dev/null 2>&1; then
+      echo "STATUS: failed"
+      echo "ERROR: Homebrew not installed. Install brew first (https://brew.sh) then re-run."
+      echo "=== END ==="
+      exit 1
+    fi
+    brew install --cask docker
+    ;;
+  Linux)
+    echo "STEP: docker-get-script"
+    curl -fsSL https://get.docker.com | sh
+    echo "STEP: usermod-docker-group"
+    sudo usermod -aG docker "$USER"
+    echo "NOTE: you may need to log out and back in for docker group membership to take effect"
+    ;;
+  *)
+    echo "STATUS: failed"
+    echo "ERROR: Unsupported platform: $(uname -s)"
+    echo "=== END ==="
+    exit 1
+    ;;
+esac
+
+if ! command -v docker >/dev/null 2>&1; then
+  echo "STATUS: failed"
+  echo "ERROR: docker not found on PATH after install"
+  echo "=== END ==="
+  exit 1
+fi
+
+echo "STATUS: installed"
+echo "DOCKER_VERSION: $(docker --version 2>/dev/null || echo unknown)"
+echo "=== END ==="
diff --git a/setup/install-gchat.sh b/setup/install-gchat.sh
new file mode 100755
index 0000000..b9166f1
--- /dev/null
+++ b/setup/install-gchat.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-gchat — bundles the preflight + install commands
+# from the /add-gchat skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Google Chat adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @chat-adapter/gchat package;
+# builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_GCHAT ==="
+
+needs_install=false
+[[ -f src/channels/gchat.ts ]] || needs_install=true
+grep -q "import './gchat.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/gchat"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/gchat ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/gchat.ts > src/channels/gchat.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './gchat.js';" src/channels/index.ts; then
+  printf "import './gchat.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/gchat@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-github.sh b/setup/install-github.sh
new file mode 100755
index 0000000..cb28bfc
--- /dev/null
+++ b/setup/install-github.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-github — bundles the preflight + install commands
+# from the /add-github skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the GitHub adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @chat-adapter/github package;
+# builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_GITHUB ==="
+
+needs_install=false
+[[ -f src/channels/github.ts ]] || needs_install=true
+grep -q "import './github.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/github"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/github ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/github.ts > src/channels/github.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './github.js';" src/channels/index.ts; then
+  printf "import './github.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/github@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-imessage.sh b/setup/install-imessage.sh
new file mode 100755
index 0000000..864e127
--- /dev/null
+++ b/setup/install-imessage.sh
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+# Setup helper: install-imessage — bundles the preflight + install commands
+# from the /add-imessage skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the iMessage adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned chat-adapter-imessage package;
+# builds. Local vs remote mode pick stays in the skill — this script only
+# handles the deterministic install. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_IMESSAGE ==="
+
+needs_install=false
+[[ -f src/channels/imessage.ts ]] || needs_install=true
+grep -q "import './imessage.js';" src/channels/index.ts || needs_install=true
+grep -q '"chat-adapter-imessage"' package.json || needs_install=true
+[[ -d node_modules/chat-adapter-imessage ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/imessage.ts > src/channels/imessage.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './imessage.js';" src/channels/index.ts; then
+  printf "import './imessage.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install chat-adapter-imessage@0.1.1
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-linear.sh b/setup/install-linear.sh
new file mode 100755
index 0000000..f8788be
--- /dev/null
+++ b/setup/install-linear.sh
@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+# Setup helper: install-linear — bundles the preflight + install commands
+# from the /add-linear skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Linear adapter in from the `channels` branch; appends the
+# self-registration import; patches src/channels/chat-sdk-bridge.ts to add
+# catch-all forwarding (Linear OAuth apps can't be @-mentioned, so the
+# onNewMention handler never fires — the bridge needs a catchAll path);
+# installs the pinned @chat-adapter/linear package; builds. All steps are
+# safe to re-run.
+#
+# Note: the bridge patch's onNewMessage handler passes `false` for isMention
+# (current trunk signature requires the arg). The /add-linear SKILL's
+# snippet omits the arg — this script uses the full signature so TypeScript
+# builds cleanly.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_LINEAR ==="
+
+needs_install=false
+[[ -f src/channels/linear.ts ]] || needs_install=true
+grep -q "import './linear.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/linear"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/linear ]] || needs_install=true
+grep -q 'catchAll' src/channels/chat-sdk-bridge.ts || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/linear.ts > src/channels/linear.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './linear.js';" src/channels/index.ts; then
+  printf "import './linear.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: patch-bridge-catchall-field"
+if ! grep -q 'catchAll?: boolean;' src/channels/chat-sdk-bridge.ts; then
+  awk '
+    /^export interface ChatSdkBridgeConfig \{/ { in_iface = 1 }
+    in_iface && /^\}/ && !inserted {
+      print "  /**"
+      print "   * Forward ALL messages in unsubscribed threads, not just @-mentions."
+      print "   * Use for platforms where the bot identity can'\''t be @-mentioned (e.g."
+      print "   * Linear OAuth apps). The thread is auto-subscribed on first message."
+      print "   */"
+      print "  catchAll?: boolean;"
+      inserted = 1
+      in_iface = 0
+    }
+    { print }
+  ' src/channels/chat-sdk-bridge.ts > src/channels/chat-sdk-bridge.ts.tmp \
+    && mv src/channels/chat-sdk-bridge.ts.tmp src/channels/chat-sdk-bridge.ts
+fi
+
+echo "STEP: patch-bridge-catchall-handler"
+if ! grep -q 'if (config.catchAll) {' src/channels/chat-sdk-bridge.ts; then
+  awk '
+    /      \/\/ DMs — apply engage rules too/ && !inserted {
+      print "      // Catch-all for platforms where @-mention isn'\''t possible (e.g. Linear"
+      print "      // OAuth apps). Forward every unsubscribed message and auto-subscribe."
+      print "      if (config.catchAll) {"
+      print "        chat.onNewMessage(/.*/, async (thread, message) => {"
+      print "          const channelId = adapter.channelIdFromThreadId(thread.id);"
+      print "          await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message, false));"
+      print "          await thread.subscribe();"
+      print "        });"
+      print "      }"
+      print ""
+      inserted = 1
+    }
+    { print }
+  ' src/channels/chat-sdk-bridge.ts > src/channels/chat-sdk-bridge.ts.tmp \
+    && mv src/channels/chat-sdk-bridge.ts.tmp src/channels/chat-sdk-bridge.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/linear@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-matrix.sh b/setup/install-matrix.sh
new file mode 100755
index 0000000..c985473
--- /dev/null
+++ b/setup/install-matrix.sh
@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+# Setup helper: install-matrix — bundles the preflight + install commands
+# from the /add-matrix skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Matrix adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @beeper/chat-adapter-matrix
+# package; patches the adapter's published dist so its matrix-js-sdk/lib
+# imports carry .js extensions (required under Node 22 strict ESM); builds.
+# All steps are safe to re-run — re-run this script after any pnpm install
+# that touches the adapter.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_MATRIX ==="
+
+needs_install=false
+[[ -f src/channels/matrix.ts ]] || needs_install=true
+grep -q "import './matrix.js';" src/channels/index.ts || needs_install=true
+grep -q '"@beeper/chat-adapter-matrix"' package.json || needs_install=true
+[[ -d node_modules/@beeper/chat-adapter-matrix ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/matrix.ts > src/channels/matrix.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './matrix.js';" src/channels/index.ts; then
+  printf "import './matrix.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @beeper/chat-adapter-matrix@0.2.0
+
+echo "STEP: patch-esm-extensions"
+node -e '
+  const fs = require("fs"), path = require("path");
+  const root = "node_modules/.pnpm";
+  const dir = fs.readdirSync(root).find(d => d.startsWith("@beeper+chat-adapter-matrix@"));
+  if (!dir) { console.log("Matrix adapter not installed"); process.exit(0); }
+  const f = path.join(root, dir, "node_modules/@beeper/chat-adapter-matrix/dist/index.js");
+  fs.writeFileSync(f, fs.readFileSync(f, "utf8").replace(
+    /from "(matrix-js-sdk\/lib\/[^"]+?)(?<!\.js)"/g, "from \"$1.js\""
+  ));
+  console.log("Patched", f);
+'
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-node.sh b/setup/install-node.sh
new file mode 100755
index 0000000..e100ccd
--- /dev/null
+++ b/setup/install-node.sh
@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+# Setup helper: install-node — bundles Node 22 install into one idempotent
+# script so /new-setup can run it without needing `curl | sudo -E bash -` in
+# the allowlist (that pattern is inherently unmatchable — bash reads from
+# stdin, so pre-approval can't inspect what's being executed).
+#
+# The script itself is the allowlisted unit; the pipes and sudo live inside
+# it. Pure bash by design — runs before Node exists on the host.
+set -euo pipefail
+
+echo "=== NANOCLAW SETUP: INSTALL_NODE ==="
+
+if command -v node >/dev/null 2>&1; then
+  echo "STATUS: already-installed"
+  echo "NODE_VERSION: $(node --version)"
+  echo "=== END ==="
+  exit 0
+fi
+
+case "$(uname -s)" in
+  Darwin)
+    echo "STEP: brew-install-node"
+    if ! command -v brew >/dev/null 2>&1; then
+      echo "STATUS: failed"
+      echo "ERROR: Homebrew not installed. Install brew first (https://brew.sh) then re-run."
+      echo "=== END ==="
+      exit 1
+    fi
+    brew install node@22
+    ;;
+  Linux)
+    echo "STEP: nodesource-setup"
+    curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
+    echo "STEP: apt-install-nodejs"
+    sudo apt-get install -y nodejs
+    ;;
+  *)
+    echo "STATUS: failed"
+    echo "ERROR: Unsupported platform: $(uname -s)"
+    echo "=== END ==="
+    exit 1
+    ;;
+esac
+
+if ! command -v node >/dev/null 2>&1; then
+  echo "STATUS: failed"
+  echo "ERROR: node not found on PATH after install"
+  echo "=== END ==="
+  exit 1
+fi
+
+echo "STATUS: installed"
+echo "NODE_VERSION: $(node --version)"
+echo "=== END ==="
diff --git a/setup/install-resend.sh b/setup/install-resend.sh
new file mode 100755
index 0000000..9f18a9f
--- /dev/null
+++ b/setup/install-resend.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-resend — bundles the preflight + install commands
+# from the /add-resend skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Resend adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @resend/chat-sdk-adapter
+# package; builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_RESEND ==="
+
+needs_install=false
+[[ -f src/channels/resend.ts ]] || needs_install=true
+grep -q "import './resend.js';" src/channels/index.ts || needs_install=true
+grep -q '"@resend/chat-sdk-adapter"' package.json || needs_install=true
+[[ -d node_modules/@resend/chat-sdk-adapter ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/resend.ts > src/channels/resend.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './resend.js';" src/channels/index.ts; then
+  printf "import './resend.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @resend/chat-sdk-adapter@0.1.1
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-slack.sh b/setup/install-slack.sh
new file mode 100755
index 0000000..55d5e85
--- /dev/null
+++ b/setup/install-slack.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-slack — bundles the preflight + install commands
+# from the /add-slack skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Slack adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @chat-adapter/slack package;
+# builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_SLACK ==="
+
+needs_install=false
+[[ -f src/channels/slack.ts ]] || needs_install=true
+grep -q "import './slack.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/slack"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/slack ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/slack.ts > src/channels/slack.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './slack.js';" src/channels/index.ts; then
+  printf "import './slack.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/slack@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-teams.sh b/setup/install-teams.sh
new file mode 100755
index 0000000..4b8c216
--- /dev/null
+++ b/setup/install-teams.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-teams — bundles the preflight + install commands
+# from the /add-teams skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Teams adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @chat-adapter/teams package;
+# builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_TEAMS ==="
+
+needs_install=false
+[[ -f src/channels/teams.ts ]] || needs_install=true
+grep -q "import './teams.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/teams"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/teams ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/teams.ts > src/channels/teams.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './teams.js';" src/channels/index.ts; then
+  printf "import './teams.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/teams@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-telegram.sh b/setup/install-telegram.sh
new file mode 100755
index 0000000..307dba2
--- /dev/null
+++ b/setup/install-telegram.sh
@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+# Setup helper: install-telegram — bundles the preflight + install commands
+# from the /add-telegram skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials and pairing.
+#
+# Copies the Telegram adapter, helpers, tests, and the pair-telegram setup
+# step in from the `channels` branch; appends the self-registration import;
+# registers the `pair-telegram` entry in the setup STEPS map; installs the
+# pinned @chat-adapter/telegram package; builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_TELEGRAM ==="
+
+CHANNEL_FILES=(
+  src/channels/telegram.ts
+  src/channels/telegram-pairing.ts
+  src/channels/telegram-pairing.test.ts
+  src/channels/telegram-markdown-sanitize.ts
+  src/channels/telegram-markdown-sanitize.test.ts
+  setup/pair-telegram.ts
+)
+
+needs_install=false
+for f in "${CHANNEL_FILES[@]}"; do
+  [[ -f "$f" ]] || needs_install=true
+done
+grep -q "import './telegram.js';" src/channels/index.ts || needs_install=true
+grep -q "'pair-telegram':" setup/index.ts || needs_install=true
+grep -q '"@chat-adapter/telegram"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/telegram ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+for f in "${CHANNEL_FILES[@]}"; do
+  git show "origin/channels:$f" > "$f"
+done
+
+echo "STEP: register-import"
+if ! grep -q "import './telegram.js';" src/channels/index.ts; then
+  printf "import './telegram.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: register-setup-step"
+if ! grep -q "'pair-telegram':" setup/index.ts; then
+  awk '
+    { print }
+    /register: \(\) => import/ && !inserted {
+      print "  '\''pair-telegram'\'': () => import('\''./pair-telegram.js'\''),"
+      inserted = 1
+    }
+  ' setup/index.ts > setup/index.ts.tmp && mv setup/index.ts.tmp setup/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/telegram@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-webex.sh b/setup/install-webex.sh
new file mode 100755
index 0000000..adf52fc
--- /dev/null
+++ b/setup/install-webex.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-webex — bundles the preflight + install commands
+# from the /add-webex skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to credentials.
+#
+# Copies the Webex adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @bitbasti/chat-adapter-webex
+# package; builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_WEBEX ==="
+
+needs_install=false
+[[ -f src/channels/webex.ts ]] || needs_install=true
+grep -q "import './webex.js';" src/channels/index.ts || needs_install=true
+grep -q '"@bitbasti/chat-adapter-webex"' package.json || needs_install=true
+[[ -d node_modules/@bitbasti/chat-adapter-webex ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/webex.ts > src/channels/webex.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './webex.js';" src/channels/index.ts; then
+  printf "import './webex.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @bitbasti/chat-adapter-webex@0.1.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-whatsapp-cloud.sh b/setup/install-whatsapp-cloud.sh
new file mode 100755
index 0000000..70e8e02
--- /dev/null
+++ b/setup/install-whatsapp-cloud.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Setup helper: install-whatsapp-cloud — bundles the preflight + install
+# commands from the /add-whatsapp-cloud skill into one idempotent script so
+# /new-setup can run them programmatically before continuing to credentials.
+#
+# Copies the WhatsApp Cloud adapter in from the `channels` branch; appends the
+# self-registration import; installs the pinned @chat-adapter/whatsapp package;
+# builds. All steps are safe to re-run.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_WHATSAPP_CLOUD ==="
+
+needs_install=false
+[[ -f src/channels/whatsapp-cloud.ts ]] || needs_install=true
+grep -q "import './whatsapp-cloud.js';" src/channels/index.ts || needs_install=true
+grep -q '"@chat-adapter/whatsapp"' package.json || needs_install=true
+[[ -d node_modules/@chat-adapter/whatsapp ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+git show origin/channels:src/channels/whatsapp-cloud.ts > src/channels/whatsapp-cloud.ts
+
+echo "STEP: register-import"
+if ! grep -q "import './whatsapp-cloud.js';" src/channels/index.ts; then
+  printf "import './whatsapp-cloud.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @chat-adapter/whatsapp@4.26.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/install-whatsapp.sh b/setup/install-whatsapp.sh
new file mode 100755
index 0000000..1c62d65
--- /dev/null
+++ b/setup/install-whatsapp.sh
@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+# Setup helper: install-whatsapp — bundles the preflight + install commands
+# from the /add-whatsapp skill into one idempotent script so /new-setup can
+# run them programmatically before continuing to QR/pairing-code auth.
+#
+# Copies the native Baileys WhatsApp adapter, its whatsapp-auth and groups
+# setup steps in from the `channels` branch; appends the self-registration
+# import; registers `groups` and `whatsapp-auth` entries in the setup STEPS
+# map; installs the pinned @whiskeysockets/baileys + qrcode + pino packages;
+# builds. All steps are safe to re-run. QR/pairing-code authentication
+# stays in the skill — this script only handles the deterministic install.
+set -euo pipefail
+
+PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+echo "=== NANOCLAW SETUP: INSTALL_WHATSAPP ==="
+
+CHANNEL_FILES=(
+  src/channels/whatsapp.ts
+  setup/whatsapp-auth.ts
+  setup/groups.ts
+)
+
+needs_install=false
+for f in "${CHANNEL_FILES[@]}"; do
+  [[ -f "$f" ]] || needs_install=true
+done
+grep -q "import './whatsapp.js';" src/channels/index.ts || needs_install=true
+grep -q "groups: " setup/index.ts || needs_install=true
+grep -q "'whatsapp-auth':" setup/index.ts || needs_install=true
+grep -q '"@whiskeysockets/baileys"' package.json || needs_install=true
+grep -q '"qrcode"' package.json || needs_install=true
+grep -q '"pino"' package.json || needs_install=true
+[[ -d node_modules/@whiskeysockets/baileys ]] || needs_install=true
+
+if ! $needs_install; then
+  echo "STATUS: already-installed"
+  echo "=== END ==="
+  exit 0
+fi
+
+echo "STEP: fetch-channels-branch"
+git fetch origin channels
+
+echo "STEP: copy-files"
+for f in "${CHANNEL_FILES[@]}"; do
+  git show "origin/channels:$f" > "$f"
+done
+
+echo "STEP: register-import"
+if ! grep -q "import './whatsapp.js';" src/channels/index.ts; then
+  printf "import './whatsapp.js';\n" >> src/channels/index.ts
+fi
+
+echo "STEP: register-setup-steps"
+if ! grep -q "'whatsapp-auth':" setup/index.ts; then
+  awk '
+    { print }
+    /register: \(\) => import/ && !inserted {
+      print "  groups: () => import('\''./groups.js'\''),"
+      print "  '\''whatsapp-auth'\'': () => import('\''./whatsapp-auth.js'\''),"
+      inserted = 1
+    }
+  ' setup/index.ts > setup/index.ts.tmp && mv setup/index.ts.tmp setup/index.ts
+fi
+
+echo "STEP: pnpm-install"
+pnpm install @whiskeysockets/baileys@6.17.16 qrcode@1.5.4 @types/qrcode@1.5.6 pino@9.6.0
+
+echo "STEP: pnpm-build"
+pnpm run build
+
+echo "STATUS: installed"
+echo "=== END ==="
diff --git a/setup/lib/agent-ping.test.ts b/setup/lib/agent-ping.test.ts
new file mode 100644
index 0000000..5f2be2c
--- /dev/null
+++ b/setup/lib/agent-ping.test.ts
@@ -0,0 +1,30 @@
+import { describe, expect, it } from 'vitest';
+
+import { classifyPingResult } from './agent-ping.js';
+
+describe('classifyPingResult', () => {
+  it('treats a normal text reply as ok', () => {
+    expect(classifyPingResult(0, 'pong\n')).toBe('ok');
+  });
+
+  it('detects Anthropic auth errors printed as a chat reply', () => {
+    expect(
+      classifyPingResult(
+        0,
+        'Failed to authenticate. API Error: 401 {"type":"error","error":{"type":"authentication_error","message":"Invalid bearer token"}}',
+      ),
+    ).toBe('auth_error');
+  });
+
+  it('detects auth errors on stderr too', () => {
+    expect(classifyPingResult(1, '', 'Authentication error')).toBe('auth_error');
+  });
+
+  it('preserves socket errors', () => {
+    expect(classifyPingResult(2, '')).toBe('socket_error');
+  });
+
+  it('treats empty output as no reply', () => {
+    expect(classifyPingResult(0, '')).toBe('no_reply');
+  });
+});
diff --git a/setup/lib/agent-ping.ts b/setup/lib/agent-ping.ts
new file mode 100644
index 0000000..49c5fe2
--- /dev/null
+++ b/setup/lib/agent-ping.ts
@@ -0,0 +1,66 @@
+/**
+ * Round-trip check against the CLI Unix socket.
+ *
+ * Shared by `setup/verify.ts` (end-of-run health check) and `setup/auto.ts`
+ * (confirm the freshly-wired agent actually responds before prompting the
+ * user to chat with it).
+ *
+ * Exit-code contract follows `scripts/chat.ts`:
+ *   0  → got a reply on stdout
+ *   2  → socket unreachable (service not running or wrong checkout)
+ *   3  → no reply before chat.ts's own 120s hard stop
+ * This wrapper also guards with its own timeout in case chat.ts hangs.
+ */
+import { spawn } from 'child_process';
+
+export type PingResult = 'ok' | 'no_reply' | 'socket_error' | 'auth_error';
+
+export function classifyPingResult(exitCode: number | null, stdout: string, stderr = ''): PingResult {
+  const output = `${stdout}\n${stderr}`;
+  if (
+    /Invalid bearer token/i.test(output) ||
+    /authentication[_ ]error/i.test(output) ||
+    /Failed to authenticate/i.test(output)
+  ) {
+    return 'auth_error';
+  }
+  if (exitCode === 2) return 'socket_error';
+  if (exitCode === 0 && stdout.trim().length > 0) return 'ok';
+  return 'no_reply';
+}
+
+export function pingCliAgent(timeoutMs = 30_000): Promise<PingResult> {
+  return new Promise((resolve) => {
+    const child = spawn('pnpm', ['run', 'chat', 'ping'], {
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    let stdout = '';
+    let stderr = '';
+    let settled = false;
+    const timer = setTimeout(() => {
+      if (settled) return;
+      settled = true;
+      child.kill('SIGKILL');
+      resolve('no_reply');
+    }, timeoutMs);
+
+    child.stdout.on('data', (chunk: Buffer) => {
+      stdout += chunk.toString('utf-8');
+    });
+    child.stderr.on('data', (chunk: Buffer) => {
+      stderr += chunk.toString('utf-8');
+    });
+    child.on('close', (code) => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      resolve(classifyPingResult(code, stdout, stderr));
+    });
+    child.on('error', () => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      resolve('socket_error');
+    });
+  });
+}
diff --git a/setup/lib/bright-select.ts b/setup/lib/bright-select.ts
new file mode 100644
index 0000000..94c4838
--- /dev/null
+++ b/setup/lib/bright-select.ts
@@ -0,0 +1,119 @@
+/**
+ * A drop-in alternative to `@clack/prompts`' `p.select` that renders
+ * unselected option labels at full brightness instead of dim gray.
+ *
+ * Why this exists: clack styles inactive options with `styleText("dim", …)`
+ * inline in its render function. There is no configuration hook to override
+ * it, and the feedback was clear — non-selected options in the setup flow
+ * were "too light, need stronger font weight". So we write our own render
+ * against `@clack/core`'s `SelectPrompt`, keeping the visual shell of clack
+ * (diamond header, `│` gutter, cyan in-progress / green on submit) but
+ * leaving the label un-dimmed. Only the bullet and hint stay dim, which
+ * gives enough contrast for the cursor to read as "active".
+ *
+ * Not a full clack-feature clone: no search, no maxItems paging, no custom
+ * bar characters. Just the bits the NanoClaw setup menus actually use.
+ */
+import { SelectPrompt } from '@clack/core';
+import { isCancel } from '@clack/prompts';
+import { styleText } from 'node:util';
+
+const BULLET_ACTIVE = '●';
+const BULLET_INACTIVE = '○';
+const BAR = '│';
+const CAP_BOT = '└';
+const DIAMOND = '◆';
+const DIAMOND_CANCEL = '■';
+const DIAMOND_SUBMIT = '◇';
+
+type PromptState = 'initial' | 'active' | 'error' | 'cancel' | 'submit';
+
+function stateColor(state: PromptState): 'cyan' | 'green' | 'red' | 'yellow' {
+  switch (state) {
+    case 'submit':
+      return 'green';
+    case 'cancel':
+      return 'red';
+    case 'error':
+      return 'yellow';
+    default:
+      return 'cyan';
+  }
+}
+
+function headerIcon(state: PromptState): string {
+  switch (state) {
+    case 'submit':
+      return styleText('green', DIAMOND_SUBMIT);
+    case 'cancel':
+      return styleText('red', DIAMOND_CANCEL);
+    default:
+      return styleText('cyan', DIAMOND);
+  }
+}
+
+export interface BrightSelectOption<T> {
+  value: T;
+  label?: string;
+  hint?: string;
+}
+
+export interface BrightSelectOptions<T> {
+  message: string;
+  options: BrightSelectOption<T>[];
+  initialValue?: T;
+}
+
+/**
+ * Matches the return shape of `p.select` — resolves to the selected value
+ * on submit, or to clack's cancel symbol on Ctrl-C / Esc. Callers pass
+ * the result through `ensureAnswer(...)` the same way they do for
+ * `p.select`.
+ */
+export function brightSelect<T>(
+  opts: BrightSelectOptions<T>,
+): Promise<T | symbol> {
+  const { message, options, initialValue } = opts;
+
+  return new SelectPrompt({
+    options: options as Array<{ value: T; label?: string; hint?: string }>,
+    initialValue,
+    render() {
+      const st = this.state as PromptState;
+      const color = stateColor(st);
+      const bar = styleText(color, BAR);
+      const grayBar = styleText('gray', BAR);
+
+      const lines: string[] = [];
+      lines.push(grayBar);
+      lines.push(`${headerIcon(st)}  ${message}`);
+
+      if (st === 'submit' || st === 'cancel') {
+        const selected =
+          options.find((o) => o.value === this.value)?.label ??
+          String(this.value ?? '');
+        const shown =
+          st === 'cancel'
+            ? styleText(['strikethrough', 'dim'], selected)
+            : styleText('dim', selected);
+        lines.push(`${grayBar}  ${shown}`);
+        return lines.join('\n');
+      }
+
+      const cursor = (this as unknown as { cursor: number }).cursor;
+      options.forEach((opt, idx) => {
+        const label = opt.label ?? String(opt.value);
+        const hint = opt.hint ? ` ${styleText('dim', `(${opt.hint})`)}` : '';
+        const marker =
+          idx === cursor
+            ? styleText('green', BULLET_ACTIVE)
+            : styleText('dim', BULLET_INACTIVE);
+        lines.push(`${bar}  ${marker} ${label}${hint}`);
+      });
+      lines.push(styleText(color, CAP_BOT));
+      return lines.join('\n');
+    },
+  }).prompt() as Promise<T | symbol>;
+}
+
+export { isCancel };
diff --git a/setup/lib/browser.ts b/setup/lib/browser.ts
new file mode 100644
index 0000000..9d801fa
--- /dev/null
+++ b/setup/lib/browser.ts
@@ -0,0 +1,51 @@
+/**
+ * Browser-open helpers shared across channel setup flows.
+ *
+ * `openUrl` is best-effort — silent on failure, so headless/SSH/WSL
+ * environments where `open`/`xdg-open` isn't wired up don't crash the
+ * setup. The URL should always be visible in the clack note that calls
+ * this so the user can copy-paste if the auto-open doesn't land.
+ *
+ * `confirmThenOpen` pauses for the operator before triggering the open —
+ * the browser tends to steal focus when it pops, and a split-second
+ * "wait what just happened" moment is worse than letting the user hit
+ * Enter when they're ready.
+ */
+import { spawn } from 'child_process';
+
+import * as p from '@clack/prompts';
+
+import { ensureAnswer } from './runner.js';
+
+/** Best-effort open of a URL in the user's default browser. Silent on failure. */
+export function openUrl(url: string): void {
+  try {
+    const cmd = process.platform === 'darwin' ? 'open' : 'xdg-open';
+    const child = spawn(cmd, [url], { stdio: 'ignore', detached: true });
+    child.on('error', () => {
+      // Headless / no browser / unknown command — URL is printed in the
+      // calling note so the user can copy-paste.
+    });
+    child.unref();
+  } catch {
+    // swallow — URL is visible in the note.
+  }
+}
+
+/**
+ * Gate a browser-open on a confirm so the user is ready for their browser
+ * to take focus. Proceeds on cancel as well — the user can always copy the
+ * URL from the note that precedes the prompt.
+ */
+export async function confirmThenOpen(
+  url: string,
+  message = 'Press Enter to open your browser',
+): Promise<void> {
+  ensureAnswer(
+    await p.confirm({
+      message,
+      initialValue: true,
+    }),
+  );
+  openUrl(url);
+}
diff --git a/setup/lib/channels-remote.sh b/setup/lib/channels-remote.sh
new file mode 100644
index 0000000..6da0159
--- /dev/null
+++ b/setup/lib/channels-remote.sh
@@ -0,0 +1,38 @@
+# channels-remote.sh — resolve the git remote that carries the `channels`
+# branch. Source this file and call `resolve_channels_remote`; echoes the
+# remote name (e.g. `origin` or `upstream`).
+#
+# Typical fork setups keep the upstream nanoclaw repo under a remote named
+# `upstream`, with `origin` pointing at the user's fork. The channels branch
+# only lives upstream, so a hardcoded `git fetch origin channels` fails for
+# forks. This helper walks `git remote -v`, picks the remote whose URL points
+# at qwibitai/nanoclaw, and prints its name.
+#
+# Fallback: if no existing remote matches, add `upstream` pointing at
+# github.com/qwibitai/nanoclaw and return that — keeps forks without an
+# explicit upstream configured working on the first try.
+#
+# Explicit override: set NANOCLAW_CHANNELS_REMOTE=<name> to skip detection.
+
+resolve_channels_remote() {
+  if [ -n "${NANOCLAW_CHANNELS_REMOTE:-}" ]; then
+    printf '%s' "$NANOCLAW_CHANNELS_REMOTE"
+    return 0
+  fi
+
+  local remote url
+  while IFS=$'\t' read -r remote url; do
+    case "$url" in
+      *qwibitai/nanoclaw*)
+        printf '%s' "$remote"
+        return 0
+        ;;
+    esac
+  done < <(git remote -v 2>/dev/null | awk '$3 == "(fetch)" { print $1"\t"$2 }')
+
+  # No matching remote — add `upstream` and use it. Silent on failure so
+  # callers see the eventual `git fetch` error rather than a cryptic
+  # remote-add failure.
+  git remote add upstream https://github.com/qwibitai/nanoclaw.git 2>/dev/null || true
+  printf '%s' "upstream"
+}
diff --git a/setup/lib/claude-assist.ts b/setup/lib/claude-assist.ts
new file mode 100644
index 0000000..1651a9c
--- /dev/null
+++ b/setup/lib/claude-assist.ts
@@ -0,0 +1,442 @@
+/**
+ * Offer Claude-assisted debugging when a setup step fails.
+ *
+ * Flow:
+ *   1. Check `claude` is on PATH and has a working credential. If not,
+ *      silently skip — pre-auth failures can't use this path.
+ *   2. Ask the user for consent ("Want me to ask Claude for a fix?").
+ *   3. Build a minimal prompt: the one-paragraph situation, the failing
+ *      step's name/message/hint, and a short list of *file references*
+ *      (not contents) so Claude can Read what it needs on its own.
+ *   4. Spawn `claude -p --output-format text` with a 2-minute timeout and
+ *      a spinner that shows elapsed time.
+ *   5. Parse `REASON:` / `COMMAND:` out of the response. Show the reason
+ *      in a clack note, then hand off to `setup/run-suggested.sh` for
+ *      editable pre-fill + exec.
+ *
+ * Skippable with NANOCLAW_SKIP_CLAUDE_ASSIST=1 for CI/scripted runs.
+ */
+import { execSync, spawn } from 'child_process';
+import fs from 'fs';
+import path from 'path';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import { ensureAnswer } from './runner.js';
+import { fitToWidth } from './theme.js';
+
+export interface AssistContext {
+  stepName: string;
+  msg: string;
+  hint?: string;
+  /** Absolute path to the per-step raw log, if the caller has one. */
+  rawLogPath?: string;
+}
+
+/**
+ * File-path hints per step. Claude reads these on its own via its Read tool
+ * rather than us stuffing contents into the prompt. Keys are step names as
+ * they appear in fail() calls; values are repo-relative paths.
+ */
+const STEP_FILES: Record<string, string[]> = {
+  bootstrap: ['setup.sh', 'setup/install-node.sh', 'nanoclaw.sh'],
+  environment: ['setup/environment.ts'],
+  container: [
+    'setup/container.ts',
+    'setup/install-docker.sh',
+    'container/Dockerfile',
+  ],
+  onecli: ['setup/onecli.ts'],
+  auth: [
+    'setup/auth.ts',
+    'setup/register-claude-token.sh',
+    'setup/install-claude.sh',
+  ],
+  mounts: ['setup/mounts.ts'],
+  service: ['setup/service.ts'],
+  'cli-agent': ['setup/cli-agent.ts', 'scripts/init-cli-agent.ts'],
+  timezone: ['setup/timezone.ts', 'setup/lib/tz-from-claude.ts'],
+  channel: ['setup/auto.ts'],
+  verify: ['setup/verify.ts'],
+  // Channel-specific sub-steps:
+  'telegram-install': ['setup/add-telegram.sh', 'setup/channels/telegram.ts'],
+  'telegram-validate': ['setup/channels/telegram.ts'],
+  'pair-telegram': ['setup/pair-telegram.ts', 'setup/channels/telegram.ts'],
+  'discord-install': ['setup/add-discord.sh', 'setup/channels/discord.ts'],
+  'slack-install': ['setup/add-slack.sh', 'setup/channels/slack.ts'],
+  'slack-validate': ['setup/channels/slack.ts'],
+  'imessage-install': ['setup/add-imessage.sh', 'setup/channels/imessage.ts'],
+  'imessage': ['setup/channels/imessage.ts'],
+  'teams-install': ['setup/add-teams.sh', 'setup/channels/teams.ts'],
+  'teams-manifest': ['setup/lib/teams-manifest.ts', 'setup/channels/teams.ts'],
+  'init-first-agent': [
+    'scripts/init-first-agent.ts',
+    'setup/channels/telegram.ts',
+    'setup/channels/discord.ts',
+  ],
+};
+
+const BIG_PICTURE_FILES = ['README.md', 'setup/auto.ts'];
+
+/**
+ * Returns `true` if the user ran a Claude-suggested fix command; callers
+ * can use that signal to offer a retry instead of aborting outright.
+ * Returns `false` for every other outcome (skipped, declined, no command,
+ * Claude unreachable, user chose not to run).
+ */
+export async function offerClaudeAssist(
+  ctx: AssistContext,
+  projectRoot: string = process.cwd(),
+): Promise<boolean> {
+  if (process.env.NANOCLAW_SKIP_CLAUDE_ASSIST === '1') return false;
+  if (!isClaudeUsable()) return false;
+
+  const want = ensureAnswer(
+    await p.confirm({
+      message: 'Want me to ask Claude to diagnose this?',
+      initialValue: true,
+    }),
+  );
+  if (!want) return false;
+
+  const prompt = buildPrompt(ctx, projectRoot);
+  const response = await queryClaudeUnderSpinner(prompt, projectRoot);
+  if (!response) return false;
+
+  const parsed = parseResponse(response);
+  if (!parsed) {
+    p.log.warn("Claude responded but I couldn't parse a command out of it.");
+    p.log.message(k.dim(response.trim().slice(0, 500)));
+    return false;
+  }
+
+  p.note(
+    `${parsed.reason}\n\n${k.cyan('$')} ${parsed.command}`,
+    "Claude's suggestion",
+  );
+
+  const run = ensureAnswer(
+    await p.confirm({
+      message: 'Run this command? (you can edit it before executing)',
+      initialValue: true,
+    }),
+  );
+  if (!run) return false;
+
+  await runSuggested(parsed.command, projectRoot);
+  return true;
+}
+
+function isClaudeUsable(): boolean {
+  try {
+    execSync('command -v claude', { stdio: 'ignore' });
+  } catch {
+    return false;
+  }
+  // Availability without auth is half the story; a real query will still
+  // fail if the token isn't registered. We try first and surface the error
+  // rather than pre-checking auth with a separate round trip.
+  return true;
+}
+
+function buildPrompt(ctx: AssistContext, projectRoot: string): string {
+  const stepRefs = STEP_FILES[ctx.stepName] ?? [];
+  const references = [
+    ...BIG_PICTURE_FILES,
+    ...stepRefs,
+    'logs/setup.log',
+    ctx.rawLogPath
+      ? path.relative(projectRoot, ctx.rawLogPath)
+      : 'logs/setup-steps/',
+  ].filter((v, i, a) => a.indexOf(v) === i);
+
+  const hintLine = ctx.hint ? `Hint shown to the user: ${ctx.hint}\n` : '';
+
+  return [
+    "I'm trying to set up NanoClaw on my machine and ran into an issue",
+    'during the setup flow. Please read the referenced files to understand',
+    'the flow and the step that failed, look at the logs to see what went',
+    'wrong, then suggest a single bash command I can run to fix it.',
+    '',
+    `Failed step: ${ctx.stepName}`,
+    `Error shown to the user: ${ctx.msg}`,
+    hintLine,
+    'References (read as needed with your Read tool):',
+    ...references.map((r) => `  - ${r}`),
+    '',
+    'Respond in EXACTLY this format, nothing before or after:',
+    '',
+    'REASON: <one short line describing the root cause>',
+    'COMMAND: <single bash command, one line, no backticks>',
+    '',
+    'If no safe single command can fix it, respond with:',
+    'REASON: <why>',
+    'COMMAND: none',
+  ].join('\n');
+}
+
+/**
+ * Fixed-height scrolling window for Claude's progress.
+ *
+ * Clack's spinner only owns one line, so long tool-use breadcrumbs wrap
+ * and blow out the gutter. Instead we manage a 4-line window ourselves:
+ * a spinner header + 3 lines showing the most recent tool actions. On
+ * each update we use raw ANSI (cursor up, clear line) to redraw in
+ * place. When the query finishes we clear the whole block and emit a
+ * single `p.log.success` / `p.log.error` so the flow continues in
+ * standard clack style.
+ */
+const WINDOW_SIZE = 3;
+const SPINNER_FRAMES = ['◒', '◐', '◓', '◑'];
+const HIDE_CURSOR = '\x1b[?25l';
+const SHOW_CURSOR = '\x1b[?25h';
+
+async function queryClaudeUnderSpinner(
+  prompt: string,
+  projectRoot: string,
+): Promise<string | null> {
+  const out = process.stdout;
+  const start = Date.now();
+  const actions: string[] = [];
+  let frameIdx = 0;
+
+  const redraw = (): void => {
+    // Move cursor back to the start of the block (WINDOW_SIZE + 1 = header + window).
+    out.write(`\x1b[${WINDOW_SIZE + 1}A`);
+
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    const icon = SPINNER_FRAMES[frameIdx % SPINNER_FRAMES.length];
+    const suffix = ` (${elapsed}s)`;
+    const header = fitToWidth('Asking Claude to diagnose…', suffix);
+    out.write(`\x1b[2K${k.cyan(icon)}  ${header}${k.dim(suffix)}\n`);
+
+    for (let i = 0; i < WINDOW_SIZE; i++) {
+      const idx = actions.length - WINDOW_SIZE + i;
+      const action = idx >= 0 ? actions[idx] : '';
+      out.write('\x1b[2K');
+      if (action) {
+        out.write(`${k.gray('│')}  ${k.dim(`▸ ${fitToWidth(action, '')}`)}`);
+      } else {
+        out.write(k.gray('│'));
+      }
+      out.write('\n');
+    }
+  };
+
+  const clearBlock = (): void => {
+    out.write(`\x1b[${WINDOW_SIZE + 1}A`);
+    for (let i = 0; i < WINDOW_SIZE + 1; i++) {
+      out.write('\x1b[2K\n');
+    }
+    out.write(`\x1b[${WINDOW_SIZE + 1}A`);
+  };
+
+  // Seed the block: move cursor to a fresh line, then write (header + window)
+  // blank lines so `redraw()`'s cursor-up math lands correctly. Hide the
+  // cursor for the duration so the redraw doesn't flicker.
+  out.write(HIDE_CURSOR);
+  for (let i = 0; i < WINDOW_SIZE + 1; i++) out.write('\n');
+  redraw();
+
+  // If the user Ctrl-C's during the query, we never reach `finish()` —
+  // add an exit hook so the cursor comes back regardless.
+  const restoreCursorOnExit = (): void => {
+    out.write(SHOW_CURSOR);
+  };
+  process.once('exit', restoreCursorOnExit);
+
+  const frameTick = setInterval(() => {
+    frameIdx++;
+    redraw();
+  }, 250);
+
+  return new Promise((resolve) => {
+    let lineBuf = '';
+    let finalText = '';
+    let stderr = '';
+    let settled = false;
+
+    const finish = (
+      kind: 'ok' | 'error',
+      payload: string | null,
+    ): void => {
+      clearInterval(frameTick);
+      clearBlock();
+      out.write(SHOW_CURSOR);
+      process.off('exit', restoreCursorOnExit);
+      const elapsed = Math.round((Date.now() - start) / 1000);
+      const suffix = ` (${elapsed}s)`;
+      if (kind === 'ok') {
+        p.log.success(`${fitToWidth('Claude replied.', suffix)}${k.dim(suffix)}`);
+        resolve(payload);
+      } else {
+        p.log.error(
+          `${fitToWidth("Claude couldn't help here.", suffix)}${k.dim(suffix)}`,
+        );
+        const tail = stderr.trim().split('\n').slice(-3).join('\n');
+        if (tail) p.log.message(k.dim(tail));
+        resolve(null);
+      }
+    };
+
+    // No hard timeout — debugging can take a long time, and the cost of
+    // cutting Claude off mid-investigation is worse than letting the
+    // spinner run. The user can Ctrl-C if they want to abort.
+    //
+    // Resume the same session on repeat invocations so Claude carries
+    // context across failures in one setup run.
+    const claudeArgs = [
+      '-p',
+      '--output-format',
+      'stream-json',
+      '--verbose',
+      '--permission-mode',
+      'bypassPermissions',
+    ];
+    if (claudeSessionId) {
+      claudeArgs.push('--resume', claudeSessionId);
+    }
+    const child = spawn('claude', claudeArgs, {
+      cwd: projectRoot,
+      stdio: ['pipe', 'pipe', 'pipe'],
+    });
+
+    child.stdout.on('data', (c: Buffer) => {
+      lineBuf += c.toString('utf-8');
+      let idx: number;
+      while ((idx = lineBuf.indexOf('\n')) !== -1) {
+        const line = lineBuf.slice(0, idx);
+        lineBuf = lineBuf.slice(idx + 1);
+        if (!line.trim()) continue;
+        try {
+          const event = JSON.parse(line) as StreamEvent;
+          // Capture the session id on the very first claude invocation of
+          // this process so later calls can --resume it.
+          if (
+            !claudeSessionId &&
+            event.type === 'system' &&
+            event.subtype === 'init' &&
+            typeof event.session_id === 'string'
+          ) {
+            claudeSessionId = event.session_id;
+          }
+          handleStreamEvent(event, {
+            setAction: (a) => {
+              actions.push(a);
+              redraw();
+            },
+            appendText: (t) => {
+              finalText += t;
+            },
+          });
+        } catch {
+          // Malformed or non-JSON line — ignore.
+        }
+      }
+    });
+    child.stderr.on('data', (c: Buffer) => {
+      stderr += c.toString('utf-8');
+    });
+    child.on('close', (code) => {
+      if (settled) return;
+      settled = true;
+      if (code === 0 && finalText.trim()) finish('ok', finalText);
+      else finish('error', null);
+    });
+    child.on('error', () => {
+      if (settled) return;
+      settled = true;
+      finish('error', null);
+    });
+
+    child.stdin.end(prompt);
+  });
+}
+
+// Minimal shape of the stream-json events we care about. Claude emits
+// many more, but we only read tool_use blocks (for breadcrumbs), text
+// blocks (to reassemble the final REASON/COMMAND answer), and the
+// session_id on the init event so follow-up invocations can resume the
+// same conversation.
+interface StreamEvent {
+  type: string;
+  subtype?: string;
+  session_id?: string;
+  message?: {
+    content?: Array<
+      | { type: 'text'; text: string }
+      | { type: 'tool_use'; name: string; input: Record<string, unknown> }
+    >;
+  };
+}
+
+// The session id from the first claude-assist invocation in this process.
+// Subsequent invocations pass `--resume <id>` so Claude sees prior failures
+// as conversation history instead of treating each failure in isolation.
+let claudeSessionId: string | null = null;
+
+function handleStreamEvent(
+  event: StreamEvent,
+  cb: { setAction: (a: string) => void; appendText: (t: string) => void },
+): void {
+  if (event.type !== 'assistant') return;
+  const blocks = event.message?.content ?? [];
+  for (const block of blocks) {
+    if (block.type === 'text') {
+      cb.appendText(block.text);
+    } else if (block.type === 'tool_use') {
+      cb.setAction(formatToolUse(block.name, block.input));
+    }
+  }
+}
+
+function formatToolUse(name: string, input: Record<string, unknown>): string {
+  const truncate = (v: string, n: number): string =>
+    v.length > n ? v.slice(0, n) + '…' : v;
+  if (name === 'Read') {
+    const f = String(input.file_path ?? '');
+    return `Reading ${shortenPath(f)}`;
+  }
+  if (name === 'Bash') {
+    const cmd = String(input.command ?? '').replace(/\s+/g, ' ').trim();
+    return `Running ${truncate(cmd, 60)}`;
+  }
+  if (name === 'Grep') return `Searching for "${truncate(String(input.pattern ?? ''), 40)}"`;
+  if (name === 'Glob') return `Finding ${truncate(String(input.pattern ?? ''), 40)}`;
+  return `Using ${name}`;
+}
+
+function shortenPath(abs: string): string {
+  const root = process.cwd();
+  return abs.startsWith(`${root}/`) ? abs.slice(root.length + 1) : abs;
+}
+
+function parseResponse(
+  raw: string,
+): { reason: string; command: string } | null {
+  // Accept the fields anywhere in the output — Claude sometimes wraps the
+  // answer in a trailing explanation we can safely ignore.
+  const reasonMatch = raw.match(/^\s*REASON:\s*(.+?)\s*$/m);
+  const commandMatch = raw.match(/^\s*COMMAND:\s*(.+?)\s*$/m);
+  if (!reasonMatch || !commandMatch) return null;
+  const command = commandMatch[1].trim();
+  if (!command || command.toLowerCase() === 'none') return null;
+  return { reason: reasonMatch[1].trim(), command };
+}
+
+function runSuggested(command: string, projectRoot: string): Promise<void> {
+  const script = path.join(projectRoot, 'setup/run-suggested.sh');
+  if (!fs.existsSync(script)) {
+    p.log.error(`Missing helper: ${script}`);
+    return Promise.resolve();
+  }
+  return new Promise((resolve) => {
+    const child = spawn('bash', [script, command], {
+      cwd: projectRoot,
+      stdio: 'inherit',
+    });
+    child.on('close', () => resolve());
+    child.on('error', () => resolve());
+  });
+}
diff --git a/setup/lib/claude-handoff.ts b/setup/lib/claude-handoff.ts
new file mode 100644
index 0000000..9c931f2
--- /dev/null
+++ b/setup/lib/claude-handoff.ts
@@ -0,0 +1,194 @@
+/**
+ * User-initiated handoff to interactive Claude, parallel to claude-assist.ts.
+ *
+ * claude-assist is for failures: it runs `claude -p` non-interactively, parses
+ * a suggested command, and offers to run it. This module is for the opposite
+ * case — the user is mid-flow, not stuck on an error, and wants Claude to
+ * walk them through something the driver can't fully automate (Azure portal
+ * clickthrough, writing a manifest, tunneling a port, etc.).
+ *
+ * Flow:
+ *   1. Build a handoff prompt from the caller's context: channel, current
+ *      step, completed steps, collected values (secrets redacted), relevant
+ *      files to read.
+ *   2. Spawn `claude --append-system-prompt "<context>"
+ *      --permission-mode acceptEdits` with `stdio: 'inherit'` so Claude owns
+ *      the terminal.
+ *   3. When Claude exits (user types /exit, Ctrl-D, or closes the session),
+ *      control returns to the setup driver. The driver can then re-offer the
+ *      same step (e.g., "How did that go?" select).
+ *
+ * Also exports a small helper for text/password prompts: `validateWithHelpEscape`
+ * wraps a validate callback so typing `?` triggers the handoff instead of
+ * attempting to parse it as a real answer.
+ */
+import { execSync, spawn } from 'child_process';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+export interface HandoffContext {
+  /** Channel this handoff is happening in (e.g., 'teams'). */
+  channel: string;
+  /** Short name of the current step the user is stuck on. */
+  step: string;
+  /** Human-readable summary of what the user was trying to do at this step. */
+  stepDescription: string;
+  /** Checklist of sub-steps already completed (displayed as `✓ <item>`). */
+  completedSteps?: string[];
+  /**
+   * Key/value pairs of values collected so far. Callers should redact
+   * secrets before passing (e.g., show last 4 chars). Used to give Claude
+   * the state of the operator's progress.
+   */
+  collectedValues?: Record<string, string>;
+  /**
+   * Repo-relative paths Claude should consider reading. Always gets
+   * logs/setup.log and the relevant SKILL.md appended by the builder.
+   */
+  files?: string[];
+}
+
+/**
+ * Spawn interactive Claude with context pre-loaded as a system-prompt
+ * append. Returns when Claude exits.
+ *
+ * Silently no-ops (returns `false`) if `claude` isn't on PATH — setup runs
+ * where the binary is guaranteed to exist (we install it in the auth step),
+ * but an ultra-early flow failure could technically reach this before that
+ * install, and crashing the handoff would be worse than the handoff not
+ * firing.
+ */
+export async function offerClaudeHandoff(ctx: HandoffContext): Promise<boolean> {
+  if (!isClaudeUsable()) {
+    p.log.warn(
+      "Claude isn't installed yet — can't hand you off here. Finish setup first, then retry.",
+    );
+    return false;
+  }
+
+  const systemPrompt = buildSystemPrompt(ctx);
+
+  p.note(
+    [
+      "I'm handing you off to Claude in interactive mode.",
+      "It has the context of where you are in setup.",
+      "",
+      k.dim("Type /exit (or press Ctrl-D) when you're ready to come back to setup."),
+    ].join('\n'),
+    'Handing off to Claude',
+  );
+
+  return new Promise<boolean>((resolve) => {
+    const child = spawn(
+      'claude',
+      [
+        '--append-system-prompt',
+        systemPrompt,
+        '--permission-mode',
+        'acceptEdits',
+      ],
+      { stdio: 'inherit' },
+    );
+    child.on('close', () => {
+      p.log.success("Back from Claude. Let's continue.");
+      resolve(true);
+    });
+    child.on('error', () => {
+      p.log.error("Couldn't launch Claude. Continuing without handoff.");
+      resolve(false);
+    });
+  });
+}
+
+/**
+ * Sentinel returned by `validateWithHelpEscape` when the user types `?`.
+ * The caller compares against this to decide whether to trigger a handoff.
+ */
+export const HELP_ESCAPE_SENTINEL = '__NANOCLAW_HELP_ESCAPE__';
+
+/**
+ * Wrap a clack `validate` callback so typing `?` short-circuits validation
+ * and returns the HELP_ESCAPE_SENTINEL. Caller should check for the sentinel
+ * after awaiting the prompt and trigger offerClaudeHandoff if matched.
+ *
+ * Usage:
+ *   const answer = await p.text({
+ *     message: 'Paste your Azure App ID',
+ *     validate: validateWithHelpEscape((v) => {
+ *       if (!/^[0-9a-f-]{36}$/.test(v)) return 'Expected a UUID';
+ *       return undefined;
+ *     }),
+ *   });
+ *   if (answer === HELP_ESCAPE_SENTINEL) { await offerClaudeHandoff(ctx); ... }
+ */
+export function validateWithHelpEscape(
+  inner?: (value: string) => string | Error | undefined,
+): (value: string) => string | Error | undefined {
+  return (value: string) => {
+    if ((value ?? '').trim() === '?') {
+      // Returning undefined lets clack accept the `?` as the "answer". The
+      // caller sees a literal "?" and should compare + escape to handoff.
+      return undefined;
+    }
+    return inner ? inner(value) : undefined;
+  };
+}
+
+/**
+ * True if the value returned by a text/password prompt should trigger a
+ * handoff. Abstracts the sentinel check so callers don't have to import it
+ * directly at every site.
+ */
+export function isHelpEscape(value: unknown): boolean {
+  return typeof value === 'string' && value.trim() === '?';
+}
+
+function isClaudeUsable(): boolean {
+  try {
+    execSync('command -v claude', { stdio: 'ignore' });
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+function buildSystemPrompt(ctx: HandoffContext): string {
+  const lines: string[] = [
+    `The user is running NanoClaw's interactive \`setup:auto\` flow to wire the ${ctx.channel} channel.`,
+    `They got stuck at the step: "${ctx.step}" (${ctx.stepDescription}) and asked for help.`,
+    '',
+    "Your job: help them complete this specific step and get back to setup.",
+    "You can read files, run commands (with acceptEdits permissions), search the web,",
+    "and explain concepts. Be concise. When they're ready to resume, tell them to type",
+    "/exit and they'll return to the setup flow at the same step.",
+    '',
+  ];
+
+  if (ctx.completedSteps && ctx.completedSteps.length > 0) {
+    lines.push('Steps they have already completed:');
+    for (const s of ctx.completedSteps) lines.push(`  ✓ ${s}`);
+    lines.push('');
+  }
+
+  if (ctx.collectedValues && Object.keys(ctx.collectedValues).length > 0) {
+    lines.push('Values collected so far (secrets redacted):');
+    for (const [k, v] of Object.entries(ctx.collectedValues)) {
+      lines.push(`  ${k}: ${v}`);
+    }
+    lines.push('');
+  }
+
+  const files = [
+    ...(ctx.files ?? []),
+    'logs/setup.log',
+    'logs/setup-steps/',
+    `.claude/skills/add-${ctx.channel}/SKILL.md`,
+    `setup/channels/${ctx.channel}.ts`,
+  ].filter((v, i, a) => a.indexOf(v) === i);
+
+  lines.push('Relevant files (read as needed with the Read tool):');
+  for (const f of files) lines.push(`  - ${f}`);
+
+  return lines.join('\n');
+}
diff --git a/setup/lib/diagnostics.sh b/setup/lib/diagnostics.sh
new file mode 100644
index 0000000..23629d7
--- /dev/null
+++ b/setup/lib/diagnostics.sh
@@ -0,0 +1,61 @@
+# diagnostics.sh — shared PostHog emitter for bash-side setup code.
+#
+# Source this file after $PROJECT_ROOT is set:
+#
+#   source "$PROJECT_ROOT/setup/lib/diagnostics.sh"
+#   ph_event bootstrap_completed status=success platform=macos
+#
+# All emits are fire-and-forget (background curl, 3s max timeout); they
+# never fail the caller. Honors NANOCLAW_NO_DIAGNOSTICS=1. The distinct_id
+# is persisted at data/install-id so the bash + node halves of setup use
+# the same id and events from one install join into a single funnel.
+
+NANOCLAW_PH_KEY='phc_fx1Hhx9ucz8GuaJC8LVZWO8u03yXZZJJ6ObS4yplnaP'
+NANOCLAW_PH_URL='https://us.i.posthog.com/capture/'
+
+# Resolve or create the persisted install id. Echoes the id (lowercase uuid).
+# Creates data/install-id on first use. Safe to call pre-Node: uses only
+# bash + uuidgen/urandom fallback + mkdir.
+ph_install_id() {
+  local root="${NANOCLAW_PROJECT_ROOT:-${PROJECT_ROOT:-$PWD}}"
+  local f="$root/data/install-id"
+  if [ ! -s "$f" ]; then
+    mkdir -p "$(dirname "$f")" 2>/dev/null || return 0
+    local id
+    id=$(uuidgen 2>/dev/null \
+      || cat /proc/sys/kernel/random/uuid 2>/dev/null \
+      || printf 'fallback-%s-%s' "$(date +%s)" "$$")
+    printf '%s' "$id" | tr 'A-Z' 'a-z' > "$f" 2>/dev/null || return 0
+  fi
+  cat "$f" 2>/dev/null
+}
+
+# Emit a PostHog event. First arg is the event name; remaining args are
+# `key=value` pairs merged into properties. Values are JSON-escaped for
+# quotes and backslashes; keep them short and alphanumeric-ish.
+ph_event() {
+  [ "${NANOCLAW_NO_DIAGNOSTICS:-}" = "1" ] && return 0
+  local event=$1
+  shift
+  local id
+  id=$(ph_install_id)
+  [ -z "$id" ] && return 0
+
+  local props='' first=1 kv k v
+  for kv in "$@"; do
+    k="${kv%%=*}"
+    v="${kv#*=}"
+    v=${v//\\/\\\\}
+    v=${v//\"/\\\"}
+    if [ "$first" = "1" ]; then first=0; else props+=','; fi
+    props+="\"$k\":\"$v\""
+  done
+
+  local payload
+  payload=$(printf '{"api_key":"%s","event":"%s","distinct_id":"%s","properties":{%s}}' \
+    "$NANOCLAW_PH_KEY" "$event" "$id" "$props")
+
+  curl -sS --max-time 3 -X POST "$NANOCLAW_PH_URL" \
+    -H 'Content-Type: application/json' \
+    -d "$payload" >/dev/null 2>&1 &
+}
diff --git a/setup/lib/diagnostics.ts b/setup/lib/diagnostics.ts
new file mode 100644
index 0000000..30605a7
--- /dev/null
+++ b/setup/lib/diagnostics.ts
@@ -0,0 +1,70 @@
+/**
+ * Thin PostHog emitter shared across setup:auto code. Fire-and-forget —
+ * never throws, never blocks. Reuses data/install-id (same file bash
+ * uses in setup/lib/diagnostics.sh) so events from the bash and node
+ * halves of a single install join into one funnel.
+ *
+ * Honors NANOCLAW_NO_DIAGNOSTICS=1.
+ */
+import { randomUUID } from 'crypto';
+import fs from 'fs';
+import path from 'path';
+
+const POSTHOG_KEY = 'phc_fx1Hhx9ucz8GuaJC8LVZWO8u03yXZZJJ6ObS4yplnaP';
+const POSTHOG_URL = 'https://us.i.posthog.com/capture/';
+const INSTALL_ID_PATH = path.join('data', 'install-id');
+
+let cached: string | null = null;
+
+export function installId(): string {
+  if (cached) return cached;
+  try {
+    const existing = fs.readFileSync(INSTALL_ID_PATH, 'utf-8').trim();
+    if (existing) {
+      cached = existing;
+      return existing;
+    }
+  } catch {
+    // fall through to create
+  }
+  const id = randomUUID().toLowerCase();
+  try {
+    fs.mkdirSync(path.dirname(INSTALL_ID_PATH), { recursive: true });
+    fs.writeFileSync(INSTALL_ID_PATH, id);
+  } catch {
+    // best-effort; still return the id so the event fires
+  }
+  cached = id;
+  return id;
+}
+
+export function emit(
+  event: string,
+  props: Record<string, string | number | boolean | undefined> = {},
+): void {
+  if (process.env.NANOCLAW_NO_DIAGNOSTICS === '1') return;
+
+  const cleaned: Record<string, unknown> = { platform: process.platform };
+  for (const [k, v] of Object.entries(props)) {
+    if (v === undefined) continue;
+    cleaned[k] = v;
+  }
+
+  const body = JSON.stringify({
+    api_key: POSTHOG_KEY,
+    event,
+    distinct_id: installId(),
+    properties: cleaned,
+  });
+
+  const ctrl = new AbortController();
+  const timer = setTimeout(() => ctrl.abort(), 3000);
+  void fetch(POSTHOG_URL, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body,
+    signal: ctrl.signal,
+  })
+    .catch(() => {})
+    .finally(() => clearTimeout(timer));
+}
diff --git a/setup/lib/install-slug.sh b/setup/lib/install-slug.sh
new file mode 100644
index 0000000..736d339
--- /dev/null
+++ b/setup/lib/install-slug.sh
@@ -0,0 +1,37 @@
+# install-slug.sh — shell mirror of setup/lib/install-slug.ts.
+#
+# Source this file after $PROJECT_ROOT is set:
+#
+#   source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+#   label=$(launchd_label)        # com.nanoclaw-v2-<slug>
+#   unit=$(systemd_unit)          # nanoclaw-v2-<slug>
+#   image=$(container_image_base) # nanoclaw-agent-v2-<slug>
+#
+# Slug is sha1(PROJECT_ROOT)[:8] — must match the TS helper exactly so both
+# halves of setup name things consistently.
+
+_nanoclaw_install_slug() {
+  local root="${NANOCLAW_PROJECT_ROOT:-${PROJECT_ROOT:-$PWD}}"
+  if command -v shasum >/dev/null 2>&1; then
+    printf '%s' "$root" | shasum | cut -c 1-8
+  elif command -v sha1sum >/dev/null 2>&1; then
+    printf '%s' "$root" | sha1sum | cut -c 1-8
+  else
+    # Fallback: hash the path with something deterministic-ish. Not ideal —
+    # but shasum is present on every modern macOS/Linux, so this is just
+    # belt-and-braces against a truly minimal system.
+    printf '%s' "$root" | od -An -tx1 | tr -d ' \n' | cut -c 1-8
+  fi
+}
+
+launchd_label() {
+  printf 'com.nanoclaw-v2-%s' "$(_nanoclaw_install_slug)"
+}
+
+systemd_unit() {
+  printf 'nanoclaw-v2-%s' "$(_nanoclaw_install_slug)"
+}
+
+container_image_base() {
+  printf 'nanoclaw-agent-v2-%s' "$(_nanoclaw_install_slug)"
+}
diff --git a/setup/lib/role-prompt.ts b/setup/lib/role-prompt.ts
new file mode 100644
index 0000000..7344ac1
--- /dev/null
+++ b/setup/lib/role-prompt.ts
@@ -0,0 +1,43 @@
+/**
+ * Shared "who's connecting this channel?" prompt used by the channel setup
+ * drivers before they hand off to scripts/init-first-agent.ts.
+ *
+ * Default: owner. Self-hosted NanoClaw is almost always a single-operator
+ * deployment, and granting the same human owner status on every channel
+ * they wire up matches what you'd want 99% of the time. The prompt
+ * surfaces admin/member for the edge cases (shared instance, collaborators
+ * with limited access), but hitting Enter assigns owner.
+ */
+import { brightSelect } from './bright-select.js';
+import { ensureAnswer } from './runner.js';
+
+export type OperatorRole = 'owner' | 'admin' | 'member';
+
+export async function askOperatorRole(
+  channelLabel: string,
+): Promise<OperatorRole> {
+  const choice = ensureAnswer(
+    await brightSelect<OperatorRole>({
+      message: `How should this ${channelLabel} account be registered?`,
+      initialValue: 'owner',
+      options: [
+        {
+          value: 'owner',
+          label: 'Owner',
+          hint: 'full access — recommended for your own account',
+        },
+        {
+          value: 'admin',
+          label: 'Admin',
+          hint: 'can manage the agent for this channel',
+        },
+        {
+          value: 'member',
+          label: 'Member',
+          hint: 'can chat with the agent but nothing more',
+        },
+      ],
+    }),
+  );
+  return choice;
+}
diff --git a/setup/lib/runner.ts b/setup/lib/runner.ts
new file mode 100644
index 0000000..c1599e4
--- /dev/null
+++ b/setup/lib/runner.ts
@@ -0,0 +1,417 @@
+/**
+ * Step runner + abort helpers for setup:auto.
+ *
+ * Responsibilities:
+ *   - Stream-parse setup-step status blocks (`=== NANOCLAW SETUP: … ===`)
+ *   - Spawn children with output tee'd to a per-step raw log (level 3)
+ *   - Wrap each run in a clack spinner with live elapsed time (level 1)
+ *   - Append a structured entry to the progression log (level 2) via
+ *     `setup/logs.ts` when the run ends
+ *   - Abort helpers (`fail`, `ensureAnswer`) used by step orchestrators
+ *
+ * See docs/setup-flow.md for the three-level output contract.
+ */
+import { spawn, spawnSync } from 'child_process';
+import fs from 'fs';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import * as setupLog from '../logs.js';
+import { offerClaudeAssist } from './claude-assist.js';
+import { emit as phEmit } from './diagnostics.js';
+import { fitToWidth } from './theme.js';
+
+export type Fields = Record<string, string>;
+export type Block = { type: string; fields: Fields };
+
+export type StepResult = {
+  ok: boolean;
+  exitCode: number;
+  blocks: Block[];
+  transcript: string;
+  /** The last block with a STATUS field (the terminal/result block). */
+  terminal: Block | null;
+};
+
+export type QuietChildResult = {
+  ok: boolean;
+  exitCode: number;
+  transcript: string;
+  terminal: Block | null;
+  blocks: Block[];
+};
+
+export type SpinnerLabels = {
+  running: string;
+  done: string;
+  skipped?: string;
+  failed?: string;
+};
+
+/**
+ * Streaming parser for `=== NANOCLAW SETUP: TYPE ===` blocks. Emits each
+ * block as it closes so the UI can react mid-stream (e.g. render a pairing
+ * code card as soon as pair-telegram emits it, rather than after the step
+ * has finished).
+ */
+export class StatusStream {
+  private lineBuf = '';
+  private current: Block | null = null;
+  readonly blocks: Block[] = [];
+  transcript = '';
+
+  constructor(private readonly onBlock: (block: Block) => void) {}
+
+  write(chunk: string): void {
+    this.transcript += chunk;
+    this.lineBuf += chunk;
+    let idx: number;
+    while ((idx = this.lineBuf.indexOf('\n')) !== -1) {
+      const line = this.lineBuf.slice(0, idx);
+      this.lineBuf = this.lineBuf.slice(idx + 1);
+      this.processLine(line);
+    }
+  }
+
+  private processLine(line: string): void {
+    const start = line.match(/^=== NANOCLAW SETUP: (\S+) ===/);
+    if (start) {
+      this.current = { type: start[1], fields: {} };
+      return;
+    }
+    if (line.startsWith('=== END ===')) {
+      if (this.current) {
+        this.blocks.push(this.current);
+        this.onBlock(this.current);
+        this.current = null;
+      }
+      return;
+    }
+    if (!this.current) return;
+    const colon = line.indexOf(':');
+    if (colon === -1) return;
+    const key = line.slice(0, colon).trim();
+    const value = line.slice(colon + 1).trim();
+    if (key) this.current.fields[key] = value;
+  }
+}
+
+/**
+ * Spawn a setup step as a child process. Output is tee'd to the provided
+ * raw log file (level 3) and parsed for status blocks (level 2 summary).
+ * The onBlock callback fires per status block as they close so the UI can
+ * react mid-stream.
+ *
+ * `onLine`, if provided, fires for every line from stdout + stderr (minus
+ * status-block control lines) so callers can render a rolling tail. Status
+ * block lines are still parsed by the `StatusStream` — they're just
+ * excluded from the line feed so they don't fill the user-facing window
+ * with `=== NANOCLAW SETUP: …` noise.
+ */
+export function spawnStep(
+  stepName: string,
+  extra: string[],
+  onBlock: (block: Block) => void,
+  rawLogPath: string,
+  onLine?: (line: string) => void,
+): Promise<StepResult> {
+  return new Promise((resolve) => {
+    const args = ['exec', 'tsx', 'setup/index.ts', '--step', stepName];
+    if (extra.length > 0) args.push('--', ...extra);
+
+    const child = spawn('pnpm', args, { stdio: ['ignore', 'pipe', 'pipe'] });
+    const stream = new StatusStream(onBlock);
+    const raw = fs.createWriteStream(rawLogPath, { flags: 'w' });
+    raw.write(`# ${stepName} — ${new Date().toISOString()}\n\n`);
+
+    // Per-line forwarder for the optional onLine callback. We keep our own
+    // buffer (separate from StatusStream's) so the parser still gets raw
+    // chunks and isn't forced through a line-by-line path it doesn't need.
+    let lineBuf = '';
+    const pushLines = (chunk: string): void => {
+      if (!onLine) return;
+      lineBuf += chunk;
+      let idx: number;
+      while ((idx = lineBuf.indexOf('\n')) !== -1) {
+        const line = lineBuf.slice(0, idx).replace(/\r/g, '');
+        lineBuf = lineBuf.slice(idx + 1);
+        if (line.startsWith('=== NANOCLAW SETUP:')) continue;
+        if (line.startsWith('=== END ===')) continue;
+        if (line.trim()) onLine(line);
+      }
+    };
+
+    child.stdout.on('data', (chunk: Buffer) => {
+      const s = chunk.toString('utf-8');
+      stream.write(s);
+      raw.write(chunk);
+      pushLines(s);
+    });
+    child.stderr.on('data', (chunk: Buffer) => {
+      const s = chunk.toString('utf-8');
+      stream.transcript += s;
+      raw.write(chunk);
+      pushLines(s);
+    });
+
+    child.on('close', (code) => {
+      raw.end();
+      const terminal =
+        [...stream.blocks].reverse().find((b) => b.fields.STATUS) ?? null;
+      const status = terminal?.fields.STATUS;
+      const ok = code === 0 && (status === 'success' || status === 'skipped');
+      resolve({
+        ok,
+        exitCode: code ?? 1,
+        blocks: stream.blocks,
+        transcript: stream.transcript,
+        terminal,
+      });
+    });
+  });
+}
+
+export function spawnQuiet(
+  cmd: string,
+  args: string[],
+  rawLogPath: string,
+  envOverride?: NodeJS.ProcessEnv,
+): Promise<QuietChildResult> {
+  return new Promise((resolve) => {
+    const child = spawn(cmd, args, {
+      stdio: ['ignore', 'pipe', 'pipe'],
+      env: envOverride ? { ...process.env, ...envOverride } : process.env,
+    });
+    let transcript = '';
+    const raw = fs.createWriteStream(rawLogPath, { flags: 'w' });
+    raw.write(`# ${[cmd, ...args].join(' ')} — ${new Date().toISOString()}\n\n`);
+    const blocks: Block[] = [];
+    const stream = new StatusStream((b) => blocks.push(b));
+    child.stdout.on('data', (c: Buffer) => {
+      const s = c.toString('utf-8');
+      transcript += s;
+      stream.write(s);
+      raw.write(c);
+    });
+    child.stderr.on('data', (c: Buffer) => {
+      transcript += c.toString('utf-8');
+      raw.write(c);
+    });
+    child.on('close', (code) => {
+      raw.end();
+      const terminal =
+        [...blocks].reverse().find((b) => b.fields.STATUS) ?? null;
+      resolve({ ok: code === 0, exitCode: code ?? 1, transcript, terminal, blocks });
+    });
+  });
+}
+
+/** Run a step under a clack spinner. Teed to a per-step raw log + progression entry at the end. */
+export async function runQuietStep(
+  stepName: string,
+  labels: SpinnerLabels,
+  extra: string[] = [],
+): Promise<StepResult & { rawLog: string; durationMs: number }> {
+  const rawLog = setupLog.stepRawLog(stepName);
+  const start = Date.now();
+  phEmit('step_started', { step: stepName });
+  const result = await runUnderSpinner(labels, () =>
+    spawnStep(stepName, extra, () => {}, rawLog),
+  );
+  const durationMs = Date.now() - start;
+  writeStepEntry(stepName, result, durationMs, rawLog);
+  phEmit('step_completed', {
+    step: stepName,
+    status: outcomeStatus(result),
+    duration_ms: durationMs,
+  });
+  return { ...result, rawLog, durationMs };
+}
+
+/** Run an arbitrary child under a spinner. Same raw-log + progression treatment as runQuietStep. */
+export async function runQuietChild(
+  logName: string,
+  cmd: string,
+  args: string[],
+  labels: SpinnerLabels,
+  opts?: {
+    /** Extra fields to merge into the progression entry (on top of any status-block fields). */
+    extraFields?: Record<string, string | number | boolean>;
+    /** Environment overrides to pass to the child process. */
+    env?: NodeJS.ProcessEnv;
+  },
+): Promise<QuietChildResult & { rawLog: string; durationMs: number }> {
+  const rawLog = setupLog.stepRawLog(logName);
+  const start = Date.now();
+  phEmit('step_started', { step: logName });
+  const result = await runUnderSpinner(labels, () =>
+    spawnQuiet(cmd, args, rawLog, opts?.env),
+  );
+  const durationMs = Date.now() - start;
+
+  const blockFields = summariseTerminalFields(result.terminal);
+  const fields = { ...blockFields, ...(opts?.extraFields ?? {}) };
+  const rawStatus = result.terminal?.fields.STATUS;
+  const status: 'success' | 'skipped' | 'failed' = !result.ok
+    ? 'failed'
+    : rawStatus === 'skipped'
+      ? 'skipped'
+      : 'success';
+  setupLog.step(logName, status, durationMs, fields, rawLog);
+  phEmit('step_completed', { step: logName, status, duration_ms: durationMs });
+  return { ...result, rawLog, durationMs };
+}
+
+/** Collapse a step run into the three-way status used by diagnostics + progression log. */
+function outcomeStatus(result: StepResult): 'success' | 'skipped' | 'failed' {
+  const rawStatus = result.terminal?.fields.STATUS;
+  if (!result.ok) return 'failed';
+  return rawStatus === 'skipped' ? 'skipped' : 'success';
+}
+
+/** Turn a step's terminal-block fields into a concise progression-log entry. */
+export function writeStepEntry(
+  stepName: string,
+  result: StepResult,
+  durationMs: number,
+  rawLog: string,
+): void {
+  const rawStatus = result.terminal?.fields.STATUS;
+  const logStatus: 'success' | 'skipped' | 'failed' = !result.ok
+    ? 'failed'
+    : rawStatus === 'skipped'
+      ? 'skipped'
+      : 'success';
+  const fields = summariseTerminalFields(result.terminal);
+  setupLog.step(stepName, logStatus, durationMs, fields, rawLog);
+}
+
+/** Strip STATUS + LOG (redundant) and any oversize values from the terminal block's fields. */
+export function summariseTerminalFields(block: Block | null): Record<string, string> {
+  if (!block) return {};
+  const out: Record<string, string> = {};
+  for (const [k, v] of Object.entries(block.fields)) {
+    if (k === 'STATUS' || k === 'LOG') continue;
+    if (v.length > 120) continue; // keep it skimmable; full value lives in the raw log
+    out[k] = v;
+  }
+  return out;
+}
+
+async function runUnderSpinner<
+  T extends { ok: boolean; transcript: string; terminal?: Block | null },
+>(
+  labels: SpinnerLabels,
+  work: () => Promise<T>,
+): Promise<T> {
+  const s = p.spinner();
+  const start = Date.now();
+  s.start(fitToWidth(labels.running, ' (999s)'));
+  const tick = setInterval(() => {
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    const suffix = ` (${elapsed}s)`;
+    s.message(`${fitToWidth(labels.running, suffix)}${k.dim(suffix)}`);
+  }, 1000);
+
+  const result = await work();
+
+  clearInterval(tick);
+  const elapsed = Math.round((Date.now() - start) / 1000);
+  const suffix = ` (${elapsed}s)`;
+  if (result.ok) {
+    const isSkipped = result.terminal?.fields.STATUS === 'skipped';
+    const msg = isSkipped && labels.skipped ? labels.skipped : labels.done;
+    // Bold the outcome so the step's headline reads stronger than the prose
+    // body copy around it. The trailing `(Ns)` timing stays dim.
+    s.stop(`${k.bold(fitToWidth(msg, suffix))}${k.dim(suffix)}`);
+  } else {
+    const failMsg = labels.failed ?? labels.running.replace(/…$/, ' failed');
+    s.stop(`${k.bold(fitToWidth(failMsg, suffix))}${k.dim(suffix)}`, 1);
+    dumpTranscriptOnFailure(result.transcript);
+  }
+  return result;
+}
+
+export function dumpTranscriptOnFailure(transcript: string): void {
+  const lines = transcript.split('\n').filter((l) => {
+    if (l.startsWith('=== NANOCLAW SETUP:')) return false;
+    if (l.startsWith('=== END ===')) return false;
+    return true;
+  });
+  const tail = lines.slice(-40).join('\n').trimEnd();
+  if (tail) {
+    console.log();
+    console.log(k.dim(tail));
+    console.log();
+  }
+}
+
+/**
+ * Abort the setup run with a user-facing error, logging the abort to the
+ * progression log. Takes the step name explicitly so callers are clear
+ * about which step they're failing from — no hidden module state.
+ *
+ * Before aborting we offer Claude-assisted debugging. Callers must
+ * `await fail(...)` so the offer can actually run before we call
+ * process.exit. The return type is `Promise<never>`; control-flow
+ * narrowing still works after `await`.
+ */
+export async function fail(
+  stepName: string,
+  msg: string,
+  hint?: string,
+  rawLogPath?: string,
+): Promise<never> {
+  setupLog.abort(stepName, msg);
+  phEmit('setup_aborted', { step: stepName, reason: msg });
+  p.log.error(msg);
+  if (hint) p.log.message(k.dim(hint));
+  p.log.message(k.dim('Logs: logs/setup.log · Raw: logs/setup-steps/'));
+
+  const ranFix = await offerClaudeAssist({ stepName, msg, hint, rawLogPath });
+
+  // If the user just ran a Claude-suggested fix, offer to resume the flow
+  // at the step that failed instead of aborting. We re-exec via spawnSync
+  // and pass NANOCLAW_SKIP with every step that already completed so the
+  // child skips them and picks up where we left off.
+  if (ranFix) {
+    const retry = ensureAnswer(
+      await p.confirm({
+        message: `Fix applied. Retry the ${stepName} step?`,
+        initialValue: true,
+      }),
+    );
+    if (retry) {
+      const existingSkip = (process.env.NANOCLAW_SKIP ?? '')
+        .split(',')
+        .map((s) => s.trim())
+        .filter(Boolean);
+      const skipList = [
+        ...new Set([...existingSkip, ...setupLog.completedStepNames()]),
+      ].join(',');
+      p.log.step(`Retrying from ${stepName}…`);
+      const result = spawnSync('pnpm', ['--silent', 'run', 'setup:auto'], {
+        stdio: 'inherit',
+        env: { ...process.env, NANOCLAW_SKIP: skipList },
+      });
+      process.exit(result.status ?? 0);
+    }
+  }
+
+  p.cancel('Setup aborted.');
+  process.exit(1);
+}
+
+/**
+ * Unwrap a clack prompt result. If the user cancelled (Ctrl-C / Esc), exit
+ * gracefully. Cancel is exit 0 — it's not an abort worth logging to the
+ * progression log, since the operator initiated it deliberately.
+ */
+export function ensureAnswer<T>(value: T | symbol): T {
+  if (p.isCancel(value)) {
+    p.cancel('Setup cancelled.');
+    process.exit(0);
+  }
+  return value as T;
+}
diff --git a/setup/lib/teams-manifest.ts b/setup/lib/teams-manifest.ts
new file mode 100644
index 0000000..c40837a
--- /dev/null
+++ b/setup/lib/teams-manifest.ts
@@ -0,0 +1,271 @@
+/**
+ * Build the Teams app package zip that the operator sideloads from the Teams
+ * "Manage your apps" screen.
+ *
+ * A Teams app package is a zip containing:
+ *   - manifest.json  — declares the bot, scopes, required permissions
+ *   - outline.png    — 32×32 transparent outline icon
+ *   - color.png      — 192×192 full-color icon
+ *
+ * Icons are generated in-process using a minimal PNG encoder so we don't
+ * need ImageMagick or vendor binary icon blobs into the repo. The outline
+ * icon is a simple rounded square outline; the color icon is a brand-blue
+ * filled square with a small white "N" blocked in by pixel setting. Good
+ * enough for a working sideload — teams admins who care can replace the
+ * icons later.
+ *
+ * The manifest is pinned to schema v1.16 to match the skill doc.
+ */
+import { execSync } from 'child_process';
+import fs from 'fs';
+import path from 'path';
+import zlib from 'zlib';
+
+const MANIFEST_SCHEMA =
+  'https://developer.microsoft.com/en-us/json-schemas/teams/v1.16/MicrosoftTeams.schema.json';
+const MANIFEST_VERSION = '1.16';
+
+export interface ManifestOptions {
+  /** The Azure AD app ID (same value used for `bots[0].botId`). */
+  appId: string;
+  /** Short bot name shown in Teams (<= 30 chars). */
+  shortName: string;
+  /** Long bot description. */
+  longDescription: string;
+  /** Developer website URL (required by schema — any reachable URL works). */
+  websiteUrl: string;
+  /** Out-dir for the generated zip + loose files. */
+  outDir: string;
+}
+
+export interface ManifestResult {
+  zipPath: string;
+  manifestPath: string;
+  outlinePath: string;
+  colorPath: string;
+}
+
+/** Build the full app package zip and return the paths. */
+export function buildTeamsAppPackage(opts: ManifestOptions): ManifestResult {
+  fs.mkdirSync(opts.outDir, { recursive: true });
+
+  const manifestPath = path.join(opts.outDir, 'manifest.json');
+  const outlinePath = path.join(opts.outDir, 'outline.png');
+  const colorPath = path.join(opts.outDir, 'color.png');
+  const zipPath = path.join(opts.outDir, 'teams-app-package.zip');
+
+  fs.writeFileSync(manifestPath, renderManifest(opts));
+  fs.writeFileSync(outlinePath, encodeOutlineIcon());
+  fs.writeFileSync(colorPath, encodeColorIcon());
+
+  // Fresh zip every run — idempotent, no stale files.
+  try {
+    fs.unlinkSync(zipPath);
+  } catch {
+    // noop if missing
+  }
+  execSync(`zip -j -q "${zipPath}" "${manifestPath}" "${outlinePath}" "${colorPath}"`, {
+    stdio: ['ignore', 'ignore', 'inherit'],
+  });
+
+  return { zipPath, manifestPath, outlinePath, colorPath };
+}
+
+function renderManifest(opts: ManifestOptions): string {
+  const manifest = {
+    $schema: MANIFEST_SCHEMA,
+    manifestVersion: MANIFEST_VERSION,
+    version: '1.0.0',
+    id: opts.appId,
+    packageName: 'com.nanoclaw.bot',
+    developer: {
+      name: 'NanoClaw',
+      websiteUrl: opts.websiteUrl,
+      privacyUrl: opts.websiteUrl,
+      termsOfUseUrl: opts.websiteUrl,
+    },
+    name: {
+      short: opts.shortName.slice(0, 30),
+      full: `${opts.shortName} Assistant`,
+    },
+    description: {
+      short: 'Your personal assistant in Teams.',
+      full: opts.longDescription,
+    },
+    icons: { outline: 'outline.png', color: 'color.png' },
+    accentColor: '#4A90D9',
+    bots: [
+      {
+        botId: opts.appId,
+        scopes: ['personal', 'team', 'groupchat'],
+        supportsFiles: false,
+        isNotificationOnly: false,
+      },
+    ],
+    permissions: ['identity', 'messageTeamMembers'],
+    validDomains: [new URL(opts.websiteUrl).host],
+  };
+  return JSON.stringify(manifest, null, 2) + '\n';
+}
+
+// ─── Minimal PNG encoder (solid color, no external deps) ──────────────────
+
+const PNG_SIG = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
+
+// Precompute the CRC-32 table per the PNG spec. Node doesn't expose CRC32
+// directly (zlib.crc32 isn't part of the public API), so we roll our own.
+const CRC_TABLE = (() => {
+  const table = new Uint32Array(256);
+  for (let n = 0; n < 256; n++) {
+    let c = n;
+    for (let k = 0; k < 8; k++) {
+      c = c & 1 ? 0xedb88320 ^ (c >>> 1) : c >>> 1;
+    }
+    table[n] = c >>> 0;
+  }
+  return table;
+})();
+
+function crc32(buf: Buffer): number {
+  let c = 0xffffffff;
+  for (let i = 0; i < buf.length; i++) {
+    c = CRC_TABLE[(c ^ buf[i]) & 0xff] ^ (c >>> 8);
+  }
+  return (c ^ 0xffffffff) >>> 0;
+}
+
+function chunk(type: string, data: Buffer): Buffer {
+  const len = Buffer.alloc(4);
+  len.writeUInt32BE(data.length, 0);
+  const typeBuf = Buffer.from(type, 'ascii');
+  const crcBuf = Buffer.alloc(4);
+  crcBuf.writeUInt32BE(crc32(Buffer.concat([typeBuf, data])), 0);
+  return Buffer.concat([len, typeBuf, data, crcBuf]);
+}
+
+/**
+ * Encode a solid-color RGBA image as a PNG. `pixels` is a width*height*4
+ * byte array (R, G, B, A per pixel, row-major, top-to-bottom).
+ */
+function encodePng(width: number, height: number, pixels: Uint8Array): Buffer {
+  // IHDR
+  const ihdr = Buffer.alloc(13);
+  ihdr.writeUInt32BE(width, 0);
+  ihdr.writeUInt32BE(height, 4);
+  ihdr[8] = 8; // bit depth
+  ihdr[9] = 6; // color type: RGBA
+  ihdr[10] = 0; // compression
+  ihdr[11] = 0; // filter
+  ihdr[12] = 0; // interlace
+
+  // IDAT: scanlines with filter byte 0 (None) prepended per row.
+  const rowBytes = width * 4;
+  const raw = Buffer.alloc(height * (rowBytes + 1));
+  for (let y = 0; y < height; y++) {
+    raw[y * (rowBytes + 1)] = 0;
+    for (let x = 0; x < rowBytes; x++) {
+      raw[y * (rowBytes + 1) + 1 + x] = pixels[y * rowBytes + x];
+    }
+  }
+  const idat = zlib.deflateSync(raw);
+
+  return Buffer.concat([
+    PNG_SIG,
+    chunk('IHDR', ihdr),
+    chunk('IDAT', idat),
+    chunk('IEND', Buffer.alloc(0)),
+  ]);
+}
+
+/**
+ * Outline icon: 32×32 transparent background with a simple white rounded-
+ * square outline. Teams renders it against a colored background so the
+ * outline needs to be visible on both light and dark.
+ */
+function encodeOutlineIcon(): Buffer {
+  const size = 32;
+  const pixels = new Uint8Array(size * size * 4);
+  const inset = 4;
+  const stroke = 2;
+  for (let y = 0; y < size; y++) {
+    for (let x = 0; x < size; x++) {
+      const onBorder =
+        ((x >= inset && x < inset + stroke) || (x >= size - inset - stroke && x < size - inset)) &&
+        y >= inset &&
+        y < size - inset;
+      const onTopBot =
+        ((y >= inset && y < inset + stroke) || (y >= size - inset - stroke && y < size - inset)) &&
+        x >= inset &&
+        x < size - inset;
+      const i = (y * size + x) * 4;
+      if (onBorder || onTopBot) {
+        pixels[i] = 255;
+        pixels[i + 1] = 255;
+        pixels[i + 2] = 255;
+        pixels[i + 3] = 255;
+      } else {
+        pixels[i] = 0;
+        pixels[i + 1] = 0;
+        pixels[i + 2] = 0;
+        pixels[i + 3] = 0; // transparent
+      }
+    }
+  }
+  return encodePng(size, size, pixels);
+}
+
+/**
+ * Color icon: 192×192 brand-blue filled square with a white "N" shape drawn
+ * with simple bars (left vertical, right vertical, diagonal from top-right
+ * to bottom-left). Crude but recognizable at a glance.
+ */
+function encodeColorIcon(): Buffer {
+  const size = 192;
+  const pixels = new Uint8Array(size * size * 4);
+  // Brand blue #4A90D9
+  const BG_R = 0x4a;
+  const BG_G = 0x90;
+  const BG_B = 0xd9;
+  const thickness = 24;
+  const margin = 40;
+  const leftBarX = margin;
+  const rightBarX = size - margin - thickness;
+  for (let y = 0; y < size; y++) {
+    for (let x = 0; x < size; x++) {
+      const i = (y * size + x) * 4;
+      pixels[i] = BG_R;
+      pixels[i + 1] = BG_G;
+      pixels[i + 2] = BG_B;
+      pixels[i + 3] = 255;
+    }
+  }
+  // Vertical bars
+  for (let y = margin; y < size - margin; y++) {
+    for (let dx = 0; dx < thickness; dx++) {
+      setWhite(pixels, size, leftBarX + dx, y);
+      setWhite(pixels, size, rightBarX + dx, y);
+    }
+  }
+  // Diagonal from top-right of left bar to bottom-left of right bar
+  const diagSteps = size - margin * 2;
+  for (let s = 0; s < diagSteps; s++) {
+    const t = s / (diagSteps - 1);
+    const cx = Math.round(leftBarX + thickness + t * (rightBarX - leftBarX - thickness));
+    const cy = Math.round(margin + t * (size - margin * 2 - 1));
+    for (let dx = -Math.floor(thickness / 2); dx < Math.ceil(thickness / 2); dx++) {
+      for (let dy = -2; dy <= 2; dy++) {
+        setWhite(pixels, size, cx + dx, cy + dy);
+      }
+    }
+  }
+  return encodePng(size, size, pixels);
+}
+
+function setWhite(pixels: Uint8Array, size: number, x: number, y: number): void {
+  if (x < 0 || x >= size || y < 0 || y >= size) return;
+  const i = (y * size + x) * 4;
+  pixels[i] = 255;
+  pixels[i + 1] = 255;
+  pixels[i + 2] = 255;
+  pixels[i + 3] = 255;
+}
diff --git a/setup/lib/theme.ts b/setup/lib/theme.ts
new file mode 100644
index 0000000..35b5ca3
--- /dev/null
+++ b/setup/lib/theme.ts
@@ -0,0 +1,121 @@
+/**
+ * NanoClaw brand palette for the terminal.
+ *
+ * Colors pulled from assets/nanoclaw-logo.png:
+ *   brand cyan  ≈ #2BB7CE  — the "Claw" wordmark + mascot body
+ *   brand navy  ≈ #171B3B  — the dark logo background + outlines
+ *
+ * Rendering gates:
+ *   - No TTY (piped / redirected) → plain text, no ANSI
+ *   - NO_COLOR set               → plain text, no ANSI
+ *   - COLORTERM truecolor/24bit  → 24-bit ANSI (exact brand cyan)
+ *   - Otherwise                  → kleur's 16-color cyan (closest fallback)
+ */
+import k from 'kleur';
+
+const USE_ANSI = Boolean(process.stdout.isTTY) && !process.env.NO_COLOR;
+const TRUECOLOR =
+  USE_ANSI &&
+  (process.env.COLORTERM === 'truecolor' || process.env.COLORTERM === '24bit');
+
+export function brand(s: string): string {
+  if (!USE_ANSI) return s;
+  if (TRUECOLOR) return `\x1b[38;2;43;183;206m${s}\x1b[0m`;
+  return k.cyan(s);
+}
+
+export function brandBold(s: string): string {
+  if (!USE_ANSI) return s;
+  if (TRUECOLOR) return `\x1b[1;38;2;43;183;206m${s}\x1b[0m`;
+  return k.bold(k.cyan(s));
+}
+
+export function brandChip(s: string): string {
+  if (!USE_ANSI) return s;
+  if (TRUECOLOR) {
+    return `\x1b[48;2;43;183;206m\x1b[38;2;23;27;59m\x1b[1m${s}\x1b[0m`;
+  }
+  return k.bgCyan(k.black(k.bold(s)));
+}
+
+/**
+ * Wrap text so it fits inside clack's gutter without the terminal's soft
+ * wrap breaking the `│ …` bar on long lines. Works on a single string with
+ * embedded `\n`s; each logical line is wrapped independently.
+ *
+ * The `gutter` argument is the total horizontal overhead clack adds for
+ * the component the text lives in (e.g. 4 for `p.log.*`'s `│  ` prefix;
+ * 6-ish for `p.note`'s box). Caller picks it; we just subtract from
+ * `process.stdout.columns` and hard-wrap at word boundaries.
+ */
+export function wrapForGutter(text: string, gutter: number): string {
+  const cols = process.stdout.columns ?? 80;
+  const width = Math.max(30, cols - gutter);
+  return text
+    .split('\n')
+    .map((line) => wrapLine(line, width))
+    .join('\n');
+}
+
+/**
+ * Wrap multi-line explanatory prose to the clack gutter. Previously
+ * dimmed its output (hence the name) — that made body copy hard to read
+ * against dark terminals. Dim is now reserved for preview/debug blocks
+ * (failure transcript tails, claude-assist streams); prose renders at
+ * the terminal's regular weight.
+ */
+export function dimWrap(text: string, gutter: number): string {
+  return wrapForGutter(text, gutter);
+}
+
+const ANSI_RE = /\x1b\[[0-9;]*m/g;
+
+function visibleLength(s: string): number {
+  return s.replace(ANSI_RE, '').length;
+}
+
+/**
+ * Truncate a label so the final line — base + reserved suffix — fits in
+ * the terminal width. Use on spinner labels that get an elapsed counter
+ * appended: if the total exceeds terminal width, clack's cursor-up
+ * redraw math breaks and each tick stacks a copy of the line instead
+ * of replacing it.
+ *
+ * `suffix` is the reserved space for what we'll append after `fit()`
+ * returns (e.g. ` (999s)` or a tool-use breadcrumb). We don't include
+ * it in the output — caller appends it.
+ */
+export function fitToWidth(base: string, suffix: string): string {
+  const cols = process.stdout.columns ?? 80;
+  // Overhead we reserve before sizing the label:
+  //   spinner icon (1) + 2 padding spaces = 3
+  //   clack's animated ellipsis after the label = up to 3 (". " -> "...")
+  //   1-char safety margin so wide-char glyphs don't tip over the edge
+  // Total reserved budget = 7 cols plus the caller's suffix.
+  const budget = Math.max(20, cols - 7 - visibleLength(suffix));
+  return base.length > budget ? base.slice(0, budget - 1) + '…' : base;
+}
+
+function wrapLine(line: string, width: number): string {
+  if (visibleLength(line) <= width) return line;
+  const words = line.split(' ');
+  const rows: string[] = [];
+  let cur = '';
+  let curLen = 0;
+  for (const word of words) {
+    const wLen = visibleLength(word);
+    if (curLen === 0) {
+      cur = word;
+      curLen = wLen;
+    } else if (curLen + 1 + wLen <= width) {
+      cur += ' ' + word;
+      curLen += 1 + wLen;
+    } else {
+      rows.push(cur);
+      cur = word;
+      curLen = wLen;
+    }
+  }
+  if (cur) rows.push(cur);
+  return rows.join('\n');
+}
diff --git a/setup/lib/tz-from-claude.ts b/setup/lib/tz-from-claude.ts
new file mode 100644
index 0000000..5486fbb
--- /dev/null
+++ b/setup/lib/tz-from-claude.ts
@@ -0,0 +1,126 @@
+/**
+ * Headless Claude fallback for timezone resolution.
+ *
+ * When the user answers the UTC-confirmation prompt with something that
+ * isn't a valid IANA zone ("NYC", "Jerusalem time", "eastern"), spawn
+ * `claude -p` with a narrow prompt asking for a single IANA string and
+ * validate the reply with `isValidTimezone` before returning it.
+ *
+ * Gated on claude being on PATH — if the user did the paste-OAuth or
+ * paste-API auth path they may not have the CLI installed. Returns null
+ * in that case so the caller can ask them to try again with a canonical
+ * zone string.
+ */
+import { execSync, spawn } from 'child_process';
+
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import { isValidTimezone } from '../../src/timezone.js';
+import { fitToWidth } from './theme.js';
+
+export function claudeCliAvailable(): boolean {
+  try {
+    execSync('command -v claude', { stdio: 'ignore' });
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Ask headless Claude to map a free-text location/timezone description to
+ * a valid IANA zone. Shows a spinner with elapsed time. Returns the
+ * resolved zone string on success, or null if the CLI is missing, Claude
+ * errored, or the reply wasn't a valid IANA zone.
+ */
+export async function resolveTimezoneViaClaude(
+  input: string,
+): Promise<string | null> {
+  if (!claudeCliAvailable()) return null;
+
+  const prompt = buildPrompt(input);
+
+  const s = p.spinner();
+  const start = Date.now();
+  const label = 'Looking up that timezone…';
+  s.start(fitToWidth(label, ' (999s)'));
+  const tick = setInterval(() => {
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    const suffix = ` (${elapsed}s)`;
+    s.message(`${fitToWidth(label, suffix)}${k.dim(suffix)}`);
+  }, 1000);
+
+  const reply = await queryClaude(prompt);
+
+  clearInterval(tick);
+  const elapsed = Math.round((Date.now() - start) / 1000);
+  const suffix = ` (${elapsed}s)`;
+
+  const resolved = reply ? extractTimezone(reply) : null;
+  if (resolved) {
+    s.stop(
+      `${fitToWidth(`Interpreted as ${resolved}.`, suffix)}${k.dim(suffix)}`,
+    );
+    return resolved;
+  }
+  s.stop(
+    `${fitToWidth("Couldn't interpret that as a timezone.", suffix)}${k.dim(
+      suffix,
+    )}`,
+    1,
+  );
+  return null;
+}
+
+function buildPrompt(input: string): string {
+  return [
+    'Convert the user\'s description of where they are into a single IANA',
+    'timezone identifier (e.g. "America/New_York", "Europe/London",',
+    '"Asia/Jerusalem"). Respond with ONLY the IANA string on a single line,',
+    'nothing else — no prose, no quotes, no punctuation. If you cannot',
+    'determine a zone with reasonable confidence, reply with exactly:',
+    'UNKNOWN',
+    '',
+    `User's description: ${input}`,
+  ].join('\n');
+}
+
+function queryClaude(prompt: string): Promise<string | null> {
+  return new Promise((resolve) => {
+    const child = spawn('claude', ['-p', '--output-format', 'text'], {
+      stdio: ['pipe', 'pipe', 'pipe'],
+    });
+    let stdout = '';
+    let settled = false;
+    const settle = (value: string | null): void => {
+      if (settled) return;
+      settled = true;
+      resolve(value);
+    };
+
+    child.stdout.on('data', (c: Buffer) => {
+      stdout += c.toString('utf-8');
+    });
+    child.on('close', (code) => {
+      settle(code === 0 && stdout.trim() ? stdout : null);
+    });
+    child.on('error', () => settle(null));
+
+    child.stdin.end(prompt);
+  });
+}
+
+function extractTimezone(reply: string): string | null {
+  // Claude occasionally prefixes with a backtick or wraps in quotes despite
+  // instructions; take the first line that looks like a zone.
+  const lines = reply
+    .split('\n')
+    .map((l) => l.trim().replace(/^["'`]+|["'`]+$/g, ''))
+    .filter(Boolean);
+  for (const line of lines) {
+    if (line === 'UNKNOWN') return null;
+    if (isValidTimezone(line)) return line;
+  }
+  return null;
+}
diff --git a/setup/lib/windowed-runner.ts b/setup/lib/windowed-runner.ts
new file mode 100644
index 0000000..875aba6
--- /dev/null
+++ b/setup/lib/windowed-runner.ts
@@ -0,0 +1,229 @@
+/**
+ * Windowed step runner: shows a fixed-height rolling tail of a long step's
+ * output so the user can see it's making progress, plus a stall detector
+ * that interrupts with a "keep waiting or ask for help?" prompt when the
+ * output stream goes silent for too long.
+ *
+ * Used for the container build (3–10 minutes on a fresh machine, no user
+ * feedback with a plain spinner). Models the UI on claude-assist.ts's
+ * 3-line action window — a single-line spinner header sitting above three
+ * gutter-prefixed lines of the most recent output, redrawn in place via
+ * ANSI cursor controls.
+ *
+ * Stall detection: a silence timer resets on every new line. When it hits
+ * STALL_THRESHOLD_MS we pause the render, show `offerClaudeAssist` with
+ * the step's raw log, and either resume (user said "keep waiting") or
+ * let the step run its course while giving them the exit path.
+ */
+import * as p from '@clack/prompts';
+import k from 'kleur';
+
+import { offerClaudeAssist } from './claude-assist.js';
+import { emit as phEmit } from './diagnostics.js';
+import type { StepResult, SpinnerLabels } from './runner.js';
+import { dumpTranscriptOnFailure, spawnStep, writeStepEntry } from './runner.js';
+import * as setupLog from '../logs.js';
+import { fitToWidth } from './theme.js';
+
+const WINDOW_SIZE = 3;
+const SPINNER_FRAMES = ['◒', '◐', '◓', '◑'];
+const HIDE_CURSOR = '\x1b[?25l';
+const SHOW_CURSOR = '\x1b[?25h';
+const STALL_THRESHOLD_MS = 60_000;
+
+/**
+ * Run a step with a 3-line rolling tail + stall detector. Same signature
+ * shape as `runQuietStep` (so auto.ts can swap them), but tails the
+ * child's stdout/stderr into a fixed-height window.
+ */
+export async function runWindowedStep(
+  stepName: string,
+  labels: SpinnerLabels,
+  extra: string[] = [],
+): Promise<StepResult & { rawLog: string; durationMs: number }> {
+  const rawLog = setupLog.stepRawLog(stepName);
+  const start = Date.now();
+  phEmit('step_started', { step: stepName });
+
+  const result = await runUnderWindow(stepName, labels, extra, rawLog);
+
+  const durationMs = Date.now() - start;
+  writeStepEntry(stepName, result, durationMs, rawLog);
+  phEmit('step_completed', {
+    step: stepName,
+    status: outcomeStatus(result),
+    duration_ms: durationMs,
+  });
+  return { ...result, rawLog, durationMs };
+}
+
+function outcomeStatus(result: StepResult): 'success' | 'skipped' | 'failed' {
+  const rawStatus = result.terminal?.fields.STATUS;
+  if (!result.ok) return 'failed';
+  return rawStatus === 'skipped' ? 'skipped' : 'success';
+}
+
+/**
+ * The core render + spawn loop. Kept separate from `runWindowedStep` so
+ * the logging bookkeeping (writeStepEntry, phEmit) lives with the
+ * public-facing wrapper and this function stays focused on terminal IO.
+ */
+async function runUnderWindow(
+  stepName: string,
+  labels: SpinnerLabels,
+  extra: string[],
+  rawLog: string,
+): Promise<StepResult> {
+  const out = process.stdout;
+  const start = Date.now();
+  const actions: string[] = [];
+  let frameIdx = 0;
+  let lastLineAt = Date.now();
+  let stallPromptActive = false;
+  let handledStall = false;
+
+  const redraw = (): void => {
+    if (stallPromptActive) return;
+    out.write(`\x1b[${WINDOW_SIZE + 1}A`);
+    const elapsed = Math.round((Date.now() - start) / 1000);
+    const icon = SPINNER_FRAMES[frameIdx % SPINNER_FRAMES.length];
+    const suffix = ` (${elapsed}s)`;
+    const header = fitToWidth(labels.running, suffix);
+    out.write(`\x1b[2K${k.cyan(icon)}  ${header}${k.dim(suffix)}\n`);
+
+    for (let i = 0; i < WINDOW_SIZE; i++) {
+      const idx = actions.length - WINDOW_SIZE + i;
+      const action = idx >= 0 ? actions[idx] : '';
+      out.write('\x1b[2K');
+      if (action) {
+        out.write(`${k.gray('│')}  ${k.dim(fitToWidth(action, ''))}`);
+      } else {
+        out.write(k.gray('│'));
+      }
+      out.write('\n');
+    }
+  };
+
+  const clearBlock = (): void => {
+    out.write(`\x1b[${WINDOW_SIZE + 1}A`);
+    for (let i = 0; i < WINDOW_SIZE + 1; i++) {
+      out.write('\x1b[2K\n');
+    }
+    out.write(`\x1b[${WINDOW_SIZE + 1}A`);
+  };
+
+  out.write(HIDE_CURSOR);
+  for (let i = 0; i < WINDOW_SIZE + 1; i++) out.write('\n');
+  redraw();
+
+  const restoreCursorOnExit = (): void => {
+    out.write(SHOW_CURSOR);
+  };
+  process.once('exit', restoreCursorOnExit);
+
+  const frameTick = setInterval(() => {
+    frameIdx++;
+    redraw();
+  }, 250);
+
+  const stallCheck = setInterval(() => {
+    if (handledStall || stallPromptActive) return;
+    if (Date.now() - lastLineAt < STALL_THRESHOLD_MS) return;
+    handledStall = true;
+    void handleStall(stepName, rawLog, {
+      pauseRender: () => {
+        stallPromptActive = true;
+        clearBlock();
+        out.write(SHOW_CURSOR);
+      },
+      resumeRender: () => {
+        out.write(HIDE_CURSOR);
+        for (let i = 0; i < WINDOW_SIZE + 1; i++) out.write('\n');
+        stallPromptActive = false;
+        lastLineAt = Date.now();
+        redraw();
+      },
+    });
+  }, 5_000);
+
+  const onLine = (line: string): void => {
+    lastLineAt = Date.now();
+    // Strip ANSI escape sequences — Docker Buildx writes color codes that
+    // mangle the rolling window layout when replayed in a narrow cell.
+    // eslint-disable-next-line no-control-regex
+    const clean = line.replace(/\x1b\[[0-9;?]*[A-Za-z]/g, '').trim();
+    if (clean) actions.push(clean);
+    redraw();
+  };
+
+  const result = await spawnStep(stepName, extra, () => {}, rawLog, onLine);
+
+  clearInterval(frameTick);
+  clearInterval(stallCheck);
+  clearBlock();
+  out.write(SHOW_CURSOR);
+  process.off('exit', restoreCursorOnExit);
+
+  const elapsed = Math.round((Date.now() - start) / 1000);
+  const suffix = ` (${elapsed}s)`;
+  if (result.ok) {
+    const isSkipped = result.terminal?.fields.STATUS === 'skipped';
+    const msg = isSkipped && labels.skipped ? labels.skipped : labels.done;
+    p.log.success(`${fitToWidth(msg, suffix)}${k.dim(suffix)}`);
+  } else {
+    const failMsg = labels.failed ?? labels.running.replace(/…$/, ' failed');
+    p.log.error(`${fitToWidth(failMsg, suffix)}${k.dim(suffix)}`);
+    dumpTranscriptOnFailure(result.transcript);
+  }
+  return result;
+}
+
+async function handleStall(
+  stepName: string,
+  rawLog: string,
+  render: { pauseRender: () => void; resumeRender: () => void },
+): Promise<void> {
+  render.pauseRender();
+  p.log.warn(
+    `This looks stuck — no output from the ${stepName} step for the last 60 seconds.`,
+  );
+  phEmit('step_stalled', { step: stepName });
+
+  const { ensureAnswer } = await import('./runner.js');
+  const { brightSelect } = await import('./bright-select.js');
+
+  const choice = ensureAnswer(
+    await brightSelect<'wait' | 'help'>({
+      message: "What now?",
+      options: [
+        {
+          value: 'wait',
+          label: "Keep waiting",
+          hint: "large images can take 5–10 minutes",
+        },
+        {
+          value: 'help',
+          label: 'Ask Claude to take a look',
+          hint: 'reads the raw build log and suggests a fix',
+        },
+      ],
+    }),
+  );
+
+  if (choice === 'help') {
+    // offerClaudeAssist runs its own spinner and may propose a fix command.
+    // We don't attempt to restart the stalled build from here — if Claude
+    // proposes a command the user accepts, they can retry setup afterwards.
+    await offerClaudeAssist({
+      stepName,
+      msg: `The ${stepName} step has produced no output for 60 seconds.`,
+      hint: 'It may be hung on a slow network pull or a failing Dockerfile step.',
+      rawLogPath: rawLog,
+    });
+    // Keep the spinner going — the underlying process is still running,
+    // and cancelling it here would race with Claude's investigation. The
+    // user can Ctrl-C if they want to bail.
+  }
+
+  render.resumeRender();
+}
diff --git a/setup/logs.ts b/setup/logs.ts
new file mode 100644
index 0000000..7e37beb
--- /dev/null
+++ b/setup/logs.ts
@@ -0,0 +1,144 @@
+/**
+ * Three-level setup logging primitives. See docs/setup-flow.md for the
+ * contract and design rationale.
+ *
+ *   Level 1: clack UI in setup/auto.ts (not here)
+ *   Level 2: logs/setup.log — structured, append-only progression log
+ *   Level 3: logs/setup-steps/NN-name.log — raw stdout+stderr per step
+ *
+ * Usage from auto.ts:
+ *
+ *   import * as setupLog from './logs.js';
+ *
+ *   const rawLog = setupLog.stepRawLog('container');
+ *   const { ok, durationMs, terminal } =
+ *     await spawnIntoRawLog('...', rawLog);
+ *   setupLog.step('container', ok ? 'success' : 'failed', durationMs,
+ *     { RUNTIME: 'docker', BUILD_OK: terminal.fields.BUILD_OK },
+ *     rawLog);
+ *
+ * nanoclaw.sh emits the bootstrap entry directly via a bash helper so
+ * the format stays consistent without needing IPC between bash and tsx.
+ */
+import fs from 'fs';
+import path from 'path';
+
+const LOGS_DIR = 'logs';
+const STEPS_DIR = path.join(LOGS_DIR, 'setup-steps');
+const PROGRESS_LOG = path.join(LOGS_DIR, 'setup.log');
+
+export const progressLogPath = PROGRESS_LOG;
+export const stepsDir = STEPS_DIR;
+
+// Track steps that finished cleanly in this run. Used by fail() to build
+// a NANOCLAW_SKIP list when re-executing after a Claude-assisted fix, so
+// the retry picks up at the failing step instead of redoing every step
+// before it.
+const completedInRun = new Set<string>();
+
+export function completedStepNames(): string[] {
+  return [...completedInRun];
+}
+
+/** Wipe prior logs and write a header. Called once per fresh run (by nanoclaw.sh or as a fallback by auto.ts if invoked standalone). */
+export function reset(meta: Record<string, string>): void {
+  if (fs.existsSync(STEPS_DIR)) {
+    fs.rmSync(STEPS_DIR, { recursive: true, force: true });
+  }
+  fs.mkdirSync(STEPS_DIR, { recursive: true });
+  if (fs.existsSync(PROGRESS_LOG)) fs.unlinkSync(PROGRESS_LOG);
+  header(meta);
+}
+
+/** Append a run-start header to the progression log. Idempotent: creates the file if missing. */
+export function header(meta: Record<string, string>): void {
+  fs.mkdirSync(LOGS_DIR, { recursive: true });
+  const ts = new Date().toISOString();
+  const lines = [`## ${ts} · setup:auto started`];
+  for (const [k, v] of Object.entries(meta)) {
+    lines.push(`  ${k}: ${v}`);
+  }
+  lines.push('');
+  fs.appendFileSync(PROGRESS_LOG, lines.join('\n') + '\n');
+}
+
+/** Append one step entry to the progression log. */
+export function step(
+  name: string,
+  status: 'success' | 'skipped' | 'failed' | 'aborted' | 'interactive',
+  durationMs: number,
+  fields: Record<string, string | number | boolean | undefined>,
+  rawRel?: string,
+): void {
+  fs.mkdirSync(LOGS_DIR, { recursive: true });
+  const ts = new Date().toISOString();
+  const dur = formatDuration(durationMs);
+  const lines = [`=== [${ts}] ${name} [${dur}] → ${status} ===`];
+  for (const [k, v] of Object.entries(fields)) {
+    if (v === undefined || v === null || v === '') continue;
+    lines.push(`  ${k.toLowerCase()}: ${String(v)}`);
+  }
+  if (rawRel) lines.push(`  raw: ${rawRel}`);
+  lines.push('');
+  fs.appendFileSync(PROGRESS_LOG, lines.join('\n') + '\n');
+
+  if (status === 'success' || status === 'skipped') {
+    completedInRun.add(name);
+  }
+}
+
+/** A user answered a prompt. Logs as its own entry because the setup path depends on it. */
+export function userInput(key: string, value: string): void {
+  fs.mkdirSync(LOGS_DIR, { recursive: true });
+  const ts = new Date().toISOString();
+  fs.appendFileSync(
+    PROGRESS_LOG,
+    `=== [${ts}] user-input → ${key} ===\n  value: ${value}\n\n`,
+  );
+}
+
+/** Append the success footer. */
+export function complete(totalMs: number): void {
+  const ts = new Date().toISOString();
+  fs.appendFileSync(
+    PROGRESS_LOG,
+    `## ${ts} · completed (total ${formatDurationTotal(totalMs)})\n`,
+  );
+}
+
+/** Append the failure footer. Keep error short — full context lives in the failing step's raw log. */
+export function abort(stepName: string, error: string): void {
+  const ts = new Date().toISOString();
+  fs.appendFileSync(
+    PROGRESS_LOG,
+    `## ${ts} · aborted at ${stepName} (${error})\n`,
+  );
+}
+
+/**
+ * Return the next raw-log path for a given step name. Numbering is derived
+ * from the count of existing NN-*.log files in STEPS_DIR, so bootstrap's
+ * pre-existing 01-bootstrap.log (written by nanoclaw.sh before this module
+ * is loaded) counts toward the sequence.
+ */
+export function stepRawLog(name: string): string {
+  fs.mkdirSync(STEPS_DIR, { recursive: true });
+  const existing = fs
+    .readdirSync(STEPS_DIR)
+    .filter((n) => /^\d+-.+\.log$/.test(n));
+  const nextIdx = existing.length + 1;
+  const num = String(nextIdx).padStart(2, '0');
+  const safeName = name.replace(/[^a-z0-9-]/gi, '-').toLowerCase();
+  return path.join(STEPS_DIR, `${num}-${safeName}.log`);
+}
+
+function formatDuration(ms: number): string {
+  if (ms < 1000) return `${Math.round(ms)}ms`;
+  return `${(ms / 1000).toFixed(1)}s`;
+}
+
+function formatDurationTotal(ms: number): string {
+  const mins = Math.floor(ms / 60000);
+  const secs = Math.round((ms % 60000) / 1000);
+  return mins > 0 ? `${mins}m${secs}s` : `${secs}s`;
+}
diff --git a/setup/onecli.ts b/setup/onecli.ts
index 226d302..3ceb1e8 100644
--- a/setup/onecli.ts
+++ b/setup/onecli.ts
@@ -1,13 +1,15 @@
 /**
  * Step: onecli — Install + configure the OneCLI gateway and CLI.
  *
- * Aggregates what the old /setup + /init-onecli skills ran as loose shell
- * commands. Idempotent: skips install if `onecli` already works, and safely
- * re-applies PATH, api-host, and .env updates.
+ * Two modes:
+ *   (default) run the OneCLI installer, configure api-host, write .env.
+ *   --reuse   skip the installer; reuse the onecli instance already running
+ *             on the host. Required for users who have other apps bound to
+ *             an existing gateway, since re-running the installer rebinds
+ *             the listener and breaks those consumers.
  *
- * Emits ONECLI_URL so /new-setup SKILL.md can forward it downstream (e.g. as
- * ${ONECLI_URL} in status messages). Polls /health to give downstream steps
- * (auth, service) a ready gateway.
+ * Emits ONECLI_URL and polls /health so downstream steps (auth, service)
+ * get a ready gateway.
  */
 import { execFileSync, execSync } from 'child_process';
 import fs from 'fs';
@@ -37,15 +39,27 @@ function onecliVersion(): string | null {
   }
 }
 
-function getApiHost(): string | null {
+/**
+ * Ask the installed onecli CLI for its configured api-host. Returns null if
+ * onecli isn't on PATH, errors, or has no api-host configured.
+ *
+ * Tolerates both JSON output (onecli 1.3+) and older raw-text output.
+ */
+export function getOnecliApiHost(): string | null {
   try {
     const out = execFileSync('onecli', ['config', 'get', 'api-host'], {
       encoding: 'utf-8',
       env: childEnv(),
       stdio: ['ignore', 'pipe', 'ignore'],
     }).trim();
-    const parsed = JSON.parse(out) as { value?: unknown };
-    return typeof parsed.value === 'string' && parsed.value ? parsed.value : null;
+    try {
+      const parsed = JSON.parse(out) as { data?: unknown; value?: unknown };
+      const val = parsed.data ?? parsed.value;
+      if (typeof val === 'string' && val.trim()) return val.trim();
+    } catch {
+      // not JSON — fall through to URL extraction
+    }
+    return extractUrlFromOutput(out);
   } catch {
     return null;
   }
@@ -83,25 +97,137 @@ function writeEnvOnecliUrl(url: string): void {
   fs.writeFileSync(envFile, content);
 }
 
+// Last-known-good CLI release. Used only if BOTH the upstream installer
+// and the redirect-based version probe fail. Bump deliberately when a
+// new CLI release ships.
+const ONECLI_CLI_FALLBACK_VERSION = '1.3.0';
+const ONECLI_CLI_REPO = 'onecli/onecli-cli';
+
 function installOnecli(): { stdout: string; ok: boolean } {
-  // OneCLI's own install script handles gateway + CLI + PATH.
-  // We run the two canonical installers in sequence and capture stdout so
-  // we can extract the printed URL as a fallback to `onecli config get`.
   let stdout = '';
+
+  // Gateway install (docker-compose based, no rate-limit concerns).
+  const gw = runInstall('curl -fsSL onecli.sh/install | sh');
+  stdout += gw.stdout;
+  if (!gw.ok) {
+    log.error('OneCLI gateway install failed', { stderr: gw.stderr });
+    return { stdout: stdout + (gw.stderr ?? ''), ok: false };
+  }
+
+  // CLI install. The upstream script calls the GitHub releases API
+  // (api.github.com) to resolve the latest tag — which 403s anonymous
+  // callers after 60 requests/hour per IP. Try upstream first; on failure
+  // resolve the version ourselves (via HTTP redirect, which isn't
+  // API-throttled) and download the release archive directly.
+  const upstream = runInstall('curl -fsSL onecli.sh/cli/install | sh');
+  stdout += upstream.stdout;
+  if (upstream.ok) return { stdout, ok: true };
+
+  log.warn('Upstream CLI installer failed — falling back to direct download', {
+    stderr: upstream.stderr,
+  });
+  stdout += (upstream.stderr ?? '') + '\n';
+
+  const fallback = installOnecliCliDirect();
+  stdout += fallback.stdout;
+  if (!fallback.ok) {
+    log.error('OneCLI CLI install failed (both upstream and direct fallback)');
+    return { stdout, ok: false };
+  }
+  return { stdout, ok: true };
+}
+
+function runInstall(cmd: string): { stdout: string; stderr?: string; ok: boolean } {
   try {
-    stdout += execSync('curl -fsSL onecli.sh/install | sh', {
-      encoding: 'utf-8',
-      stdio: ['ignore', 'pipe', 'pipe'],
-    });
-    stdout += execSync('curl -fsSL onecli.sh/cli/install | sh', {
+    const stdout = execSync(cmd, {
       encoding: 'utf-8',
       stdio: ['ignore', 'pipe', 'pipe'],
     });
     return { stdout, ok: true };
   } catch (err) {
     const e = err as { stdout?: string; stderr?: string };
-    log.error('OneCLI install failed', { stderr: e.stderr });
-    return { stdout: stdout + (e.stdout ?? '') + (e.stderr ?? ''), ok: false };
+    return { stdout: e.stdout ?? '', stderr: e.stderr, ok: false };
+  }
+}
+
+/**
+ * Reinstate the OneCLI CLI install without hitting GitHub's rate-limited
+ * releases API. Resolves the version via the HTTP redirect from
+ * /releases/latest → /releases/tag/vX.Y.Z, then downloads the archive
+ * directly. Falls back to ONECLI_CLI_FALLBACK_VERSION if the redirect
+ * probe also fails.
+ */
+function installOnecliCliDirect(): { stdout: string; ok: boolean } {
+  const lines: string[] = [];
+  const append = (s: string): void => {
+    lines.push(s);
+  };
+
+  const osName =
+    process.platform === 'darwin' ? 'darwin' : process.platform === 'linux' ? 'linux' : null;
+  if (!osName) {
+    append(`Unsupported platform: ${process.platform}`);
+    return { stdout: lines.join('\n'), ok: false };
+  }
+  const arch =
+    process.arch === 'x64' ? 'amd64' : process.arch === 'arm64' ? 'arm64' : null;
+  if (!arch) {
+    append(`Unsupported arch: ${process.arch}`);
+    return { stdout: lines.join('\n'), ok: false };
+  }
+
+  let version: string | null = null;
+  try {
+    const redirect = execSync(
+      `curl -fsSL -o /dev/null -w '%{url_effective}' https://github.com/${ONECLI_CLI_REPO}/releases/latest`,
+      { encoding: 'utf-8', stdio: ['ignore', 'pipe', 'pipe'] },
+    ).trim();
+    const m = redirect.match(/\/tag\/v?([^/]+)$/);
+    if (m) version = m[1];
+  } catch {
+    // redirect probe failed — we'll pin the fallback
+  }
+  if (!version) {
+    version = ONECLI_CLI_FALLBACK_VERSION;
+    append(`Version probe failed; installing pinned fallback ${version}.`);
+  } else {
+    append(`Resolved onecli CLI ${version} via release redirect.`);
+  }
+
+  const archive = `onecli_${version}_${osName}_${arch}.tar.gz`;
+  const url = `https://github.com/${ONECLI_CLI_REPO}/releases/download/v${version}/${archive}`;
+  const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'onecli-'));
+  const archivePath = path.join(tmpDir, archive);
+
+  try {
+    append(`Downloading ${url}`);
+    execSync(
+      `curl -fsSL -o ${JSON.stringify(archivePath)} ${JSON.stringify(url)}`,
+      { stdio: ['ignore', 'pipe', 'pipe'] },
+    );
+    execSync(`tar -xzf ${JSON.stringify(archivePath)} -C ${JSON.stringify(tmpDir)}`, {
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+
+    let installDir = '/usr/local/bin';
+    try {
+      fs.accessSync(installDir, fs.constants.W_OK);
+    } catch {
+      installDir = LOCAL_BIN;
+      fs.mkdirSync(installDir, { recursive: true });
+    }
+    const binSrc = path.join(tmpDir, 'onecli');
+    const binDest = path.join(installDir, 'onecli');
+    fs.copyFileSync(binSrc, binDest);
+    fs.chmodSync(binDest, 0o755);
+    append(`onecli ${version} installed to ${binDest}.`);
+    return { stdout: lines.join('\n'), ok: true };
+  } catch (err) {
+    const e = err as { stdout?: string; stderr?: string; message?: string };
+    append(`Direct install failed: ${e.stderr ?? e.message ?? String(err)}`);
+    return { stdout: lines.join('\n'), ok: false };
+  } finally {
+    fs.rmSync(tmpDir, { recursive: true, force: true });
   }
 }
 
@@ -120,63 +246,92 @@ async function pollHealth(url: string, timeoutMs: number): Promise<boolean> {
   return false;
 }
 
-export async function run(_args: string[]): Promise<void> {
+export async function run(args: string[]): Promise<void> {
+  const reuse = args.includes('--reuse');
   ensureShellProfilePath();
 
-  let installOutput = '';
-  let present = !!onecliVersion();
-  if (!present) {
-    log.info('Installing OneCLI gateway and CLI');
-    const res = installOnecli();
-    installOutput = res.stdout;
-    if (!res.ok) {
+  if (reuse) {
+    // Reuse-mode: don't touch the running gateway at all. Just verify it
+    // exists, read its api-host, write ONECLI_URL to .env, and move on.
+    const version = onecliVersion();
+    if (!version) {
       emitStatus('ONECLI', {
         INSTALLED: false,
         STATUS: 'failed',
-        ERROR: 'install_failed',
+        ERROR: 'onecli_not_found_for_reuse',
+        HINT: 'onecli not on PATH. Re-run setup and choose "install fresh".',
         LOG: 'logs/setup.log',
       });
       process.exit(1);
     }
-    present = !!onecliVersion();
-    if (!present) {
+    const url = getOnecliApiHost();
+    if (!url) {
       emitStatus('ONECLI', {
-        INSTALLED: false,
+        INSTALLED: true,
         STATUS: 'failed',
-        ERROR: 'onecli_not_on_path_after_install',
-        HINT: 'Open a new shell or run `export PATH="$HOME/.local/bin:$PATH"` and retry.',
+        ERROR: 'onecli_api_host_not_configured',
+        HINT: 'Existing onecli has no api-host set. Run `onecli config set api-host <url>` or re-run setup with install-fresh.',
         LOG: 'logs/setup.log',
       });
       process.exit(1);
     }
+    writeEnvOnecliUrl(url);
+    log.info('Reusing existing OneCLI', { url });
+    const healthy = await pollHealth(url, 5000);
+    emitStatus('ONECLI', {
+      INSTALLED: true,
+      REUSED: true,
+      ONECLI_URL: url,
+      HEALTHY: healthy,
+      STATUS: 'success',
+      LOG: 'logs/setup.log',
+    });
+    return;
   }
 
-  let url = getApiHost();
-  if (!url && installOutput) {
-    url = extractUrlFromOutput(installOutput);
-    if (url) {
-      try {
-        execFileSync('onecli', ['config', 'set', 'api-host', url], {
-          stdio: 'ignore',
-          env: childEnv(),
-        });
-      } catch (err) {
-        log.warn('onecli config set api-host failed', { err });
-      }
-    }
+  log.info('Installing OneCLI gateway and CLI');
+  const res = installOnecli();
+  if (!res.ok) {
+    emitStatus('ONECLI', {
+      INSTALLED: false,
+      STATUS: 'failed',
+      ERROR: 'install_failed',
+      LOG: 'logs/setup.log',
+    });
+    process.exit(1);
+  }
+  if (!onecliVersion()) {
+    emitStatus('ONECLI', {
+      INSTALLED: false,
+      STATUS: 'failed',
+      ERROR: 'onecli_not_on_path_after_install',
+      HINT: 'Open a new shell or run `export PATH="$HOME/.local/bin:$PATH"` and retry.',
+      LOG: 'logs/setup.log',
+    });
+    process.exit(1);
   }
 
+  const url = extractUrlFromOutput(res.stdout);
   if (!url) {
     emitStatus('ONECLI', {
       INSTALLED: true,
       STATUS: 'failed',
       ERROR: 'could_not_resolve_api_host',
-      HINT: 'Run `onecli config get api-host` to inspect the gateway URL.',
+      HINT: 'Inspect logs/setup.log for the install output.',
       LOG: 'logs/setup.log',
     });
     process.exit(1);
   }
 
+  try {
+    execFileSync('onecli', ['config', 'set', 'api-host', url], {
+      stdio: 'ignore',
+      env: childEnv(),
+    });
+  } catch (err) {
+    log.warn('onecli config set api-host failed', { err });
+  }
+
   writeEnvOnecliUrl(url);
   log.info('Wrote ONECLI_URL to .env', { url });
 
diff --git a/setup/pair-telegram.ts b/setup/pair-telegram.ts
new file mode 100644
index 0000000..f3f9bf8
--- /dev/null
+++ b/setup/pair-telegram.ts
@@ -0,0 +1,116 @@
+/**
+ * Step: pair-telegram — issue a one-time pairing code and wait for the
+ * operator to send the code from the chat they want to register.
+ *
+ * Emits machine-readable status blocks only. The parent driver
+ * (`setup:auto`) renders the code / attempt / success UI with clack. Running
+ * this step directly will look sparse — that's intentional.
+ *
+ * Blocks emitted:
+ *   PAIR_TELEGRAM_CODE       { CODE, REASON=initial|regenerated }
+ *   PAIR_TELEGRAM_ATTEMPT    { CANDIDATE }
+ *   PAIR_TELEGRAM (final)    { STATUS=success, CODE, INTENT, PLATFORM_ID,
+ *                              IS_GROUP, PAIRED_USER_ID }
+ *                         or { STATUS=failed, CODE, ERROR }
+ *
+ * Depends on src/channels/telegram-pairing.js, which setup/add-telegram.sh
+ * copies in from the `channels` branch before this step runs. setup/ is
+ * excluded from the host tsconfig, so this file's import resolves only at
+ * runtime — tsc won't complain on branches that haven't run add-telegram yet.
+ */
+import path from 'path';
+
+import {
+  createPairing,
+  waitForPairing,
+  type PairingIntent,
+} from '../src/channels/telegram-pairing.js';
+import { DATA_DIR } from '../src/config.js';
+import { initDb } from '../src/db/connection.js';
+import { runMigrations } from '../src/db/migrations/index.js';
+
+import { emitStatus } from './status.js';
+
+function parseArgs(args: string[]): PairingIntent {
+  let intent: PairingIntent = 'main';
+  for (let i = 0; i < args.length; i++) {
+    if (args[i] === '--intent') {
+      const raw = args[++i] || 'main';
+      if (raw === 'main') {
+        intent = 'main';
+      } else if (raw.startsWith('wire-to:')) {
+        intent = { kind: 'wire-to', folder: raw.slice('wire-to:'.length) };
+      } else if (raw.startsWith('new-agent:')) {
+        intent = { kind: 'new-agent', folder: raw.slice('new-agent:'.length) };
+      } else {
+        throw new Error(`Unknown intent: ${raw}`);
+      }
+    }
+  }
+  return intent;
+}
+
+function intentToString(intent: PairingIntent): string {
+  if (intent === 'main') return 'main';
+  return `${intent.kind}:${intent.folder}`;
+}
+
+export async function run(args: string[]): Promise<void> {
+  const intent = parseArgs(args);
+
+  // Pairing stores state under DATA_DIR; the DB isn't strictly needed for the
+  // pairing primitive itself, but the inbound interceptor running inside the
+  // live service needs migrations applied. Touch it here so a fresh install
+  // doesn't fail on the first code match.
+  const db = initDb(path.join(DATA_DIR, 'v2.db'));
+  runMigrations(db);
+
+  const MAX_REGENERATIONS = 5;
+  let record = await createPairing(intent);
+  emitStatus('PAIR_TELEGRAM_CODE', {
+    CODE: record.code,
+    REASON: 'initial',
+  });
+
+  for (let regen = 0; regen <= MAX_REGENERATIONS; regen++) {
+    try {
+      const consumed = await waitForPairing(record.code, {
+        onAttempt: (a) => {
+          emitStatus('PAIR_TELEGRAM_ATTEMPT', {
+            CANDIDATE: a.candidate,
+          });
+        },
+      });
+
+      emitStatus('PAIR_TELEGRAM', {
+        STATUS: 'success',
+        CODE: record.code,
+        INTENT: intentToString(consumed.intent),
+        PLATFORM_ID: consumed.consumed!.platformId,
+        IS_GROUP: consumed.consumed!.isGroup,
+        PAIRED_USER_ID: consumed.consumed!.adminUserId
+          ? `telegram:${consumed.consumed!.adminUserId}`
+          : '',
+      });
+      return;
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      const invalidated = /invalidated by wrong code/.test(message);
+      if (invalidated && regen < MAX_REGENERATIONS) {
+        record = await createPairing(intent);
+        emitStatus('PAIR_TELEGRAM_CODE', {
+          CODE: record.code,
+          REASON: 'regenerated',
+        });
+        continue;
+      }
+      const reason = invalidated ? 'max-regenerations-exceeded' : message;
+      emitStatus('PAIR_TELEGRAM', {
+        STATUS: 'failed',
+        CODE: record.code,
+        ERROR: reason,
+      });
+      process.exit(2);
+    }
+  }
+}
diff --git a/setup/peer-cleanup.ts b/setup/peer-cleanup.ts
new file mode 100644
index 0000000..10b22b9
--- /dev/null
+++ b/setup/peer-cleanup.ts
@@ -0,0 +1,186 @@
+/**
+ * Detect and clean up unhealthy NanoClaw peer services.
+ *
+ * Runs as a setup preflight before we install our own service. A crash-looping
+ * peer install (typically the legacy v1 `com.nanoclaw` plist) silently trashes
+ * this install's containers on every respawn because its `cleanupOrphans()`
+ * reaps anything matching `nanoclaw-`. We scope our reaper by label now, but
+ * we still need to stop the peer from killing us on its way down.
+ *
+ * A peer is "unhealthy" when:
+ *   - launchd: `state != running` AND `runs > UNHEALTHY_RUNS_THRESHOLD`
+ *   - systemd: unit is in `failed` state, OR `activating` with many restarts
+ *
+ * Healthy peers are left alone — multiple installs can coexist fine now that
+ * container-reaper is label-scoped.
+ */
+import { execFileSync } from 'child_process';
+import fs from 'fs';
+import os from 'os';
+import path from 'path';
+
+import { getLaunchdLabel, getSystemdUnit } from '../src/install-slug.js';
+import { log } from '../src/log.js';
+
+const UNHEALTHY_RUNS_THRESHOLD = 10;
+
+export interface PeerStatus {
+  label: string;
+  configPath: string;
+  state: string;
+  runs: number;
+  unhealthy: boolean;
+}
+
+export interface PeerCleanupResult {
+  checked: PeerStatus[];
+  unloaded: PeerStatus[];
+  failures: Array<{ label: string; err: string }>;
+}
+
+/**
+ * Scan for peer NanoClaw services and unload any that are crash-looping.
+ * Returns a summary suitable for emitStatus / setup-log reporting.
+ */
+export function cleanupUnhealthyPeers(projectRoot: string = process.cwd()): PeerCleanupResult {
+  const platform = os.platform();
+  if (platform === 'darwin') {
+    return cleanupLaunchdPeers(projectRoot);
+  }
+  if (platform === 'linux') {
+    return cleanupSystemdPeers(projectRoot);
+  }
+  return { checked: [], unloaded: [], failures: [] };
+}
+
+// ---- launchd (macOS) --------------------------------------------------------
+
+function cleanupLaunchdPeers(projectRoot: string): PeerCleanupResult {
+  const ownLabel = getLaunchdLabel(projectRoot);
+  const agentsDir = path.join(os.homedir(), 'Library', 'LaunchAgents');
+  const result: PeerCleanupResult = { checked: [], unloaded: [], failures: [] };
+
+  let plists: string[];
+  try {
+    plists = fs
+      .readdirSync(agentsDir)
+      .filter((f) => /^com\.nanoclaw.*\.plist$/.test(f))
+      .map((f) => path.join(agentsDir, f));
+  } catch {
+    return result;
+  }
+
+  const uid = process.getuid?.() ?? 0;
+
+  for (const plistPath of plists) {
+    const label = path.basename(plistPath, '.plist');
+    if (label === ownLabel) continue;
+
+    const status = probeLaunchdPeer(label, plistPath, uid);
+    if (!status) continue;
+    result.checked.push(status);
+
+    if (!status.unhealthy) continue;
+
+    try {
+      execFileSync('launchctl', ['unload', plistPath], { stdio: 'pipe' });
+      log.info('Unloaded unhealthy peer launchd service', {
+        label,
+        state: status.state,
+        runs: status.runs,
+        plistPath,
+      });
+      result.unloaded.push(status);
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      log.warn('Failed to unload peer launchd service', { label, err: message });
+      result.failures.push({ label, err: message });
+    }
+  }
+
+  return result;
+}
+
+function probeLaunchdPeer(label: string, plistPath: string, uid: number): PeerStatus | null {
+  let output: string;
+  try {
+    output = execFileSync('launchctl', ['print', `gui/${uid}/${label}`], {
+      stdio: ['ignore', 'pipe', 'pipe'],
+      encoding: 'utf-8',
+    });
+  } catch {
+    // Not loaded → not currently a threat. Skip silently.
+    return null;
+  }
+
+  const state = /^\s*state\s*=\s*(.+?)\s*$/m.exec(output)?.[1] ?? 'unknown';
+  const runsStr = /^\s*runs\s*=\s*(\d+)/m.exec(output)?.[1];
+  const runs = runsStr ? parseInt(runsStr, 10) : 0;
+
+  const unhealthy = state !== 'running' && runs > UNHEALTHY_RUNS_THRESHOLD;
+  return { label, configPath: plistPath, state, runs, unhealthy };
+}
+
+// ---- systemd (Linux) --------------------------------------------------------
+
+function cleanupSystemdPeers(projectRoot: string): PeerCleanupResult {
+  const ownUnit = getSystemdUnit(projectRoot);
+  const unitDir = path.join(os.homedir(), '.config', 'systemd', 'user');
+  const result: PeerCleanupResult = { checked: [], unloaded: [], failures: [] };
+
+  let units: string[];
+  try {
+    units = fs
+      .readdirSync(unitDir)
+      .filter((f) => /^nanoclaw.*\.service$/.test(f))
+      .map((f) => f.replace(/\.service$/, ''));
+  } catch {
+    return result;
+  }
+
+  for (const unit of units) {
+    if (unit === ownUnit) continue;
+
+    const status = probeSystemdPeer(unit);
+    if (!status) continue;
+    result.checked.push(status);
+
+    if (!status.unhealthy) continue;
+
+    try {
+      execFileSync('systemctl', ['--user', 'disable', '--now', `${unit}.service`], { stdio: 'pipe' });
+      log.info('Disabled unhealthy peer systemd unit', {
+        unit,
+        state: status.state,
+        runs: status.runs,
+      });
+      result.unloaded.push(status);
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      log.warn('Failed to disable peer systemd unit', { unit, err: message });
+      result.failures.push({ label: unit, err: message });
+    }
+  }
+
+  return result;
+}
+
+function probeSystemdPeer(unit: string): PeerStatus | null {
+  const unitPath = path.join(os.homedir(), '.config', 'systemd', 'user', `${unit}.service`);
+  try {
+    const output = execFileSync(
+      'systemctl',
+      ['--user', 'show', '--property=ActiveState,NRestarts', `${unit}.service`],
+      { stdio: ['ignore', 'pipe', 'pipe'], encoding: 'utf-8' },
+    );
+    const activeState = /^ActiveState=(.+)$/m.exec(output)?.[1]?.trim() ?? 'unknown';
+    const restartsStr = /^NRestarts=(\d+)/m.exec(output)?.[1];
+    const runs = restartsStr ? parseInt(restartsStr, 10) : 0;
+
+    const unhealthy =
+      activeState === 'failed' || (activeState !== 'active' && runs > UNHEALTHY_RUNS_THRESHOLD);
+    return { label: unit, configPath: unitPath, state: activeState, runs, unhealthy };
+  } catch {
+    return null;
+  }
+}
diff --git a/setup/probe.sh b/setup/probe.sh
index 6f40fff..f4cbf3f 100755
--- a/setup/probe.sh
+++ b/setup/probe.sh
@@ -19,7 +19,13 @@ START_S=$(date +%s)
 
 PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 LOCAL_BIN="$HOME/.local/bin"
-AGENT_IMAGE="nanoclaw-agent:latest"
+
+# Per-checkout install names (match setup/lib/install-slug.ts).
+# shellcheck source=setup/lib/install-slug.sh
+source "$PROJECT_ROOT/setup/lib/install-slug.sh"
+LAUNCHD_LABEL=$(launchd_label)
+SYSTEMD_UNIT=$(systemd_unit)
+AGENT_IMAGE="$(container_image_base):latest"
 
 export PATH="$LOCAL_BIN:$PATH"
 
@@ -144,7 +150,7 @@ probe_service_status() {
     macos)
       command_exists launchctl || { echo "not_configured"; return; }
       local line
-      line=$(with_timeout launchctl list 2>/dev/null | grep "com.nanoclaw") || {
+      line=$(with_timeout launchctl list 2>/dev/null | grep "$LAUNCHD_LABEL") || {
         echo "not_configured"; return; }
       local pid
       pid=$(echo "$line" | awk '{print $1}')
@@ -156,7 +162,7 @@ probe_service_status() {
       ;;
     linux|wsl)
       command_exists systemctl || { echo "not_configured"; return; }
-      if with_timeout systemctl --user is-active nanoclaw >/dev/null 2>&1; then
+      if with_timeout systemctl --user is-active "$SYSTEMD_UNIT" >/dev/null 2>&1; then
         echo "running"
       elif with_timeout systemctl --user cat nanoclaw >/dev/null 2>&1; then
         echo "stopped"
diff --git a/setup/register-claude-token.sh b/setup/register-claude-token.sh
new file mode 100755
index 0000000..e0adfc6
--- /dev/null
+++ b/setup/register-claude-token.sh
@@ -0,0 +1,106 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Register a Claude subscription OAuth token with OneCLI — the *only* auth
+# path that needs a TTY break in the flow. Paste-based paths (existing
+# OAuth token / API key) are handled in-process by setup/auto.ts using
+# clack prompts, then onecli secrets create is invoked directly from TS.
+#
+# Flow:
+#   1. Run `claude setup-token` under a PTY (via script(1)) so the browser
+#      OAuth dance works and its token is captured into a tempfile.
+#   2. Regex the sk-ant-oat…AA token out of the ANSI-stripped capture.
+#   3. Register it with OneCLI.
+#
+# Env overrides:
+#   SECRET_NAME   OneCLI secret name   (default: Anthropic)
+#   HOST_PATTERN  OneCLI host pattern  (default: api.anthropic.com)
+
+# Prefer bash 4+ (for `read -e -i` readline preload). macOS ships 3.2 in
+# /bin/bash, but Homebrew users usually have 5.x first on PATH. The
+# readline preload is optional — on 3.x we fall back to a plain prompt.
+
+SECRET_NAME="${SECRET_NAME:-Anthropic}"
+HOST_PATTERN="${HOST_PATTERN:-api.anthropic.com}"
+
+command -v onecli >/dev/null \
+  || { echo "onecli not found. Install it first (see /setup §4)." >&2; exit 1; }
+
+if ! command -v claude >/dev/null 2>&1; then
+  echo "Claude Code CLI not found — installing it now (needed for subscription sign-in)…"
+  SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+  if ! bash "$SCRIPT_DIR/install-claude.sh"; then
+    echo >&2
+    echo "Couldn't install the Claude Code CLI automatically." >&2
+    echo "Install it manually with" >&2
+    echo "  curl -fsSL https://claude.ai/install.sh | bash" >&2
+    echo "and re-run setup." >&2
+    exit 1
+  fi
+  # install-claude.sh PATH additions are scoped to its own subshell; redo
+  # them here so the rest of this script can see the fresh `claude` binary.
+  if [ -d "$HOME/.local/bin" ] && [[ ":$PATH:" != *":$HOME/.local/bin:"* ]]; then
+    export PATH="$HOME/.local/bin:$PATH"
+  fi
+  hash -r 2>/dev/null || true
+fi
+
+command -v script >/dev/null \
+  || { echo "script(1) is required for PTY capture." >&2; exit 1; }
+
+tmpfile=$(mktemp -t claude-setup-token.XXXXXX)
+trap 'rm -f "$tmpfile"' EXIT
+
+cat <<'EOF'
+A browser window will open for you to sign in with your Claude account.
+When you finish, we'll save the token to your OneCLI vault automatically.
+
+Press Enter to continue, or edit the command first.
+
+EOF
+
+cmd="claude setup-token"
+if [ "${BASH_VERSINFO[0]:-0}" -ge 4 ]; then
+  # bash 4+: pre-fill the readline buffer so Enter literally submits.
+  read -r -e -i "$cmd" -p "$ " cmd </dev/tty
+else
+  # bash 3.x (macOS default /bin/bash): no readline preload. Fall back.
+  echo "$ $cmd"
+  read -r -p "Press Enter to run, Ctrl-C to abort. " _ </dev/tty
+fi
+
+# `script` arg order differs between BSD (macOS) and util-linux.
+if script --version 2>/dev/null | grep -q util-linux; then
+  script -q -c "$cmd" "$tmpfile"
+else
+  # BSD script: command is argv after the file, so let it word-split.
+  # shellcheck disable=SC2086
+  script -q "$tmpfile" $cmd
+fi
+
+# Strip ANSI codes + newlines (TTY wraps the token mid-string), then match
+# the sk-ant-oat…AA token. perl because BSD grep caps {n,m} at 255.
+token=$(sed $'s/\x1b\\[[0-9;]*[a-zA-Z]//g' "$tmpfile" \
+        | tr -d '\n\r' \
+        | perl -ne 'print "$1\n" while /(sk-ant-oat[A-Za-z0-9_-]{80,500}AA)/g' \
+        | tail -1 || true)
+
+if [ -z "$token" ]; then
+  keep=$(mktemp -t claude-setup-token-log.XXXXXX)
+  cp "$tmpfile" "$keep"
+  echo >&2
+  echo "No sk-ant-oat…AA token found. Raw log: $keep" >&2
+  exit 1
+fi
+
+echo
+echo "Got token: ${token:0:16}…${token: -4}"
+echo "Saving it to your OneCLI vault as '${SECRET_NAME}' (host: ${HOST_PATTERN})…"
+
+onecli secrets create \
+  --name "$SECRET_NAME" \
+  --type anthropic \
+  --value "$token" \
+  --host-pattern "$HOST_PATTERN"
+
+echo "Done."
diff --git a/setup/register.ts b/setup/register.ts
index a308add..7bd5ae3 100644
--- a/setup/register.ts
+++ b/setup/register.ts
@@ -20,6 +20,7 @@ import {
 import { isValidGroupFolder } from '../src/group-folder.js';
 import { initGroupFilesystem } from '../src/group-init.js';
 import { log } from '../src/log.js';
+import { namespacedPlatformId } from '../src/platform-id.js';
 import { resolveSession, writeSessionMessage } from '../src/session-manager.js';
 import { emitStatus } from './status.js';
 
@@ -112,12 +113,10 @@ export async function run(args: string[]): Promise<void> {
     process.exit(4);
   }
 
-  // Chat SDK adapters prefix platform IDs with the channel type
-  // (e.g. "telegram:123", "discord:guild:channel"). Normalize here so
-  // the stored ID always matches what the adapter sends at runtime.
-  if (!parsed.platformId.startsWith(`${parsed.channel}:`)) {
-    parsed.platformId = `${parsed.channel}:${parsed.platformId}`;
-  }
+  // Normalize platform_id to the same shape the adapter will emit at runtime,
+  // so the router's (channel_type, platform_id) lookup matches what we store.
+  // Chat SDK adapters prefix, native adapters (WhatsApp/iMessage/Signal) don't.
+  parsed.platformId = namespacedPlatformId(parsed.channel, parsed.platformId);
 
   log.info('Registering channel', parsed);
 
@@ -167,19 +166,22 @@ export async function run(args: string[]): Promise<void> {
   if (!existing) {
     newlyWired = true;
     const mgaId = generateId('mga');
-    const triggerRules = parsed.trigger
-      ? JSON.stringify({
-          pattern: parsed.trigger,
-          requiresTrigger: parsed.requiresTrigger,
-        })
-      : null;
+    // Mirrors scripts/init-first-agent.ts:wireIfMissing so both setup paths
+    // create rows with the same shape. Groups default to 'mention' (bot only
+    // responds when addressed); DMs default to 'pattern'/'.' (respond to
+    // every message). An explicit --trigger overrides the pattern regex.
+    const isGroup = messagingGroup.is_group === 1;
+    const engageMode: 'pattern' | 'mention' = isGroup && !parsed.trigger ? 'mention' : 'pattern';
+    const engagePattern: string | null = engageMode === 'pattern' ? parsed.trigger || '.' : null;
     createMessagingGroupAgent({
       id: mgaId,
       messaging_group_id: messagingGroup.id,
       agent_group_id: agentGroup.id,
-      trigger_rules: triggerRules,
-      response_scope: 'all',
-      session_mode: parsed.sessionMode,
+      engage_mode: engageMode,
+      engage_pattern: engagePattern,
+      sender_scope: 'all',
+      ignored_message_policy: 'drop',
+      session_mode: parsed.sessionMode as 'shared' | 'per-thread' | 'agent-shared',
       priority: 0,
       created_at: new Date().toISOString(),
     });
diff --git a/setup/run-suggested.sh b/setup/run-suggested.sh
new file mode 100755
index 0000000..3cd47b5
--- /dev/null
+++ b/setup/run-suggested.sh
@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+# Run a command suggested by claude-assist, giving the user a chance to
+# edit it first. Same pattern as setup/register-claude-token.sh: bash 4+
+# pre-fills readline so Enter literally submits; bash 3.x (macOS default
+# /bin/bash) shows the command and waits for Enter.
+#
+# This script is the allowlisted unit — the `eval` happens inside. The
+# caller has already shown the command to the user and gotten confirmation.
+
+set -u
+
+CMD="${1:-}"
+if [ -z "$CMD" ]; then
+  echo "run-suggested: no command provided" >&2
+  exit 1
+fi
+
+echo
+if [ "${BASH_VERSINFO[0]:-0}" -ge 4 ]; then
+  # Pre-fill readline; user can edit before pressing Enter.
+  read -r -e -i "$CMD" -p "$ " cmd </dev/tty
+else
+  # Fallback: display + Enter-to-run.
+  echo "$ $CMD"
+  read -r -p "Press Enter to run, Ctrl-C to abort. " _ </dev/tty
+  cmd="$CMD"
+fi
+
+if [ -z "$cmd" ]; then
+  echo "run-suggested: empty command after edit, skipping." >&2
+  exit 0
+fi
+
+echo
+eval "$cmd"
diff --git a/setup/service.test.ts b/setup/service.test.ts
index 9168fe1..9bc899e 100644
--- a/setup/service.test.ts
+++ b/setup/service.test.ts
@@ -1,6 +1,8 @@
 import { describe, it, expect } from 'vitest';
 import path from 'path';
 
+import { getLaunchdLabel } from '../src/install-slug.js';
+
 /**
  * Tests for service configuration generation.
  *
@@ -14,12 +16,13 @@ function generatePlist(
   projectRoot: string,
   homeDir: string,
 ): string {
+  const label = getLaunchdLabel(projectRoot);
   return `<?xml version="1.0" encoding="UTF-8"?>
 <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
 <plist version="1.0">
 <dict>
     <key>Label</key>
-    <string>com.nanoclaw</string>
+    <string>${label}</string>
     <key>ProgramArguments</key>
     <array>
         <string>${nodePath}</string>
@@ -73,13 +76,11 @@ WantedBy=${isSystem ? 'multi-user.target' : 'default.target'}`;
 }
 
 describe('plist generation', () => {
-  it('contains the correct label', () => {
-    const plist = generatePlist(
-      '/usr/local/bin/node',
-      '/home/user/nanoclaw',
-      '/home/user',
-    );
-    expect(plist).toContain('<string>com.nanoclaw</string>');
+  it('contains the slug-scoped label', () => {
+    const projectRoot = '/home/user/nanoclaw';
+    const plist = generatePlist('/usr/local/bin/node', projectRoot, '/home/user');
+    expect(plist).toContain(`<string>${getLaunchdLabel(projectRoot)}</string>`);
+    expect(plist).toMatch(/<string>com\.nanoclaw-v2-[0-9a-f]{8}<\/string>/);
   });
 
   it('uses the correct node path', () => {
diff --git a/setup/service.ts b/setup/service.ts
index bc85d16..777c0c5 100644
--- a/setup/service.ts
+++ b/setup/service.ts
@@ -10,7 +10,10 @@ import os from 'os';
 import path from 'path';
 
 import { log } from '../src/log.js';
+import { getLaunchdLabel, getSystemdUnit } from '../src/install-slug.js';
+import { cleanupUnhealthyPeers } from './peer-cleanup.js';
 import {
+  commandExists,
   getPlatform,
   getNodePath,
   getServiceManager,
@@ -51,6 +54,19 @@ export async function run(_args: string[]): Promise<void> {
 
   fs.mkdirSync(path.join(projectRoot, 'logs'), { recursive: true });
 
+  // Peer preflight — a crash-looping peer install (most often the legacy v1
+  // `com.nanoclaw` plist) will keep trashing this install's containers on
+  // every respawn via its own cleanupOrphans. Detect and unload any peer
+  // that's unhealthy before we install our service. Healthy peers are left
+  // alone now that container reaping is install-label-scoped.
+  const peerReport = cleanupUnhealthyPeers(projectRoot);
+  if (peerReport.unloaded.length > 0) {
+    log.warn('Unloaded unhealthy peer NanoClaw services', {
+      count: peerReport.unloaded.length,
+      labels: peerReport.unloaded.map((p) => p.label),
+    });
+  }
+
   if (platform === 'macos') {
     setupLaunchd(projectRoot, nodePath, homeDir);
   } else if (platform === 'linux') {
@@ -73,11 +89,14 @@ function setupLaunchd(
   nodePath: string,
   homeDir: string,
 ): void {
+  // Per-checkout service label so multiple NanoClaw installs can coexist
+  // without clobbering each other's plist.
+  const label = getLaunchdLabel(projectRoot);
   const plistPath = path.join(
     homeDir,
     'Library',
     'LaunchAgents',
-    'com.nanoclaw.plist',
+    `${label}.plist`,
   );
   fs.mkdirSync(path.dirname(plistPath), { recursive: true });
 
@@ -86,7 +105,7 @@ function setupLaunchd(
 <plist version="1.0">
 <dict>
     <key>Label</key>
-    <string>com.nanoclaw</string>
+    <string>${label}</string>
     <key>ProgramArguments</key>
     <array>
         <string>${nodePath}</string>
@@ -115,26 +134,44 @@ function setupLaunchd(
   fs.writeFileSync(plistPath, plist);
   log.info('Wrote launchd plist', { plistPath });
 
+  // Unload first to force launchd to drop any cached plist and re-read from
+  // disk. Bare `launchctl load` on an already-loaded plist errors with
+  // "already loaded" and keeps the ORIGINAL plist's ProgramArguments /
+  // WorkingDirectory in memory — even if the file on disk changed. That
+  // bit us when the plist target shifted between installs: kickstart kept
+  // relaunching the old binary and the CLI socket landed in the wrong dir.
+  // unload succeeds whether or not the service was previously loaded; the
+  // failure case is "Could not find specified service" which is harmless.
+  try {
+    execSync(`launchctl unload ${JSON.stringify(plistPath)}`, {
+      stdio: 'ignore',
+    });
+    log.info('launchctl unload succeeded');
+  } catch {
+    log.info('launchctl unload noop (plist was not previously loaded)');
+  }
+
   try {
     execSync(`launchctl load ${JSON.stringify(plistPath)}`, {
       stdio: 'ignore',
     });
     log.info('launchctl load succeeded');
-  } catch {
-    log.warn('launchctl load failed (may already be loaded)');
+  } catch (err) {
+    log.error('launchctl load failed', { err });
   }
 
   // Verify
   let serviceLoaded = false;
   try {
     const output = execSync('launchctl list', { encoding: 'utf-8' });
-    serviceLoaded = output.includes('com.nanoclaw');
+    serviceLoaded = output.includes(label);
   } catch {
     // launchctl list failed
   }
 
   emitStatus('SETUP_SERVICE', {
     SERVICE_TYPE: 'launchd',
+    SERVICE_LABEL: label,
     NODE_PATH: nodePath,
     PROJECT_PATH: projectRoot,
     PLIST_PATH: plistPath,
@@ -207,13 +244,15 @@ function setupSystemd(
   homeDir: string,
 ): void {
   const runningAsRoot = isRoot();
+  const unitName = getSystemdUnit(projectRoot);
+  const unitFileName = `${unitName}.service`;
 
   // Root uses system-level service, non-root uses user-level
   let unitPath: string;
   let systemctlPrefix: string;
 
   if (runningAsRoot) {
-    unitPath = '/etc/systemd/system/nanoclaw.service';
+    unitPath = `/etc/systemd/system/${unitFileName}`;
     systemctlPrefix = 'systemctl';
     log.info('Running as root — installing system-level systemd unit');
   } else {
@@ -229,7 +268,7 @@ function setupSystemd(
     }
     const unitDir = path.join(homeDir, '.config', 'systemd', 'user');
     fs.mkdirSync(unitDir, { recursive: true });
-    unitPath = path.join(unitDir, 'nanoclaw.service');
+    unitPath = path.join(unitDir, unitFileName);
     systemctlPrefix = 'systemctl --user';
   }
 
@@ -255,12 +294,34 @@ WantedBy=${runningAsRoot ? 'multi-user.target' : 'default.target'}`;
   fs.writeFileSync(unitPath, unit);
   log.info('Wrote systemd unit', { unitPath });
 
-  // Detect stale docker group before starting (user systemd only)
-  const dockerGroupStale = !runningAsRoot && checkDockerGroupStale();
+  // Detect stale docker group before starting (user systemd only). The user
+  // systemd manager is a long-running process whose group list is frozen at
+  // login, so `usermod -aG docker` mid-session doesn't reach it. Rather than
+  // require the user to log out + back in, punch a POSIX ACL onto the socket
+  // that grants the current user rw directly. This is temporary — the socket
+  // is recreated by dockerd on restart (and by then the user has relogged, so
+  // normal group perms apply again).
+  let dockerGroupStale = !runningAsRoot && checkDockerGroupStale();
   if (dockerGroupStale) {
     log.warn(
       'Docker group not active in systemd session — user was likely added to docker group mid-session',
     );
+    if (commandExists('setfacl')) {
+      const user = execSync('whoami', { encoding: 'utf-8' }).trim();
+      try {
+        execSync(`sudo setfacl -m u:${user}:rw /var/run/docker.sock`, {
+          stdio: 'inherit',
+        });
+        log.info(
+          'Applied temporary ACL to /var/run/docker.sock (resets on docker restart or reboot)',
+        );
+        dockerGroupStale = false;
+      } catch (err) {
+        log.warn('Failed to apply setfacl workaround', { err });
+      }
+    } else {
+      log.warn('setfacl not installed — cannot apply automatic workaround');
+    }
   }
 
   // Kill orphaned nanoclaw processes to avoid channel connection conflicts
@@ -288,21 +349,26 @@ WantedBy=${runningAsRoot ? 'multi-user.target' : 'default.target'}`;
   }
 
   try {
-    execSync(`${systemctlPrefix} enable nanoclaw`, { stdio: 'ignore' });
+    execSync(`${systemctlPrefix} enable ${unitName}`, { stdio: 'ignore' });
   } catch (err) {
     log.error('systemctl enable failed', { err });
   }
 
+  // restart (not start) so a previously-running instance picks up edits to
+  // the unit file. `start` on an active unit is a no-op, which would leave
+  // the old ExecStart / WorkingDirectory in effect even after daemon-reload.
+  // `restart` on a stopped unit is equivalent to `start`, so this is safe
+  // as a first-install path too.
   try {
-    execSync(`${systemctlPrefix} start nanoclaw`, { stdio: 'ignore' });
+    execSync(`${systemctlPrefix} restart ${unitName}`, { stdio: 'ignore' });
   } catch (err) {
-    log.error('systemctl start failed', { err });
+    log.error('systemctl restart failed', { err });
   }
 
   // Verify
   let serviceLoaded = false;
   try {
-    execSync(`${systemctlPrefix} is-active nanoclaw`, { stdio: 'ignore' });
+    execSync(`${systemctlPrefix} is-active ${unitName}`, { stdio: 'ignore' });
     serviceLoaded = true;
   } catch {
     // Not active
@@ -310,6 +376,7 @@ WantedBy=${runningAsRoot ? 'multi-user.target' : 'default.target'}`;
 
   emitStatus('SETUP_SERVICE', {
     SERVICE_TYPE: runningAsRoot ? 'systemd-system' : 'systemd-user',
+    SERVICE_UNIT: unitName,
     NODE_PATH: nodePath,
     PROJECT_PATH: projectRoot,
     UNIT_PATH: unitPath,
diff --git a/setup/set-env.ts b/setup/set-env.ts
new file mode 100644
index 0000000..5ee4b4e
--- /dev/null
+++ b/setup/set-env.ts
@@ -0,0 +1,77 @@
+/**
+ * Step: set-env — Write or update a KEY=VALUE in .env, with optional sync to
+ * data/env/env (the container-mounted copy).
+ *
+ * Usage:
+ *   pnpm exec tsx setup/index.ts --step set-env -- \
+ *     --key TELEGRAM_BOT_TOKEN --value "<token>" [--sync-container]
+ *
+ * Exists so channel-install flows don't have to invent grep/sed/rm pipelines
+ * (which can't be allowlisted tightly — sed can read any file, and each
+ * segment of an && chain is matched separately).
+ *
+ * Logs the key but never the value.
+ */
+import fs from 'fs';
+import path from 'path';
+
+import { log } from '../src/log.js';
+import { emitStatus } from './status.js';
+
+export async function run(args: string[]): Promise<void> {
+  const keyIdx = args.indexOf('--key');
+  const valueIdx = args.indexOf('--value');
+  const syncContainer = args.includes('--sync-container');
+
+  if (keyIdx === -1 || !args[keyIdx + 1]) {
+    throw new Error('--key <KEY> is required');
+  }
+  if (valueIdx === -1 || args[valueIdx + 1] === undefined) {
+    throw new Error('--value <VALUE> is required');
+  }
+
+  const key = args[keyIdx + 1];
+  const value = args[valueIdx + 1];
+
+  if (!/^[A-Z][A-Z0-9_]*$/.test(key)) {
+    throw new Error(`Invalid env key: ${key} (must be UPPER_SNAKE_CASE)`);
+  }
+
+  const projectRoot = process.cwd();
+  const envFile = path.join(projectRoot, '.env');
+
+  let content = '';
+  if (fs.existsSync(envFile)) {
+    content = fs.readFileSync(envFile, 'utf-8');
+  }
+
+  const lineRegex = new RegExp(`^${key}=.*$`, 'm');
+  const newLine = `${key}=${value}`;
+  const existed = lineRegex.test(content);
+
+  if (existed) {
+    content = content.replace(lineRegex, newLine);
+  } else {
+    const sep = content && !content.endsWith('\n') ? '\n' : '';
+    content = content + sep + newLine + '\n';
+  }
+
+  fs.writeFileSync(envFile, content);
+  log.info('Updated .env', { key, existed });
+
+  let synced = false;
+  if (syncContainer) {
+    const dataEnvDir = path.join(projectRoot, 'data', 'env');
+    fs.mkdirSync(dataEnvDir, { recursive: true });
+    fs.copyFileSync(envFile, path.join(dataEnvDir, 'env'));
+    synced = true;
+    log.info('Synced .env to container mount', { path: 'data/env/env' });
+  }
+
+  emitStatus('SET_ENV', {
+    KEY: key,
+    EXISTED: existed,
+    SYNCED_TO_CONTAINER: synced,
+    STATUS: 'success',
+  });
+}
diff --git a/setup/signal-auth.ts b/setup/signal-auth.ts
new file mode 100644
index 0000000..ce289db
--- /dev/null
+++ b/setup/signal-auth.ts
@@ -0,0 +1,182 @@
+/**
+ * Step: signal-auth — link this host to an existing Signal account via
+ * signal-cli's QR-code flow.
+ *
+ * signal-cli `link` opens a bi-directional handshake with the Signal
+ * servers: it prints one line containing a linking URL (`sgnl://linkdevice?…`
+ * or older `tsdevice://linkdevice?…`), then blocks until either the user
+ * scans it from an existing Signal install, or the code expires. On
+ * success, a secondary account is created under the user's signal-cli
+ * data directory, associated with the phone number of the scanner.
+ *
+ * Methods:
+ *   (no args)                    Spawn signal-cli link, emit SIGNAL_AUTH_QR
+ *                                with the URL, wait for completion.
+ *
+ * Block schema (parent parses these):
+ *   SIGNAL_AUTH_QR       { QR: "<sgnl:// or tsdevice:// url>" }   — one-shot
+ *   SIGNAL_AUTH          { STATUS: success, ACCOUNT: +<digits> }  — terminal
+ *                        { STATUS: skipped, ACCOUNT, REASON: already-authenticated }
+ *                        { STATUS: failed, ERROR: <reason> }
+ *
+ * STATUS values match the runner's vocabulary (success/skipped/failed) so
+ * spawnStep recognises them and sets `ok` correctly; Signal-specific UI
+ * lives in setup/channels/signal.ts.
+ *
+ * If one or more accounts are already linked (discovered via
+ * `signal-cli -o json listAccounts`), the step emits SIGNAL_AUTH
+ * STATUS=skipped with the first account so the driver can reuse it.
+ * Selecting a different existing account is a driver concern.
+ */
+import { spawn, spawnSync } from 'child_process';
+
+import { emitStatus } from './status.js';
+
+const LINK_TIMEOUT_MS = 180_000;
+const DEFAULT_DEVICE_NAME = 'NanoClaw';
+
+interface SignalAccount {
+  account?: string;
+  registered?: boolean;
+}
+
+function cliPath(): string {
+  return process.env.SIGNAL_CLI_PATH || 'signal-cli';
+}
+
+/**
+ * Query signal-cli for currently linked accounts. Empty array if none
+ * configured, no binary, or the call fails for any other reason.
+ */
+function listAccounts(): string[] {
+  const cli = cliPath();
+  try {
+    const res = spawnSync(cli, ['-o', 'json', 'listAccounts'], {
+      encoding: 'utf-8',
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    if (res.status !== 0) return [];
+    const parsed = JSON.parse(res.stdout || '[]') as SignalAccount[];
+    return parsed
+      .filter((a) => a.registered !== false)
+      .map((a) => a.account ?? '')
+      .filter(Boolean);
+  } catch {
+    return [];
+  }
+}
+
+export async function run(_args: string[]): Promise<void> {
+  const cli = cliPath();
+
+  // Verify signal-cli exists before we commit to the long-running link.
+  // The driver checks too, but this keeps the step honest when run alone.
+  const probe = spawnSync(cli, ['--version'], {
+    stdio: ['ignore', 'pipe', 'pipe'],
+  });
+  if (probe.error || probe.status !== 0) {
+    emitStatus('SIGNAL_AUTH', {
+      STATUS: 'failed',
+      ERROR: 'signal-cli not found. Install signal-cli first.',
+    });
+    return;
+  }
+
+  const existing = listAccounts();
+  if (existing.length > 0) {
+    emitStatus('SIGNAL_AUTH', {
+      STATUS: 'skipped',
+      ACCOUNT: existing[0],
+      REASON: 'already-authenticated',
+    });
+    return;
+  }
+
+  await new Promise<void>((resolve) => {
+    let settled = false;
+    let qrEmitted = false;
+
+    const finish = (block: Record<string, string | number | boolean>, code: number): void => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      emitStatus('SIGNAL_AUTH', block);
+      resolve();
+      setTimeout(() => process.exit(code), 500);
+    };
+
+    const timer = setTimeout(() => {
+      try {
+        child.kill('SIGTERM');
+      } catch {
+        /* ignore */
+      }
+      finish({ STATUS: 'failed', ERROR: 'qr_timeout' }, 1);
+    }, LINK_TIMEOUT_MS);
+
+    const child = spawn(cli, ['link', '--name', DEFAULT_DEVICE_NAME], {
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+
+    // stdout carries the URL on the first line; subsequent lines may print
+    // status like "Associated with: +1555…". We don't strictly need to parse
+    // the number — listAccounts after exit is the source of truth — but the
+    // URL match drives the QR emit, which is the whole point.
+    let stdoutBuf = '';
+    const handleStdout = (chunk: Buffer): void => {
+      stdoutBuf += chunk.toString('utf-8');
+      let idx: number;
+      while ((idx = stdoutBuf.indexOf('\n')) !== -1) {
+        const line = stdoutBuf.slice(0, idx).trim();
+        stdoutBuf = stdoutBuf.slice(idx + 1);
+        if (!line) continue;
+        // Match both modern (sgnl://) and legacy (tsdevice://) schemes.
+        if (/^(sgnl|tsdevice):\/\/linkdevice\?/.test(line) && !qrEmitted) {
+          qrEmitted = true;
+          emitStatus('SIGNAL_AUTH_QR', { QR: line });
+        }
+      }
+    };
+    child.stdout.on('data', handleStdout);
+
+    // Capture stderr for the transcript / log — signal-cli writes warnings
+    // and errors there. We don't emit on partial stderr lines since a
+    // successful link can still produce noise.
+    let stderrBuf = '';
+    child.stderr.on('data', (chunk: Buffer) => {
+      stderrBuf += chunk.toString('utf-8');
+    });
+
+    child.on('error', (err) => {
+      finish({ STATUS: 'failed', ERROR: `spawn error: ${err.message}` }, 1);
+    });
+
+    child.on('close', (code) => {
+      // After a successful link, signal-cli exits 0 and the newly linked
+      // account shows up in listAccounts. Use that as the source of truth
+      // rather than scraping stdout — more robust across signal-cli versions.
+      if (code === 0) {
+        const post = listAccounts();
+        if (post.length === 0) {
+          finish(
+            { STATUS: 'failed', ERROR: 'link exited 0 but no account registered' },
+            1,
+          );
+          return;
+        }
+        finish({ STATUS: 'success', ACCOUNT: post[0] }, 0);
+        return;
+      }
+
+      // Non-zero exit. Surface the last non-empty stderr line for context;
+      // signal-cli's own error messages are usually informative.
+      const lastErr =
+        stderrBuf
+          .split('\n')
+          .map((l) => l.trim())
+          .filter(Boolean)
+          .slice(-1)[0] ?? `signal-cli link exited with code ${code}`;
+      finish({ STATUS: 'failed', ERROR: lastErr }, 1);
+    });
+  });
+}
diff --git a/setup/verify.test.ts b/setup/verify.test.ts
new file mode 100644
index 0000000..1e09acd
--- /dev/null
+++ b/setup/verify.test.ts
@@ -0,0 +1,55 @@
+import { describe, expect, it } from 'vitest';
+
+import { determineVerifyStatus } from './verify.js';
+
+const healthyBase = {
+  service: 'running' as const,
+  credentials: 'configured',
+  anyChannelConfigured: false,
+  registeredGroups: 1,
+  agentPing: 'ok' as const,
+};
+
+describe('determineVerifyStatus', () => {
+  it('accepts a working CLI-only install', () => {
+    expect(determineVerifyStatus(healthyBase)).toBe('success');
+  });
+
+  it('accepts a messaging-channel install when CLI ping is skipped', () => {
+    expect(
+      determineVerifyStatus({
+        ...healthyBase,
+        anyChannelConfigured: true,
+        agentPing: 'skipped',
+      }),
+    ).toBe('success');
+  });
+
+  it('fails when neither CLI nor messaging channels are usable', () => {
+    expect(
+      determineVerifyStatus({
+        ...healthyBase,
+        agentPing: 'skipped',
+      }),
+    ).toBe('failed');
+  });
+
+  it('fails when the CLI agent does not respond', () => {
+    expect(
+      determineVerifyStatus({
+        ...healthyBase,
+        anyChannelConfigured: true,
+        agentPing: 'no_reply',
+      }),
+    ).toBe('failed');
+  });
+
+  it('fails when no agent groups are registered', () => {
+    expect(
+      determineVerifyStatus({
+        ...healthyBase,
+        registeredGroups: 0,
+      }),
+    ).toBe('failed');
+  });
+});
diff --git a/setup/verify.ts b/setup/verify.ts
index 566cc9b..30a5408 100644
--- a/setup/verify.ts
+++ b/setup/verify.ts
@@ -14,6 +14,8 @@ import Database from 'better-sqlite3';
 import { DATA_DIR } from '../src/config.js';
 import { readEnvFile } from '../src/env.js';
 import { log } from '../src/log.js';
+import { pingCliAgent, type PingResult } from './lib/agent-ping.js';
+import { getLaunchdLabel, getSystemdUnit } from '../src/install-slug.js';
 import {
   getPlatform,
   getServiceManager,
@@ -29,19 +31,38 @@ export async function run(_args: string[]): Promise<void> {
 
   log.info('Starting verification');
 
-  // 1. Check service status
-  let service = 'not_found';
+  // 1. Check service status + detect checkout mismatch.
+  //
+  // Why the mismatch matters: the host binds `<DATA_DIR>/cli.sock` relative
+  // to the project root it was started from. If the running service is from
+  // a sibling checkout (common for developers with multiple clones), this
+  // repo's `data/cli.sock` won't exist — AGENT_PING would return a
+  // misleading `socket_error`. Surface the mismatch directly instead.
+  let service:
+    | 'not_found'
+    | 'stopped'
+    | 'running'
+    | 'running_other_checkout' = 'not_found';
+  let runningFromPath: string | null = null;
   const mgr = getServiceManager();
 
+  const launchdLabel = getLaunchdLabel(projectRoot);
+  const systemdUnit = getSystemdUnit(projectRoot);
+
   if (mgr === 'launchd') {
     try {
       const output = execSync('launchctl list', { encoding: 'utf-8' });
-      if (output.includes('com.nanoclaw')) {
-        // Check if it has a PID (actually running)
-        const line = output.split('\n').find((l) => l.includes('com.nanoclaw'));
-        if (line) {
-          const pidField = line.trim().split(/\s+/)[0];
-          service = pidField !== '-' && pidField ? 'running' : 'stopped';
+      const line = output.split('\n').find((l) => l.includes(launchdLabel));
+      if (line) {
+        const pidField = line.trim().split(/\s+/)[0];
+        if (pidField !== '-' && pidField) {
+          service = 'running';
+          const pid = Number(pidField);
+          if (Number.isInteger(pid) && pid > 0) {
+            runningFromPath = resolveBinaryScript(pid);
+          }
+        } else {
+          service = 'stopped';
         }
       }
     } catch {
@@ -50,14 +71,26 @@ export async function run(_args: string[]): Promise<void> {
   } else if (mgr === 'systemd') {
     const prefix = isRoot() ? 'systemctl' : 'systemctl --user';
     try {
-      execSync(`${prefix} is-active nanoclaw`, { stdio: 'ignore' });
+      execSync(`${prefix} is-active ${systemdUnit}`, { stdio: 'ignore' });
       service = 'running';
+      try {
+        const pidStr = execSync(
+          `${prefix} show ${systemdUnit} -p MainPID --value`,
+          { encoding: 'utf-8' },
+        ).trim();
+        const pid = Number(pidStr);
+        if (Number.isInteger(pid) && pid > 0) {
+          runningFromPath = resolveBinaryScript(pid);
+        }
+      } catch {
+        // couldn't read MainPID; leave runningFromPath null
+      }
     } catch {
       try {
         const output = execSync(`${prefix} list-unit-files`, {
           encoding: 'utf-8',
         });
-        if (output.includes('nanoclaw')) {
+        if (output.includes(systemdUnit)) {
           service = 'stopped';
         }
       } catch {
@@ -74,26 +107,31 @@ export async function run(_args: string[]): Promise<void> {
         if (raw && Number.isInteger(pid) && pid > 0) {
           process.kill(pid, 0);
           service = 'running';
+          runningFromPath = resolveBinaryScript(pid);
         }
       } catch {
         service = 'stopped';
       }
     }
   }
-  log.info('Service status', { service });
+
+  if (
+    service === 'running' &&
+    runningFromPath &&
+    !isPathInside(runningFromPath, projectRoot)
+  ) {
+    service = 'running_other_checkout';
+  }
+
+  log.info('Service status', { service, runningFromPath });
 
   // 2. Check container runtime
   let containerRuntime = 'none';
   try {
-    execSync('command -v container', { stdio: 'ignore' });
-    containerRuntime = 'apple-container';
+    execSync('docker info', { stdio: 'ignore' });
+    containerRuntime = 'docker';
   } catch {
-    try {
-      execSync('docker info', { stdio: 'ignore' });
-      containerRuntime = 'docker';
-    } catch {
-      // No runtime
-    }
+    // Docker not running
   }
 
   // 3. Check credentials
@@ -180,14 +218,24 @@ export async function run(_args: string[]): Promise<void> {
     mountAllowlist = 'configured';
   }
 
-  // Determine overall status
-  const status =
-    service === 'running' &&
-    credentials !== 'missing' &&
-    anyChannelConfigured &&
-    registeredGroups > 0
-      ? 'success'
-      : 'failed';
+  // 7. End-to-end: ping the CLI agent and confirm it replies. Only run if
+  // everything upstream looks healthy, since a broken socket would just hang.
+  let agentPing: 'ok' | 'no_reply' | 'socket_error' | 'auth_error' | 'skipped' = 'skipped';
+  if (service === 'running' && registeredGroups > 0) {
+    log.info('Pinging CLI agent');
+    agentPing = await pingCliAgent();
+    log.info('Agent ping result', { agentPing });
+  }
+
+  // Determine overall status. A CLI-only install is valid when the local
+  // agent round-trip succeeds; messaging app credentials are optional.
+  const status = determineVerifyStatus({
+    service,
+    credentials,
+    anyChannelConfigured,
+    registeredGroups,
+    agentPing,
+  });
 
   log.info('Verification complete', { status, channelAuth });
 
@@ -199,9 +247,55 @@ export async function run(_args: string[]): Promise<void> {
     CHANNEL_AUTH: JSON.stringify(channelAuth),
     REGISTERED_GROUPS: registeredGroups,
     MOUNT_ALLOWLIST: mountAllowlist,
+    AGENT_PING: agentPing,
     STATUS: status,
     LOG: 'logs/setup.log',
   });
 
   if (status === 'failed') process.exit(1);
 }
+
+export function determineVerifyStatus(input: {
+  service: 'not_found' | 'stopped' | 'running' | 'running_other_checkout';
+  credentials: string;
+  anyChannelConfigured: boolean;
+  registeredGroups: number;
+  agentPing: PingResult | 'skipped';
+}): 'success' | 'failed' {
+  const cliAgentResponds = input.agentPing === 'ok';
+  const hasUsableChannel = input.anyChannelConfigured || cliAgentResponds;
+
+  return input.service === 'running' &&
+    input.credentials !== 'missing' &&
+    hasUsableChannel &&
+    input.registeredGroups > 0 &&
+    (cliAgentResponds || input.agentPing === 'skipped')
+    ? 'success'
+    : 'failed';
+}
+
+/**
+ * Given a PID, resolve the script path the process is executing (i.e. the
+ * first `.js` / `.ts` / `.mjs` arg after `node`). Returns null on any
+ * error — callers should treat null as "couldn't tell" and skip the
+ * mismatch check rather than flag a false positive.
+ */
+function resolveBinaryScript(pid: number): string | null {
+  try {
+    // BSD ps (macOS) and util-linux both honour `-o command=` (full argv,
+    // no header). Node argv: "node /path/to/dist/index.js ...".
+    const out = execSync(`ps -p ${pid} -o command=`, {
+      encoding: 'utf-8',
+    }).trim();
+    const tokens = out.split(/\s+/);
+    const script = tokens.find((t) => /\.(js|mjs|cjs|ts)$/.test(t));
+    return script ?? null;
+  } catch {
+    return null;
+  }
+}
+
+function isPathInside(candidate: string, parent: string): boolean {
+  const rel = path.relative(parent, candidate);
+  return rel !== '' && !rel.startsWith('..') && !path.isAbsolute(rel);
+}
diff --git a/setup/whatsapp-auth.ts b/setup/whatsapp-auth.ts
new file mode 100644
index 0000000..47bfc6e
--- /dev/null
+++ b/setup/whatsapp-auth.ts
@@ -0,0 +1,221 @@
+/**
+ * Step: whatsapp-auth — standalone WhatsApp (Baileys) authentication.
+ *
+ * Forked from the channels-branch version so setup:auto's driver can render
+ * the terminal UX itself (inside clack) instead of the step dumping a raw QR
+ * to stdout. The browser method has been dropped — one less moving part and
+ * it kept biting headless/SSH users.
+ *
+ * Methods:
+ *   --method qr (default)          Emit each rotating QR as a status block
+ *                                  with the raw QR string. Driver renders.
+ *   --method pairing-code --phone  Request a pairing code. Emitted in a
+ *                                  status block once the Baileys call returns.
+ *
+ * Block schema (parent parses these):
+ *   WHATSAPP_AUTH_QR             { QR: "<raw>" }              — repeats
+ *   WHATSAPP_AUTH_PAIRING_CODE   { CODE: "XXXX-XXXX" }        — one-shot
+ *   WHATSAPP_AUTH                { STATUS: success }          — terminal
+ *                                { STATUS: skipped, AUTH_DIR, REASON }
+ *                                { STATUS: failed, ERROR: <reason> }
+ *
+ * STATUS values are kept in the runner's vocabulary (success/skipped/failed)
+ * so `spawnStep` recognises them and sets `ok` correctly; WhatsApp-specific
+ * UI text (e.g. "WhatsApp linked") lives in the driver's block handler.
+ *
+ * On success, credentials land in store/auth/ and the process exits 0.
+ */
+import fs from 'fs';
+import path from 'path';
+import { createRequire } from 'module';
+// Named import (not default) — pino's d.ts under NodeNext resolves the
+// default export to `typeof pino` (namespace), which isn't callable. The
+// named `pino` export resolves to the callable function.
+import { pino } from 'pino';
+
+import {
+  makeWASocket,
+  Browsers,
+  DisconnectReason,
+  fetchLatestWaWebVersion,
+  makeCacheableSignalKeyStore,
+  useMultiFileAuthState,
+} from '@whiskeysockets/baileys';
+import { emitStatus } from './status.js';
+
+const AUTH_DIR = path.join(process.cwd(), 'store', 'auth');
+const PAIRING_CODE_FILE = path.join(process.cwd(), 'store', 'pairing-code.txt');
+const baileysLogger = pino({ level: 'silent' });
+
+// Baileys v6 bug: getPlatformId sends charCode (49) instead of enum value (1).
+// Fixed in Baileys 7.x but not backported. Without this patch pairing codes
+// fail with "couldn't link device" because WhatsApp receives an invalid
+// platform id. createRequire because proto is not a named ESM export.
+const _require = createRequire(import.meta.url);
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+const { proto } = _require('@whiskeysockets/baileys') as { proto: any };
+try {
+  const _generics = _require(
+    '@whiskeysockets/baileys/lib/Utils/generics',
+  ) as Record<string, unknown>;
+  _generics.getPlatformId = (browser: string): string => {
+    const platformType =
+      proto.DeviceProps.PlatformType[
+        browser.toUpperCase() as keyof typeof proto.DeviceProps.PlatformType
+      ];
+    return platformType ? platformType.toString() : '1';
+  };
+} catch {
+  // If CJS require fails, QR auth still works; only pairing code may be affected.
+}
+
+type AuthMethod = 'qr' | 'pairing-code';
+
+function parseArgs(args: string[]): { method: AuthMethod; phone?: string } {
+  let method: AuthMethod = 'qr';
+  let phone: string | undefined;
+
+  for (let i = 0; i < args.length; i++) {
+    switch (args[i]) {
+      case '--method': {
+        const raw = args[++i];
+        if (raw === 'qr' || raw === 'pairing-code') {
+          method = raw;
+        } else {
+          console.error(`Unknown --method: ${raw} (expected 'qr' or 'pairing-code')`);
+          process.exit(1);
+        }
+        break;
+      }
+      case '--phone':
+        phone = args[++i];
+        break;
+    }
+  }
+
+  if (method === 'pairing-code' && !phone) {
+    console.error('--phone is required for pairing-code method');
+    process.exit(1);
+  }
+
+  return { method, phone };
+}
+
+export async function run(args: string[]): Promise<void> {
+  const { method, phone } = parseArgs(args);
+
+  if (fs.existsSync(path.join(AUTH_DIR, 'creds.json'))) {
+    emitStatus('WHATSAPP_AUTH', {
+      STATUS: 'skipped',
+      REASON: 'already-authenticated',
+      AUTH_DIR,
+    });
+    return;
+  }
+
+  fs.mkdirSync(AUTH_DIR, { recursive: true });
+
+  return new Promise<void>((resolve) => {
+    const timeout = setTimeout(() => {
+      emitStatus('WHATSAPP_AUTH', { STATUS: 'failed', ERROR: 'timeout' });
+      process.exit(1);
+    }, 120_000);
+
+    let succeeded = false;
+    function succeed(): void {
+      if (succeeded) return;
+      succeeded = true;
+      clearTimeout(timeout);
+      try {
+        if (fs.existsSync(PAIRING_CODE_FILE)) fs.unlinkSync(PAIRING_CODE_FILE);
+      } catch {
+        // ignore — the pairing code file is best-effort cleanup
+      }
+      emitStatus('WHATSAPP_AUTH', { STATUS: 'success' });
+      resolve();
+      // Give a moment for creds to flush before exiting.
+      setTimeout(() => process.exit(0), 1000);
+    }
+
+    async function connectSocket(isReconnect = false): Promise<void> {
+      const { state, saveCreds } = await useMultiFileAuthState(AUTH_DIR);
+      const { version } = await fetchLatestWaWebVersion({}).catch(() => ({
+        version: undefined,
+      }));
+
+      const sock = makeWASocket({
+        version,
+        auth: {
+          creds: state.creds,
+          keys: makeCacheableSignalKeyStore(state.keys, baileysLogger),
+        },
+        printQRInTerminal: false,
+        logger: baileysLogger,
+        browser: Browsers.macOS('Chrome'),
+      });
+
+      // Request pairing code only on first connect (not reconnect after 515).
+      if (
+        !isReconnect &&
+        method === 'pairing-code' &&
+        phone &&
+        !state.creds.registered
+      ) {
+        setTimeout(async () => {
+          try {
+            const code = await sock.requestPairingCode(phone);
+            fs.writeFileSync(PAIRING_CODE_FILE, code, 'utf-8');
+            emitStatus('WHATSAPP_AUTH_PAIRING_CODE', { CODE: code });
+          } catch (err) {
+            const message = err instanceof Error ? err.message : String(err);
+            emitStatus('WHATSAPP_AUTH', { STATUS: 'failed', ERROR: message });
+            process.exit(1);
+          }
+        }, 3000);
+      }
+
+      sock.ev.on('connection.update', (update) => {
+        const { connection, lastDisconnect, qr } = update;
+
+        // QR method: emit each rotation as a block. Parent renders.
+        if (qr && method === 'qr') {
+          emitStatus('WHATSAPP_AUTH_QR', { QR: qr });
+        }
+
+        if (connection === 'open') {
+          succeed();
+          sock.end(undefined);
+        }
+
+        if (connection === 'close') {
+          const reason = (
+            lastDisconnect?.error as { output?: { statusCode?: number } }
+          )?.output?.statusCode;
+          if (reason === DisconnectReason.loggedOut) {
+            clearTimeout(timeout);
+            emitStatus('WHATSAPP_AUTH', {
+              STATUS: 'failed',
+              ERROR: 'logged_out',
+            });
+            process.exit(1);
+          } else if (reason === DisconnectReason.timedOut) {
+            clearTimeout(timeout);
+            emitStatus('WHATSAPP_AUTH', {
+              STATUS: 'failed',
+              ERROR: 'qr_timeout',
+            });
+            process.exit(1);
+          } else if (reason === 515) {
+            // 515 = stream error after pairing succeeds but before registration
+            // completes. Reconnect to finish the handshake.
+            connectSocket(true);
+          }
+        }
+      });
+
+      sock.ev.on('creds.update', saveCreds);
+    }
+
+    connectSocket();
+  });
+}
diff --git a/src/channels/adapter.ts b/src/channels/adapter.ts
index 55efde1..82247a1 100644
--- a/src/channels/adapter.ts
+++ b/src/channels/adapter.ts
@@ -5,23 +5,19 @@
  * Two patterns: native adapters (implement directly) or Chat SDK bridge (wrap a Chat SDK adapter).
  */
 
-/** Configuration for a registered conversation (messaging group + agent wiring). */
-export interface ConversationConfig {
-  platformId: string;
-  agentGroupId: string;
-  triggerPattern?: string; // regex string (for native channels)
-  requiresTrigger: boolean;
-  sessionMode: 'shared' | 'per-thread' | 'agent-shared';
-}
-
 /** Passed to the adapter at setup time. */
 export interface ChannelSetup {
-  /** Known conversations from central DB. */
-  conversations: ConversationConfig[];
-
   /** Called when an inbound message arrives from the platform. */
   onInbound(platformId: string, threadId: string | null, message: InboundMessage): void | Promise<void>;
 
+  /**
+   * Called by admin-transport adapters (CLI) that want to route a message to
+   * an arbitrary channel/platform and optionally redirect replies elsewhere.
+   * Regular chat adapters should use `onInbound`; `onInboundEvent` skips the
+   * adapter-channel-type injection so the caller can target any wired mg.
+   */
+  onInboundEvent(event: InboundEvent): void | Promise<void>;
+
   /** Called when the adapter discovers metadata about a conversation. */
   onMetadata(platformId: string, name?: string, isGroup?: boolean): void;
 
@@ -29,12 +25,66 @@ export interface ChannelSetup {
   onAction(questionId: string, selectedOption: string, userId: string): void;
 }
 
+/** Delivery address used for reply-to overrides and (normally) the inbound's own origin. */
+export interface DeliveryAddress {
+  channelType: string;
+  platformId: string;
+  threadId: string | null;
+}
+
+/**
+ * Full inbound event handed to the router.
+ *
+ * `channelType` + `platformId` + `threadId` identify which messaging group /
+ * session receives the message. `replyTo`, when set, overrides where the
+ * agent's reply is delivered — used by the CLI admin transport when the
+ * operator wants a message routed to one channel but replies echoed back to
+ * their terminal. Agents cannot set `replyTo`; it is a router-layer concept
+ * set only by external adapters carrying operator intent.
+ */
+export interface InboundEvent {
+  channelType: string;
+  platformId: string;
+  threadId: string | null;
+  message: {
+    id: string;
+    kind: 'chat' | 'chat-sdk';
+    content: string; // JSON blob
+    timestamp: string;
+    /**
+     * Platform-confirmed bot-mention signal forwarded from the adapter.
+     * See InboundMessage.isMention for the full explanation.
+     */
+    isMention?: boolean;
+    /** True when the source is a group/channel thread, false for DMs. */
+    isGroup?: boolean;
+  };
+  replyTo?: DeliveryAddress;
+}
+
 /** Inbound message from adapter to host. */
 export interface InboundMessage {
   id: string;
   kind: 'chat' | 'chat-sdk';
   content: unknown; // JS object — host will JSON.stringify before writing to session DB
   timestamp: string;
+  /**
+   * Platform-confirmed signal that this message is a mention of the bot.
+   *
+   * Set by adapters that know the platform's own mention semantics — e.g.
+   * the Chat SDK bridge sets it true from `onNewMention` / `onDirectMessage`
+   * and forwards `message.isMention` from `onSubscribedMessage`. Use this
+   * in the router instead of agent-name regex matching, which breaks on
+   * platforms where the mention text is the bot's platform username (e.g.
+   * Telegram's `@nanoclaw_v2_refactr_1_bot`) rather than the agent_group
+   * display name (e.g. `@Andy`).
+   *
+   * Adapters that don't set it (native / legacy) leave it undefined — the
+   * router falls back to text-match against agent_group_name.
+   */
+  isMention?: boolean;
+  /** True when the source is a group/channel thread, false for DMs. */
+  isGroup?: boolean;
 }
 
 /** A file attachment to deliver alongside a message. */
@@ -85,7 +135,17 @@ export interface ChannelAdapter {
   // Optional
   setTyping?(platformId: string, threadId: string | null): Promise<void>;
   syncConversations?(): Promise<ConversationInfo[]>;
-  updateConversations?(conversations: ConversationConfig[]): void;
+
+  /**
+   * Subscribe the bot to a thread so follow-up messages route via the
+   * platform's "subscribed message" path (onSubscribedMessage in Chat SDK).
+   * Called by the router when a mention-sticky wiring first engages in a
+   * thread. Idempotent: calling twice on the same thread is a no-op.
+   *
+   * Platforms without a subscription concept can omit this; the router
+   * treats absence as a no-op.
+   */
+  subscribe?(platformId: string, threadId: string): Promise<void>;
 
   /**
    * Open (or fetch) a DM with this user, returning the platform_id of the
diff --git a/src/channels/channel-registry.test.ts b/src/channels/channel-registry.test.ts
index 0abbf9d..27ee660 100644
--- a/src/channels/channel-registry.test.ts
+++ b/src/channels/channel-registry.test.ts
@@ -10,7 +10,6 @@ import type { ChannelAdapter, ChannelSetup, InboundMessage, OutboundMessage } fr
 // Mock container runner
 vi.mock('../container-runner.js', () => ({
   wakeContainer: vi.fn().mockResolvedValue(undefined),
-  resetContainerIdleTimer: vi.fn(),
   isContainerRunning: vi.fn().mockReturnValue(false),
   getActiveContainerCount: vi.fn().mockReturnValue(0),
   killContainer: vi.fn(),
@@ -65,8 +64,6 @@ function createMockAdapter(
     },
 
     async setTyping() {},
-
-    updateConversations() {},
   };
 }
 
@@ -108,6 +105,7 @@ describe('channel registry', () => {
     await initChannelAdapters(() => ({
       conversations: [],
       onInbound: () => {},
+      onInboundEvent: () => {},
       onMetadata: () => {},
       onAction: () => {},
     }));
@@ -149,8 +147,10 @@ describe('channel + router integration', () => {
       id: 'mga-1',
       messaging_group_id: 'mg-1',
       agent_group_id: 'ag-1',
-      trigger_rules: null,
-      response_scope: 'all',
+      engage_mode: 'pattern',
+      engage_pattern: '.',
+      sender_scope: 'all',
+      ignored_message_policy: 'drop',
       session_mode: 'shared',
       priority: 0,
       created_at: now(),
@@ -209,6 +209,7 @@ describe('channel + router integration', () => {
     await initChannelAdapters(() => ({
       conversations: [],
       onInbound: () => {},
+      onInboundEvent: () => {},
       onMetadata: () => {},
       onAction: () => {},
     }));
diff --git a/src/channels/chat-sdk-bridge.test.ts b/src/channels/chat-sdk-bridge.test.ts
index e71ccb2..7e3c4ff 100644
--- a/src/channels/chat-sdk-bridge.test.ts
+++ b/src/channels/chat-sdk-bridge.test.ts
@@ -2,13 +2,47 @@ import { describe, expect, it } from 'vitest';
 
 import type { Adapter } from 'chat';
 
-import { createChatSdkBridge } from './chat-sdk-bridge.js';
+import { createChatSdkBridge, splitForLimit } from './chat-sdk-bridge.js';
 
 function stubAdapter(partial: Partial<Adapter>): Adapter {
   return { name: 'stub', ...partial } as unknown as Adapter;
 }
 
+describe('splitForLimit', () => {
+  it('returns a single chunk when text fits', () => {
+    expect(splitForLimit('short text', 100)).toEqual(['short text']);
+  });
+
+  it('splits on paragraph boundaries when available', () => {
+    const text = 'para one line one\npara one line two\n\npara two line one\npara two line two';
+    const chunks = splitForLimit(text, 40);
+    expect(chunks.length).toBeGreaterThan(1);
+    for (const c of chunks) expect(c.length).toBeLessThanOrEqual(40);
+  });
+
+  it('falls back to line boundaries when no paragraph fits', () => {
+    const text = 'alpha\nbravo\ncharlie\ndelta\necho\nfoxtrot';
+    const chunks = splitForLimit(text, 15);
+    expect(chunks.length).toBeGreaterThan(1);
+    for (const c of chunks) expect(c.length).toBeLessThanOrEqual(15);
+  });
+
+  it('hard-cuts when no whitespace is available', () => {
+    const text = 'a'.repeat(100);
+    const chunks = splitForLimit(text, 30);
+    expect(chunks.length).toBe(Math.ceil(100 / 30));
+    for (const c of chunks) expect(c.length).toBeLessThanOrEqual(30);
+    expect(chunks.join('')).toBe(text);
+  });
+});
+
 describe('createChatSdkBridge', () => {
+  // The bridge is now transport-only: forward inbound events, relay outbound
+  // ops. All per-wiring engage / accumulate / drop / subscribe decisions live
+  // in the router (src/router.ts routeInbound / evaluateEngage) and are
+  // exercised by host-core.test.ts end-to-end. These tests only cover the
+  // bridge's narrow, platform-adjacent surface.
+
   it('omits openDM when the underlying Chat SDK adapter has none', () => {
     const bridge = createChatSdkBridge({
       adapter: stubAdapter({}),
@@ -35,4 +69,12 @@ describe('createChatSdkBridge', () => {
     expect(openDMCalls).toEqual(['user-42']);
     expect(platformId).toBe('stub:user-42');
   });
+
+  it('exposes subscribe (lets the router initiate thread subscription on mention-sticky engage)', () => {
+    const bridge = createChatSdkBridge({
+      adapter: stubAdapter({}),
+      supportsThreads: true,
+    });
+    expect(typeof bridge.subscribe).toBe('function');
+  });
 });
diff --git a/src/channels/chat-sdk-bridge.ts b/src/channels/chat-sdk-bridge.ts
index 30ba0e8..18ab2cb 100644
--- a/src/channels/chat-sdk-bridge.ts
+++ b/src/channels/chat-sdk-bridge.ts
@@ -21,7 +21,7 @@ import { SqliteStateAdapter } from '../state-sqlite.js';
 import { registerWebhookAdapter } from '../webhook-server.js';
 import { getAskQuestionRender } from '../db/sessions.js';
 import { normalizeOptions, type NormalizedOption } from './ask-question.js';
-import type { ChannelAdapter, ChannelSetup, ConversationConfig, InboundMessage } from './adapter.js';
+import type { ChannelAdapter, ChannelSetup, InboundMessage } from './adapter.js';
 
 /** Adapter with optional gateway support (e.g., Discord). */
 interface GatewayAdapter extends Adapter {
@@ -63,6 +63,58 @@ export interface ChatSdkBridgeConfig {
    * quirk (e.g. Telegram's legacy Markdown parse mode).
    */
   transformOutboundText?: (text: string) => string;
+  /**
+   * Maximum text length the underlying adapter accepts in a single message.
+   * When set, the bridge splits outbound text longer than this on paragraph
+   * → line → hard-char boundaries and posts multiple messages. Without this,
+   * adapters like Discord (2000) and Telegram (4096) silently truncate
+   * mid-response. The returned id is the first chunk's id so subsequent edits
+   * and reactions still target the head of the reply.
+   */
+  maxTextLength?: number;
+}
+
+/**
+ * Split `text` into chunks no larger than `limit`, preferring paragraph
+ * breaks, then line breaks, then a hard character cut as a last resort.
+ * Preserves code fences only structurally — a fenced block that straddles a
+ * chunk boundary will render as two independent blocks on the receiving
+ * platform, which is the same behavior as manually re-opening a fence.
+ */
+/**
+ * Decode the actual option value from a button callback. Buttons are encoded
+ * with an integer index (to keep under Telegram's 64-byte callback_data cap),
+ * and the real value is looked up via `getAskQuestionRender(questionId)`.
+ * Falls back to treating the tail as a literal value so old in-flight cards
+ * (encoded before this shortening landed) still resolve.
+ */
+function resolveSelectedOption(
+  render: { options: NormalizedOption[] } | undefined,
+  eventValue: string | undefined,
+  tail: string | undefined,
+): string {
+  const candidate = eventValue ?? tail ?? '';
+  if (render && /^\d+$/.test(candidate)) {
+    const idx = Number(candidate);
+    if (render.options[idx]) return render.options[idx].value;
+  }
+  return candidate;
+}
+
+export function splitForLimit(text: string, limit: number): string[] {
+  if (text.length <= limit) return [text];
+  const chunks: string[] = [];
+  let remaining = text;
+  while (remaining.length > limit) {
+    let cut = remaining.lastIndexOf('\n\n', limit);
+    if (cut <= 0) cut = remaining.lastIndexOf('\n', limit);
+    if (cut <= 0) cut = remaining.lastIndexOf(' ', limit);
+    if (cut <= 0) cut = limit;
+    chunks.push(remaining.slice(0, cut).trimEnd());
+    remaining = remaining.slice(cut).trimStart();
+  }
+  if (remaining.length > 0) chunks.push(remaining);
+  return chunks;
 }
 
 export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter {
@@ -71,18 +123,13 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
   let chat: Chat;
   let state: SqliteStateAdapter;
   let setupConfig: ChannelSetup;
-  let conversations: Map<string, ConversationConfig>;
   let gatewayAbort: AbortController | null = null;
 
-  function buildConversationMap(configs: ConversationConfig[]): Map<string, ConversationConfig> {
-    const map = new Map<string, ConversationConfig>();
-    for (const conv of configs) {
-      map.set(conv.platformId, conv);
-    }
-    return map;
-  }
-
-  async function messageToInbound(message: ChatMessage): Promise<InboundMessage> {
+  async function messageToInbound(
+    message: ChatMessage,
+    isMention: boolean,
+    isGroup?: boolean,
+  ): Promise<InboundMessage> {
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     const serialized = message.toJSON() as Record<string, any>;
 
@@ -138,6 +185,8 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
       kind: 'chat-sdk',
       content: serialized,
       timestamp: message.metadata.dateSent.toISOString(),
+      isMention,
+      isGroup,
     };
   }
 
@@ -148,7 +197,6 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
 
     async setup(hostConfig: ChannelSetup) {
       setupConfig = hostConfig;
-      conversations = buildConversationMap(hostConfig.conversations);
 
       state = new SqliteStateAdapter();
 
@@ -160,23 +208,35 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
         logger: 'silent',
       });
 
-      // Subscribed threads — forward all messages
+      // Four SDK dispatch paths — bridge just forwards. All per-wiring
+      // engage / accumulate / drop / subscribe decisions live in the host
+      // router (src/router.ts routeInbound / evaluateEngage). The bridge
+      // only resolves channel ids and sets the platform-confirmed isMention
+      // flag that routeInbound evaluates; the router calls back into
+      // bridge.subscribe(...) when a mention-sticky wiring engages.
+
+      // Subscribed threads — every message in a thread we've previously
+      // engaged. Carry the SDK's `message.isMention` through so mention-mode
+      // wirings still fire on in-thread mentions.
       chat.onSubscribedMessage(async (thread, message) => {
         const channelId = adapter.channelIdFromThreadId(thread.id);
-        await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message));
+        await setupConfig.onInbound(
+          channelId,
+          thread.id,
+          await messageToInbound(message, message.isMention === true, true),
+        );
       });
 
-      // @mention in unsubscribed thread — forward + subscribe
+      // @mention in an unsubscribed thread — SDK-confirmed bot mention.
       chat.onNewMention(async (thread, message) => {
         const channelId = adapter.channelIdFromThreadId(thread.id);
-        await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message));
-        await thread.subscribe();
+        await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message, true, true));
       });
 
-      // DMs — always forward + subscribe. Pass thread.id so sub-thread
-      // context carries through to delivery (Slack users can open threads
-      // inside a DM). The router collapses DM sub-threads to one session
-      // (is_group=0 short-circuits the per-thread escalation).
+      // DMs — by definition addressed to the bot. Thread id flows through
+      // so sub-thread context reaches delivery (Slack users can open threads
+      // inside a DM). Router collapses DM sub-threads to one session via
+      // is_group=0 short-circuit.
       chat.onDirectMessage(async (thread, message) => {
         const channelId = adapter.channelIdFromThreadId(thread.id);
         log.info('Inbound DM received', {
@@ -185,8 +245,22 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
           sender: (message.author as any)?.fullName ?? (message.author as any)?.userId ?? 'unknown',
           threadId: thread.id,
         });
-        await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message));
-        await thread.subscribe();
+        await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message, true, false));
+      });
+
+      // Plain messages in unsubscribed threads.
+      //
+      // Chat SDK dispatch (handling-events.mdx §"Handler dispatch order") is
+      // exclusive: subscribed → onSubscribedMessage; unsubscribed+mention →
+      // onNewMention; unsubscribed+pattern-match → onNewMessage. Registering
+      // with `/./` lets the router see every plain message on every
+      // unsubscribed thread the bot can see. The router short-circuits via
+      // getMessagingGroupWithAgentCount (~1 DB read) for unwired channels,
+      // so forwarding every one is cheap enough to not need a bridge-side
+      // flood gate.
+      chat.onNewMessage(/./, async (thread, message) => {
+        const channelId = adapter.channelIdFromThreadId(thread.id);
+        await setupConfig.onInbound(channelId, thread.id, await messageToInbound(message, false, true));
       });
 
       // Handle button clicks (ask_user_question)
@@ -195,11 +269,15 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
         const parts = event.actionId.split(':');
         if (parts.length < 3) return;
         const questionId = parts[1];
-        const selectedOption = event.value || '';
+        const tail = parts.slice(2).join(':');
         const userId = event.user?.userId || '';
 
         // Resolve render metadata BEFORE dispatching onAction (which deletes the row).
         const render = getAskQuestionRender(questionId);
+        // New format: button id/value is an integer index into options (kept
+        // short to fit Telegram's 64-byte callback_data cap). Old format:
+        // the full value is embedded in actionId/value directly.
+        const selectedOption = resolveSelectedOption(render, event.value, tail);
         const title = render?.title ?? '❓ Question';
         const matched = render?.options.find((o) => o.value === selectedOption);
         const selectedLabel = matched?.selectedLabel ?? selectedOption ?? '(clicked)';
@@ -303,8 +381,13 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
           children: [
             CardText(question),
             Actions(
-              options.map((opt) =>
-                Button({ id: `ncq:${questionId}:${opt.value}`, label: opt.label, value: opt.value }),
+              // Encode button id/value with the option index rather than the
+              // full value. Telegram caps callback_data at 64 bytes, and
+              // long values (e.g. ISO datetimes, URLs) push the JSON payload
+              // well past that. The onAction handlers resolve the index back
+              // to the real value via getAskQuestionRender(questionId).
+              options.map((opt, idx) =>
+                Button({ id: `ncq:${questionId}:${idx}`, label: opt.label, value: String(idx) }),
               ),
             ),
           ],
@@ -325,13 +408,23 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
           data: f.data,
           filename: f.filename,
         }));
-        if (fileUploads && fileUploads.length > 0) {
-          const result = await adapter.postMessage(tid, { markdown: text, files: fileUploads });
-          return result?.id;
-        } else {
-          const result = await adapter.postMessage(tid, { markdown: text });
-          return result?.id;
+        // Split if over the adapter's max length. Files ride on the first
+        // chunk so the head of the reply still carries them.
+        const chunks =
+          config.maxTextLength && text.length > config.maxTextLength
+            ? splitForLimit(text, config.maxTextLength)
+            : [text];
+        let firstId: string | undefined;
+        for (let i = 0; i < chunks.length; i++) {
+          const chunk = chunks[i];
+          const attachFiles = i === 0 && fileUploads && fileUploads.length > 0;
+          const result = await adapter.postMessage(
+            tid,
+            attachFiles ? { markdown: chunk, files: fileUploads } : { markdown: chunk },
+          );
+          if (i === 0) firstId = result?.id;
         }
+        return firstId;
       } else if (message.files && message.files.length > 0) {
         // Files only, no text
         const fileUploads = message.files.map((f: { data: Buffer; filename: string }) => ({
@@ -358,8 +451,13 @@ export function createChatSdkBridge(config: ChatSdkBridgeConfig): ChannelAdapter
       return true;
     },
 
-    updateConversations(configs: ConversationConfig[]) {
-      conversations = buildConversationMap(configs);
+    async subscribe(_platformId: string, threadId: string) {
+      // Chat SDK's subscription state lives on the StateAdapter (not on the
+      // Chat instance itself). SqliteStateAdapter.subscribe is idempotent —
+      // a second call on an already-subscribed thread is a no-op. threadId
+      // is the SDK's thread id, which is what the router already has from
+      // the original inbound event.
+      await state.subscribe(threadId);
     },
   };
 
@@ -441,18 +539,21 @@ async function handleForwardedEvent(
     // type 3 = MessageComponent (button/select)
     if (interaction.type === 3) {
       const customId = (interaction.data as Record<string, unknown>)?.custom_id as string;
-      const user = (interaction.member as Record<string, unknown>)?.user as Record<string, string> | undefined;
+      // In guilds the clicker is at interaction.member.user; in DMs it's interaction.user directly.
+      const user =
+        ((interaction.member as Record<string, unknown>)?.user as Record<string, string> | undefined) ??
+        (interaction.user as Record<string, string> | undefined);
       const interactionId = interaction.id as string;
       const interactionToken = interaction.token as string;
 
       // Parse the selected option from custom_id
       let questionId: string | undefined;
-      let selectedOption: string | undefined;
+      let tail: string | undefined;
       if (customId?.startsWith('ncq:')) {
         const colonIdx = customId.indexOf(':', 4); // after "ncq:"
         if (colonIdx !== -1) {
           questionId = customId.slice(4, colonIdx);
-          selectedOption = customId.slice(colonIdx + 1);
+          tail = customId.slice(colonIdx + 1);
         }
       }
 
@@ -461,6 +562,9 @@ async function handleForwardedEvent(
         ((interaction.message as Record<string, unknown>)?.embeds as Array<Record<string, unknown>>) || [];
       const originalDescription = (originalEmbeds[0]?.description as string) || '';
       const render = questionId ? getAskQuestionRender(questionId) : undefined;
+      // Discord custom_id mirrors the new index-based encoding (see Button
+      // construction). Decode back to the real option value for downstream.
+      const selectedOption = resolveSelectedOption(render, tail, tail);
       const cardTitle = render?.title ?? ((originalEmbeds[0]?.title as string) || '❓ Question');
       const matchedOpt = render?.options.find((o) => o.value === selectedOption);
       const selectedLabel = matchedOpt?.selectedLabel ?? selectedOption ?? customId;
diff --git a/src/channels/cli.ts b/src/channels/cli.ts
index c84952c..ad78bea 100644
--- a/src/channels/cli.ts
+++ b/src/channels/cli.ts
@@ -7,19 +7,31 @@
  * the normal router/delivery path like any other adapter — `/clear` and
  * other session-level commands work identically.
  *
- * MVP shape:
- *   - One hardcoded messaging_group: `cli/local`. Wired to one agent via
- *     the setup flow (see `scripts/init-first-agent.ts`). Multi-agent
- *     support can add per-agent messaging_groups later without breaking
- *     the wire protocol.
- *   - Single connected client at a time. A second connection closes the
- *     first with a "superseded" notice.
- *   - Wire format: one JSON object per line.
- *       Client → server: { "text": "user message" }
- *       Server → client: { "text": "agent reply" }
- *   - deliver() silently no-ops when no client is connected. The outbound
- *     row is already in outbound.db, so the message isn't lost — it just
- *     doesn't reach this run's terminal. Reconnect to see subsequent replies.
+ * Wire format: one JSON object per line.
+ *
+ *   Client → server:
+ *     { "text": "user message" }                          # default — talk to cli/local
+ *     { "text": "...", "to": {"channelType": "discord",
+ *                             "platformId": "discord:@me:149...",
+ *                             "threadId": null} }         # route to a specific mg
+ *     { "text": "...", "to": {...}, "reply_to": {...} }   # + redirect replies
+ *   Server → client:
+ *     { "text": "agent reply" }
+ *
+ * The `to` and `reply_to` addressing is how admin transports (the bootstrap
+ * script) inject messages targeting any wired channel. `reply_to` is a
+ * router-layer concept — agents cannot set it; it is carried only on
+ * inbound events from CLI clients that hold operator privilege (the socket
+ * is chmod 0600, so "connected to this socket" ≈ "is the owner").
+ *
+ * Single-client chat semantics: one connected terminal at a time. A second
+ * "chat" connection closes the first with a "superseded" notice. Admin
+ * route-opcode connections (`to` set) are one-shot and do NOT evict an
+ * active chat client.
+ *
+ * deliver() silently no-ops when no client is connected. The outbound row
+ * is already in outbound.db, so the message isn't lost — it just doesn't
+ * reach this run's terminal. Reconnect to see subsequent replies.
  */
 import fs from 'fs';
 import net from 'net';
@@ -27,12 +39,7 @@ import path from 'path';
 
 import { DATA_DIR } from '../config.js';
 import { log } from '../log.js';
-import type {
-  ChannelAdapter,
-  ChannelSetup,
-  InboundMessage,
-  OutboundMessage,
-} from './adapter.js';
+import type { ChannelAdapter, ChannelSetup, DeliveryAddress, InboundEvent, OutboundMessage } from './adapter.js';
 import { registerChannelAdapter } from './channel-registry.js';
 
 const PLATFORM_ID = 'local';
@@ -129,16 +136,25 @@ function createAdapter(): ChannelAdapter {
   };
 
   function handleConnection(socket: net.Socket, config: ChannelSetup): void {
-    if (client) {
-      try {
-        client.write(JSON.stringify({ text: '[superseded by a newer client]' }) + '\n');
-        client.end();
-      } catch {
-        // swallow
+    // Defer the chat-slot swap until we see the first line — if it turns out
+    // to be a routed (`to`-bearing) one-shot, we leave the existing chat
+    // client in place. Only plain chat connections participate in supersede.
+    let claimedChatSlot = false;
+
+    const claimChatSlot = () => {
+      if (claimedChatSlot) return;
+      claimedChatSlot = true;
+      if (client && client !== socket) {
+        try {
+          client.write(JSON.stringify({ text: '[superseded by a newer client]' }) + '\n');
+          client.end();
+        } catch {
+          // swallow
+        }
       }
-    }
-    client = socket;
-    log.info('CLI client connected');
+      client = socket;
+      log.info('CLI client connected');
+    };
 
     let buffer = '';
     socket.on('data', (chunk) => {
@@ -148,13 +164,13 @@ function createAdapter(): ChannelAdapter {
         const line = buffer.slice(0, idx).trim();
         buffer = buffer.slice(idx + 1);
         if (!line) continue;
-        void handleLine(line, config);
+        void handleLine(line, config, claimChatSlot);
       }
     });
 
     socket.on('close', () => {
       if (client === socket) client = null;
-      log.info('CLI client disconnected');
+      if (claimedChatSlot) log.info('CLI client disconnected');
     });
 
     socket.on('error', (err) => {
@@ -162,8 +178,14 @@ function createAdapter(): ChannelAdapter {
     });
   }
 
-  async function handleLine(line: string, config: ChannelSetup): Promise<void> {
-    let payload: { text?: unknown };
+  async function handleLine(line: string, config: ChannelSetup, claimChatSlot: () => void): Promise<void> {
+    let payload: {
+      text?: unknown;
+      to?: unknown;
+      reply_to?: unknown;
+      sender?: unknown;
+      senderId?: unknown;
+    };
     try {
       payload = JSON.parse(line);
     } catch (err) {
@@ -172,23 +194,73 @@ function createAdapter(): ChannelAdapter {
     }
     if (typeof payload.text !== 'string' || payload.text.length === 0) return;
 
-    const inbound: InboundMessage = {
-      id: `cli-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
-      kind: 'chat',
-      timestamp: new Date().toISOString(),
-      content: {
-        text: payload.text,
-        sender: 'cli',
-        senderId: `cli:${PLATFORM_ID}`,
-      },
-    };
+    const to = parseAddress(payload.to);
+    const replyTo = parseAddress(payload.reply_to);
+
+    if (to) {
+      // Routed message — admin transport. Build a full InboundEvent targeting
+      // `to`'s channel/platform, and let `reply_to` (if any) redirect replies.
+      // Does NOT claim the chat slot, so an active terminal chat isn't evicted.
+      const event: InboundEvent = {
+        channelType: to.channelType,
+        platformId: to.platformId,
+        threadId: to.threadId,
+        message: {
+          id: `cli-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
+          kind: 'chat',
+          timestamp: new Date().toISOString(),
+          content: JSON.stringify({
+            text: payload.text,
+            sender: typeof payload.sender === 'string' ? payload.sender : 'cli',
+            senderId: typeof payload.senderId === 'string' ? payload.senderId : `cli:${PLATFORM_ID}`,
+          }),
+        },
+        replyTo: replyTo ?? undefined,
+      };
+      try {
+        await config.onInboundEvent(event);
+      } catch (err) {
+        log.error('CLI: onInboundEvent threw', { err });
+      }
+      return;
+    }
+
+    // Plain chat — claim the slot (evicting any prior client) and route via
+    // the standard onInbound path (adapter injects its own channelType).
+    claimChatSlot();
     try {
-      await config.onInbound(PLATFORM_ID, null, inbound);
+      await config.onInbound(PLATFORM_ID, null, {
+        id: `cli-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
+        kind: 'chat',
+        timestamp: new Date().toISOString(),
+        content: {
+          text: payload.text,
+          sender: 'cli',
+          senderId: `cli:${PLATFORM_ID}`,
+        },
+      });
     } catch (err) {
       log.error('CLI: onInbound threw', { err });
     }
   }
 
+  function parseAddress(raw: unknown): DeliveryAddress | null {
+    if (!raw || typeof raw !== 'object') return null;
+    const obj = raw as Record<string, unknown>;
+    if (typeof obj.channelType !== 'string' || typeof obj.platformId !== 'string') return null;
+    const threadId =
+      obj.threadId === null || obj.threadId === undefined
+        ? null
+        : typeof obj.threadId === 'string'
+          ? obj.threadId
+          : null;
+    return {
+      channelType: obj.channelType,
+      platformId: obj.platformId,
+      threadId,
+    };
+  }
+
   return adapter;
 }
 
diff --git a/src/claude-md-compose.ts b/src/claude-md-compose.ts
new file mode 100644
index 0000000..c0519e4
--- /dev/null
+++ b/src/claude-md-compose.ts
@@ -0,0 +1,205 @@
+/**
+ * CLAUDE.md composition for agent groups.
+ *
+ * Replaces the per-group "written once at init, owned by the group" pattern
+ * with a host-regenerated entry point that imports:
+ *   - a shared base (`container/CLAUDE.md` mounted RO at `/app/CLAUDE.md`)
+ *   - optional per-skill fragments (skills that ship `instructions.md`)
+ *   - optional per-MCP-server fragments (inline `instructions` field in
+ *     `container.json`)
+ *   - per-group agent memory (`CLAUDE.local.md`, auto-loaded by Claude Code)
+ *
+ * Runs on every spawn from `container-runner.buildMounts()`. Deterministic —
+ * same inputs produce the same CLAUDE.md, and stale fragments are pruned.
+ *
+ * See `docs/claude-md-composition.md` for the full design.
+ */
+import fs from 'fs';
+import path from 'path';
+
+import { GROUPS_DIR } from './config.js';
+import { readContainerConfig } from './container-config.js';
+import { log } from './log.js';
+import type { AgentGroup } from './types.js';
+
+// Symlink targets are container paths — dangling on host (hence the readlink
+// dance instead of existsSync), valid inside the container via RO mounts.
+const SHARED_CLAUDE_MD_CONTAINER_PATH = '/app/CLAUDE.md';
+const SHARED_SKILLS_CONTAINER_BASE = '/app/skills';
+const SHARED_MCP_TOOLS_CONTAINER_BASE = '/app/src/mcp-tools';
+
+// Host-side source paths used to discover fragment sources at compose time.
+// Resolved at call time (process.cwd() = project root) so tests can swap cwd.
+const MCP_TOOLS_HOST_SUBPATH = path.join('container', 'agent-runner', 'src', 'mcp-tools');
+
+const COMPOSED_HEADER = '<!-- Composed at spawn — do not edit. Edit CLAUDE.local.md for per-group content. -->';
+
+/**
+ * Regenerate `groups/<folder>/CLAUDE.md` from the shared base, enabled skill
+ * fragments, and MCP server fragments declared in `container.json`. Creates
+ * an empty `CLAUDE.local.md` if missing.
+ */
+export function composeGroupClaudeMd(group: AgentGroup): void {
+  const groupDir = path.resolve(GROUPS_DIR, group.folder);
+  if (!fs.existsSync(groupDir)) {
+    fs.mkdirSync(groupDir, { recursive: true });
+  }
+
+  const sharedLink = path.join(groupDir, '.claude-shared.md');
+  syncSymlink(sharedLink, SHARED_CLAUDE_MD_CONTAINER_PATH);
+
+  const fragmentsDir = path.join(groupDir, '.claude-fragments');
+  if (!fs.existsSync(fragmentsDir)) {
+    fs.mkdirSync(fragmentsDir, { recursive: true });
+  }
+
+  // Desired fragment set.
+  const config = readContainerConfig(group.folder);
+  const desired = new Map<string, { type: 'symlink' | 'inline'; content: string }>();
+
+  // Skill fragments — every skill that ships an `instructions.md`.
+  // TODO (shared-source refactor): respect `container.json` skill selection.
+  const skillsHostDir = path.join(process.cwd(), 'container', 'skills');
+  if (fs.existsSync(skillsHostDir)) {
+    for (const skillName of fs.readdirSync(skillsHostDir)) {
+      const hostFragment = path.join(skillsHostDir, skillName, 'instructions.md');
+      if (fs.existsSync(hostFragment)) {
+        desired.set(`skill-${skillName}.md`, {
+          type: 'symlink',
+          content: `${SHARED_SKILLS_CONTAINER_BASE}/${skillName}/instructions.md`,
+        });
+      }
+    }
+  }
+
+  // Built-in module fragments — every MCP tool source file that ships a
+  // sibling `<name>.instructions.md`. These describe how the agent should
+  // use that module's MCP tools (schedule_task, install_packages, etc.).
+  // Always included — these are built-in, not toggleable.
+  const mcpToolsHostDir = path.join(process.cwd(), MCP_TOOLS_HOST_SUBPATH);
+  if (fs.existsSync(mcpToolsHostDir)) {
+    for (const entry of fs.readdirSync(mcpToolsHostDir)) {
+      const match = entry.match(/^(.+)\.instructions\.md$/);
+      if (!match) continue;
+      const moduleName = match[1];
+      desired.set(`module-${moduleName}.md`, {
+        type: 'symlink',
+        content: `${SHARED_MCP_TOOLS_CONTAINER_BASE}/${entry}`,
+      });
+    }
+  }
+
+  // MCP server fragments — inline instructions from container.json for
+  // user-added external MCP servers.
+  for (const [name, mcp] of Object.entries(config.mcpServers)) {
+    if (mcp.instructions) {
+      desired.set(`mcp-${name}.md`, {
+        type: 'inline',
+        content: mcp.instructions,
+      });
+    }
+  }
+
+  // Reconcile: drop stale, write desired.
+  for (const existing of fs.readdirSync(fragmentsDir)) {
+    if (!desired.has(existing)) {
+      fs.unlinkSync(path.join(fragmentsDir, existing));
+    }
+  }
+  for (const [name, frag] of desired) {
+    const fragPath = path.join(fragmentsDir, name);
+    if (frag.type === 'symlink') {
+      syncSymlink(fragPath, frag.content);
+    } else {
+      writeAtomic(fragPath, frag.content);
+    }
+  }
+
+  // Composed entry — imports only.
+  const imports = ['@./.claude-shared.md'];
+  for (const name of [...desired.keys()].sort()) {
+    imports.push(`@./.claude-fragments/${name}`);
+  }
+  const body = [COMPOSED_HEADER, ...imports, ''].join('\n');
+  writeAtomic(path.join(groupDir, 'CLAUDE.md'), body);
+
+  const localFile = path.join(groupDir, 'CLAUDE.local.md');
+  if (!fs.existsSync(localFile)) {
+    fs.writeFileSync(localFile, '');
+  }
+}
+
+/**
+ * One-time cutover from the `groups/global/CLAUDE.md` + `.claude-global.md`
+ * pattern. Idempotent — safe to run on every host startup.
+ *
+ * For each group dir:
+ *   - remove `.claude-global.md` symlink if present
+ *   - rename `CLAUDE.md` → `CLAUDE.local.md` (only if `CLAUDE.local.md`
+ *     doesn't already exist — preserves pre-cutover content as per-group
+ *     memory; after the first spawn regenerates `CLAUDE.md`, this branch
+ *     is skipped because `CLAUDE.local.md` now exists)
+ *
+ * Globally:
+ *   - delete `groups/global/` (content already in `container/CLAUDE.md`)
+ */
+export function migrateGroupsToClaudeLocal(): void {
+  if (!fs.existsSync(GROUPS_DIR)) return;
+
+  const actions: string[] = [];
+
+  for (const entry of fs.readdirSync(GROUPS_DIR, { withFileTypes: true })) {
+    if (!entry.isDirectory()) continue;
+    if (entry.name === 'global') continue;
+
+    const groupDir = path.join(GROUPS_DIR, entry.name);
+
+    const oldGlobalLink = path.join(groupDir, '.claude-global.md');
+    try {
+      fs.lstatSync(oldGlobalLink);
+      fs.unlinkSync(oldGlobalLink);
+      actions.push(`${entry.name}/.claude-global.md removed`);
+    } catch {
+      /* already gone */
+    }
+
+    const claudeMd = path.join(groupDir, 'CLAUDE.md');
+    const claudeLocal = path.join(groupDir, 'CLAUDE.local.md');
+    if (fs.existsSync(claudeMd) && !fs.existsSync(claudeLocal)) {
+      fs.renameSync(claudeMd, claudeLocal);
+      actions.push(`${entry.name}/CLAUDE.md → CLAUDE.local.md`);
+    }
+  }
+
+  const globalDir = path.join(GROUPS_DIR, 'global');
+  if (fs.existsSync(globalDir)) {
+    fs.rmSync(globalDir, { recursive: true, force: true });
+    actions.push('groups/global/ removed');
+  }
+
+  if (actions.length > 0) {
+    log.info('Migrated groups to CLAUDE.local.md model', { actions });
+  }
+}
+
+function syncSymlink(linkPath: string, target: string): void {
+  let currentTarget: string | null = null;
+  try {
+    currentTarget = fs.readlinkSync(linkPath);
+  } catch {
+    /* missing */
+  }
+  if (currentTarget === target) return;
+  try {
+    fs.unlinkSync(linkPath);
+  } catch {
+    /* missing */
+  }
+  fs.symlinkSync(target, linkPath);
+}
+
+function writeAtomic(filePath: string, content: string): void {
+  const tmp = `${filePath}.tmp-${process.pid}`;
+  fs.writeFileSync(tmp, content);
+  fs.renameSync(tmp, filePath);
+}
diff --git a/src/command-gate.ts b/src/command-gate.ts
new file mode 100644
index 0000000..a0c1979
--- /dev/null
+++ b/src/command-gate.ts
@@ -0,0 +1,63 @@
+/**
+ * Host-side command gate. Classifies inbound slash commands and gates
+ * them before they reach the container.
+ *
+ * - Filtered commands: dropped silently (never reach the container)
+ * - Admin commands: checked against user_roles; denied senders get a
+ *   "Permission denied" response written directly to messages_out
+ * - Normal messages: pass through unchanged
+ */
+import { getDb, hasTable } from './db/connection.js';
+
+export type GateResult = { action: 'pass' } | { action: 'filter' } | { action: 'deny'; command: string };
+
+const FILTERED_COMMANDS = new Set(['/help', '/login', '/logout', '/doctor', '/config', '/remote-control']);
+const ADMIN_COMMANDS = new Set(['/clear', '/compact', '/context', '/cost', '/files']);
+
+/**
+ * Classify a message and decide whether it should reach the container.
+ * Returns 'pass' for normal messages and authorized admin commands,
+ * 'filter' for silently-dropped commands, 'deny' for unauthorized
+ * admin commands.
+ */
+export function gateCommand(content: string, userId: string | null, agentGroupId: string): GateResult {
+  let text: string;
+  try {
+    const parsed = JSON.parse(content);
+    text = (parsed.text || '').trim();
+  } catch {
+    text = content.trim();
+  }
+
+  if (!text.startsWith('/')) return { action: 'pass' };
+
+  const command = text.split(/\s/)[0].toLowerCase();
+
+  if (FILTERED_COMMANDS.has(command)) return { action: 'filter' };
+
+  if (ADMIN_COMMANDS.has(command)) {
+    if (isAdmin(userId, agentGroupId)) {
+      return { action: 'pass' };
+    }
+    return { action: 'deny', command };
+  }
+
+  // Unknown slash commands pass through (the agent/SDK handles them)
+  return { action: 'pass' };
+}
+
+function isAdmin(userId: string | null, agentGroupId: string): boolean {
+  if (!userId) return false;
+  if (!hasTable(getDb(), 'user_roles')) return true; // no permissions module = allow all
+  const db = getDb();
+  const row = db
+    .prepare(
+      `SELECT 1 FROM user_roles
+       WHERE user_id = ?
+         AND (role = 'owner' OR role = 'admin')
+         AND (agent_group_id IS NULL OR agent_group_id = ?)
+       LIMIT 1`,
+    )
+    .get(userId, agentGroupId);
+  return row != null;
+}
diff --git a/src/config.ts b/src/config.ts
index ef1ba9e..a82d4f5 100644
--- a/src/config.ts
+++ b/src/config.ts
@@ -2,16 +2,15 @@ import os from 'os';
 import path from 'path';
 
 import { readEnvFile } from './env.js';
+import { getContainerImageBase, getDefaultContainerImage, getInstallSlug } from './install-slug.js';
 import { isValidTimezone } from './timezone.js';
 
 // Read config values from .env (falls back to process.env).
-const envConfig = readEnvFile(['ASSISTANT_NAME', 'ASSISTANT_HAS_OWN_NUMBER', 'ONECLI_URL', 'TZ']);
+const envConfig = readEnvFile(['ASSISTANT_NAME', 'ASSISTANT_HAS_OWN_NUMBER', 'ONECLI_URL', 'ONECLI_API_KEY', 'TZ']);
 
 export const ASSISTANT_NAME = process.env.ASSISTANT_NAME || envConfig.ASSISTANT_NAME || 'Andy';
 export const ASSISTANT_HAS_OWN_NUMBER =
   (process.env.ASSISTANT_HAS_OWN_NUMBER || envConfig.ASSISTANT_HAS_OWN_NUMBER) === 'true';
-export const POLL_INTERVAL = 2000;
-export const SCHEDULER_POLL_INTERVAL = 60000;
 
 // Absolute paths needed for container mounts
 const PROJECT_ROOT = process.cwd();
@@ -24,12 +23,19 @@ export const STORE_DIR = path.resolve(PROJECT_ROOT, 'store');
 export const GROUPS_DIR = path.resolve(PROJECT_ROOT, 'groups');
 export const DATA_DIR = path.resolve(PROJECT_ROOT, 'data');
 
-export const CONTAINER_IMAGE = process.env.CONTAINER_IMAGE || 'nanoclaw-agent:latest';
+// Per-checkout image tag so two installs on the same host don't share
+// `nanoclaw-agent:latest` and clobber each other on rebuild.
+export const CONTAINER_IMAGE_BASE = process.env.CONTAINER_IMAGE_BASE || getContainerImageBase(PROJECT_ROOT);
+export const CONTAINER_IMAGE = process.env.CONTAINER_IMAGE || getDefaultContainerImage(PROJECT_ROOT);
+// Install slug — stamped onto every spawned container via --label so
+// cleanupOrphans only reaps containers from this install, not peers.
+export const INSTALL_SLUG = getInstallSlug(PROJECT_ROOT);
+export const CONTAINER_INSTALL_LABEL = `nanoclaw-install=${INSTALL_SLUG}`;
 export const CONTAINER_TIMEOUT = parseInt(process.env.CONTAINER_TIMEOUT || '1800000', 10);
 export const CONTAINER_MAX_OUTPUT_SIZE = parseInt(process.env.CONTAINER_MAX_OUTPUT_SIZE || '10485760', 10); // 10MB default
 export const ONECLI_URL = process.env.ONECLI_URL || envConfig.ONECLI_URL;
+export const ONECLI_API_KEY = process.env.ONECLI_API_KEY || envConfig.ONECLI_API_KEY;
 export const MAX_MESSAGES_PER_PROMPT = Math.max(1, parseInt(process.env.MAX_MESSAGES_PER_PROMPT || '10', 10) || 10);
-export const IPC_POLL_INTERVAL = 1000;
 export const IDLE_TIMEOUT = parseInt(process.env.IDLE_TIMEOUT || '1800000', 10); // 30min default — how long to keep container alive after last result
 export const MAX_CONCURRENT_CONTAINERS = Math.max(1, parseInt(process.env.MAX_CONCURRENT_CONTAINERS || '5', 10) || 5);
 
diff --git a/src/container-config.ts b/src/container-config.ts
index e1366e3..d972842 100644
--- a/src/container-config.ts
+++ b/src/container-config.ts
@@ -1,15 +1,8 @@
 /**
  * Per-group container config, stored as a plain JSON file at
- * `groups/<folder>/container.json`. Replaces the former
- * `agent_groups.container_config` DB column.
- *
- * Shape:
- *   {
- *     mcpServers:      { [name]: { command, args, env } }
- *     packages:        { apt: string[], npm: string[] }
- *     imageTag?:       string                    // set by buildAgentGroupImage on rebuild
- *     additionalMounts?: Array<{hostPath, containerPath, readonly}>
- *   }
+ * `groups/<folder>/container.json`. Mounted read-only inside the container
+ * at `/workspace/agent/container.json` — the runner reads it at startup but
+ * cannot modify it. Config changes go through the self-mod approval flow.
  *
  * All fields are optional — a missing file or a partial file both resolve
  * to sensible defaults. Writes are atomic-enough (write-then-rename is not
@@ -25,6 +18,10 @@ export interface McpServerConfig {
   command: string;
   args?: string[];
   env?: Record<string, string>;
+  // Optional always-in-context guidance. When set, the host writes the
+  // content to `.claude-fragments/mcp-<name>.md` at spawn and imports it
+  // into the composed CLAUDE.md.
+  instructions?: string;
 }
 
 export interface AdditionalMountConfig {
@@ -38,6 +35,18 @@ export interface ContainerConfig {
   packages: { apt: string[]; npm: string[] };
   imageTag?: string;
   additionalMounts: AdditionalMountConfig[];
+  /** Which skills to enable — array of skill names or "all" (default). */
+  skills: string[] | 'all';
+  /** Agent provider name (e.g. "claude", "opencode"). Default: "claude". */
+  provider?: string;
+  /** Agent group display name (used in transcript archiving). */
+  groupName?: string;
+  /** Assistant display name (used in system prompt / responses). */
+  assistantName?: string;
+  /** Agent group ID — set by the host, read by the runner. */
+  agentGroupId?: string;
+  /** Max messages per prompt. Falls back to code default if unset. */
+  maxMessagesPerPrompt?: number;
 }
 
 function emptyConfig(): ContainerConfig {
@@ -45,6 +54,7 @@ function emptyConfig(): ContainerConfig {
     mcpServers: {},
     packages: { apt: [], npm: [] },
     additionalMounts: [],
+    skills: 'all',
   };
 }
 
@@ -71,6 +81,12 @@ export function readContainerConfig(folder: string): ContainerConfig {
       },
       imageTag: raw.imageTag,
       additionalMounts: raw.additionalMounts ?? [],
+      skills: raw.skills ?? 'all',
+      provider: raw.provider,
+      groupName: raw.groupName,
+      assistantName: raw.assistantName,
+      agentGroupId: raw.agentGroupId,
+      maxMessagesPerPrompt: raw.maxMessagesPerPrompt,
     };
   } catch (err) {
     console.error(`[container-config] failed to parse ${p}: ${String(err)}`);
diff --git a/src/container-runner.test.ts b/src/container-runner.test.ts
new file mode 100644
index 0000000..cd18a72
--- /dev/null
+++ b/src/container-runner.test.ts
@@ -0,0 +1,32 @@
+import { describe, expect, it } from 'vitest';
+
+import { resolveProviderName } from './container-runner.js';
+
+describe('resolveProviderName', () => {
+  it('prefers session over group and container.json', () => {
+    expect(resolveProviderName('codex', 'opencode', 'claude')).toBe('codex');
+  });
+
+  it('falls back to group when session is null', () => {
+    expect(resolveProviderName(null, 'codex', 'claude')).toBe('codex');
+  });
+
+  it('falls back to container.json when session and group are null', () => {
+    expect(resolveProviderName(null, null, 'opencode')).toBe('opencode');
+  });
+
+  it('defaults to claude when nothing is set', () => {
+    expect(resolveProviderName(null, null, undefined)).toBe('claude');
+  });
+
+  it('lowercases the resolved name', () => {
+    expect(resolveProviderName('CODEX', null, null)).toBe('codex');
+    expect(resolveProviderName(null, 'OpenCode', null)).toBe('opencode');
+    expect(resolveProviderName(null, null, 'Claude')).toBe('claude');
+  });
+
+  it('treats empty string as unset (falls through)', () => {
+    expect(resolveProviderName('', 'codex', null)).toBe('codex');
+    expect(resolveProviderName(null, '', 'opencode')).toBe('opencode');
+  });
+});
diff --git a/src/container-runner.ts b/src/container-runner.ts
index c3fb24f..029b5fe 100644
--- a/src/container-runner.ts
+++ b/src/container-runner.ts
@@ -9,9 +9,19 @@ import path from 'path';
 
 import { OneCLI } from '@onecli-sh/sdk';
 
-import { CONTAINER_IMAGE, DATA_DIR, GROUPS_DIR, IDLE_TIMEOUT, ONECLI_URL, TIMEZONE } from './config.js';
+import {
+  CONTAINER_IMAGE,
+  CONTAINER_IMAGE_BASE,
+  CONTAINER_INSTALL_LABEL,
+  DATA_DIR,
+  GROUPS_DIR,
+  ONECLI_API_KEY,
+  ONECLI_URL,
+  TIMEZONE,
+} from './config.js';
 import { readContainerConfig, writeContainerConfig } from './container-config.js';
 import { CONTAINER_RUNTIME_BIN, hostGatewayArgs, readonlyMountArgs, stopContainer } from './container-runtime.js';
+import { composeGroupClaudeMd } from './claude-md-compose.js';
 import { getAgentGroup } from './db/agent-groups.js';
 import { getDb, hasTable } from './db/connection.js';
 import { initGroupFilesystem } from './group-init.js';
@@ -27,6 +37,7 @@ import {
   type VolumeMount,
 } from './providers/provider-container-registry.js';
 import {
+  heartbeatPath,
   markContainerRunning,
   markContainerStopped,
   sessionDir,
@@ -34,7 +45,7 @@ import {
 } from './session-manager.js';
 import type { AgentGroup, Session } from './types.js';
 
-const onecli = new OneCLI({ url: ONECLI_URL });
+const onecli = new OneCLI({ url: ONECLI_URL, apiKey: ONECLI_API_KEY });
 
 /** Active containers tracked by session ID. */
 const activeContainers = new Map<string, { process: ChildProcess; containerName: string }>();
@@ -96,20 +107,42 @@ async function spawnContainer(session: Session): Promise<void> {
   }
   writeSessionRouting(agentGroup.id, session.id);
 
+  // Read container config once — threaded through provider resolution,
+  // buildMounts, and buildContainerArgs so we don't re-read the file.
+  const containerConfig = readContainerConfig(agentGroup.folder);
+
+  // Ensure container.json has the agent group identity fields the runner needs.
+  // Written at spawn time so the runner can read them from the RO mount.
+  ensureRuntimeFields(containerConfig, agentGroup);
+
   // Resolve the effective provider + any host-side contribution it declares
   // (extra mounts, env passthrough). Computed once and threaded through both
   // buildMounts and buildContainerArgs so side effects (mkdir, etc.) fire once.
-  const { provider, contribution } = resolveProviderContribution(session, agentGroup);
+  const { provider, contribution } = resolveProviderContribution(session, agentGroup, containerConfig);
 
-  const mounts = buildMounts(agentGroup, session, contribution);
+  const mounts = buildMounts(agentGroup, session, containerConfig, contribution);
   const containerName = `nanoclaw-v2-${agentGroup.folder}-${Date.now()}`;
   // OneCLI agent identifier is always the agent group id — stable across
   // sessions and reversible via getAgentGroup() for approval routing.
   const agentIdentifier = agentGroup.id;
-  const args = await buildContainerArgs(mounts, containerName, agentGroup, provider, contribution, agentIdentifier);
+  const args = await buildContainerArgs(
+    mounts,
+    containerName,
+    agentGroup,
+    containerConfig,
+    provider,
+    contribution,
+    agentIdentifier,
+  );
 
   log.info('Spawning container', { sessionId: session.id, agentGroup: agentGroup.name, containerName });
 
+  // Clear any orphan heartbeat from a previous container instance — the
+  // sweep's ceiling check treats a missing file as "fresh spawn, give grace"
+  // (host-sweep.ts line 87). Without this, the stale mtime can trigger an
+  // immediate kill before the new container touches the file itself.
+  fs.rmSync(heartbeatPath(agentGroup.id, session.id), { force: true });
+
   const container = spawn(CONTAINER_RUNTIME_BIN, args, { stdio: ['ignore', 'pipe', 'pipe'] });
 
   activeContainers.set(session.id, { process: container, containerName });
@@ -125,22 +158,12 @@ async function spawnContainer(session: Session): Promise<void> {
   // stdout is unused in v2 (all IO is via session DB)
   container.stdout?.on('data', () => {});
 
-  // Idle timeout: kill container after IDLE_TIMEOUT of no activity
-  let idleTimer = setTimeout(() => killContainer(session.id, 'idle timeout'), IDLE_TIMEOUT);
-
-  const resetIdle = () => {
-    clearTimeout(idleTimer);
-    idleTimer = setTimeout(() => killContainer(session.id, 'idle timeout'), IDLE_TIMEOUT);
-  };
-
-  // Reset idle timer when the host detects new messages_out (called by delivery.ts)
-  const entry = activeContainers.get(session.id);
-  if (entry) {
-    (entry as { resetIdle?: () => void }).resetIdle = resetIdle;
-  }
+  // No host-side idle timeout. Stale/stuck detection is driven by the host
+  // sweep reading heartbeat mtime + processing_ack claim age + container_state
+  // (see src/host-sweep.ts). This avoids killing long-running legitimate work
+  // on a wall-clock timer.
 
   container.on('close', (code) => {
-    clearTimeout(idleTimer);
     activeContainers.delete(session.id);
     markContainerStopped(session.id);
     stopTypingRefresh(session.id);
@@ -148,7 +171,6 @@ async function spawnContainer(session: Session): Promise<void> {
   });
 
   container.on('error', (err) => {
-    clearTimeout(idleTimer);
     activeContainers.delete(session.id);
     markContainerStopped(session.id);
     stopTypingRefresh(session.id);
@@ -156,12 +178,6 @@ async function spawnContainer(session: Session): Promise<void> {
   });
 }
 
-/** Reset the idle timer for a session's container (called when messages_out are delivered). */
-export function resetContainerIdleTimer(sessionId: string): void {
-  const entry = activeContainers.get(sessionId) as { resetIdle?: () => void } | undefined;
-  entry?.resetIdle?.();
-}
-
 /** Kill a container for a session. */
 export function killContainer(sessionId: string, reason: string): void {
   const entry = activeContainers.get(sessionId);
@@ -175,11 +191,31 @@ export function killContainer(sessionId: string, reason: string): void {
   }
 }
 
+/**
+ * Resolve the provider name for a session using the precedence documented in
+ * the provider-install skills:
+ *
+ *   sessions.agent_provider
+ *     → agent_groups.agent_provider
+ *     → container.json `provider`
+ *     → 'claude'
+ *
+ * Pure so the precedence can be unit-tested without a DB or filesystem.
+ */
+export function resolveProviderName(
+  sessionProvider: string | null | undefined,
+  agentGroupProvider: string | null | undefined,
+  containerConfigProvider: string | null | undefined,
+): string {
+  return (sessionProvider || agentGroupProvider || containerConfigProvider || 'claude').toLowerCase();
+}
+
 function resolveProviderContribution(
   session: Session,
   agentGroup: AgentGroup,
+  containerConfig: import('./container-config.js').ContainerConfig,
 ): { provider: string; contribution: ProviderContainerContribution } {
-  const provider = (session.agent_provider || agentGroup.agent_provider || 'claude').toLowerCase();
+  const provider = resolveProviderName(session.agent_provider, agentGroup.agent_provider, containerConfig.provider);
   const fn = getProviderContainerConfig(provider);
   const contribution = fn
     ? fn({
@@ -194,15 +230,24 @@ function resolveProviderContribution(
 function buildMounts(
   agentGroup: AgentGroup,
   session: Session,
+  containerConfig: import('./container-config.js').ContainerConfig,
   providerContribution: ProviderContainerContribution,
 ): VolumeMount[] {
+  const projectRoot = process.cwd();
+
   // Per-group filesystem state lives forever after first creation. Init is
   // idempotent: it only writes paths that don't already exist, so this call
-  // is a no-op for groups that have spawned before. Pulling in upstream
-  // built-in skill or agent-runner source updates is an explicit operation
-  // (host-mediated tools), not something the spawn path does silently.
+  // is a no-op for groups that have spawned before.
   initGroupFilesystem(agentGroup);
 
+  // Sync skill symlinks based on container.json selection before mounting.
+  const claudeDir = path.join(DATA_DIR, 'v2-sessions', agentGroup.id, '.claude-shared');
+  syncSkillSymlinks(claudeDir, containerConfig);
+
+  // Compose CLAUDE.md fresh every spawn from the shared base, enabled skill
+  // fragments, and MCP server instructions. See `claude-md-compose.ts`.
+  composeGroupClaudeMd(agentGroup);
+
   const mounts: VolumeMount[] = [];
   const sessDir = sessionDir(agentGroup.id, session.id);
   const groupDir = path.resolve(GROUPS_DIR, agentGroup.folder);
@@ -210,28 +255,60 @@ function buildMounts(
   // Session folder at /workspace (contains inbound.db, outbound.db, outbox/, .claude/)
   mounts.push({ hostPath: sessDir, containerPath: '/workspace', readonly: false });
 
-  // Agent group folder at /workspace/agent
+  // Agent group folder at /workspace/agent (RW for working files + CLAUDE.local.md)
   mounts.push({ hostPath: groupDir, containerPath: '/workspace/agent', readonly: false });
 
-  // Global memory directory — always read-only. Edits to global config
-  // happen through the approval flow, not by handing one workspace RW.
+  // container.json — nested RO mount on top of RW group dir so the agent
+  // can read its config but cannot modify it.
+  const containerJsonPath = path.join(groupDir, 'container.json');
+  if (fs.existsSync(containerJsonPath)) {
+    mounts.push({ hostPath: containerJsonPath, containerPath: '/workspace/agent/container.json', readonly: true });
+  }
+
+  // Composer-managed CLAUDE.md artifacts — nested RO mounts. These are
+  // regenerated from the shared base + fragments on every spawn; any
+  // agent-side writes would be clobbered, so enforce read-only. Only
+  // CLAUDE.local.md (per-group memory) remains RW via the group-dir mount.
+  // `.claude-shared.md` is a symlink whose target (`/app/CLAUDE.md`) is
+  // already RO-mounted, so writes through it fail regardless — no need for
+  // a nested mount there.
+  const composedClaudeMd = path.join(groupDir, 'CLAUDE.md');
+  if (fs.existsSync(composedClaudeMd)) {
+    mounts.push({ hostPath: composedClaudeMd, containerPath: '/workspace/agent/CLAUDE.md', readonly: true });
+  }
+  const fragmentsDir = path.join(groupDir, '.claude-fragments');
+  if (fs.existsSync(fragmentsDir)) {
+    mounts.push({ hostPath: fragmentsDir, containerPath: '/workspace/agent/.claude-fragments', readonly: true });
+  }
+
+  // Global memory directory — always read-only.
   const globalDir = path.join(GROUPS_DIR, 'global');
   if (fs.existsSync(globalDir)) {
     mounts.push({ hostPath: globalDir, containerPath: '/workspace/global', readonly: true });
   }
 
+  // Shared CLAUDE.md — read-only, imported by the composed entry point via
+  // the `.claude-shared.md` symlink inside the group dir.
+  const sharedClaudeMd = path.join(process.cwd(), 'container', 'CLAUDE.md');
+  if (fs.existsSync(sharedClaudeMd)) {
+    mounts.push({ hostPath: sharedClaudeMd, containerPath: '/app/CLAUDE.md', readonly: true });
+  }
+
   // Per-group .claude-shared at /home/node/.claude (Claude state, settings,
-  // skills — initialized once at group creation, persistent thereafter)
-  const claudeDir = path.join(DATA_DIR, 'v2-sessions', agentGroup.id, '.claude-shared');
+  // skill symlinks)
   mounts.push({ hostPath: claudeDir, containerPath: '/home/node/.claude', readonly: false });
 
-  // Per-group agent-runner source at /app/src (initialized once at group
-  // creation, persistent thereafter — agents can modify their runner)
-  const groupRunnerDir = path.join(DATA_DIR, 'v2-sessions', agentGroup.id, 'agent-runner-src');
-  mounts.push({ hostPath: groupRunnerDir, containerPath: '/app/src', readonly: false });
+  // Shared agent-runner source — read-only, same code for all groups.
+  const agentRunnerSrc = path.join(projectRoot, 'container', 'agent-runner', 'src');
+  mounts.push({ hostPath: agentRunnerSrc, containerPath: '/app/src', readonly: true });
 
-  // Additional mounts from container config (groups/<folder>/container.json)
-  const containerConfig = readContainerConfig(agentGroup.folder);
+  // Shared skills — read-only, symlinks in .claude-shared/skills/ point here.
+  const skillsSrc = path.join(projectRoot, 'container', 'skills');
+  if (fs.existsSync(skillsSrc)) {
+    mounts.push({ hostPath: skillsSrc, containerPath: '/app/skills', readonly: true });
+  }
+
+  // Additional mounts from container config
   if (containerConfig.additionalMounts && containerConfig.additionalMounts.length > 0) {
     const validated = validateAdditionalMounts(containerConfig.additionalMounts, agentGroup.name);
     mounts.push(...validated);
@@ -245,29 +322,110 @@ function buildMounts(
   return mounts;
 }
 
+/**
+ * Sync skill symlinks in .claude-shared/skills/ to match the container.json
+ * selection. Each symlink points to a container path (/app/skills/<name>)
+ * so it's dangling on the host but valid inside the container.
+ */
+function syncSkillSymlinks(claudeDir: string, containerConfig: import('./container-config.js').ContainerConfig): void {
+  const skillsDir = path.join(claudeDir, 'skills');
+  if (!fs.existsSync(skillsDir)) {
+    fs.mkdirSync(skillsDir, { recursive: true });
+  }
+
+  // Determine desired skill set
+  const projectRoot = process.cwd();
+  const sharedSkillsDir = path.join(projectRoot, 'container', 'skills');
+  let desired: string[];
+  if (containerConfig.skills === 'all') {
+    // Recompute from shared dir — newly-added upstream skills appear automatically
+    desired = fs.existsSync(sharedSkillsDir)
+      ? fs.readdirSync(sharedSkillsDir).filter((e) => {
+          try {
+            return fs.statSync(path.join(sharedSkillsDir, e)).isDirectory();
+          } catch {
+            return false;
+          }
+        })
+      : [];
+  } else {
+    desired = containerConfig.skills;
+  }
+
+  const desiredSet = new Set(desired);
+
+  // Remove symlinks not in the desired set
+  for (const entry of fs.readdirSync(skillsDir)) {
+    const entryPath = path.join(skillsDir, entry);
+    let isSymlink = false;
+    try {
+      isSymlink = fs.lstatSync(entryPath).isSymbolicLink();
+    } catch {
+      continue;
+    }
+    if (isSymlink && !desiredSet.has(entry)) {
+      fs.unlinkSync(entryPath);
+    }
+  }
+
+  // Create symlinks for desired skills (container path targets)
+  for (const skill of desired) {
+    const linkPath = path.join(skillsDir, skill);
+    let exists = false;
+    try {
+      fs.lstatSync(linkPath);
+      exists = true;
+    } catch {
+      /* missing */
+    }
+    if (!exists) {
+      fs.symlinkSync(`/app/skills/${skill}`, linkPath);
+    }
+  }
+}
+
+/**
+ * Ensure container.json has the runtime identity fields the runner needs.
+ * Written at spawn time so they're always current even if the DB values
+ * change (e.g. group rename). Only writes if values differ to avoid
+ * unnecessary file churn.
+ */
+function ensureRuntimeFields(
+  containerConfig: import('./container-config.js').ContainerConfig,
+  agentGroup: AgentGroup,
+): void {
+  let dirty = false;
+  if (containerConfig.agentGroupId !== agentGroup.id) {
+    containerConfig.agentGroupId = agentGroup.id;
+    dirty = true;
+  }
+  if (containerConfig.groupName !== agentGroup.name) {
+    containerConfig.groupName = agentGroup.name;
+    dirty = true;
+  }
+  if (containerConfig.assistantName !== agentGroup.name) {
+    containerConfig.assistantName = agentGroup.name;
+    dirty = true;
+  }
+  if (dirty) {
+    writeContainerConfig(agentGroup.folder, containerConfig);
+  }
+}
+
 async function buildContainerArgs(
   mounts: VolumeMount[],
   containerName: string,
   agentGroup: AgentGroup,
+  containerConfig: import('./container-config.js').ContainerConfig,
   provider: string,
   providerContribution: ProviderContainerContribution,
   agentIdentifier?: string,
 ): Promise<string[]> {
-  const args: string[] = ['run', '--rm', '--name', containerName];
+  const args: string[] = ['run', '--rm', '--name', containerName, '--label', CONTAINER_INSTALL_LABEL];
 
-  // Environment
+  // Environment — only vars read by code we don't own.
+  // Everything NanoClaw-specific is in container.json (read by runner at startup).
   args.push('-e', `TZ=${TIMEZONE}`);
-  args.push('-e', `AGENT_PROVIDER=${provider}`);
-  // Two-DB split: container reads inbound.db, writes outbound.db
-  args.push('-e', 'SESSION_INBOUND_DB_PATH=/workspace/inbound.db');
-  args.push('-e', 'SESSION_OUTBOUND_DB_PATH=/workspace/outbound.db');
-  args.push('-e', 'SESSION_HEARTBEAT_PATH=/workspace/.heartbeat');
-
-  if (agentGroup.name) {
-    args.push('-e', `NANOCLAW_ASSISTANT_NAME=${agentGroup.name}`);
-  }
-  args.push('-e', `NANOCLAW_AGENT_GROUP_ID=${agentGroup.id}`);
-  args.push('-e', `NANOCLAW_AGENT_GROUP_NAME=${agentGroup.name}`);
 
   // Provider-contributed env vars (e.g. XDG_DATA_HOME, OPENCODE_*, NO_PROXY).
   if (providerContribution.env) {
@@ -276,39 +434,8 @@ async function buildContainerArgs(
     }
   }
 
-  // Users allowed to run admin commands (e.g. /clear) inside this container.
-  // Computed at wake time: owners + global admins + admins scoped to this
-  // agent group. Role changes take effect on next container spawn.
-  //
-  // SQL inlined to keep core independent of the permissions module — we
-  // guard on the `user_roles` table directly. If the permissions module
-  // isn't installed, the table doesn't exist and the set stays empty; the
-  // formatter treats an empty admin set as permissionless mode (every
-  // sender is admin).
-  const adminUserIds = new Set<string>();
-  if (hasTable(getDb(), 'user_roles')) {
-    const db = getDb();
-    const owners = db
-      .prepare("SELECT user_id FROM user_roles WHERE role = 'owner' AND agent_group_id IS NULL")
-      .all() as Array<{ user_id: string }>;
-    const globalAdmins = db
-      .prepare("SELECT user_id FROM user_roles WHERE role = 'admin' AND agent_group_id IS NULL")
-      .all() as Array<{ user_id: string }>;
-    const scopedAdmins = db
-      .prepare("SELECT user_id FROM user_roles WHERE role = 'admin' AND agent_group_id = ?")
-      .all(agentGroup.id) as Array<{ user_id: string }>;
-    for (const r of owners) adminUserIds.add(r.user_id);
-    for (const r of globalAdmins) adminUserIds.add(r.user_id);
-    for (const r of scopedAdmins) adminUserIds.add(r.user_id);
-  }
-  if (adminUserIds.size > 0) {
-    args.push('-e', `NANOCLAW_ADMIN_USER_IDS=${Array.from(adminUserIds).join(',')}`);
-  }
-
   // OneCLI gateway — injects HTTPS_PROXY + certs so container API calls
   // are routed through the agent vault for credential injection.
-  // Must ensureAgent first for non-admin groups, otherwise applyContainerConfig
-  // rejects the unknown agent identifier and returns false.
   try {
     if (agentIdentifier) {
       await onecli.ensureAgent({ name: agentGroup.name, identifier: agentIdentifier });
@@ -343,16 +470,7 @@ async function buildContainerArgs(
     }
   }
 
-  // Pass additional MCP servers from container config (groups/<folder>/container.json)
-  const containerConfig = readContainerConfig(agentGroup.folder);
-  if (containerConfig.mcpServers && Object.keys(containerConfig.mcpServers).length > 0) {
-    args.push('-e', `NANOCLAW_MCP_SERVERS=${JSON.stringify(containerConfig.mcpServers)}`);
-  }
-
   // Override entrypoint: run v2 entry point directly via Bun (no tsc, no stdin).
-  // The image's ENTRYPOINT (tini → entrypoint.sh) handles the stdin-piped
-  // invocation path; the host-spawned sessions don't need stdin because all
-  // IO flows through the mounted session DBs.
   args.push('--entrypoint', 'bash');
 
   // Use per-agent-group image if one has been built, otherwise base image
@@ -391,7 +509,7 @@ export async function buildAgentGroupImage(agentGroupId: string): Promise<void>
   }
   dockerfile += 'USER node\n';
 
-  const imageTag = `nanoclaw-agent:${agentGroupId}`;
+  const imageTag = `${CONTAINER_IMAGE_BASE}:${agentGroupId}`;
 
   log.info('Building per-agent-group image', { agentGroupId, imageTag, apt: aptPackages, npm: npmPackages });
 
diff --git a/src/container-runtime.test.ts b/src/container-runtime.test.ts
index 47d9744..f6f6e8a 100644
--- a/src/container-runtime.test.ts
+++ b/src/container-runtime.test.ts
@@ -24,6 +24,7 @@ import {
   ensureContainerRuntimeRunning,
   cleanupOrphans,
 } from './container-runtime.js';
+import { CONTAINER_INSTALL_LABEL } from './config.js';
 import { log } from './log.js';
 
 beforeEach(() => {
@@ -84,6 +85,17 @@ describe('ensureContainerRuntimeRunning', () => {
 // --- cleanupOrphans ---
 
 describe('cleanupOrphans', () => {
+  it('filters ps by the install label so peers are not reaped', () => {
+    mockExecSync.mockReturnValueOnce('');
+
+    cleanupOrphans();
+
+    expect(mockExecSync).toHaveBeenCalledWith(
+      `${CONTAINER_RUNTIME_BIN} ps --filter label=${CONTAINER_INSTALL_LABEL} --format '{{.Names}}'`,
+      expect.any(Object),
+    );
+  });
+
   it('stops orphaned nanoclaw containers', () => {
     // docker ps returns container names, one per line
     mockExecSync.mockReturnValueOnce('nanoclaw-group1-111\nnanoclaw-group2-222\n');
diff --git a/src/container-runtime.ts b/src/container-runtime.ts
index 5e68426..82ddb5e 100644
--- a/src/container-runtime.ts
+++ b/src/container-runtime.ts
@@ -5,6 +5,7 @@
 import { execSync } from 'child_process';
 import os from 'os';
 
+import { CONTAINER_INSTALL_LABEL } from './config.js';
 import { log } from './log.js';
 
 /** The container runtime binary name. */
@@ -56,13 +57,22 @@ export function ensureContainerRuntimeRunning(): void {
   }
 }
 
-/** Kill orphaned NanoClaw containers from previous runs. */
+/**
+ * Kill orphaned NanoClaw containers from THIS install's previous runs.
+ *
+ * Scoped by label `nanoclaw-install=<slug>` so a crash-looping peer install
+ * cannot reap our containers, and we cannot reap theirs. The label is
+ * stamped onto every container at spawn time — see container-runner.ts.
+ */
 export function cleanupOrphans(): void {
   try {
-    const output = execSync(`${CONTAINER_RUNTIME_BIN} ps --filter name=nanoclaw- --format '{{.Names}}'`, {
-      stdio: ['pipe', 'pipe', 'pipe'],
-      encoding: 'utf-8',
-    });
+    const output = execSync(
+      `${CONTAINER_RUNTIME_BIN} ps --filter label=${CONTAINER_INSTALL_LABEL} --format '{{.Names}}'`,
+      {
+        stdio: ['pipe', 'pipe', 'pipe'],
+        encoding: 'utf-8',
+      },
+    );
     const orphans = output.trim().split('\n').filter(Boolean);
     for (const name of orphans) {
       try {
diff --git a/src/db/db-v2.test.ts b/src/db/db-v2.test.ts
index f8689eb..e0cebdf 100644
--- a/src/db/db-v2.test.ts
+++ b/src/db/db-v2.test.ts
@@ -178,8 +178,10 @@ describe('messaging group agents', () => {
     id: 'mga-1',
     messaging_group_id: 'mg-1',
     agent_group_id: 'ag-1',
-    trigger_rules: null,
-    response_scope: 'all' as const,
+    engage_mode: 'pattern' as const,
+    engage_pattern: '.',
+    sender_scope: 'all' as const,
+    ignored_message_policy: 'drop' as const,
     session_mode: 'shared' as const,
     priority: 0,
     created_at: now(),
@@ -229,7 +231,8 @@ describe('messaging group agents', () => {
   });
 
   it('auto-creates an agent_destinations row for the wiring', async () => {
-    const { getDestinationByTarget, getDestinations } = await import('../modules/agent-to-agent/db/agent-destinations.js');
+    const { getDestinationByTarget, getDestinations } =
+      await import('../modules/agent-to-agent/db/agent-destinations.js');
     createMessagingGroupAgent(mga());
 
     const dest = getDestinationByTarget('ag-1', 'channel', 'mg-1');
diff --git a/src/db/messaging-groups.ts b/src/db/messaging-groups.ts
index 0c0ba22..33c8715 100644
--- a/src/db/messaging-groups.ts
+++ b/src/db/messaging-groups.ts
@@ -37,6 +37,37 @@ export function getMessagingGroupByPlatform(channelType: string, platformId: str
     .get(channelType, platformId) as MessagingGroup | undefined;
 }
 
+/**
+ * Combined lookup for the router's fast-drop path. Returns the messaging
+ * group (if it exists) and a count of wired agents in one query — lets
+ * `routeInbound` short-circuit messages for unwired / unknown channels
+ * with a single DB read instead of four (mg lookup, sender upsert, agents
+ * lookup, dropped_messages insert).
+ *
+ * Returns `null` when no messaging_groups row exists for this channel.
+ * Returns `{ mg, agentCount: 0 }` when the row exists but has no wired
+ * agents. Uses the `UNIQUE(channel_type, platform_id)` index plus the
+ * `UNIQUE(messaging_group_id, agent_group_id)` index for the JOIN — both
+ * covered by existing SQLite auto-indexes from the UNIQUE constraints.
+ */
+export function getMessagingGroupWithAgentCount(
+  channelType: string,
+  platformId: string,
+): { mg: MessagingGroup; agentCount: number } | null {
+  const row = getDb()
+    .prepare(
+      `SELECT mg.*, COUNT(mga.id) AS agent_count
+         FROM messaging_groups mg
+    LEFT JOIN messaging_group_agents mga ON mga.messaging_group_id = mg.id
+        WHERE mg.channel_type = ? AND mg.platform_id = ?
+     GROUP BY mg.id`,
+    )
+    .get(channelType, platformId) as (MessagingGroup & { agent_count: number }) | undefined;
+  if (!row) return null;
+  const { agent_count, ...mg } = row;
+  return { mg: mg as MessagingGroup, agentCount: agent_count };
+}
+
 export function getAllMessagingGroups(): MessagingGroup[] {
   return getDb().prepare('SELECT * FROM messaging_groups ORDER BY name').all() as MessagingGroup[];
 }
@@ -69,6 +100,20 @@ export function deleteMessagingGroup(id: string): void {
   getDb().prepare('DELETE FROM messaging_groups WHERE id = ?').run(id);
 }
 
+/**
+ * Mark a messaging group as denied by the owner (channel-registration flow).
+ * Future mentions on this channel silently drop until an admin explicitly
+ * wires it via `createMessagingGroupAgent`, which implicitly clears the
+ * denied state by making `agentCount > 0` — the router's denied-channel
+ * check sits on the `agentCount === 0` branch.
+ *
+ * Passing null unsets the flag (used by tests or a future "unblock channel"
+ * admin command).
+ */
+export function setMessagingGroupDeniedAt(id: string, deniedAt: string | null): void {
+  getDb().prepare('UPDATE messaging_groups SET denied_at = ? WHERE id = ?').run(deniedAt, id);
+}
+
 // ── Messaging Group Agents ──
 
 /**
@@ -87,8 +132,16 @@ export function deleteMessagingGroup(id: string): void {
 export function createMessagingGroupAgent(mga: MessagingGroupAgent): void {
   getDb()
     .prepare(
-      `INSERT INTO messaging_group_agents (id, messaging_group_id, agent_group_id, trigger_rules, response_scope, session_mode, priority, created_at)
-       VALUES (@id, @messaging_group_id, @agent_group_id, @trigger_rules, @response_scope, @session_mode, @priority, @created_at)`,
+      `INSERT INTO messaging_group_agents (
+         id, messaging_group_id, agent_group_id,
+         engage_mode, engage_pattern, sender_scope, ignored_message_policy,
+         session_mode, priority, created_at
+       )
+       VALUES (
+         @id, @messaging_group_id, @agent_group_id,
+         @engage_mode, @engage_pattern, @sender_scope, @ignored_message_policy,
+         @session_mode, @priority, @created_at
+       )`,
     )
     .run(mga);
 
@@ -160,7 +213,12 @@ export function getMessagingGroupAgent(id: string): MessagingGroupAgent | undefi
 
 export function updateMessagingGroupAgent(
   id: string,
-  updates: Partial<Pick<MessagingGroupAgent, 'trigger_rules' | 'response_scope' | 'session_mode' | 'priority'>>,
+  updates: Partial<
+    Pick<
+      MessagingGroupAgent,
+      'engage_mode' | 'engage_pattern' | 'sender_scope' | 'ignored_message_policy' | 'session_mode' | 'priority'
+    >
+  >,
 ): void {
   const fields: string[] = [];
   const values: Record<string, unknown> = { id };
diff --git a/src/db/migrations/010-engage-modes.ts b/src/db/migrations/010-engage-modes.ts
new file mode 100644
index 0000000..e7bff99
--- /dev/null
+++ b/src/db/migrations/010-engage-modes.ts
@@ -0,0 +1,103 @@
+/**
+ * Replace `trigger_rules` (opaque JSON) + `response_scope` (conflated axis)
+ * with four explicit orthogonal columns on messaging_group_agents:
+ *
+ *   engage_mode            'pattern' | 'mention' | 'mention-sticky'
+ *   engage_pattern         regex string (required when engage_mode='pattern';
+ *                          '.' means "match everything" — the "always" flavor)
+ *   sender_scope           'all' | 'known'
+ *   ignored_message_policy 'drop' | 'accumulate'
+ *
+ * Backfill rules (applied per-row, reading the old JSON):
+ *   - If trigger_rules.pattern is a non-empty string → engage_mode='pattern',
+ *     engage_pattern = that value
+ *   - Else if trigger_rules.requiresTrigger === false OR response_scope='all'
+ *     → engage_mode='pattern', engage_pattern='.'
+ *   - Else (requires trigger but no pattern specified) → engage_mode='mention'
+ *   - sender_scope: 'known' when response_scope was 'allowlisted', 'all' otherwise
+ *   - ignored_message_policy: 'drop' (conservative default; no old-schema analog)
+ */
+import type Database from 'better-sqlite3';
+import type { Migration } from './index.js';
+
+import { log } from '../../log.js';
+
+interface LegacyRow {
+  id: string;
+  trigger_rules: string | null;
+  response_scope: string | null;
+}
+
+function backfill(row: LegacyRow): {
+  engage_mode: 'pattern' | 'mention' | 'mention-sticky';
+  engage_pattern: string | null;
+  sender_scope: 'all' | 'known';
+  ignored_message_policy: 'drop' | 'accumulate';
+} {
+  let parsed: Record<string, unknown> = {};
+  if (row.trigger_rules) {
+    try {
+      parsed = JSON.parse(row.trigger_rules) as Record<string, unknown>;
+    } catch {
+      // Invalid JSON falls through to conservative defaults.
+    }
+  }
+
+  const pattern = typeof parsed.pattern === 'string' && parsed.pattern.length > 0 ? (parsed.pattern as string) : null;
+  const requiresTrigger = parsed.requiresTrigger;
+
+  let engage_mode: 'pattern' | 'mention' | 'mention-sticky' = 'mention';
+  let engage_pattern: string | null = null;
+  if (pattern) {
+    engage_mode = 'pattern';
+    engage_pattern = pattern;
+  } else if (requiresTrigger === false || row.response_scope === 'all') {
+    engage_mode = 'pattern';
+    engage_pattern = '.';
+  }
+
+  const sender_scope: 'all' | 'known' = row.response_scope === 'allowlisted' ? 'known' : 'all';
+
+  return { engage_mode, engage_pattern, sender_scope, ignored_message_policy: 'drop' };
+}
+
+export const migration010: Migration = {
+  version: 10,
+  name: 'engage-modes',
+  up: (db: Database.Database) => {
+    // Add the four new columns alongside the existing two. SQLite ALTER ADD
+    // is cheap and non-rewriting.
+    db.exec(`
+      ALTER TABLE messaging_group_agents ADD COLUMN engage_mode            TEXT;
+      ALTER TABLE messaging_group_agents ADD COLUMN engage_pattern         TEXT;
+      ALTER TABLE messaging_group_agents ADD COLUMN sender_scope           TEXT;
+      ALTER TABLE messaging_group_agents ADD COLUMN ignored_message_policy TEXT;
+    `);
+
+    // Backfill existing rows in JS (parsing JSON per-row is painful in pure SQL).
+    const rows = db
+      .prepare('SELECT id, trigger_rules, response_scope FROM messaging_group_agents')
+      .all() as LegacyRow[];
+    const update = db.prepare(
+      `UPDATE messaging_group_agents
+         SET engage_mode            = ?,
+             engage_pattern         = ?,
+             sender_scope           = ?,
+             ignored_message_policy = ?
+       WHERE id = ?`,
+    );
+    for (const row of rows) {
+      const v = backfill(row);
+      update.run(v.engage_mode, v.engage_pattern, v.sender_scope, v.ignored_message_policy, row.id);
+    }
+
+    // Drop the legacy columns. DROP COLUMN requires SQLite 3.35+ (2021); our
+    // better-sqlite3 ships a current build.
+    db.exec(`
+      ALTER TABLE messaging_group_agents DROP COLUMN trigger_rules;
+      ALTER TABLE messaging_group_agents DROP COLUMN response_scope;
+    `);
+
+    log.info('engage-modes migration: backfilled rows', { count: rows.length });
+  },
+};
diff --git a/src/db/migrations/011-pending-sender-approvals.ts b/src/db/migrations/011-pending-sender-approvals.ts
new file mode 100644
index 0000000..2331a6e
--- /dev/null
+++ b/src/db/migrations/011-pending-sender-approvals.ts
@@ -0,0 +1,40 @@
+/**
+ * Unknown-sender approval flow. When `unknown_sender_policy = 'request_approval'`
+ * a non-member message triggers a card to the most appropriate admin. An
+ * in-flight entry in this table dedups concurrent attempts from the same
+ * sender; the row is cleared on approve / deny.
+ *
+ * Previously this migration also rebuilt `messaging_groups` to flip the
+ * column DEFAULT from `'strict'` to `'request_approval'`. Removed: the
+ * rebuild failed SQLite's foreign-key integrity check at DROP time on live
+ * DBs with existing FK references (sessions, user_dms, etc.), and `PRAGMA
+ * foreign_keys` / `defer_foreign_keys` can't be toggled inside the
+ * implicit migration transaction. The default-flip was cosmetic anyway —
+ * every `createMessagingGroup` callsite passes `unknown_sender_policy`
+ * explicitly, and the router's auto-create path was updated to hardcode
+ * `'request_approval'` directly (see src/router.ts:123).
+ */
+import type Database from 'better-sqlite3';
+import type { Migration } from './index.js';
+
+export const migration011: Migration = {
+  version: 11,
+  name: 'pending-sender-approvals',
+  up: (db: Database.Database) => {
+    db.exec(`
+      CREATE TABLE IF NOT EXISTS pending_sender_approvals (
+        id                   TEXT PRIMARY KEY,
+        messaging_group_id   TEXT NOT NULL REFERENCES messaging_groups(id),
+        agent_group_id       TEXT NOT NULL REFERENCES agent_groups(id),
+        sender_identity      TEXT NOT NULL,      -- namespaced user id (channel_type:handle)
+        sender_name          TEXT,
+        original_message     TEXT NOT NULL,      -- JSON serialized InboundEvent
+        approver_user_id     TEXT NOT NULL,
+        created_at           TEXT NOT NULL,
+        UNIQUE(messaging_group_id, sender_identity)
+      );
+      CREATE INDEX IF NOT EXISTS idx_pending_sender_approvals_mg
+        ON pending_sender_approvals(messaging_group_id);
+    `);
+  },
+};
diff --git a/src/db/migrations/012-channel-registration.ts b/src/db/migrations/012-channel-registration.ts
new file mode 100644
index 0000000..eca8911
--- /dev/null
+++ b/src/db/migrations/012-channel-registration.ts
@@ -0,0 +1,48 @@
+/**
+ * Unknown-channel registration flow.
+ *
+ * When a channel that isn't wired to any agent group receives a mention or
+ * DM, the router escalates to the owner for approval before wiring. Approve
+ * creates a `messaging_group_agents` row (with conservative defaults) and
+ * replays the triggering event. Deny marks the channel denied forever
+ * (stored as a timestamp on `messaging_groups.denied_at`) so future
+ * messages on that channel drop silently without re-prompting.
+ *
+ * Two changes:
+ *   1. `messaging_groups.denied_at TEXT NULL` — set on deny, checked in the
+ *      router before re-escalating. ALTER TABLE ADD COLUMN is FK-safe
+ *      unlike the table rebuild that bit us in migration 011.
+ *   2. `pending_channel_approvals` table. PRIMARY KEY on
+ *      `messaging_group_id` gives free in-flight dedup — a second mention
+ *      while the card is pending is silently dropped by INSERT OR IGNORE,
+ *      preventing card spam.
+ */
+import type Database from 'better-sqlite3';
+import type { Migration } from './index.js';
+
+export const migration012: Migration = {
+  version: 12,
+  name: 'channel-registration',
+  up: (db: Database.Database) => {
+    // 1. Add denied_at to messaging_groups. Idempotent guard in case the
+    //    column was added by some other path before this migration ran.
+    const cols = db.prepare("PRAGMA table_info('messaging_groups')").all() as Array<{ name: string }>;
+    if (!cols.some((c) => c.name === 'denied_at')) {
+      db.exec(`ALTER TABLE messaging_groups ADD COLUMN denied_at TEXT`);
+    }
+
+    // 2. pending_channel_approvals.
+    db.exec(`
+      CREATE TABLE IF NOT EXISTS pending_channel_approvals (
+        messaging_group_id   TEXT PRIMARY KEY REFERENCES messaging_groups(id),
+        agent_group_id       TEXT NOT NULL REFERENCES agent_groups(id),
+                             -- The agent the approved wiring will target.
+                             -- Picked at request time (currently: earliest
+                             -- agent_group by created_at).
+        original_message     TEXT NOT NULL,      -- JSON serialized InboundEvent
+        approver_user_id     TEXT NOT NULL,
+        created_at           TEXT NOT NULL
+      );
+    `);
+  },
+};
diff --git a/src/db/migrations/013-approval-render-metadata.ts b/src/db/migrations/013-approval-render-metadata.ts
new file mode 100644
index 0000000..3a1af28
--- /dev/null
+++ b/src/db/migrations/013-approval-render-metadata.ts
@@ -0,0 +1,27 @@
+/**
+ * Persist ask_question render metadata (title + options_json) on
+ * `pending_channel_approvals` and `pending_sender_approvals`, mirroring the
+ * columns migration 003 / module-approvals-title-options added to
+ * `pending_approvals`.
+ *
+ * Before this, `getAskQuestionRender` hardcoded the title + option labels
+ * for these two tables in the DB-access layer — duplicating wording that
+ * also lived in the approval modules and causing a visible drift between
+ * the initial card title ("📣 Bot mentioned in new chat" / "💬 New direct
+ * message", chosen per event) and the post-click render ("📣 Channel
+ * registration", constant). Storing the render metadata alongside the row
+ * lets both sides read from the same source.
+ */
+import type Database from 'better-sqlite3';
+import type { Migration } from './index.js';
+
+export const migration013: Migration = {
+  version: 13,
+  name: 'approval-render-metadata',
+  up(db: Database.Database) {
+    db.exec(`ALTER TABLE pending_channel_approvals ADD COLUMN title TEXT NOT NULL DEFAULT ''`);
+    db.exec(`ALTER TABLE pending_channel_approvals ADD COLUMN options_json TEXT NOT NULL DEFAULT '[]'`);
+    db.exec(`ALTER TABLE pending_sender_approvals ADD COLUMN title TEXT NOT NULL DEFAULT ''`);
+    db.exec(`ALTER TABLE pending_sender_approvals ADD COLUMN options_json TEXT NOT NULL DEFAULT '[]'`);
+  },
+};
diff --git a/src/db/migrations/index.ts b/src/db/migrations/index.ts
index 3a87797..b46e678 100644
--- a/src/db/migrations/index.ts
+++ b/src/db/migrations/index.ts
@@ -6,6 +6,10 @@ import { migration002 } from './002-chat-sdk-state.js';
 import { moduleAgentToAgentDestinations } from './module-agent-to-agent-destinations.js';
 import { migration008 } from './008-dropped-messages.js';
 import { migration009 } from './009-drop-pending-credentials.js';
+import { migration010 } from './010-engage-modes.js';
+import { migration011 } from './011-pending-sender-approvals.js';
+import { migration012 } from './012-channel-registration.js';
+import { migration013 } from './013-approval-render-metadata.js';
 import { moduleApprovalsPendingApprovals } from './module-approvals-pending-approvals.js';
 import { moduleApprovalsTitleOptions } from './module-approvals-title-options.js';
 
@@ -23,6 +27,10 @@ const migrations: Migration[] = [
   moduleApprovalsTitleOptions,
   migration008,
   migration009,
+  migration010,
+  migration011,
+  migration012,
+  migration013,
 ];
 
 export function runMigrations(db: Database.Database): void {
@@ -52,8 +60,8 @@ export function runMigrations(db: Database.Database): void {
   for (const m of pending) {
     db.transaction(() => {
       m.up(db);
-      const next =
-        (db.prepare('SELECT COALESCE(MAX(version), 0) + 1 AS v FROM schema_version').get() as { v: number }).v;
+      const next = (db.prepare('SELECT COALESCE(MAX(version), 0) + 1 AS v FROM schema_version').get() as { v: number })
+        .v;
       db.prepare('INSERT INTO schema_version (version, name, applied) VALUES (?, ?, ?)').run(
         next,
         m.name,
diff --git a/src/db/migrations/module-approvals-pending-approvals.ts b/src/db/migrations/module-approvals-pending-approvals.ts
index 91aa08e..699e305 100644
--- a/src/db/migrations/module-approvals-pending-approvals.ts
+++ b/src/db/migrations/module-approvals-pending-approvals.ts
@@ -3,8 +3,8 @@ import type { Migration } from './index.js';
 /**
  * `pending_approvals` table — host-side records for any approval-requiring
  * request. Used by:
- *   - install_packages / request_rebuild / add_mcp_server  (session-bound,
- *     `session_id` set, status stays at default 'pending' until handled)
+ *   - install_packages / add_mcp_server  (session-bound, `session_id` set,
+ *     status stays at default 'pending' until handled)
  *   - OneCLI credential approvals from the SDK `configureManualApproval`
  *     callback (session_id may be null, action='onecli_credential').
  *
diff --git a/src/db/schema.ts b/src/db/schema.ts
index 044d717..8433035 100644
--- a/src/db/schema.ts
+++ b/src/db/schema.ts
@@ -19,27 +19,39 @@ CREATE TABLE agent_groups (
 
 -- Platform groups/channels. unknown_sender_policy governs what happens
 -- when a sender we've never seen before posts in this chat.
+-- The column DEFAULT is "strict" (inherited from migration 001), but it
+-- only matters if something inserts without specifying the field, which no
+-- current callsite does. Router auto-create hardcodes "request_approval"
+-- (see src/router.ts:151); setup scripts pick per context.
 CREATE TABLE messaging_groups (
   id                    TEXT PRIMARY KEY,
   channel_type          TEXT NOT NULL,
   platform_id           TEXT NOT NULL,
   name                  TEXT,
   is_group              INTEGER DEFAULT 0,
-  unknown_sender_policy TEXT NOT NULL DEFAULT 'strict', -- 'strict' | 'request_approval' | 'public'
+  unknown_sender_policy TEXT NOT NULL DEFAULT 'strict',
+                        -- 'strict' | 'request_approval' | 'public'
   created_at            TEXT NOT NULL,
   UNIQUE(channel_type, platform_id)
 );
 
--- Which agent groups handle which messaging groups
+-- Which agent groups handle which messaging groups.
+-- engage_mode / engage_pattern / sender_scope / ignored_message_policy are
+-- the four orthogonal axes that together replace v1's opaque trigger_rules
+-- JSON + response_scope enum. See docs/v1-vs-v2/ACTION-ITEMS.md item 1.
 CREATE TABLE messaging_group_agents (
-  id                 TEXT PRIMARY KEY,
-  messaging_group_id TEXT NOT NULL REFERENCES messaging_groups(id),
-  agent_group_id     TEXT NOT NULL REFERENCES agent_groups(id),
-  trigger_rules      TEXT,
-  response_scope     TEXT DEFAULT 'all',
-  session_mode       TEXT DEFAULT 'shared',
-  priority           INTEGER DEFAULT 0,
-  created_at         TEXT NOT NULL,
+  id                     TEXT PRIMARY KEY,
+  messaging_group_id     TEXT NOT NULL REFERENCES messaging_groups(id),
+  agent_group_id         TEXT NOT NULL REFERENCES agent_groups(id),
+  engage_mode            TEXT NOT NULL DEFAULT 'mention',
+                         -- 'pattern' | 'mention' | 'mention-sticky'
+  engage_pattern         TEXT,   -- regex; required when engage_mode='pattern';
+                                 -- '.' means "match every message" (the "always" flavor)
+  sender_scope           TEXT NOT NULL DEFAULT 'all',    -- 'all' | 'known'
+  ignored_message_policy TEXT NOT NULL DEFAULT 'drop',   -- 'drop' | 'accumulate'
+  session_mode           TEXT DEFAULT 'shared',
+  priority               INTEGER DEFAULT 0,
+  created_at             TEXT NOT NULL,
   UNIQUE(messaging_group_id, agent_group_id)
 );
 
@@ -116,6 +128,22 @@ CREATE TABLE pending_questions (
   options_json   TEXT NOT NULL,
   created_at     TEXT NOT NULL
 );
+
+-- Pending approvals for unknown senders (unknown_sender_policy='request_approval').
+-- In-flight dedup via UNIQUE(messaging_group_id, sender_identity): a second
+-- message from the same unknown sender while a card is pending is silently
+-- dropped instead of spamming the admin.
+CREATE TABLE pending_sender_approvals (
+  id                 TEXT PRIMARY KEY,
+  messaging_group_id TEXT NOT NULL REFERENCES messaging_groups(id),
+  agent_group_id     TEXT NOT NULL REFERENCES agent_groups(id),
+  sender_identity    TEXT NOT NULL,    -- namespaced user id (channel_type:handle)
+  sender_name        TEXT,
+  original_message   TEXT NOT NULL,    -- JSON of the original InboundEvent
+  approver_user_id   TEXT NOT NULL,
+  created_at         TEXT NOT NULL,
+  UNIQUE(messaging_group_id, sender_identity)
+);
 `;
 
 /**
@@ -138,6 +166,8 @@ CREATE TABLE IF NOT EXISTS messages_in (
   recurrence     TEXT,
   series_id      TEXT,
   tries          INTEGER DEFAULT 0,
+  trigger        INTEGER NOT NULL DEFAULT 1,
+                 -- 0 = accumulated context (don't wake), 1 = wake agent
   platform_id    TEXT,
   channel_type   TEXT,
   thread_id      TEXT,
@@ -213,4 +243,16 @@ CREATE TABLE IF NOT EXISTS session_state (
   value      TEXT NOT NULL,
   updated_at TEXT NOT NULL
 );
+
+-- Current tool-in-flight state. Single-row table (id=1). Container writes on
+-- PreToolUse and clears on PostToolUse / PostToolUseFailure. Host reads in the
+-- sweep to extend the stuck-tolerance window when Bash is running with a
+-- declared timeout > 60s (long-running scripts shouldn't be flagged as stuck).
+CREATE TABLE IF NOT EXISTS container_state (
+  id                       INTEGER PRIMARY KEY CHECK (id = 1),
+  current_tool             TEXT,
+  tool_declared_timeout_ms INTEGER,
+  tool_started_at          TEXT,
+  updated_at               TEXT NOT NULL
+);
 `;
diff --git a/src/db/session-db.ts b/src/db/session-db.ts
index 05104cf..48e9297 100644
--- a/src/db/session-db.ts
+++ b/src/db/session-db.ts
@@ -95,13 +95,19 @@ export function insertMessage(
     content: string;
     processAfter: string | null;
     recurrence: string | null;
+    /**
+     * 1 = wake the agent (default); 0 = accumulate as context only.
+     * Host countDueMessages gates on this; container reads everything.
+     */
+    trigger?: 0 | 1;
   },
 ): void {
   db.prepare(
-    `INSERT INTO messages_in (id, seq, kind, timestamp, status, platform_id, channel_type, thread_id, content, process_after, recurrence, series_id)
-     VALUES (@id, @seq, @kind, @timestamp, 'pending', @platformId, @channelType, @threadId, @content, @processAfter, @recurrence, @id)`,
+    `INSERT INTO messages_in (id, seq, kind, timestamp, status, platform_id, channel_type, thread_id, content, process_after, recurrence, series_id, trigger)
+     VALUES (@id, @seq, @kind, @timestamp, 'pending', @platformId, @channelType, @threadId, @content, @processAfter, @recurrence, @id, @trigger)`,
   ).run({
     ...message,
+    trigger: message.trigger ?? 1,
     seq: nextEvenSeq(db),
   });
 }
@@ -112,6 +118,7 @@ export function countDueMessages(db: Database.Database): number {
       .prepare(
         `SELECT COUNT(*) as count FROM messages_in
        WHERE status = 'pending'
+         AND trigger = 1
          AND (process_after IS NULL OR datetime(process_after) <= datetime('now'))`,
       )
       .get() as { count: number }
@@ -132,10 +139,10 @@ export function getMessageForRetry(
   db: Database.Database,
   messageId: string,
   status: string,
-): { id: string; tries: number } | undefined {
-  return db.prepare('SELECT id, tries FROM messages_in WHERE id = ? AND status = ?').get(messageId, status) as
-    | { id: string; tries: number }
-    | undefined;
+): { id: string; tries: number; processAfter: string | null } | undefined {
+  return db
+    .prepare('SELECT id, tries, process_after as processAfter FROM messages_in WHERE id = ? AND status = ?')
+    .get(messageId, status) as { id: string; tries: number; processAfter: string | null } | undefined;
 }
 
 export function syncProcessingAcks(inDb: Database.Database, outDb: Database.Database): void {
@@ -161,6 +168,45 @@ export function getStuckProcessingIds(outDb: Database.Database): string[] {
   ).map((r) => r.message_id);
 }
 
+export interface ProcessingClaim {
+  message_id: string;
+  status_changed: string;
+}
+
+/** Return processing_ack rows still in 'processing' with their claim timestamps. */
+export function getProcessingClaims(outDb: Database.Database): ProcessingClaim[] {
+  return outDb
+    .prepare("SELECT message_id, status_changed FROM processing_ack WHERE status = 'processing'")
+    .all() as ProcessingClaim[];
+}
+
+export interface ContainerState {
+  current_tool: string | null;
+  tool_declared_timeout_ms: number | null;
+  tool_started_at: string | null;
+}
+
+/**
+ * Read the container's current tool-in-flight state, if any. Returns null
+ * when either the table doesn't exist yet (older session DB) or no tool is
+ * active. Host sweep reads this to widen stuck-detection tolerance while
+ * Bash is running with a long declared timeout.
+ */
+export function getContainerState(outDb: Database.Database): ContainerState | null {
+  try {
+    const row = outDb
+      .prepare(
+        `SELECT current_tool, tool_declared_timeout_ms, tool_started_at
+           FROM container_state WHERE id = 1`,
+      )
+      .get() as ContainerState | undefined;
+    return row ?? null;
+  } catch {
+    // Table not present on older session DBs — treat as "no tool in flight".
+    return null;
+  }
+}
+
 // ---------------------------------------------------------------------------
 // messages_out (read-only from host)
 // ---------------------------------------------------------------------------
@@ -221,10 +267,9 @@ export function migrateDeliveredTable(db: Database.Database): void {
   }
 }
 
-// Adds series_id (groups all occurrences of a recurring task) to pre-existing
-// messages_in tables. No-op on fresh installs where the column is in the schema.
-// Backfills existing rows so cancel/pause/resume queries can rely on
-// series_id IS NOT NULL.
+// Adds columns added to messages_in after the initial v2 schema to
+// pre-existing session DBs. No-op on fresh installs where the columns are
+// in the baseline schema. Backfills existing rows so invariants hold.
 export function migrateMessagesInTable(db: Database.Database): void {
   const cols = new Set(
     (db.prepare("PRAGMA table_info('messages_in')").all() as Array<{ name: string }>).map((c) => c.name),
@@ -234,4 +279,9 @@ export function migrateMessagesInTable(db: Database.Database): void {
     db.prepare('UPDATE messages_in SET series_id = id WHERE series_id IS NULL').run();
     db.prepare('CREATE INDEX IF NOT EXISTS idx_messages_in_series ON messages_in(series_id)').run();
   }
+  if (!cols.has('trigger')) {
+    // All pre-existing rows got written with the old "every inbound wakes
+    // the agent" semantics, so backfill 1 and default 1 for new inserts.
+    db.prepare('ALTER TABLE messages_in ADD COLUMN trigger INTEGER NOT NULL DEFAULT 1').run();
+  }
 }
diff --git a/src/db/sessions.ts b/src/db/sessions.ts
index 01e48cd..504aa26 100644
--- a/src/db/sessions.ts
+++ b/src/db/sessions.ts
@@ -1,5 +1,5 @@
 import type { PendingApproval, PendingQuestion, Session } from '../types.js';
-import { getDb } from './connection.js';
+import { getDb, hasTable } from './connection.js';
 
 // ── Sessions ──
 
@@ -27,6 +27,31 @@ export function findSession(messagingGroupId: string, threadId: string | null):
     .get(messagingGroupId, 'active') as Session | undefined;
 }
 
+/**
+ * Session lookup scoped to a specific agent group. Needed when multiple
+ * agents are wired to the same messaging group + thread (fan-out) — the
+ * plain `findSession` would return whichever agent's session happened to
+ * be first and route to the wrong container.
+ */
+export function findSessionForAgent(
+  agentGroupId: string,
+  messagingGroupId: string,
+  threadId: string | null,
+): Session | undefined {
+  if (threadId) {
+    return getDb()
+      .prepare(
+        "SELECT * FROM sessions WHERE agent_group_id = ? AND messaging_group_id = ? AND thread_id = ? AND status = 'active'",
+      )
+      .get(agentGroupId, messagingGroupId, threadId) as Session | undefined;
+  }
+  return getDb()
+    .prepare(
+      "SELECT * FROM sessions WHERE agent_group_id = ? AND messaging_group_id = ? AND thread_id IS NULL AND status = 'active'",
+    )
+    .get(agentGroupId, messagingGroupId) as Session | undefined;
+}
+
 /** Find an active session scoped to an agent group (ignoring messaging group). */
 export function findSessionByAgentGroup(agentGroupId: string): Session | undefined {
   return getDb()
@@ -72,10 +97,16 @@ export function deleteSession(id: string): void {
 
 // ── Pending Questions ──
 
-export function createPendingQuestion(pq: PendingQuestion): void {
-  getDb()
+/**
+ * Insert a pending question row. Idempotent: when delivery fails and retries,
+ * the second attempt calls this with the same question_id — without `OR
+ * IGNORE` that would throw UNIQUE and prevent the retry from reaching the
+ * actual send step. Returns true if a new row was inserted.
+ */
+export function createPendingQuestion(pq: PendingQuestion): boolean {
+  const result = getDb()
     .prepare(
-      `INSERT INTO pending_questions (question_id, session_id, message_out_id, platform_id, channel_type, thread_id, title, options_json, created_at)
+      `INSERT OR IGNORE INTO pending_questions (question_id, session_id, message_out_id, platform_id, channel_type, thread_id, title, options_json, created_at)
        VALUES (@question_id, @session_id, @message_out_id, @platform_id, @channel_type, @thread_id, @title, @options_json, @created_at)`,
     )
     .run({
@@ -89,6 +120,7 @@ export function createPendingQuestion(pq: PendingQuestion): void {
       options_json: JSON.stringify(pq.options),
       created_at: pq.created_at,
     });
+  return result.changes > 0;
 }
 
 export function getPendingQuestion(questionId: string): PendingQuestion | undefined {
@@ -106,16 +138,21 @@ export function deletePendingQuestion(questionId: string): void {
 
 // ── Pending Approvals ──
 
+/**
+ * Insert a pending approval row. Idempotent for the same reason as
+ * createPendingQuestion: delivery retries with the same approval_id must not
+ * fail on UNIQUE before the send step gets a chance to succeed.
+ */
 export function createPendingApproval(
   pa: Partial<PendingApproval> &
     Pick<
       PendingApproval,
       'approval_id' | 'request_id' | 'action' | 'payload' | 'created_at' | 'title' | 'options_json'
     >,
-): void {
-  getDb()
+): boolean {
+  const result = getDb()
     .prepare(
-      `INSERT INTO pending_approvals
+      `INSERT OR IGNORE INTO pending_approvals
          (approval_id, session_id, request_id, action, payload, created_at,
           agent_group_id, channel_type, platform_id, platform_message_id, expires_at, status,
           title, options_json)
@@ -134,6 +171,7 @@ export function createPendingApproval(
       status: 'pending',
       ...pa,
     });
+  return result.changes > 0;
 }
 
 export function getPendingApproval(approvalId: string): PendingApproval | undefined {
@@ -167,6 +205,23 @@ export function getAskQuestionRender(
   const a = getDb().prepare('SELECT title, options_json FROM pending_approvals WHERE approval_id = ?').get(id) as
     | { title: string; options_json: string }
     | undefined;
-  if (!a || !a.title) return undefined;
-  return { title: a.title, options: JSON.parse(a.options_json) };
+  if (a?.title) return { title: a.title, options: JSON.parse(a.options_json) };
+
+  // Channel-registration + unknown-sender approvals persist title/options_json
+  // the same way pending_approvals does — just SELECT and return.
+  if (hasTable(getDb(), 'pending_channel_approvals')) {
+    const c = getDb()
+      .prepare('SELECT title, options_json FROM pending_channel_approvals WHERE messaging_group_id = ?')
+      .get(id) as { title: string; options_json: string } | undefined;
+    if (c?.title) return { title: c.title, options: JSON.parse(c.options_json) };
+  }
+
+  if (hasTable(getDb(), 'pending_sender_approvals')) {
+    const s = getDb().prepare('SELECT title, options_json FROM pending_sender_approvals WHERE id = ?').get(id) as
+      | { title: string; options_json: string }
+      | undefined;
+    if (s?.title) return { title: s.title, options: JSON.parse(s.options_json) };
+  }
+
+  return undefined;
 }
diff --git a/src/delivery.test.ts b/src/delivery.test.ts
index d631836..a5e1efd 100644
--- a/src/delivery.test.ts
+++ b/src/delivery.test.ts
@@ -14,7 +14,6 @@ import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest';
 
 vi.mock('./container-runner.js', () => ({
   wakeContainer: vi.fn().mockResolvedValue(undefined),
-  resetContainerIdleTimer: vi.fn(),
   isContainerRunning: vi.fn().mockReturnValue(false),
   killContainer: vi.fn(),
   buildAgentGroupImage: vi.fn().mockResolvedValue(undefined),
diff --git a/src/delivery.ts b/src/delivery.ts
index 7b1ee7d..036153a 100644
--- a/src/delivery.ts
+++ b/src/delivery.ts
@@ -23,7 +23,6 @@ import {
 import { log } from './log.js';
 import { normalizeOptions } from './channels/ask-question.js';
 import { clearOutbox, openInboundDb, openOutboundDb, readOutboxFiles } from './session-manager.js';
-import { resetContainerIdleTimer } from './container-runner.js';
 import { pauseTypingRefreshAfterDelivery, setTypingAdapter } from './modules/typing/index.js';
 import type { OutboundFile } from './channels/adapter.js';
 import type { Session } from './types.js';
@@ -193,7 +192,6 @@ async function drainSession(session: Session): Promise<void> {
         const platformMsgId = await deliverMessage(msg, session, inDb);
         markDelivered(inDb, msg.id, platformMsgId ?? null);
         deliveryAttempts.delete(msg.id);
-        resetContainerIdleTimer(session.id);
 
         // Pause the typing indicator after a real user-facing message
         // lands on the user's screen, so the client has time to visually
@@ -323,7 +321,7 @@ async function deliverMessage(
         questionId: content.questionId,
       });
     } else {
-      createPendingQuestion({
+      const inserted = createPendingQuestion({
         question_id: content.questionId,
         session_id: session.id,
         message_out_id: msg.id,
@@ -334,7 +332,9 @@ async function deliverMessage(
         options: normalizeOptions(rawOptions as never),
         created_at: new Date().toISOString(),
       });
-      log.info('Pending question created', { questionId: content.questionId, sessionId: session.id });
+      if (inserted) {
+        log.info('Pending question created', { questionId: content.questionId, sessionId: session.id });
+      }
     }
   }
 
diff --git a/src/group-init.ts b/src/group-init.ts
index 527ba6b..437d10f 100644
--- a/src/group-init.ts
+++ b/src/group-init.ts
@@ -6,18 +6,6 @@ import { initContainerConfig } from './container-config.js';
 import { log } from './log.js';
 import type { AgentGroup } from './types.js';
 
-// Container path where groups/global is mounted. The symlink we drop
-// into each group's dir resolves to this target inside the container.
-// It's a dangling symlink on the host — that's fine, host tools don't
-// follow it and the container mount makes it valid at read time.
-const GLOBAL_MEMORY_CONTAINER_PATH = '/workspace/global/CLAUDE.md';
-
-// Symlink name inside the group's dir. Claude Code's @-import only
-// follows paths inside cwd, so we can't reference /workspace/global
-// directly — we symlink into the group dir and import the symlink.
-export const GLOBAL_MEMORY_LINK_NAME = '.claude-global.md';
-export const GLOBAL_CLAUDE_IMPORT = `@./${GLOBAL_MEMORY_LINK_NAME}`;
-
 const DEFAULT_SETTINGS_JSON =
   JSON.stringify(
     {
@@ -36,13 +24,17 @@ const DEFAULT_SETTINGS_JSON =
  * every step is gated on the target not already existing, so re-running on
  * an already-initialized group is a no-op.
  *
- * Called once per group lifetime: at creation, or defensively from
- * `buildMounts()` for groups that pre-date this code path. After init, the
- * host never overwrites any of these paths automatically — agents own them.
- * To pull in upstream changes, use the host-mediated reset/refresh tools.
+ * Called once per group lifetime at creation, or defensively from
+ * `buildMounts()` for groups that pre-date this code path.
+ *
+ * Source code and skills are shared RO mounts — not copied per-group.
+ * Skill symlinks are synced at spawn time by container-runner.ts.
+ *
+ * The composed `CLAUDE.md` is NOT written here — it's regenerated on every
+ * spawn by `composeGroupClaudeMd()` (see `claude-md-compose.ts`). Initial
+ * per-group instructions (if provided) seed `CLAUDE.local.md`.
  */
 export function initGroupFilesystem(group: AgentGroup, opts?: { instructions?: string }): void {
-  const projectRoot = process.cwd();
   const initialized: string[] = [];
 
   // 1. groups/<folder>/ — group memory + working dir
@@ -52,29 +44,13 @@ export function initGroupFilesystem(group: AgentGroup, opts?: { instructions?: s
     initialized.push('groupDir');
   }
 
-  // groups/<folder>/.claude-global.md — symlink into the group dir so
-  // Claude Code's @-import can follow it. Uses lstat to avoid tripping
-  // existsSync on a dangling symlink (target only resolves inside the
-  // container).
-  const globalLinkPath = path.join(groupDir, GLOBAL_MEMORY_LINK_NAME);
-  let linkExists = false;
-  try {
-    fs.lstatSync(globalLinkPath);
-    linkExists = true;
-  } catch {
-    /* missing — recreate */
-  }
-  if (!linkExists) {
-    fs.symlinkSync(GLOBAL_MEMORY_CONTAINER_PATH, globalLinkPath);
-    initialized.push('.claude-global.md');
-  }
-
-  // groups/<folder>/CLAUDE.md — written once, then owned by the group
-  const claudeMdFile = path.join(groupDir, 'CLAUDE.md');
-  if (!fs.existsSync(claudeMdFile)) {
-    const body = [GLOBAL_CLAUDE_IMPORT, '', opts?.instructions ?? `# ${group.name}`].join('\n') + '\n';
-    fs.writeFileSync(claudeMdFile, body);
-    initialized.push('CLAUDE.md');
+  // groups/<folder>/CLAUDE.local.md — per-group agent memory, auto-loaded by
+  // Claude Code. Seeded with caller-provided instructions on first creation.
+  const claudeLocalFile = path.join(groupDir, 'CLAUDE.local.md');
+  if (!fs.existsSync(claudeLocalFile)) {
+    const body = opts?.instructions ? opts.instructions + '\n' : '';
+    fs.writeFileSync(claudeLocalFile, body);
+    initialized.push('CLAUDE.local.md');
   }
 
   // groups/<folder>/container.json — empty container config, replaces the
@@ -97,23 +73,12 @@ export function initGroupFilesystem(group: AgentGroup, opts?: { instructions?: s
     initialized.push('settings.json');
   }
 
+  // Skills directory — created empty here; symlinks are synced at spawn
+  // time by container-runner.ts based on container.json skills selection.
   const skillsDst = path.join(claudeDir, 'skills');
   if (!fs.existsSync(skillsDst)) {
-    const skillsSrc = path.join(projectRoot, 'container', 'skills');
-    if (fs.existsSync(skillsSrc)) {
-      fs.cpSync(skillsSrc, skillsDst, { recursive: true });
-      initialized.push('skills/');
-    }
-  }
-
-  // 3. data/v2-sessions/<id>/agent-runner-src/ — per-group source copy
-  const groupRunnerDir = path.join(DATA_DIR, 'v2-sessions', group.id, 'agent-runner-src');
-  if (!fs.existsSync(groupRunnerDir)) {
-    const agentRunnerSrc = path.join(projectRoot, 'container', 'agent-runner', 'src');
-    if (fs.existsSync(agentRunnerSrc)) {
-      fs.cpSync(agentRunnerSrc, groupRunnerDir, { recursive: true });
-      initialized.push('agent-runner-src/');
-    }
+    fs.mkdirSync(skillsDst, { recursive: true });
+    initialized.push('skills/');
   }
 
   if (initialized.length > 0) {
diff --git a/src/host-core.test.ts b/src/host-core.test.ts
index a8b4684..9906c4b 100644
--- a/src/host-core.test.ts
+++ b/src/host-core.test.ts
@@ -25,12 +25,11 @@ import {
   outboundDbPath,
 } from './session-manager.js';
 import { getSession, findSession } from './db/sessions.js';
-import type { InboundEvent } from './router.js';
+import type { InboundEvent } from './channels/adapter.js';
 
 // Mock container runner to prevent actual Docker spawning
 vi.mock('./container-runner.js', () => ({
   wakeContainer: vi.fn().mockResolvedValue(undefined),
-  resetContainerIdleTimer: vi.fn(),
   isContainerRunning: vi.fn().mockReturnValue(false),
   getActiveContainerCount: vi.fn().mockReturnValue(0),
   killContainer: vi.fn(),
@@ -200,8 +199,10 @@ describe('router', () => {
       id: 'mga-1',
       messaging_group_id: 'mg-1',
       agent_group_id: 'ag-1',
-      trigger_rules: null,
-      response_scope: 'all',
+      engage_mode: 'pattern',
+      engage_pattern: '.',
+      sender_scope: 'all',
+      ignored_message_policy: 'drop',
       session_mode: 'shared',
       priority: 0,
       created_at: now(),
@@ -243,26 +244,42 @@ describe('router', () => {
     expect(wakeContainer).toHaveBeenCalled();
   });
 
-  it('should auto-create messaging group for unknown platform', async () => {
+  it('auto-creates messaging group only when the bot is addressed (mention/DM)', async () => {
+    // The router's no-mg branch is escalation-gated: plain chatter on an
+    // unknown channel stays silent (no DB writes) so a bot that sits in
+    // many unwired channels doesn't bloat messaging_groups. Only explicit
+    // mentions and DMs trigger auto-create.
     const { routeInbound } = await import('./router.js');
+    const { getMessagingGroupByPlatform } = await import('./db/messaging-groups.js');
 
-    const event: InboundEvent = {
+    // Plain message on unknown channel — should NOT auto-create.
+    await routeInbound({
       channelType: 'slack',
-      platformId: 'C-NEW-CHANNEL',
+      platformId: 'C-PLAIN',
       threadId: null,
       message: {
-        id: 'msg-2',
+        id: 'msg-plain',
         kind: 'chat',
         content: JSON.stringify({ sender: 'User', text: 'Hi' }),
         timestamp: now(),
       },
-    };
+    });
+    expect(getMessagingGroupByPlatform('slack', 'C-PLAIN')).toBeUndefined();
 
-    await routeInbound(event);
-
-    const { getMessagingGroupByPlatform } = await import('./db/messaging-groups.js');
-    const mg = getMessagingGroupByPlatform('slack', 'C-NEW-CHANNEL');
-    expect(mg).toBeDefined();
+    // Mention on unknown channel — SHOULD auto-create (next step: channel-registration flow).
+    await routeInbound({
+      channelType: 'slack',
+      platformId: 'C-MENTIONED',
+      threadId: null,
+      message: {
+        id: 'msg-mentioned',
+        kind: 'chat',
+        content: JSON.stringify({ sender: 'User', text: '@bot hi' }),
+        timestamp: now(),
+        isMention: true,
+      },
+    });
+    expect(getMessagingGroupByPlatform('slack', 'C-MENTIONED')).toBeDefined();
   });
 
   it('should route multiple messages to the same session', async () => {
@@ -296,6 +313,106 @@ describe('router', () => {
 
     expect(rows).toHaveLength(2);
   });
+
+  it('fans out to every matching agent, each in its own session', async () => {
+    const { routeInbound } = await import('./router.js');
+    const { wakeContainer } = await import('./container-runner.js');
+    (wakeContainer as unknown as ReturnType<typeof vi.fn>).mockClear();
+
+    // Wire a second agent to the same messaging group.
+    createAgentGroup({
+      id: 'ag-2',
+      name: 'Secondary Agent',
+      folder: 'secondary-agent',
+      agent_provider: null,
+      created_at: now(),
+    });
+    createMessagingGroupAgent({
+      id: 'mga-2',
+      messaging_group_id: 'mg-1',
+      agent_group_id: 'ag-2',
+      engage_mode: 'pattern',
+      engage_pattern: '.',
+      sender_scope: 'all',
+      ignored_message_policy: 'drop',
+      session_mode: 'shared',
+      priority: 0,
+      created_at: now(),
+    });
+
+    await routeInbound({
+      channelType: 'discord',
+      platformId: 'chan-123',
+      threadId: null,
+      message: { id: 'msg-fan', kind: 'chat', content: JSON.stringify({ text: 'hello all' }), timestamp: now() },
+    });
+
+    // Both agents should now have their own session and be woken.
+    expect(wakeContainer).toHaveBeenCalledTimes(2);
+
+    const { getSessionsByAgentGroup } = await import('./db/sessions.js');
+    expect(getSessionsByAgentGroup('ag-1')).toHaveLength(1);
+    expect(getSessionsByAgentGroup('ag-2')).toHaveLength(1);
+  });
+
+  it('accumulates without waking when engage fails + ignored_message_policy=accumulate', async () => {
+    const { routeInbound } = await import('./router.js');
+    const { wakeContainer } = await import('./container-runner.js');
+    (wakeContainer as unknown as ReturnType<typeof vi.fn>).mockClear();
+
+    // Replace the seed row with a mention-only wiring whose accumulate
+    // policy should store context even when the message doesn't mention us.
+    const { updateMessagingGroupAgent } = await import('./db/messaging-groups.js');
+    updateMessagingGroupAgent('mga-1', {
+      engage_mode: 'mention',
+      ignored_message_policy: 'accumulate',
+    });
+
+    await routeInbound({
+      channelType: 'discord',
+      platformId: 'chan-123',
+      threadId: null,
+      message: {
+        id: 'msg-nomatch',
+        kind: 'chat',
+        content: JSON.stringify({ text: 'no mention here' }),
+        timestamp: now(),
+      },
+    });
+
+    expect(wakeContainer).not.toHaveBeenCalled();
+
+    const session = findSession('mg-1', null);
+    expect(session).toBeDefined();
+    const db = new Database(inboundDbPath('ag-1', session!.id));
+    const rows = db.prepare('SELECT id, trigger FROM messages_in').all() as Array<{
+      id: string;
+      trigger: number;
+    }>;
+    db.close();
+    expect(rows).toHaveLength(1);
+    expect(rows[0].trigger).toBe(0);
+  });
+
+  it('drops silently when engage fails + ignored_message_policy=drop', async () => {
+    const { routeInbound } = await import('./router.js');
+    const { wakeContainer } = await import('./container-runner.js');
+    (wakeContainer as unknown as ReturnType<typeof vi.fn>).mockClear();
+
+    const { updateMessagingGroupAgent } = await import('./db/messaging-groups.js');
+    updateMessagingGroupAgent('mga-1', { engage_mode: 'mention' }); // drop is the default
+
+    await routeInbound({
+      channelType: 'discord',
+      platformId: 'chan-123',
+      threadId: null,
+      message: { id: 'msg-drop', kind: 'chat', content: JSON.stringify({ text: 'ignored' }), timestamp: now() },
+    });
+
+    expect(wakeContainer).not.toHaveBeenCalled();
+    // No session should have been created for this agent.
+    expect(findSession('mg-1', null)).toBeUndefined();
+  });
 });
 
 describe('delivery', () => {
diff --git a/src/host-sweep.test.ts b/src/host-sweep.test.ts
new file mode 100644
index 0000000..eefcc8a
--- /dev/null
+++ b/src/host-sweep.test.ts
@@ -0,0 +1,146 @@
+/**
+ * Unit tests for the stuck-container decision logic introduced by
+ * ACTION-ITEMS item 9. Lives on the pure helper `decideStuckAction` so we
+ * don't have to mock the filesystem or the container runner.
+ */
+import { describe, expect, it } from 'vitest';
+
+import { ABSOLUTE_CEILING_MS, CLAIM_STUCK_MS, decideStuckAction } from './host-sweep.js';
+
+const BASE = Date.parse('2026-04-20T12:00:00.000Z');
+
+function claim(id: string, offsetMs: number) {
+  return { message_id: id, status_changed: new Date(BASE - offsetMs).toISOString() };
+}
+
+describe('decideStuckAction', () => {
+  it('returns ok when heartbeat is fresh and no claims', () => {
+    expect(
+      decideStuckAction({
+        now: BASE,
+        heartbeatMtimeMs: BASE - 5_000,
+        containerState: null,
+        claims: [],
+      }),
+    ).toEqual({ action: 'ok' });
+  });
+
+  it('returns kill-ceiling when heartbeat older than 30 min', () => {
+    const heartbeatMtimeMs = BASE - ABSOLUTE_CEILING_MS - 1_000;
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs,
+      containerState: null,
+      claims: [],
+    });
+    expect(res.action).toBe('kill-ceiling');
+    if (res.action !== 'kill-ceiling') return;
+    expect(res.ceilingMs).toBe(ABSOLUTE_CEILING_MS);
+    expect(res.heartbeatAgeMs).toBeGreaterThan(ABSOLUTE_CEILING_MS);
+  });
+
+  it('skips the ceiling check when no heartbeat file exists (fresh container not yet ticked)', () => {
+    // A freshly-spawned container hasn't produced any SDK events yet, so no
+    // heartbeat. Prior behavior treated this as infinitely stale and killed
+    // every container within seconds of spawn. With no claims either, we
+    // should conclude everything is fine.
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs: 0,
+      containerState: null,
+      claims: [],
+    });
+    expect(res.action).toBe('ok');
+  });
+
+  it('kills on claim-stuck when heartbeat is absent AND a claim has aged past tolerance', () => {
+    // Hanging fresh container: spawned, picked up a message (claim recorded
+    // in processing_ack), but never wrote a heartbeat. Falls through the
+    // skipped ceiling check into claim-stuck — which correctly fires.
+    const claimedAgeMs = CLAIM_STUCK_MS + 5_000;
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs: 0,
+      containerState: null,
+      claims: [claim('msg-1', claimedAgeMs)],
+    });
+    expect(res.action).toBe('kill-claim');
+  });
+
+  it('extends the ceiling when Bash has a declared timeout longer than 30 min', () => {
+    const twoHrMs = 2 * 60 * 60 * 1000;
+    const res = decideStuckAction({
+      now: BASE,
+      // 45 min — over the default ceiling, but under the Bash timeout
+      heartbeatMtimeMs: BASE - 45 * 60 * 1000,
+      containerState: {
+        current_tool: 'Bash',
+        tool_declared_timeout_ms: twoHrMs,
+        tool_started_at: new Date(BASE - 45 * 60 * 1000).toISOString(),
+      },
+      claims: [],
+    });
+    expect(res.action).toBe('ok');
+  });
+
+  it('returns kill-claim when a claim is past 60s and heartbeat has not moved', () => {
+    const claimedAgeMs = CLAIM_STUCK_MS + 10_000;
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs: BASE - claimedAgeMs - 5_000, // older than the claim
+      containerState: null,
+      claims: [claim('msg-1', claimedAgeMs)],
+    });
+    expect(res.action).toBe('kill-claim');
+    if (res.action !== 'kill-claim') return;
+    expect(res.messageId).toBe('msg-1');
+    expect(res.toleranceMs).toBe(CLAIM_STUCK_MS);
+  });
+
+  it('does not kill when heartbeat has been touched since the claim', () => {
+    const claimedAgeMs = CLAIM_STUCK_MS + 10_000;
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs: BASE - 2_000, // fresh, updated after the claim
+      containerState: null,
+      claims: [claim('msg-1', claimedAgeMs)],
+    });
+    expect(res.action).toBe('ok');
+  });
+
+  it('does not kill when claim age is below tolerance', () => {
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs: BASE - CLAIM_STUCK_MS - 10_000, // old, but claim is recent
+      containerState: null,
+      claims: [claim('msg-1', 5_000)],
+    });
+    expect(res.action).toBe('ok');
+  });
+
+  it('widens per-claim tolerance for a running Bash with long timeout', () => {
+    const tenMinMs = 10 * 60 * 1000;
+    const res = decideStuckAction({
+      now: BASE,
+      // 5 min since claim, over the 60s default but under the declared Bash timeout
+      heartbeatMtimeMs: BASE - 5 * 60 * 1000 - 5_000,
+      containerState: {
+        current_tool: 'Bash',
+        tool_declared_timeout_ms: tenMinMs,
+        tool_started_at: new Date(BASE - 5 * 60 * 1000).toISOString(),
+      },
+      claims: [claim('msg-1', 5 * 60 * 1000)],
+    });
+    expect(res.action).toBe('ok');
+  });
+
+  it('ignores claims with unparseable timestamps', () => {
+    const res = decideStuckAction({
+      now: BASE,
+      heartbeatMtimeMs: BASE - 5_000,
+      containerState: null,
+      claims: [{ message_id: 'x', status_changed: 'not-a-date' }],
+    });
+    expect(res.action).toBe('ok');
+  });
+});
diff --git a/src/host-sweep.ts b/src/host-sweep.ts
index 7a7688f..4dc2fb7 100644
--- a/src/host-sweep.ts
+++ b/src/host-sweep.ts
@@ -2,10 +2,29 @@
  * Host sweep — periodic maintenance of all session DBs.
  *
  * Two-DB architecture:
- *   - Reads processing_ack from outbound.db to sync message status
- *   - Writes to inbound.db (host-owned) for status updates and recurrence
- *   - Uses heartbeat file mtime for stale container detection (not DB writes)
+ *   - Reads processing_ack + container_state from outbound.db
+ *   - Writes to inbound.db (host-owned) for status updates + recurrence
+ *   - Uses heartbeat file mtime for liveness (never polls DB for it)
  *   - Never writes to outbound.db — preserves single-writer-per-file invariant
+ *
+ * Stuck / idle detection (replaces the old IDLE_TIMEOUT setTimeout + 10-min
+ * heartbeat threshold):
+ *
+ *   If the container isn't running and there are 'processing' rows left over
+ *   (e.g. it crashed mid-turn) → reset them to pending with backoff +
+ *   tries++. Existing retry machinery does the rest.
+ *
+ *   If the container IS running:
+ *     1. Absolute ceiling: heartbeat age > max(30 min, current_bash_timeout)
+ *        → kill. Covers the "alive but silent for 30 min" case. Extended
+ *        only while Bash is declared as running longer, honouring the
+ *        user's own timeout directive. Kill then resets processing rows.
+ *
+ *     2. Message-scoped stuck: for each 'processing' row, tolerance =
+ *        max(60s, current_bash_timeout_ms_if_Bash_running). If
+ *        (claim_age > tolerance) AND (heartbeat_mtime <= status_changed)
+ *        → kill + reset this message + tries++. Semantics: "container
+ *        claimed a message and went quiet past tolerance since the claim."
  */
 import type Database from 'better-sqlite3';
 import fs from 'fs';
@@ -14,22 +33,78 @@ import { getActiveSessions } from './db/sessions.js';
 import { getAgentGroup } from './db/agent-groups.js';
 import {
   countDueMessages,
-  syncProcessingAcks,
-  getStuckProcessingIds,
+  getContainerState,
   getMessageForRetry,
+  getProcessingClaims,
   markMessageFailed,
   retryWithBackoff,
+  syncProcessingAcks,
+  type ContainerState,
 } from './db/session-db.js';
 import { log } from './log.js';
 import { openInboundDb, openOutboundDb, inboundDbPath, heartbeatPath } from './session-manager.js';
-import { wakeContainer, isContainerRunning } from './container-runner.js';
+import { isContainerRunning, killContainer, wakeContainer } from './container-runner.js';
 import type { Session } from './types.js';
 
 const SWEEP_INTERVAL_MS = 60_000;
-const STALE_THRESHOLD_MS = 10 * 60 * 1000; // 10 minutes
+// Absolute idle ceiling for a running container. If the heartbeat file hasn't
+// been touched in this long, the container is either stuck or doing genuinely
+// nothing — kill and restart on the next inbound.
+export const ABSOLUTE_CEILING_MS = 30 * 60 * 1000;
+// Stuck tolerance window applied per 'processing' claim — "did we see any
+// signs of life since this message was claimed?"
+export const CLAIM_STUCK_MS = 60 * 1000;
 const MAX_TRIES = 5;
 const BACKOFF_BASE_MS = 5000;
 
+export type StuckDecision =
+  | { action: 'ok' }
+  | { action: 'kill-ceiling'; heartbeatAgeMs: number; ceilingMs: number }
+  | { action: 'kill-claim'; messageId: string; claimAgeMs: number; toleranceMs: number };
+
+/**
+ * Pure decision for whether a running container should be killed this sweep
+ * tick. Inputs are all deterministic; filesystem + DB reads happen in the
+ * caller.
+ */
+export function decideStuckAction(args: {
+  now: number;
+  heartbeatMtimeMs: number; // 0 when heartbeat file absent
+  containerState: ContainerState | null;
+  claims: Array<{ message_id: string; status_changed: string }>;
+}): StuckDecision {
+  const { now, heartbeatMtimeMs, containerState, claims } = args;
+  const declaredBashMs = bashTimeoutMs(containerState);
+
+  // Ceiling check only applies when we have an actual heartbeat timestamp.
+  // A freshly-spawned container hasn't had any SDK activity yet so no
+  // heartbeat file exists — if we treated that as infinitely stale we'd
+  // kill every container within seconds of spawn. Genuinely-dead containers
+  // that never wrote a heartbeat are caught by the separate "container
+  // process not running" cleanup path, not here. If a fresh container is
+  // hanging at the gate (claimed a message but never did anything) the
+  // claim-stuck check below handles it.
+  if (heartbeatMtimeMs !== 0) {
+    const heartbeatAge = now - heartbeatMtimeMs;
+    const ceiling = Math.max(ABSOLUTE_CEILING_MS, declaredBashMs ?? 0);
+    if (heartbeatAge > ceiling) {
+      return { action: 'kill-ceiling', heartbeatAgeMs: heartbeatAge, ceilingMs: ceiling };
+    }
+  }
+
+  const tolerance = Math.max(CLAIM_STUCK_MS, declaredBashMs ?? 0);
+  for (const claim of claims) {
+    const claimedAt = Date.parse(claim.status_changed);
+    if (Number.isNaN(claimedAt)) continue;
+    const claimAge = now - claimedAt;
+    if (claimAge <= tolerance) continue;
+    if (heartbeatMtimeMs > claimedAt) continue;
+    return { action: 'kill-claim', messageId: claim.message_id, claimAgeMs: claimAge, toleranceMs: tolerance };
+  }
+
+  return { action: 'ok' };
+}
+
 let running = false;
 
 export function startHostSweep(): void {
@@ -84,20 +159,34 @@ async function sweepSession(session: Session): Promise<void> {
       syncProcessingAcks(inDb, outDb);
     }
 
-    // 2. Check for due pending messages → wake container
+    // 2. Wake a container if work is due and nothing is running. Ordered
+    // before the crashed-container cleanup so a fresh container gets a chance
+    // to clean its own orphan processing_ack rows on startup (see
+    // container/agent-runner/src/db/connection.ts). Otherwise the reset path
+    // would keep bumping process_after into the future, dueCount would stay 0,
+    // and the wake would never fire.
     const dueCount = countDueMessages(inDb);
-
     if (dueCount > 0 && !isContainerRunning(session.id)) {
       log.info('Waking container for due messages', { sessionId: session.id, count: dueCount });
       await wakeContainer(session);
     }
 
-    // 3. Detect stale containers via heartbeat file
-    if (outDb) {
-      detectStaleContainers(inDb, outDb, session, agentGroup.id);
+    const alive = isContainerRunning(session.id);
+
+    // 3. Running-container SLA: absolute ceiling + per-claim stuck rules.
+    if (alive && outDb) {
+      enforceRunningContainerSla(inDb, outDb, session, agentGroup.id);
     }
 
-    // 4. Handle recurrence for completed messages.
+    // 4. Crashed-container cleanup: processing rows left behind get retried.
+    // Only fires when wake in step 2 didn't pick up the work (no due messages,
+    // or wake failed). resetStuckProcessingRows itself is idempotent — it
+    // skips messages already scheduled for a future retry.
+    if (!alive && outDb) {
+      resetStuckProcessingRows(inDb, outDb, session, 'container not running');
+    }
+
+    // 5. Recurrence fanout for completed recurring tasks.
     // MODULE-HOOK:scheduling-recurrence:start
     const { handleRecurrence } = await import('./modules/scheduling/recurrence.js');
     await handleRecurrence(inDb, session);
@@ -108,45 +197,90 @@ async function sweepSession(session: Session): Promise<void> {
   }
 }
 
-/**
- * Detect stale containers using heartbeat file mtime.
- * If the heartbeat is older than STALE_THRESHOLD and processing_ack has
- * 'processing' entries, the container likely crashed — reset with backoff.
- */
-function detectStaleContainers(
+function heartbeatMtimeMs(agentGroupId: string, sessionId: string): number {
+  const hbPath = heartbeatPath(agentGroupId, sessionId);
+  try {
+    return fs.statSync(hbPath).mtimeMs;
+  } catch {
+    return 0;
+  }
+}
+
+function bashTimeoutMs(state: ContainerState | null): number | null {
+  if (!state || state.current_tool !== 'Bash') return null;
+  return typeof state.tool_declared_timeout_ms === 'number' ? state.tool_declared_timeout_ms : null;
+}
+
+function enforceRunningContainerSla(
   inDb: Database.Database,
   outDb: Database.Database,
   session: Session,
   agentGroupId: string,
 ): void {
-  const hbPath = heartbeatPath(agentGroupId, session.id);
-  let heartbeatAge = Infinity;
-  try {
-    const stat = fs.statSync(hbPath);
-    heartbeatAge = Date.now() - stat.mtimeMs;
-  } catch {
-    // No heartbeat file — container may never have started, or it's very old
+  const decision = decideStuckAction({
+    now: Date.now(),
+    heartbeatMtimeMs: heartbeatMtimeMs(agentGroupId, session.id),
+    containerState: getContainerState(outDb),
+    claims: getProcessingClaims(outDb),
+  });
+
+  if (decision.action === 'ok') return;
+
+  if (decision.action === 'kill-ceiling') {
+    log.warn('Killing container past absolute ceiling', {
+      sessionId: session.id,
+      heartbeatAgeMs: decision.heartbeatAgeMs,
+      ceilingMs: decision.ceilingMs,
+    });
+    killContainer(session.id, 'absolute-ceiling');
+    resetStuckProcessingRows(inDb, outDb, session, 'absolute-ceiling');
+    return;
   }
 
-  if (heartbeatAge < STALE_THRESHOLD_MS) return; // Container is alive
+  log.warn('Killing container — message claimed then silent', {
+    sessionId: session.id,
+    messageId: decision.messageId,
+    claimAgeMs: decision.claimAgeMs,
+    toleranceMs: decision.toleranceMs,
+  });
+  killContainer(session.id, 'claim-stuck');
+  resetStuckProcessingRows(inDb, outDb, session, 'claim-stuck');
+}
 
-  // Heartbeat is stale — check for stuck processing entries
-  const processingIds = getStuckProcessingIds(outDb);
-  if (processingIds.length === 0) return;
-
-  for (const messageId of processingIds) {
-    const msg = getMessageForRetry(inDb, messageId, 'pending');
+function resetStuckProcessingRows(
+  inDb: Database.Database,
+  outDb: Database.Database,
+  session: Session,
+  reason: string,
+): void {
+  const claims = getProcessingClaims(outDb);
+  const now = Date.now();
+  for (const { message_id } of claims) {
+    const msg = getMessageForRetry(inDb, message_id, 'pending');
     if (!msg) continue;
 
+    // Already rescheduled for a future retry — don't bump tries again. The
+    // wake path (sweep step 2) will fire when process_after elapses and a
+    // fresh container will clean the orphan claim on startup.
+    if (msg.processAfter && Date.parse(msg.processAfter) > now) continue;
+
     if (msg.tries >= MAX_TRIES) {
       markMessageFailed(inDb, msg.id);
-      log.warn('Message marked as failed after max retries', { messageId: msg.id, sessionId: session.id });
+      log.warn('Message marked as failed after max retries', {
+        messageId: msg.id,
+        sessionId: session.id,
+        reason,
+      });
     } else {
       const backoffMs = BACKOFF_BASE_MS * Math.pow(2, msg.tries);
       const backoffSec = Math.floor(backoffMs / 1000);
       retryWithBackoff(inDb, msg.id, backoffSec);
-      log.info('Reset stale message with backoff', { messageId: msg.id, tries: msg.tries, backoffMs });
+      log.info('Reset stale message with backoff', {
+        messageId: msg.id,
+        tries: msg.tries,
+        backoffMs,
+        reason,
+      });
     }
   }
 }
-
diff --git a/src/index.ts b/src/index.ts
index ffb2731..ea9fba6 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -7,9 +7,9 @@
 import path from 'path';
 
 import { DATA_DIR } from './config.js';
+import { migrateGroupsToClaudeLocal } from './claude-md-compose.js';
 import { initDb } from './db/connection.js';
 import { runMigrations } from './db/migrations/index.js';
-import { getMessagingGroupsByChannel, getMessagingGroupAgents } from './db/messaging-groups.js';
 import { ensureContainerRuntimeRunning, cleanupOrphans } from './container-runtime.js';
 import { startActiveDeliveryPoll, startSweepDeliveryPoll, setDeliveryAdapter, stopDeliveryPolls } from './delivery.js';
 import { startHostSweep, stopHostSweep } from './host-sweep.js';
@@ -52,7 +52,7 @@ import './channels/index.js';
 // append registry-based modules. Imported for side effects (registrations).
 import './modules/index.js';
 
-import type { ChannelAdapter, ChannelSetup, ConversationConfig } from './channels/adapter.js';
+import type { ChannelAdapter, ChannelSetup } from './channels/adapter.js';
 import { initChannelAdapters, teardownChannelAdapters, getChannelAdapter } from './channels/channel-registry.js';
 
 async function main(): Promise<void> {
@@ -64,15 +64,16 @@ async function main(): Promise<void> {
   runMigrations(db);
   log.info('Central DB ready', { path: dbPath });
 
+  // 1b. One-time filesystem cutover — idempotent, no-op after first run.
+  migrateGroupsToClaudeLocal();
+
   // 2. Container runtime
   ensureContainerRuntimeRunning();
   cleanupOrphans();
 
   // 3. Channel adapters
   await initChannelAdapters((adapter: ChannelAdapter): ChannelSetup => {
-    const conversations = buildConversationConfigs(adapter.channelType);
     return {
-      conversations,
       onInbound(platformId, threadId, message) {
         routeInbound({
           channelType: adapter.channelType,
@@ -83,11 +84,22 @@ async function main(): Promise<void> {
             kind: message.kind,
             content: JSON.stringify(message.content),
             timestamp: message.timestamp,
+            isMention: message.isMention,
+            isGroup: message.isGroup,
           },
         }).catch((err) => {
           log.error('Failed to route inbound message', { channelType: adapter.channelType, err });
         });
       },
+      onInboundEvent(event) {
+        routeInbound(event).catch((err) => {
+          log.error('Failed to route inbound event', {
+            sourceAdapter: adapter.channelType,
+            targetChannelType: event.channelType,
+            err,
+          });
+        });
+      },
       onMetadata(platformId, name, isGroup) {
         log.info('Channel metadata discovered', {
           channelType: adapter.channelType,
@@ -150,28 +162,6 @@ async function main(): Promise<void> {
   log.info('NanoClaw running');
 }
 
-/** Build ConversationConfig[] for a channel type from the central DB. */
-function buildConversationConfigs(channelType: string): ConversationConfig[] {
-  const groups = getMessagingGroupsByChannel(channelType);
-  const configs: ConversationConfig[] = [];
-
-  for (const mg of groups) {
-    const agents = getMessagingGroupAgents(mg.id);
-    for (const agent of agents) {
-      const triggerRules = agent.trigger_rules ? JSON.parse(agent.trigger_rules) : null;
-      configs.push({
-        platformId: mg.platform_id,
-        agentGroupId: agent.agent_group_id,
-        triggerPattern: triggerRules?.pattern,
-        requiresTrigger: triggerRules?.requiresTrigger ?? false,
-        sessionMode: agent.session_mode,
-      });
-    }
-  }
-
-  return configs;
-}
-
 /** Graceful shutdown. */
 async function shutdown(signal: string): Promise<void> {
   log.info('Shutdown signal received', { signal });
diff --git a/src/install-slug.ts b/src/install-slug.ts
new file mode 100644
index 0000000..8d6443a
--- /dev/null
+++ b/src/install-slug.ts
@@ -0,0 +1,33 @@
+/**
+ * Per-checkout install identifiers. Lets two NanoClaw installs coexist on
+ * one host without clobbering each other's service registration or the
+ * shared `nanoclaw-agent:latest` docker image tag.
+ *
+ * Slug is sha1(projectRoot)[:8] — deterministic per checkout path, stable
+ * across re-runs, unique enough across installs.
+ */
+import { createHash } from 'crypto';
+
+export function getInstallSlug(projectRoot: string = process.cwd()): string {
+  return createHash('sha1').update(projectRoot).digest('hex').slice(0, 8);
+}
+
+/** launchd Label + plist basename. e.g. `com.nanoclaw-v2-ab12cd34`. */
+export function getLaunchdLabel(projectRoot?: string): string {
+  return `com.nanoclaw-v2-${getInstallSlug(projectRoot)}`;
+}
+
+/** systemd unit name (no .service suffix). e.g. `nanoclaw-v2-ab12cd34`. */
+export function getSystemdUnit(projectRoot?: string): string {
+  return `nanoclaw-v2-${getInstallSlug(projectRoot)}`;
+}
+
+/** Docker image base (no tag). e.g. `nanoclaw-agent-v2-ab12cd34`. */
+export function getContainerImageBase(projectRoot?: string): string {
+  return `nanoclaw-agent-v2-${getInstallSlug(projectRoot)}`;
+}
+
+/** Default full container image reference with `:latest` tag. */
+export function getDefaultContainerImage(projectRoot?: string): string {
+  return `${getContainerImageBase(projectRoot)}:latest`;
+}
diff --git a/src/modules/agent-to-agent/agent-route.test.ts b/src/modules/agent-to-agent/agent-route.test.ts
new file mode 100644
index 0000000..4d48f6f
--- /dev/null
+++ b/src/modules/agent-to-agent/agent-route.test.ts
@@ -0,0 +1,46 @@
+import { describe, expect, it } from 'vitest';
+
+import { isSafeAttachmentName } from './agent-route.js';
+
+/**
+ * `forwardAttachedFiles` has a filesystem side that's awkward to unit-test
+ * without mocking DATA_DIR. The guarantee worth pinning is that the
+ * filename validator rejects everything that could escape the inbox dir —
+ * `forwardAttachedFiles` runs this guard before any I/O, so traversal is
+ * impossible as long as this matrix holds.
+ */
+describe('isSafeAttachmentName', () => {
+  it('accepts plain filenames', () => {
+    expect(isSafeAttachmentName('baby-duck.png')).toBe(true);
+    expect(isSafeAttachmentName('file with spaces.pdf')).toBe(true);
+    expect(isSafeAttachmentName('report.v2.docx')).toBe(true);
+    expect(isSafeAttachmentName('.hidden')).toBe(true); // leading dot is fine, just not `.` / `..`
+  });
+
+  it('rejects empty / sentinel values', () => {
+    expect(isSafeAttachmentName('')).toBe(false);
+    expect(isSafeAttachmentName('.')).toBe(false);
+    expect(isSafeAttachmentName('..')).toBe(false);
+  });
+
+  it('rejects path separators', () => {
+    expect(isSafeAttachmentName('../evil.png')).toBe(false);
+    expect(isSafeAttachmentName('/etc/passwd')).toBe(false);
+    expect(isSafeAttachmentName('nested/file.txt')).toBe(false);
+    expect(isSafeAttachmentName('windows\\path.exe')).toBe(false);
+  });
+
+  it('rejects NUL bytes', () => {
+    expect(isSafeAttachmentName('clean\0.png')).toBe(false);
+  });
+
+  it('rejects anything path.basename would strip', () => {
+    expect(isSafeAttachmentName('a/b')).toBe(false);
+    expect(isSafeAttachmentName('./thing')).toBe(false);
+  });
+
+  it('rejects non-string input', () => {
+    expect(isSafeAttachmentName(null as unknown as string)).toBe(false);
+    expect(isSafeAttachmentName(undefined as unknown as string)).toBe(false);
+  });
+});
diff --git a/src/modules/agent-to-agent/agent-route.ts b/src/modules/agent-to-agent/agent-route.ts
index 760356c..812cb8e 100644
--- a/src/modules/agent-to-agent/agent-route.ts
+++ b/src/modules/agent-to-agent/agent-route.ts
@@ -3,9 +3,13 @@
  *
  * Outbound messages with `channel_type === 'agent'` target another agent
  * group rather than a channel. Permission is enforced via `agent_destinations` —
- * the source agent must have a row for the target. Content is copied verbatim;
- * the target's formatter looks up the source agent in its own local map to
- * display a name.
+ * the source agent must have a row for the target. Content is copied into the
+ * target's inbound DB; if the source message had `files` (from `send_file`),
+ * the actual bytes are copied from the source's outbox into the target's
+ * `inbox/<a2a-msg-id>/` directory and surfaced to the target agent as
+ * `attachments` (existing formatter convention — see formatter.ts:230).
+ * The target agent can then forward the file onward via its own `send_file`
+ * call using the absolute `/workspace/inbox/<a2a-msg-id>/<filename>` path.
  *
  * Self-messages are always allowed (used for system notes injected back into
  * an agent's own session, e.g. post-approval follow-up prompts).
@@ -14,14 +18,102 @@
  * `channel_type === 'agent'` check. When the module is absent the check in
  * core throws with a "module not installed" message so retry → mark failed.
  */
+import fs from 'fs';
+import path from 'path';
+
 import { getAgentGroup } from '../../db/agent-groups.js';
 import { getSession } from '../../db/sessions.js';
 import { wakeContainer } from '../../container-runner.js';
 import { log } from '../../log.js';
-import { resolveSession, writeSessionMessage } from '../../session-manager.js';
+import { resolveSession, sessionDir, writeSessionMessage } from '../../session-manager.js';
 import type { Session } from '../../types.js';
 import { hasDestination } from './db/agent-destinations.js';
 
+export interface ForwardedAttachment {
+  name: string;
+  filename: string;
+  type: 'file';
+  localPath: string;
+}
+
+/**
+ * Is `name` safe to use as the last segment of a path inside the target
+ * agent's inbox directory? Filenames arrive in messages_out content from
+ * the source agent — under a multi-agent setup with heterogenous providers
+ * (or a compromised / hallucinating sub-agent) they can't be trusted.
+ *
+ * Rejects:
+ *   - empty string
+ *   - `.` / `..` (traversal sentinels that path.basename returns as-is)
+ *   - anything containing a path separator (`/` or `\`) or NUL
+ *   - any value where `path.basename(name) !== name`, catching OS-specific
+ *     separators and covering drives/prefixes on Windows runtimes
+ */
+export function isSafeAttachmentName(name: string): boolean {
+  if (typeof name !== 'string' || name.length === 0) return false;
+  if (name === '.' || name === '..') return false;
+  if (/[\\/\0]/.test(name)) return false;
+  return path.basename(name) === name;
+}
+
+/**
+ * Copy file attachments from the source agent's outbox into the target
+ * agent's inbox. Returns attachments using the formatter's existing
+ * `{name, type, localPath}` convention — target agent reads `localPath`
+ * as relative to `/workspace/`, matching how channel-inbound attachments
+ * are surfaced today.
+ *
+ * Missing source files and unsafe (path-traversal) filenames are skipped
+ * with a warning rather than failing the whole route — a bad filename
+ * reference shouldn't kill the accompanying text.
+ */
+export function forwardAttachedFiles(
+  source: { agentGroupId: string; sessionId: string; messageId: string; filenames: string[] },
+  target: { agentGroupId: string; sessionId: string; messageId: string },
+): ForwardedAttachment[] {
+  if (source.filenames.length === 0) return [];
+
+  const sourceDir = path.join(sessionDir(source.agentGroupId, source.sessionId), 'outbox', source.messageId);
+  if (!fs.existsSync(sourceDir)) {
+    log.warn('agent-route: source outbox dir missing, no files forwarded', {
+      sourceMsgId: source.messageId,
+      sourceDir,
+    });
+    return [];
+  }
+
+  const targetInboxDir = path.join(sessionDir(target.agentGroupId, target.sessionId), 'inbox', target.messageId);
+  fs.mkdirSync(targetInboxDir, { recursive: true });
+
+  const attachments: ForwardedAttachment[] = [];
+  for (const filename of source.filenames) {
+    if (!isSafeAttachmentName(filename)) {
+      log.warn('agent-route: rejecting unsafe attachment filename (path traversal attempt?)', {
+        sourceMsgId: source.messageId,
+        filename,
+      });
+      continue;
+    }
+    const src = path.join(sourceDir, filename);
+    if (!fs.existsSync(src)) {
+      log.warn('agent-route: referenced file missing in source outbox, skipped', {
+        sourceMsgId: source.messageId,
+        filename,
+      });
+      continue;
+    }
+    const dst = path.join(targetInboxDir, filename);
+    fs.copyFileSync(src, dst);
+    attachments.push({
+      name: filename,
+      filename,
+      type: 'file',
+      localPath: `inbox/${target.messageId}/${filename}`,
+    });
+  }
+  return attachments;
+}
+
 export interface RoutableAgentMessage {
   id: string;
   platform_id: string | null;
@@ -45,20 +137,87 @@ export async function routeAgentMessage(msg: RoutableAgentMessage, session: Sess
     throw new Error(`target agent group ${targetAgentGroupId} not found for message ${msg.id}`);
   }
   const { session: targetSession } = resolveSession(targetAgentGroupId, null, null, 'agent-shared');
+  const a2aMsgId = `a2a-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+
+  // If the source message references files (via `send_file`), forward the
+  // bytes from the source's outbox into the target's inbox so the target
+  // agent can actually see and re-send them. Without this, agent-to-agent
+  // file attachments look like they arrive but the target has no way to
+  // read the bytes — they live in a session dir it doesn't mount.
+  const forwardedContent = forwardFileAttachments(msg, a2aMsgId, session, targetAgentGroupId, targetSession.id);
+
   writeSessionMessage(targetAgentGroupId, targetSession.id, {
-    id: `a2a-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
+    id: a2aMsgId,
     kind: 'chat',
     timestamp: new Date().toISOString(),
     platformId: session.agent_group_id,
     channelType: 'agent',
     threadId: null,
-    content: msg.content,
+    content: forwardedContent,
   });
   log.info('Agent message routed', {
     from: session.agent_group_id,
     to: targetAgentGroupId,
     targetSession: targetSession.id,
+    a2aMsgId,
+    forwardedFileCount: countForwardedFiles(forwardedContent),
   });
   const fresh = getSession(targetSession.id);
   if (fresh) await wakeContainer(fresh);
 }
+
+/**
+ * Parse source content, copy any referenced `files` from source outbox to
+ * target inbox, and return a JSON string with an `attachments` array added
+ * (formatter.ts:223 already knows how to render this shape).
+ *
+ * If the source content isn't JSON or has no files, returns the original
+ * content string unchanged — this is safe to call on every route.
+ */
+function forwardFileAttachments(
+  msg: RoutableAgentMessage,
+  a2aMsgId: string,
+  sourceSession: Session,
+  targetAgentGroupId: string,
+  targetSessionId: string,
+): string {
+  let parsed: Record<string, unknown>;
+  try {
+    parsed = JSON.parse(msg.content);
+  } catch {
+    return msg.content;
+  }
+  const files = parsed.files as unknown;
+  if (!Array.isArray(files) || files.length === 0) return msg.content;
+  const filenames = files.filter((f): f is string => typeof f === 'string');
+  if (filenames.length === 0) return msg.content;
+
+  const attachments = forwardAttachedFiles(
+    {
+      agentGroupId: sourceSession.agent_group_id,
+      sessionId: sourceSession.id,
+      messageId: msg.id,
+      filenames,
+    },
+    {
+      agentGroupId: targetAgentGroupId,
+      sessionId: targetSessionId,
+      messageId: a2aMsgId,
+    },
+  );
+
+  // Merge into any existing `attachments` (unlikely in a2a context but safe).
+  const existing = Array.isArray(parsed.attachments) ? (parsed.attachments as Record<string, unknown>[]) : [];
+  parsed.attachments = [...existing, ...attachments];
+
+  return JSON.stringify(parsed);
+}
+
+function countForwardedFiles(contentStr: string): number {
+  try {
+    const parsed = JSON.parse(contentStr);
+    return Array.isArray(parsed.attachments) ? parsed.attachments.length : 0;
+  } catch {
+    return 0;
+  }
+}
diff --git a/src/modules/approvals/agent.md b/src/modules/approvals/agent.md
index f992040..57b65d0 100644
--- a/src/modules/approvals/agent.md
+++ b/src/modules/approvals/agent.md
@@ -16,7 +16,7 @@ install_packages({
 
 - Max 20 packages per request.
 - Names must match strict regex (blocks shell injection via `vim; curl evil.com`).
-- After approval: rebuild runs automatically. You do NOT need to call `request_rebuild` separately.
+- On approval, the image rebuild and container restart happen automatically — there is no separate rebuild step for you to trigger.
 
 ### add_mcp_server
 
@@ -32,15 +32,7 @@ add_mcp_server({
 ```
 
 - Does NOT install packages. Use `install_packages` first if the command isn't already available.
-- On approval, container is killed so the next message wakes it with the new server wired up.
-
-### request_rebuild
-
-Rebuild your container image. Only useful if you've already landed `install_packages` approvals whose rebuild step failed, or if you're recovering from a bad config edit.
-
-```
-request_rebuild({ reason: "previous install_packages rebuild failed" })
-```
+- On approval, the container is killed and the next message wakes it with the new server wired up. No image rebuild — bun runs TS directly.
 
 ### How approval works
 
diff --git a/src/modules/approvals/index.ts b/src/modules/approvals/index.ts
index 2bd8446..f70a43f 100644
--- a/src/modules/approvals/index.ts
+++ b/src/modules/approvals/index.ts
@@ -12,9 +12,9 @@
  *     once the delivery adapter is set.
  *   - A shutdown callback that stops the OneCLI handler cleanly.
  *
- * Self-mod flows (install_packages, request_rebuild, add_mcp_server) moved
- * out to `src/modules/self-mod/` in PR #7 — they now register delivery
- * actions + approval handlers via this module's public API.
+ * Self-mod flows (install_packages, add_mcp_server) moved out to
+ * `src/modules/self-mod/` in PR #7 — they now register delivery actions
+ * + approval handlers via this module's public API.
  */
 import { onDeliveryAdapterReady } from '../../delivery.js';
 import { registerResponseHandler, onShutdown } from '../../response-registry.js';
diff --git a/src/modules/approvals/onecli-approvals.ts b/src/modules/approvals/onecli-approvals.ts
index 1594a82..eec05c0 100644
--- a/src/modules/approvals/onecli-approvals.ts
+++ b/src/modules/approvals/onecli-approvals.ts
@@ -20,7 +20,7 @@
 import { OneCLI, type ApprovalRequest, type ManualApprovalHandle } from '@onecli-sh/sdk';
 
 import { pickApprovalDelivery, pickApprover } from './primitive.js';
-import { ONECLI_URL } from '../../config.js';
+import { ONECLI_API_KEY, ONECLI_URL } from '../../config.js';
 import { getAgentGroup } from '../../db/agent-groups.js';
 import {
   createPendingApproval,
@@ -36,7 +36,7 @@ export const ONECLI_ACTION = 'onecli_credential';
 
 type Decision = 'approve' | 'deny';
 
-const onecli = new OneCLI({ url: ONECLI_URL });
+const onecli = new OneCLI({ url: ONECLI_URL, apiKey: ONECLI_API_KEY });
 
 interface PendingState {
   resolve: (decision: Decision) => void;
diff --git a/src/modules/approvals/picks.test.ts b/src/modules/approvals/picks.test.ts
index 508aa35..0d1784a 100644
--- a/src/modules/approvals/picks.test.ts
+++ b/src/modules/approvals/picks.test.ts
@@ -6,7 +6,11 @@
 import { beforeEach, afterEach, describe, expect, it } from 'vitest';
 
 import type { ChannelAdapter, OutboundMessage } from '../../channels/adapter.js';
-import { initChannelAdapters, registerChannelAdapter, teardownChannelAdapters } from '../../channels/channel-registry.js';
+import {
+  initChannelAdapters,
+  registerChannelAdapter,
+  teardownChannelAdapters,
+} from '../../channels/channel-registry.js';
 import { closeDb, createAgentGroup, initTestDb, runMigrations } from '../../db/index.js';
 import { createUser } from '../permissions/db/users.js';
 import { grantRole } from '../permissions/db/user-roles.js';
@@ -57,6 +61,7 @@ async function mountMockAdapter(
   await initChannelAdapters(() => ({
     conversations: [],
     onInbound: () => {},
+    onInboundEvent: () => {},
     onMetadata: () => {},
     onAction: () => {},
   }));
diff --git a/src/modules/approvals/project.md b/src/modules/approvals/project.md
index 19dae67..6a1f10f 100644
--- a/src/modules/approvals/project.md
+++ b/src/modules/approvals/project.md
@@ -4,13 +4,13 @@ Admin-gated approval flow for agent self-modification and OneCLI credential acce
 
 ### Two flows
 
-**Agent-initiated (DB-backed, fire-and-forget).** The container writes a `system`-kind outbound row with one of three actions — `install_packages`, `request_rebuild`, `add_mcp_server`. The module's delivery-action handlers validate, route to the right approver's DM, and persist a `pending_approvals` row. When the admin clicks a button, the registered response handler applies the change (config update → image rebuild → container kill) and notifies the agent via system chat.
+**Agent-initiated (DB-backed, fire-and-forget).** The container writes a `system`-kind outbound row with one of two actions — `install_packages`, `add_mcp_server`. The module's delivery-action handlers validate, route to the right approver's DM, and persist a `pending_approvals` row. When the admin clicks a button, the registered response handler applies the change (config update → image rebuild if needed → container kill) and notifies the agent via system chat.
 
 **OneCLI credential (long-poll).** The OneCLI gateway holds an HTTP connection open when it needs credential approval. `onecli-approvals.ts` delivers a card, persists a `pending_approvals` row (action = `onecli_credential`), and waits on an in-memory Promise that resolves on click or expiry timer. Survives host restart: the startup sweep edits stale cards to "Expired (host restarted)" and drops the rows.
 
 ### Wiring
 
-- **Delivery actions:** `install_packages`, `request_rebuild`, `add_mcp_server` via `registerDeliveryAction`.
+- **Delivery actions:** `install_packages`, `add_mcp_server` via `registerDeliveryAction`.
 - **Response handler:** single handler claims both agent-initiated and OneCLI approvals. OneCLI is tried first (in-memory Promise); falls through to `pending_approvals` lookup.
 - **Adapter-ready hook (`onDeliveryAdapterReady`):** starts the OneCLI manual-approval handler once the delivery adapter is set.
 - **Shutdown hook (`onShutdown`):** stops the OneCLI handler.
diff --git a/src/modules/approvals/response-handler.ts b/src/modules/approvals/response-handler.ts
index bd0c2c5..2bbdc9d 100644
--- a/src/modules/approvals/response-handler.ts
+++ b/src/modules/approvals/response-handler.ts
@@ -96,7 +96,9 @@ async function handleRegisteredApproval(
     log.info('Approval handled', { approvalId: approval.approval_id, action: approval.action, userId });
   } catch (err) {
     log.error('Approval handler threw', { approvalId: approval.approval_id, action: approval.action, err });
-    notify(`Your ${approval.action} was approved, but applying it failed: ${err instanceof Error ? err.message : String(err)}.`);
+    notify(
+      `Your ${approval.action} was approved, but applying it failed: ${err instanceof Error ? err.message : String(err)}.`,
+    );
   }
 
   deletePendingApproval(approval.approval_id);
diff --git a/src/modules/index.ts b/src/modules/index.ts
index 2df4477..0228509 100644
--- a/src/modules/index.ts
+++ b/src/modules/index.ts
@@ -22,4 +22,3 @@ import './scheduling/index.js';
 import './permissions/index.js';
 import './agent-to-agent/index.js';
 import './self-mod/index.js';
-
diff --git a/src/modules/interactive/index.ts b/src/modules/interactive/index.ts
index 5a3b8af..324adbe 100644
--- a/src/modules/interactive/index.ts
+++ b/src/modules/interactive/index.ts
@@ -46,7 +46,11 @@ async function handleInteractiveResponse(payload: ResponsePayload): Promise<bool
   });
 
   deletePendingQuestion(payload.questionId);
-  log.info('Question response routed', { questionId: payload.questionId, selectedOption: payload.value, sessionId: session.id });
+  log.info('Question response routed', {
+    questionId: payload.questionId,
+    selectedOption: payload.value,
+    sessionId: session.id,
+  });
 
   await wakeContainer(session);
   return true;
diff --git a/src/modules/permissions/access.ts b/src/modules/permissions/access.ts
index fb7c491..8365e94 100644
--- a/src/modules/permissions/access.ts
+++ b/src/modules/permissions/access.ts
@@ -1,14 +1,13 @@
 /**
- * Access control (permissions module half of src/access.ts).
+ * Access control.
  *
  * Privilege is user-level, not group-level. A user holds zero or more roles
  * (owner | admin) via `user_roles`, and is optionally "known" in specific
  * agent groups via `agent_group_members`. Admins are implicitly members of
  * the groups they administer.
  *
- * The approver-picking functions (pickApprover, pickApprovalDelivery) stay
- * in src/access.ts for now — they move into the approvals module in the
- * planned PR #7 re-tier.
+ * Approver-picking (`pickApprover`, `pickApprovalDelivery`) lives in the
+ * approvals module — see `src/modules/approvals/primitive.ts`.
  */
 import { isMember } from './db/agent-group-members.js';
 import { isAdminOfAgentGroup, isGlobalAdmin, isOwner } from './db/user-roles.js';
diff --git a/src/modules/permissions/channel-approval.test.ts b/src/modules/permissions/channel-approval.test.ts
new file mode 100644
index 0000000..da992d2
--- /dev/null
+++ b/src/modules/permissions/channel-approval.test.ts
@@ -0,0 +1,389 @@
+/**
+ * Integration tests for the unknown-channel registration flow (ACTION-ITEMS
+ * item 22).
+ *
+ * Covers:
+ *  - Mention on an unwired channel fires an owner-approval card
+ *  - DM on an unwired channel fires a card (engage_mode will default to pattern='.')
+ *  - In-flight dedup: second mention while a card is pending doesn't spam
+ *  - Approve: wiring created with correct defaults, triggering sender added
+ *    as member, replay wakes the container
+ *  - Deny: messaging_groups.denied_at set, future mentions drop silently
+ *  - Unauthorized clicker is rejected (same pattern as sender-approval)
+ *  - No-owner install: no card, no row
+ *  - No agent groups configured: no card, no row
+ */
+import fs from 'fs';
+import { beforeEach, afterEach, describe, expect, it, vi } from 'vitest';
+
+import { initTestDb, closeDb, runMigrations } from '../../db/index.js';
+import { createAgentGroup } from '../../db/agent-groups.js';
+import { createMessagingGroup, getMessagingGroupByPlatform } from '../../db/messaging-groups.js';
+import { upsertUser } from './db/users.js';
+import { grantRole } from './db/user-roles.js';
+
+// Mock container runner — prevent actual docker spawn.
+vi.mock('../../container-runner.js', () => ({
+  wakeContainer: vi.fn().mockResolvedValue(undefined),
+  isContainerRunning: vi.fn().mockReturnValue(false),
+  getActiveContainerCount: vi.fn().mockReturnValue(0),
+  killContainer: vi.fn(),
+}));
+
+// Mock delivery adapter.
+const deliverMock = vi.fn().mockResolvedValue('plat-msg-id');
+vi.mock('../../delivery.js', () => ({
+  getDeliveryAdapter: () => ({ deliver: deliverMock }),
+}));
+
+// Mock ensureUserDm — look up the owner's preconfigured DM row instead of
+// hitting a real openDM RPC.
+vi.mock('./user-dm.js', () => ({
+  ensureUserDm: vi.fn(async (userId: string) => {
+    const { getDb } = await import('../../db/connection.js');
+    const row = getDb()
+      .prepare(
+        `SELECT mg.* FROM messaging_groups mg
+           JOIN user_dms ud ON ud.messaging_group_id = mg.id
+          WHERE ud.user_id = ?`,
+      )
+      .get(userId);
+    return row;
+  }),
+}));
+
+vi.mock('../../config.js', async () => {
+  const actual = await vi.importActual('../../config.js');
+  return { ...actual, DATA_DIR: '/tmp/nanoclaw-test-channel-approval' };
+});
+
+const TEST_DIR = '/tmp/nanoclaw-test-channel-approval';
+
+function now() {
+  return new Date().toISOString();
+}
+
+beforeEach(async () => {
+  if (fs.existsSync(TEST_DIR)) fs.rmSync(TEST_DIR, { recursive: true });
+  fs.mkdirSync(TEST_DIR, { recursive: true });
+  const db = initTestDb();
+  runMigrations(db);
+
+  await import('./index.js'); // register hooks
+
+  // Base fixtures: one agent group + owner with a DM on 'telegram'.
+  createAgentGroup({ id: 'ag-1', name: 'Andy', folder: 'andy', agent_provider: null, created_at: now() });
+
+  upsertUser({ id: 'telegram:owner', kind: 'telegram', display_name: 'Owner', created_at: now() });
+  grantRole({
+    user_id: 'telegram:owner',
+    role: 'owner',
+    agent_group_id: null,
+    granted_by: null,
+    granted_at: now(),
+  });
+
+  // Pre-seed owner's DM messaging group + user_dms mapping.
+  createMessagingGroup({
+    id: 'mg-dm-owner',
+    channel_type: 'telegram',
+    platform_id: 'dm-owner',
+    name: 'Owner DM',
+    is_group: 0,
+    unknown_sender_policy: 'public',
+    created_at: now(),
+  });
+  const { getDb } = await import('../../db/connection.js');
+  getDb()
+    .prepare(
+      `INSERT INTO user_dms (user_id, channel_type, messaging_group_id, resolved_at)
+       VALUES (?, ?, ?, ?)`,
+    )
+    .run('telegram:owner', 'telegram', 'mg-dm-owner', now());
+
+  deliverMock.mockClear();
+});
+
+afterEach(() => {
+  closeDb();
+  if (fs.existsSync(TEST_DIR)) fs.rmSync(TEST_DIR, { recursive: true });
+});
+
+function groupMention(platformId: string, text = '@bot hello') {
+  return {
+    channelType: 'telegram',
+    platformId,
+    threadId: 'thread-1', // non-null → is_group=true per channel-approval default-picker logic
+    message: {
+      id: `msg-${Math.random().toString(36).slice(2, 8)}`,
+      kind: 'chat' as const,
+      content: JSON.stringify({ senderId: 'caller', senderName: 'Caller', text }),
+      timestamp: now(),
+      isMention: true,
+    },
+  };
+}
+
+function dmEvent(platformId: string, text = 'hello') {
+  return {
+    channelType: 'telegram',
+    platformId,
+    threadId: null,
+    message: {
+      id: `msg-${Math.random().toString(36).slice(2, 8)}`,
+      kind: 'chat' as const,
+      content: JSON.stringify({ senderId: 'stranger', senderName: 'Stranger', text }),
+      timestamp: now(),
+      isMention: true, // DM bridge sets isMention=true
+    },
+  };
+}
+
+describe('unknown-channel registration flow', () => {
+  it('delivers an approval card on mention into an unwired group', async () => {
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(groupMention('chat-new'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).toHaveBeenCalledTimes(1);
+    const [channel, platformId, thread, kind, content] = deliverMock.mock.calls[0];
+    expect(channel).toBe('telegram');
+    expect(platformId).toBe('dm-owner'); // delivered to owner's DM
+    expect(thread).toBeNull();
+    expect(kind).toBe('chat-sdk');
+    const payload = JSON.parse(content as string);
+    expect(payload.type).toBe('ask_question');
+    // Card names the target agent so the owner knows what they're wiring to.
+    expect(payload.question).toContain('Andy');
+
+    const { getDb } = await import('../../db/connection.js');
+    const rows = getDb().prepare('SELECT * FROM pending_channel_approvals').all() as Array<{
+      messaging_group_id: string;
+    }>;
+    expect(rows).toHaveLength(1);
+  });
+
+  it('delivers a card on DM too (non-threaded event)', async () => {
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(dmEvent('dm-new-user'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).toHaveBeenCalledTimes(1);
+    const { getDb } = await import('../../db/connection.js');
+    const count = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number }).c;
+    expect(count).toBe(1);
+  });
+
+  it('dedups a second mention while the card is pending', async () => {
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(groupMention('chat-busy'));
+    await new Promise((r) => setTimeout(r, 10));
+    await routeInbound(groupMention('chat-busy', '@bot still here'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).toHaveBeenCalledTimes(1);
+    const { getDb } = await import('../../db/connection.js');
+    const count = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number }).c;
+    expect(count).toBe(1);
+  });
+
+  it('approve → creates wiring, admits triggering sender, replays', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+    const { wakeContainer } = await import('../../container-runner.js');
+    (wakeContainer as unknown as ReturnType<typeof vi.fn>).mockClear();
+
+    await routeInbound(groupMention('chat-approve'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT messaging_group_id FROM pending_channel_approvals').get() as {
+      messaging_group_id: string;
+    };
+    expect(pending).toBeDefined();
+
+    // Owner clicks approve.
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.messaging_group_id,
+        value: 'approve',
+        userId: 'owner', // raw platform id — handler namespaces it
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // Wiring created with MVP defaults.
+    const mga = getDb()
+      .prepare('SELECT * FROM messaging_group_agents WHERE messaging_group_id = ?')
+      .get(pending.messaging_group_id) as {
+      engage_mode: string;
+      engage_pattern: string | null;
+      sender_scope: string;
+      ignored_message_policy: string;
+      agent_group_id: string;
+    };
+    expect(mga).toBeDefined();
+    expect(mga.engage_mode).toBe('mention-sticky'); // group (threadId != null)
+    expect(mga.engage_pattern).toBeNull();
+    expect(mga.sender_scope).toBe('known');
+    expect(mga.ignored_message_policy).toBe('accumulate');
+    expect(mga.agent_group_id).toBe('ag-1');
+
+    // Triggering sender auto-admitted so sender_scope='known' doesn't
+    // bounce the replay into sender-approval.
+    const member = getDb()
+      .prepare('SELECT 1 AS x FROM agent_group_members WHERE user_id = ? AND agent_group_id = ?')
+      .get('telegram:caller', 'ag-1');
+    expect(member).toBeDefined();
+
+    // Pending row cleared and container woken via replay.
+    const stillPending = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number })
+      .c;
+    expect(stillPending).toBe(0);
+    expect(wakeContainer).toHaveBeenCalled();
+  });
+
+  it('approve on a DM wires with pattern="." defaults', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+
+    await routeInbound(dmEvent('dm-approve-user'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT messaging_group_id FROM pending_channel_approvals').get() as {
+      messaging_group_id: string;
+    };
+
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.messaging_group_id,
+        value: 'approve',
+        userId: 'owner',
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    const mga = getDb()
+      .prepare('SELECT engage_mode, engage_pattern FROM messaging_group_agents WHERE messaging_group_id = ?')
+      .get(pending.messaging_group_id) as { engage_mode: string; engage_pattern: string };
+    expect(mga.engage_mode).toBe('pattern');
+    expect(mga.engage_pattern).toBe('.');
+  });
+
+  it('deny → sets denied_at; future mentions drop silently without a second card', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+
+    await routeInbound(groupMention('chat-deny'));
+    await new Promise((r) => setTimeout(r, 10));
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT messaging_group_id FROM pending_channel_approvals').get() as {
+      messaging_group_id: string;
+    };
+
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.messaging_group_id,
+        value: 'reject',
+        userId: 'owner',
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // denied_at set, pending row cleared, no wiring.
+    const mg = getMessagingGroupByPlatform('telegram', 'chat-deny');
+    expect(mg?.denied_at).not.toBeNull();
+    expect(mg?.denied_at).toBeTruthy();
+    const mgaCount = (
+      getDb()
+        .prepare('SELECT COUNT(*) AS c FROM messaging_group_agents WHERE messaging_group_id = ?')
+        .get(pending.messaging_group_id) as { c: number }
+    ).c;
+    expect(mgaCount).toBe(0);
+
+    // A follow-up mention on the denied channel: no new card, no new pending row.
+    deliverMock.mockClear();
+    await routeInbound(groupMention('chat-deny', '@bot please'));
+    await new Promise((r) => setTimeout(r, 10));
+    expect(deliverMock).not.toHaveBeenCalled();
+    const stillPending = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number })
+      .c;
+    expect(stillPending).toBe(0);
+  });
+
+  it('rejects clicks from an unauthorized user (prevents self-admit via forwarded card)', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+
+    await routeInbound(groupMention('chat-unauth'));
+    await new Promise((r) => setTimeout(r, 10));
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT messaging_group_id FROM pending_channel_approvals').get() as {
+      messaging_group_id: string;
+    };
+
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.messaging_group_id,
+        value: 'approve',
+        userId: 'random-bystander',
+        channelType: 'telegram',
+        platformId: 'dm-random',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // No wiring created, pending row preserved so a real approver can act on it.
+    const mgaCount = (
+      getDb()
+        .prepare('SELECT COUNT(*) AS c FROM messaging_group_agents WHERE messaging_group_id = ?')
+        .get(pending.messaging_group_id) as { c: number }
+    ).c;
+    expect(mgaCount).toBe(0);
+    const stillPending = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number })
+      .c;
+    expect(stillPending).toBe(1);
+  });
+});
+
+describe('no-owner / no-agent failure modes', () => {
+  it('no owner → no card, no pending row (fresh-install bootstrap path)', async () => {
+    // Wipe the owner grant set up in the outer beforeEach.
+    const { getDb } = await import('../../db/connection.js');
+    getDb().prepare('DELETE FROM user_roles').run();
+
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(groupMention('chat-noowner'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).not.toHaveBeenCalled();
+    const count = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number }).c;
+    expect(count).toBe(0);
+  });
+
+  it('no agent groups → no card, no pending row', async () => {
+    const { getDb } = await import('../../db/connection.js');
+    // Drop foreign-key-dependent rows first, then the agent group itself.
+    getDb().prepare('DELETE FROM user_roles').run();
+    getDb().prepare('DELETE FROM agent_groups').run();
+
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(groupMention('chat-noagent'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).not.toHaveBeenCalled();
+    const count = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number }).c;
+    expect(count).toBe(0);
+  });
+});
diff --git a/src/modules/permissions/channel-approval.ts b/src/modules/permissions/channel-approval.ts
new file mode 100644
index 0000000..8ab41bc
--- /dev/null
+++ b/src/modules/permissions/channel-approval.ts
@@ -0,0 +1,174 @@
+/**
+ * Unknown-channel registration flow.
+ *
+ * When the router hits an unwired messaging group AND the message was
+ * addressed to the bot (SDK-confirmed mention or DM), it calls
+ * `requestChannelApproval` instead of silently dropping. The flow:
+ *
+ *   1. Pick the target agent group we'd wire to (MVP: first by name).
+ *      Multi-agent picker is a follow-up — see ACTION-ITEMS.
+ *   2. Pick an eligible approver (owner / admin) and a reachable DM for
+ *      them, reusing the same primitives the sender-approval flow uses.
+ *   3. Deliver an Approve / Ignore card that names the target agent
+ *      explicitly so the owner knows what they're wiring to.
+ *   4. Record a `pending_channel_approvals` row holding the original event
+ *      so it can be re-routed on approve.
+ *
+ * On approve (handler in index.ts):
+ *   - Create `messaging_group_agents` with MVP defaults
+ *     (mention-sticky for groups / pattern='.' for DMs,
+ *      sender_scope='known', ignored_message_policy='accumulate')
+ *   - Add the triggering sender to `agent_group_members` so sender_scope
+ *     doesn't bounce the replayed message into a sender-approval cascade
+ *   - Delete the pending row, replay the original event
+ *
+ * On ignore:
+ *   - Set `messaging_groups.denied_at = now()` so the router stops
+ *     escalating on this channel until an admin explicitly re-wires
+ *   - Delete the pending row
+ *
+ * Dedup: `pending_channel_approvals` PK on messaging_group_id. Second
+ * mention while pending silently dropped.
+ *
+ * Failure modes (log + no row, so a future attempt can try again):
+ *   - No agent groups exist (install never set up a first agent).
+ *   - No eligible approver in user_roles (no owner yet).
+ *   - Approver has no reachable DM.
+ *   - Delivery adapter missing.
+ */
+import { normalizeOptions, type RawOption } from '../../channels/ask-question.js';
+import { getAllAgentGroups } from '../../db/agent-groups.js';
+import { getMessagingGroup } from '../../db/messaging-groups.js';
+import { getDeliveryAdapter } from '../../delivery.js';
+import { log } from '../../log.js';
+import type { InboundEvent } from '../../channels/adapter.js';
+import { pickApprovalDelivery, pickApprover } from '../approvals/primitive.js';
+import { createPendingChannelApproval, hasInFlightChannelApproval } from './db/pending-channel-approvals.js';
+
+const APPROVAL_OPTIONS: RawOption[] = [
+  { label: 'Approve', selectedLabel: '✅ Wired', value: 'approve' },
+  { label: 'Ignore', selectedLabel: '🙅 Ignored', value: 'reject' },
+];
+
+export interface RequestChannelApprovalInput {
+  messagingGroupId: string;
+  event: InboundEvent;
+}
+
+export async function requestChannelApproval(input: RequestChannelApprovalInput): Promise<void> {
+  const { messagingGroupId, event } = input;
+
+  // In-flight dedup: don't spam the owner if the same unwired channel
+  // gets more mentions / DMs while a card is already pending.
+  if (hasInFlightChannelApproval(messagingGroupId)) {
+    log.debug('Channel registration already in flight — dropping retry', {
+      messagingGroupId,
+    });
+    return;
+  }
+
+  // MVP: pick the first agent group by name. Multi-agent systems will get
+  // a richer card later (user picks the target from a list).
+  const agentGroups = getAllAgentGroups();
+  if (agentGroups.length === 0) {
+    log.warn('Channel registration skipped — no agent groups configured. Run /init-first-agent.', {
+      messagingGroupId,
+    });
+    return;
+  }
+  const target = agentGroups[0];
+
+  // pickApprover takes the target agent group's id — gets scoped admins +
+  // global admins + owners. For fresh installs with only an owner, the
+  // owner is returned.
+  const approvers = pickApprover(target.id);
+  if (approvers.length === 0) {
+    log.warn('Channel registration skipped — no owner or admin configured', {
+      messagingGroupId,
+      targetAgentGroupId: target.id,
+    });
+    return;
+  }
+
+  const originMg = getMessagingGroup(messagingGroupId);
+  const originChannelType = originMg?.channel_type ?? '';
+  const delivery = await pickApprovalDelivery(approvers, originChannelType);
+  if (!delivery) {
+    log.warn('Channel registration skipped — no DM channel for any approver', {
+      messagingGroupId,
+      targetAgentGroupId: target.id,
+    });
+    return;
+  }
+
+  const isGroup = event.message?.isGroup ?? originMg?.is_group === 1;
+
+  // Extract sender name from the event content for a human-readable card.
+  let senderName: string | undefined;
+  try {
+    const parsed = JSON.parse(event.message.content) as Record<string, unknown>;
+    senderName = (parsed.senderName ?? parsed.sender) as string | undefined;
+  } catch {
+    // non-critical — fall through to generic wording
+  }
+
+  const title = isGroup ? '📣 Bot mentioned in new chat' : '💬 New direct message';
+  const question = isGroup
+    ? senderName
+      ? `${senderName} mentioned your agent in a ${originChannelType} channel. Wire it to ${target.name} and let it engage?`
+      : `Your agent was mentioned in a ${originChannelType} channel. Wire it to ${target.name} and let it engage?`
+    : senderName
+      ? `${senderName} DM'd your agent on ${originChannelType}. Wire it to ${target.name} and let it respond?`
+      : `Someone DM'd your agent on ${originChannelType}. Wire it to ${target.name} and let it respond?`;
+  const options = normalizeOptions(APPROVAL_OPTIONS);
+
+  createPendingChannelApproval({
+    messaging_group_id: messagingGroupId,
+    agent_group_id: target.id,
+    original_message: JSON.stringify(event),
+    approver_user_id: delivery.userId,
+    created_at: new Date().toISOString(),
+    title,
+    options_json: JSON.stringify(options),
+  });
+
+  const adapter = getDeliveryAdapter();
+  if (!adapter) {
+    log.error('Channel registration row created but no delivery adapter is wired', {
+      messagingGroupId,
+    });
+    return;
+  }
+
+  try {
+    await adapter.deliver(
+      delivery.messagingGroup.channel_type,
+      delivery.messagingGroup.platform_id,
+      null,
+      'chat-sdk',
+      JSON.stringify({
+        type: 'ask_question',
+        // Use messaging_group_id as the questionId — it's unique per card
+        // (PK on pending table dedups) and lets the response handler look
+        // up the pending row directly without another index.
+        questionId: messagingGroupId,
+        title,
+        question,
+        options,
+      }),
+    );
+    log.info('Channel registration card delivered', {
+      messagingGroupId,
+      targetAgentGroupId: target.id,
+      approver: delivery.userId,
+    });
+  } catch (err) {
+    log.error('Channel registration card delivery failed', {
+      messagingGroupId,
+      err,
+    });
+  }
+}
+
+export const APPROVE_VALUE = 'approve';
+export const REJECT_VALUE = 'reject';
diff --git a/src/modules/permissions/db/pending-channel-approvals.ts b/src/modules/permissions/db/pending-channel-approvals.ts
new file mode 100644
index 0000000..d402074
--- /dev/null
+++ b/src/modules/permissions/db/pending-channel-approvals.ts
@@ -0,0 +1,56 @@
+/**
+ * CRUD for pending_channel_approvals — the in-flight state for the
+ * unknown-channel registration flow. A row exists while an owner-approval
+ * card is outstanding; it's deleted on approve (after wiring is created)
+ * or deny (after denied_at is set on the messaging_group).
+ *
+ * PRIMARY KEY on messaging_group_id gives free in-flight dedup. A second
+ * mention/DM while a card is pending resolves via
+ * `hasInFlightChannelApproval` in the request flow and drops silently
+ * instead of spamming the owner.
+ */
+import { getDb } from '../../../db/connection.js';
+
+export interface PendingChannelApproval {
+  messaging_group_id: string;
+  agent_group_id: string;
+  original_message: string;
+  approver_user_id: string;
+  created_at: string;
+  /** Card title shown at creation and re-used by getAskQuestionRender on click. */
+  title: string;
+  /** Normalized options (JSON-encoded NormalizedOption[]) — same shape persisted on pending_approvals. */
+  options_json: string;
+}
+
+export function createPendingChannelApproval(row: PendingChannelApproval): void {
+  getDb()
+    .prepare(
+      `INSERT INTO pending_channel_approvals (
+         messaging_group_id, agent_group_id, original_message,
+         approver_user_id, created_at, title, options_json
+       )
+       VALUES (
+         @messaging_group_id, @agent_group_id, @original_message,
+         @approver_user_id, @created_at, @title, @options_json
+       )`,
+    )
+    .run(row);
+}
+
+export function getPendingChannelApproval(messagingGroupId: string): PendingChannelApproval | undefined {
+  return getDb()
+    .prepare('SELECT * FROM pending_channel_approvals WHERE messaging_group_id = ?')
+    .get(messagingGroupId) as PendingChannelApproval | undefined;
+}
+
+export function hasInFlightChannelApproval(messagingGroupId: string): boolean {
+  const row = getDb()
+    .prepare('SELECT 1 AS x FROM pending_channel_approvals WHERE messaging_group_id = ?')
+    .get(messagingGroupId) as { x: number } | undefined;
+  return row !== undefined;
+}
+
+export function deletePendingChannelApproval(messagingGroupId: string): void {
+  getDb().prepare('DELETE FROM pending_channel_approvals WHERE messaging_group_id = ?').run(messagingGroupId);
+}
diff --git a/src/modules/permissions/db/pending-sender-approvals.ts b/src/modules/permissions/db/pending-sender-approvals.ts
new file mode 100644
index 0000000..4d32bf4
--- /dev/null
+++ b/src/modules/permissions/db/pending-sender-approvals.ts
@@ -0,0 +1,60 @@
+/**
+ * CRUD for pending_sender_approvals — the in-flight state for the
+ * request_approval unknown-sender flow. Rows are created when an unknown
+ * sender writes into a wired messaging group with that policy, and are
+ * deleted on admin approve (after adding the user as a member) or deny.
+ *
+ * UNIQUE(messaging_group_id, sender_identity) enforces in-flight dedup:
+ * a retry / second message from the same unknown sender while a card is
+ * still pending is silently dropped instead of spamming the admin.
+ */
+import { getDb } from '../../../db/connection.js';
+
+export interface PendingSenderApproval {
+  id: string;
+  messaging_group_id: string;
+  agent_group_id: string;
+  sender_identity: string;
+  sender_name: string | null;
+  original_message: string;
+  approver_user_id: string;
+  created_at: string;
+  /** Card title shown at creation and re-used by getAskQuestionRender on click. */
+  title: string;
+  /** Normalized options (JSON-encoded NormalizedOption[]) — same shape persisted on pending_approvals. */
+  options_json: string;
+}
+
+export function createPendingSenderApproval(row: PendingSenderApproval): void {
+  getDb()
+    .prepare(
+      `INSERT INTO pending_sender_approvals (
+         id, messaging_group_id, agent_group_id, sender_identity,
+         sender_name, original_message, approver_user_id, created_at,
+         title, options_json
+       )
+       VALUES (
+         @id, @messaging_group_id, @agent_group_id, @sender_identity,
+         @sender_name, @original_message, @approver_user_id, @created_at,
+         @title, @options_json
+       )`,
+    )
+    .run(row);
+}
+
+export function getPendingSenderApproval(id: string): PendingSenderApproval | undefined {
+  return getDb().prepare('SELECT * FROM pending_sender_approvals WHERE id = ?').get(id) as
+    | PendingSenderApproval
+    | undefined;
+}
+
+export function hasInFlightSenderApproval(messagingGroupId: string, senderIdentity: string): boolean {
+  const row = getDb()
+    .prepare('SELECT 1 AS x FROM pending_sender_approvals WHERE messaging_group_id = ? AND sender_identity = ?')
+    .get(messagingGroupId, senderIdentity) as { x: number } | undefined;
+  return row !== undefined;
+}
+
+export function deletePendingSenderApproval(id: string): void {
+  getDb().prepare('DELETE FROM pending_sender_approvals WHERE id = ?').run(id);
+}
diff --git a/src/modules/permissions/index.ts b/src/modules/permissions/index.ts
index e7cc282..83390d8 100644
--- a/src/modules/permissions/index.ts
+++ b/src/modules/permissions/index.ts
@@ -16,11 +16,27 @@
  * access gate is not registered and core defaults to allow-all.
  */
 import { recordDroppedMessage } from '../../db/dropped-messages.js';
-import { setAccessGate, setSenderResolver, type AccessGateResult, type InboundEvent } from '../../router.js';
+import { createMessagingGroupAgent, setMessagingGroupDeniedAt } from '../../db/messaging-groups.js';
+import {
+  routeInbound,
+  setAccessGate,
+  setChannelRequestGate,
+  setSenderResolver,
+  setSenderScopeGate,
+  type AccessGateResult,
+} from '../../router.js';
+import type { InboundEvent } from '../../channels/adapter.js';
+import { registerResponseHandler, type ResponsePayload } from '../../response-registry.js';
 import { log } from '../../log.js';
-import type { MessagingGroup } from '../../types.js';
+import type { MessagingGroup, MessagingGroupAgent } from '../../types.js';
 import { canAccessAgentGroup } from './access.js';
+import { requestChannelApproval } from './channel-approval.js';
+import { addMember } from './db/agent-group-members.js';
+import { deletePendingChannelApproval, getPendingChannelApproval } from './db/pending-channel-approvals.js';
+import { deletePendingSenderApproval, getPendingSenderApproval } from './db/pending-sender-approvals.js';
+import { hasAdminPrivilege } from './db/user-roles.js';
 import { getUser, upsertUser } from './db/users.js';
+import { requestSenderApproval } from './sender-approval.js';
 
 function extractAndUpsertUser(event: InboundEvent): string | null {
   let content: Record<string, unknown>;
@@ -76,11 +92,12 @@ function handleUnknownSender(
   event: InboundEvent,
 ): void {
   const parsed = safeParseContent(event.message.content);
+  const senderName = parsed.sender ?? null;
   const dropRecord = {
     channel_type: event.channelType,
     platform_id: event.platformId,
     user_id: userId,
-    sender_name: parsed.sender ?? null,
+    sender_name: senderName,
     reason: `unknown_sender_${mg.unknown_sender_policy}`,
     messaging_group_id: mg.id,
     agent_group_id: agentGroupId,
@@ -98,13 +115,27 @@ function handleUnknownSender(
   }
 
   if (mg.unknown_sender_policy === 'request_approval') {
-    log.info('MESSAGE DROPPED — unknown sender (approval flow TODO)', {
+    log.info('MESSAGE DROPPED — unknown sender (approval requested)', {
       messagingGroupId: mg.id,
       agentGroupId,
       userId,
       accessReason,
     });
     recordDroppedMessage(dropRecord);
+    // Fire-and-forget; pick-approver + delivery + row-insert are all async.
+    // If it fails it logs internally — the user's message still stays dropped
+    // either way. Requires a resolved userId (senderResolver populates users
+    // row before the gate fires); if we got here without one, there's nothing
+    // to identify for approval and we just stay in the "silent strict" branch.
+    if (userId) {
+      requestSenderApproval({
+        messagingGroupId: mg.id,
+        agentGroupId,
+        senderIdentity: userId,
+        senderName,
+        event,
+      }).catch((err) => log.error('Sender-approval flow threw', { err }));
+    }
     return;
   }
 
@@ -132,3 +163,231 @@ setAccessGate((event, userId, mg, agentGroupId): AccessGateResult => {
   handleUnknownSender(mg, userId, agentGroupId, decision.reason, event);
   return { allowed: false, reason: decision.reason };
 });
+
+/**
+ * Per-wiring sender-scope enforcement. Stricter than the messaging-group
+ * `unknown_sender_policy` — a wiring can require `sender_scope='known'`
+ * (explicit owner / admin / member) even on a 'public' messaging group.
+ *
+ * 'all' is a no-op; any sender passes. 'known' requires a userId that
+ * canAccessAgentGroup accepts (owner, admin, or group member).
+ */
+setSenderScopeGate(
+  (_event: InboundEvent, userId: string | null, _mg: MessagingGroup, agent: MessagingGroupAgent): AccessGateResult => {
+    if (agent.sender_scope === 'all') return { allowed: true };
+    if (!userId) return { allowed: false, reason: 'unknown_user_scope' };
+    const decision = canAccessAgentGroup(userId, agent.agent_group_id);
+    if (decision.allowed) return { allowed: true };
+    return { allowed: false, reason: `sender_scope_${decision.reason}` };
+  },
+);
+
+/**
+ * Response handler for the unknown-sender approval card.
+ *
+ * Claim rule: questionId matches a row in pending_sender_approvals. If no
+ * such row, return false so the next handler (approvals module, OneCLI,
+ * interactive) gets a shot.
+ *
+ * Approve: add the sender to agent_group_members + re-invoke routeInbound
+ * with the stored event. The second routing attempt clears the gate because
+ * the user is now a member.
+ *
+ * Deny: delete the row (no "deny list" — a future message re-triggers a
+ * fresh card per ACTION-ITEMS item 5 "no denial persistence").
+ */
+async function handleSenderApprovalResponse(payload: ResponsePayload): Promise<boolean> {
+  const row = getPendingSenderApproval(payload.questionId);
+  if (!row) return false;
+
+  // payload.userId is the raw platform userId (e.g. "6037840640"); namespace it
+  // with the channel type so it matches users(id) format. Then verify the
+  // clicker is the designated approver OR has owner/admin privilege over this
+  // agent group — any other click is rejected so random users can't self-admit
+  // via stolen card forwarding.
+  const clickerId = payload.userId ? `${payload.channelType}:${payload.userId}` : null;
+  const isAuthorized =
+    clickerId !== null && (clickerId === row.approver_user_id || hasAdminPrivilege(clickerId, row.agent_group_id));
+  if (!isAuthorized) {
+    log.warn('Unknown-sender approval click rejected — unauthorized clicker', {
+      approvalId: row.id,
+      clickerId,
+      expectedApprover: row.approver_user_id,
+    });
+    return true; // claim the response so it's not unclaimed-logged, but do nothing
+  }
+  const approverId = clickerId;
+  const approved = payload.value === 'approve';
+
+  if (approved) {
+    addMember({
+      user_id: row.sender_identity,
+      agent_group_id: row.agent_group_id,
+      added_by: approverId,
+      added_at: new Date().toISOString(),
+    });
+    log.info('Unknown sender approved — member added', {
+      approvalId: row.id,
+      senderIdentity: row.sender_identity,
+      agentGroupId: row.agent_group_id,
+      approverId,
+    });
+
+    // Clear the pending row BEFORE re-routing so the gate check on the
+    // second attempt doesn't see the in-flight row and short-circuit.
+    deletePendingSenderApproval(row.id);
+
+    try {
+      const event = JSON.parse(row.original_message) as InboundEvent;
+      await routeInbound(event);
+    } catch (err) {
+      log.error('Failed to replay message after sender approval', { approvalId: row.id, err });
+    }
+    return true;
+  }
+
+  log.info('Unknown sender denied', {
+    approvalId: row.id,
+    senderIdentity: row.sender_identity,
+    agentGroupId: row.agent_group_id,
+    approverId,
+  });
+  deletePendingSenderApproval(row.id);
+  return true;
+}
+
+registerResponseHandler(handleSenderApprovalResponse);
+
+// ── Unknown-channel registration flow ──
+
+setChannelRequestGate(async (mg, event) => {
+  await requestChannelApproval({ messagingGroupId: mg.id, event });
+});
+
+/**
+ * Response handler for the unknown-channel registration card.
+ *
+ * Claim rule: questionId matches a pending_channel_approvals row (keyed
+ * by messaging_group_id). If no such row, return false so downstream
+ * handlers get a shot.
+ *
+ * Approve: create the wiring with MVP defaults (mention-sticky for
+ * groups / pattern='.' for DMs; sender_scope='known';
+ * ignored_message_policy='accumulate'), add the triggering sender as a
+ * member so sender_scope doesn't immediately bounce them into a
+ * sender-approval card, then replay the original event.
+ *
+ * Deny: set `messaging_groups.denied_at = now()` so future mentions on
+ * this channel drop silently until an admin explicitly wires it.
+ */
+async function handleChannelApprovalResponse(payload: ResponsePayload): Promise<boolean> {
+  const row = getPendingChannelApproval(payload.questionId);
+  if (!row) return false;
+
+  // Click-auth: same pattern as sender-approval (see commit 68058cb).
+  // Raw platform userId → namespace with channelType → must match the
+  // designated approver OR have admin privilege over the target agent.
+  const clickerId = payload.userId ? `${payload.channelType}:${payload.userId}` : null;
+  const isAuthorized =
+    clickerId !== null && (clickerId === row.approver_user_id || hasAdminPrivilege(clickerId, row.agent_group_id));
+  if (!isAuthorized) {
+    log.warn('Channel registration click rejected — unauthorized clicker', {
+      messagingGroupId: row.messaging_group_id,
+      clickerId,
+      expectedApprover: row.approver_user_id,
+    });
+    return true; // claim but take no action
+  }
+  const approverId = clickerId;
+  const approved = payload.value === 'approve';
+
+  if (!approved) {
+    setMessagingGroupDeniedAt(row.messaging_group_id, new Date().toISOString());
+    deletePendingChannelApproval(row.messaging_group_id);
+    log.info('Channel registration denied', {
+      messagingGroupId: row.messaging_group_id,
+      agentGroupId: row.agent_group_id,
+      approverId,
+    });
+    return true;
+  }
+
+  // Rehydrate the original event to know (a) whether it was a DM or group
+  // (chooses engage_mode default), and (b) who the triggering sender was
+  // (auto-member-add so sender_scope='known' doesn't bounce the replay).
+  let event: InboundEvent;
+  try {
+    event = JSON.parse(row.original_message) as InboundEvent;
+  } catch (err) {
+    log.error('Channel registration: failed to parse stored event', {
+      messagingGroupId: row.messaging_group_id,
+      err,
+    });
+    deletePendingChannelApproval(row.messaging_group_id);
+    return true;
+  }
+
+  // Decide engage_mode from the original event. DMs (`isMention=true` &
+  // not in a group) get `pattern='.'` (always respond). Group mentions
+  // get `mention-sticky` (respond now + follow the thread).
+  //
+  // We can't read `mg.is_group` reliably here because we only auto-create
+  // the mg with `is_group=0` on first sight — the adapter hasn't told us
+  // yet whether it's actually a group. Fall back to the InboundEvent's
+  // `threadId`: a non-null threadId implies a threaded platform (Slack
+  // channel thread, Discord thread), which we treat as a group.
+  const isGroup = event.threadId !== null;
+  const engageMode: MessagingGroupAgent['engage_mode'] = isGroup ? 'mention-sticky' : 'pattern';
+  const engagePattern = isGroup ? null : '.';
+
+  const mgaId = `mga-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+  createMessagingGroupAgent({
+    id: mgaId,
+    messaging_group_id: row.messaging_group_id,
+    agent_group_id: row.agent_group_id,
+    engage_mode: engageMode,
+    engage_pattern: engagePattern,
+    sender_scope: 'known',
+    ignored_message_policy: 'accumulate',
+    session_mode: 'shared',
+    priority: 0,
+    created_at: new Date().toISOString(),
+  });
+  log.info('Channel registration approved — wiring created', {
+    messagingGroupId: row.messaging_group_id,
+    agentGroupId: row.agent_group_id,
+    mgaId,
+    engageMode,
+    approverId,
+  });
+
+  // Auto-admit the triggering sender. Without this, the replay below
+  // would bounce through sender-approval (sender_scope='known' +
+  // sender-is-not-a-member).
+  const senderUserId = extractAndUpsertUser(event);
+  if (senderUserId) {
+    addMember({
+      user_id: senderUserId,
+      agent_group_id: row.agent_group_id,
+      added_by: approverId,
+      added_at: new Date().toISOString(),
+    });
+  }
+
+  // Clear the pending row BEFORE replay so the gate check on the second
+  // attempt sees a wired channel (agentCount > 0) and takes the fan-out
+  // path normally.
+  deletePendingChannelApproval(row.messaging_group_id);
+
+  try {
+    await routeInbound(event);
+  } catch (err) {
+    log.error('Failed to replay message after channel approval', {
+      messagingGroupId: row.messaging_group_id,
+      err,
+    });
+  }
+  return true;
+}
+
+registerResponseHandler(handleChannelApprovalResponse);
diff --git a/src/modules/permissions/permissions.test.ts b/src/modules/permissions/permissions.test.ts
index d76d0d6..505c926 100644
--- a/src/modules/permissions/permissions.test.ts
+++ b/src/modules/permissions/permissions.test.ts
@@ -6,7 +6,11 @@
 import { beforeEach, afterEach, describe, expect, it } from 'vitest';
 
 import type { ChannelAdapter, OutboundMessage } from '../../channels/adapter.js';
-import { initChannelAdapters, registerChannelAdapter, teardownChannelAdapters } from '../../channels/channel-registry.js';
+import {
+  initChannelAdapters,
+  registerChannelAdapter,
+  teardownChannelAdapters,
+} from '../../channels/channel-registry.js';
 import { closeDb, createAgentGroup, createMessagingGroup, initTestDb, runMigrations } from '../../db/index.js';
 import { canAccessAgentGroup } from './access.js';
 import { addMember, isMember } from './db/agent-group-members.js';
@@ -60,6 +64,7 @@ async function mountMockAdapter(
   await initChannelAdapters(() => ({
     conversations: [],
     onInbound: () => {},
+    onInboundEvent: () => {},
     onMetadata: () => {},
     onAction: () => {},
   }));
diff --git a/src/modules/permissions/sender-approval.test.ts b/src/modules/permissions/sender-approval.test.ts
new file mode 100644
index 0000000..f037fa8
--- /dev/null
+++ b/src/modules/permissions/sender-approval.test.ts
@@ -0,0 +1,344 @@
+/**
+ * Integration tests for the unknown-sender request_approval flow
+ * (ACTION-ITEMS item 5).
+ *
+ * Covers:
+ *  - request_approval policy fires `requestSenderApproval` on first unknown
+ *    message from a sender
+ *  - In-flight dedup: second message from the same sender while pending is
+ *    silently dropped (no second card, no second row)
+ *  - Approve path: member added, original message replayed via routeInbound,
+ *    container woken
+ *  - Deny path: pending row deleted, no member added
+ */
+import fs from 'fs';
+import { beforeEach, afterEach, describe, expect, it, vi } from 'vitest';
+
+import { initTestDb, closeDb, runMigrations } from '../../db/index.js';
+import { createAgentGroup } from '../../db/agent-groups.js';
+import { createMessagingGroup, createMessagingGroupAgent } from '../../db/messaging-groups.js';
+import { upsertUser } from './db/users.js';
+import { grantRole } from './db/user-roles.js';
+
+// Mock container runner — prevent actual docker spawn.
+vi.mock('../../container-runner.js', () => ({
+  wakeContainer: vi.fn().mockResolvedValue(undefined),
+  isContainerRunning: vi.fn().mockReturnValue(false),
+  getActiveContainerCount: vi.fn().mockReturnValue(0),
+  killContainer: vi.fn(),
+}));
+
+// Mock delivery adapter — record card deliveries for assertions.
+const deliverMock = vi.fn().mockResolvedValue('plat-msg-id');
+vi.mock('../../delivery.js', () => ({
+  getDeliveryAdapter: () => ({
+    deliver: deliverMock,
+  }),
+}));
+
+// Mock ensureUserDm to return the approver's existing messaging group
+// instead of hitting a real openDM RPC.
+vi.mock('./user-dm.js', () => ({
+  ensureUserDm: vi.fn(async (userId: string) => {
+    const { getDb } = await import('../../db/connection.js');
+    const row = getDb()
+      .prepare(
+        `SELECT mg.* FROM messaging_groups mg
+           JOIN user_dms ud ON ud.messaging_group_id = mg.id
+          WHERE ud.user_id = ?`,
+      )
+      .get(userId);
+    return row;
+  }),
+}));
+
+vi.mock('../../config.js', async () => {
+  const actual = await vi.importActual('../../config.js');
+  return { ...actual, DATA_DIR: '/tmp/nanoclaw-test-sender-approval' };
+});
+
+const TEST_DIR = '/tmp/nanoclaw-test-sender-approval';
+
+function now() {
+  return new Date().toISOString();
+}
+
+beforeEach(async () => {
+  if (fs.existsSync(TEST_DIR)) fs.rmSync(TEST_DIR, { recursive: true });
+  fs.mkdirSync(TEST_DIR, { recursive: true });
+  const db = initTestDb();
+  runMigrations(db);
+
+  // Side-effect imports: register hooks (permissions module) AFTER the
+  // mocks are in place so the access gate / response handler pick up the
+  // mocked delivery + user-dm helpers.
+  await import('./index.js');
+
+  // Fixtures: agent group, messaging group with request_approval, wiring,
+  // owner + DM messaging group for approver delivery.
+  createAgentGroup({ id: 'ag-1', name: 'Agent', folder: 'agent', agent_provider: null, created_at: now() });
+
+  createMessagingGroup({
+    id: 'mg-chat',
+    channel_type: 'telegram',
+    platform_id: 'chat-123',
+    name: 'Group Chat',
+    is_group: 1,
+    unknown_sender_policy: 'request_approval',
+    created_at: now(),
+  });
+  createMessagingGroupAgent({
+    id: 'mga-1',
+    messaging_group_id: 'mg-chat',
+    agent_group_id: 'ag-1',
+    engage_mode: 'pattern',
+    engage_pattern: '.',
+    sender_scope: 'all',
+    ignored_message_policy: 'drop',
+    session_mode: 'shared',
+    priority: 0,
+    created_at: now(),
+  });
+
+  // Owner user + their DM messaging group (pickApprover + ensureUserDm target).
+  upsertUser({ id: 'telegram:owner', kind: 'telegram', display_name: 'Owner', created_at: now() });
+  grantRole({
+    user_id: 'telegram:owner',
+    role: 'owner',
+    agent_group_id: null,
+    granted_by: null,
+    granted_at: now(),
+  });
+  createMessagingGroup({
+    id: 'mg-dm-owner',
+    channel_type: 'telegram',
+    platform_id: 'dm-owner',
+    name: 'Owner DM',
+    is_group: 0,
+    unknown_sender_policy: 'public',
+    created_at: now(),
+  });
+  const { getDb } = await import('../../db/connection.js');
+  getDb()
+    .prepare(
+      `INSERT INTO user_dms (user_id, channel_type, messaging_group_id, resolved_at)
+       VALUES (?, ?, ?, ?)`,
+    )
+    .run('telegram:owner', 'telegram', 'mg-dm-owner', now());
+
+  deliverMock.mockClear();
+});
+
+afterEach(() => {
+  closeDb();
+  if (fs.existsSync(TEST_DIR)) fs.rmSync(TEST_DIR, { recursive: true });
+});
+
+function stranger(text: string) {
+  return {
+    channelType: 'telegram',
+    platformId: 'chat-123',
+    threadId: null,
+    message: {
+      id: `stranger-${Math.random().toString(36).slice(2, 8)}`,
+      kind: 'chat' as const,
+      content: JSON.stringify({
+        senderId: 'tg:stranger',
+        senderName: 'Stranger',
+        text,
+      }),
+      timestamp: now(),
+    },
+  };
+}
+
+describe('unknown-sender request_approval flow', () => {
+  it('delivers an approval card on first unknown message', async () => {
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(stranger('hi'));
+
+    // Wait for the fire-and-forget requestSenderApproval to resolve.
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).toHaveBeenCalledTimes(1);
+    const [channel, platformId, thread, kind, content] = deliverMock.mock.calls[0];
+    expect(channel).toBe('telegram');
+    expect(platformId).toBe('dm-owner'); // delivered to owner's DM
+    expect(thread).toBeNull();
+    expect(kind).toBe('chat-sdk');
+    const payload = JSON.parse(content as string);
+    expect(payload.type).toBe('ask_question');
+    expect(payload.questionId).toMatch(/^nsa-/);
+
+    const { getDb } = await import('../../db/connection.js');
+    const rows = getDb().prepare('SELECT * FROM pending_sender_approvals').all();
+    expect(rows).toHaveLength(1);
+  });
+
+  it('dedups a second message from the same stranger while pending', async () => {
+    const { routeInbound } = await import('../../router.js');
+    await routeInbound(stranger('hello'));
+    await new Promise((r) => setTimeout(r, 10));
+    await routeInbound(stranger('are you there?'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    expect(deliverMock).toHaveBeenCalledTimes(1);
+    const { getDb } = await import('../../db/connection.js');
+    const count = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_sender_approvals').get() as { c: number }).c;
+    expect(count).toBe(1);
+  });
+
+  it('approve → adds member and replays the original message', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+    const { wakeContainer } = await import('../../container-runner.js');
+    (wakeContainer as unknown as ReturnType<typeof vi.fn>).mockClear();
+
+    await routeInbound(stranger('please let me in'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT id FROM pending_sender_approvals').get() as { id: string };
+    expect(pending).toBeDefined();
+
+    // Fire the approve click through the response-handler chain.
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.id,
+        value: 'approve',
+        // Chat SDK's onAction surfaces the raw platform userId (e.g. Telegram
+        // chat id). The permissions handler namespaces it with channelType to
+        // match users(id).
+        userId: 'owner',
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // Member row added for the stranger against the wired agent group.
+    const member = getDb()
+      .prepare('SELECT 1 AS x FROM agent_group_members WHERE user_id = ? AND agent_group_id = ?')
+      .get('tg:stranger', 'ag-1');
+    expect(member).toBeDefined();
+
+    // Pending row cleared.
+    const stillPending = getDb().prepare('SELECT COUNT(*) AS c FROM pending_sender_approvals').get() as { c: number };
+    expect(stillPending.c).toBe(0);
+
+    // Message replayed + container woken.
+    expect(wakeContainer).toHaveBeenCalled();
+  });
+
+  it('deny → deletes the pending row without adding a member', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+
+    await routeInbound(stranger('hello'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT id FROM pending_sender_approvals').get() as { id: string };
+    expect(pending).toBeDefined();
+
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.id,
+        value: 'reject',
+        userId: 'owner', // raw platform id — handler namespaces with channelType
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    const count = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_sender_approvals').get() as { c: number }).c;
+    expect(count).toBe(0);
+    const member = getDb()
+      .prepare('SELECT 1 AS x FROM agent_group_members WHERE user_id = ? AND agent_group_id = ?')
+      .get('tg:stranger', 'ag-1');
+    expect(member).toBeUndefined();
+  });
+
+  it('rejects clicks from an unauthorized user (prevents self-admit via forwarded card)', async () => {
+    // Stranger triggers the approval flow; card goes to the owner.
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+
+    await routeInbound(stranger('can I play'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT id FROM pending_sender_approvals').get() as { id: string };
+    expect(pending).toBeDefined();
+
+    // A random user (not the stranger, not the owner, not an admin) tries to
+    // click the approval — e.g. they got the card forwarded. Should be
+    // rejected without admitting them.
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.id,
+        value: 'approve',
+        userId: 'random-bystander', // not owner, not admin
+        channelType: 'telegram',
+        platformId: 'dm-random',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // No member added for the stranger.
+    const member = getDb()
+      .prepare('SELECT 1 AS x FROM agent_group_members WHERE user_id = ? AND agent_group_id = ?')
+      .get('tg:stranger', 'ag-1');
+    expect(member).toBeUndefined();
+
+    // Pending row is still there — a legitimate approver can still act on it.
+    const stillPending = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_sender_approvals').get() as { c: number })
+      .c;
+    expect(stillPending).toBe(1);
+  });
+
+  it('accepts a click from a global admin even if they are not the designated approver', async () => {
+    // Pre-seed a separate admin user so we can click as them.
+    upsertUser({ id: 'telegram:admin-bob', kind: 'telegram', display_name: 'Bob', created_at: now() });
+    grantRole({
+      user_id: 'telegram:admin-bob',
+      role: 'admin',
+      agent_group_id: null,
+      granted_by: 'telegram:owner',
+      granted_at: now(),
+    });
+
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+
+    await routeInbound(stranger('knock knock'));
+    await new Promise((r) => setTimeout(r, 10));
+
+    const { getDb } = await import('../../db/connection.js');
+    const pending = getDb().prepare('SELECT id FROM pending_sender_approvals').get() as { id: string };
+    expect(pending).toBeDefined();
+
+    // Admin clicks approve (not the designated approver, which was owner).
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.id,
+        value: 'approve',
+        userId: 'admin-bob',
+        channelType: 'telegram',
+        platformId: 'dm-bob',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // Stranger admitted thanks to the admin's authority.
+    const member = getDb()
+      .prepare('SELECT 1 AS x FROM agent_group_members WHERE user_id = ? AND agent_group_id = ?')
+      .get('tg:stranger', 'ag-1');
+    expect(member).toBeDefined();
+  });
+});
diff --git a/src/modules/permissions/sender-approval.ts b/src/modules/permissions/sender-approval.ts
new file mode 100644
index 0000000..fb3e24e
--- /dev/null
+++ b/src/modules/permissions/sender-approval.ts
@@ -0,0 +1,155 @@
+/**
+ * Unknown-sender approval flow.
+ *
+ * When `messaging_groups.unknown_sender_policy = 'request_approval'` and a
+ * non-member writes into a wired chat, the access gate drops the routing
+ * attempt and calls `requestSenderApproval` to:
+ *
+ *   1. Pick an eligible approver (owner / admin of the agent group).
+ *   2. Open / reuse a DM to that approver on a reachable channel.
+ *   3. Deliver an Approve / Deny card.
+ *   4. Record a pending_sender_approvals row that holds the original message
+ *      so it can be re-routed on approve.
+ *
+ * On approve: the handler in index.ts adds an agent_group_members row for
+ * the sender and re-invokes routeInbound with the stored event — the second
+ * routing attempt passes the gate because the user is now a member.
+ *
+ * Failure modes (logged + row NOT created, so the dedup gate lets a future
+ * attempt try again):
+ *   - No eligible approver in user_roles — fresh install, no owner yet.
+ *   - Approver has no reachable DM (no user_dms row + channel can't
+ *     openDM) — e.g. owner hasn't registered on any channel we're wired to.
+ *   - Delivery adapter missing.
+ *
+ * Dedup: `pending_sender_approvals` has UNIQUE(messaging_group_id,
+ * sender_identity). A retry / rapid second message from the same unknown
+ * sender is silently dropped (no duplicate card sent).
+ */
+import { normalizeOptions, type RawOption } from '../../channels/ask-question.js';
+import { getMessagingGroup } from '../../db/messaging-groups.js';
+import { getDeliveryAdapter } from '../../delivery.js';
+import { log } from '../../log.js';
+import type { InboundEvent } from '../../channels/adapter.js';
+import { pickApprovalDelivery, pickApprover } from '../approvals/primitive.js';
+import { createPendingSenderApproval, hasInFlightSenderApproval } from './db/pending-sender-approvals.js';
+
+const APPROVAL_OPTIONS: RawOption[] = [
+  { label: 'Allow', selectedLabel: '✅ Allowed', value: 'approve' },
+  { label: 'Deny', selectedLabel: '❌ Denied', value: 'reject' },
+];
+
+function generateId(): string {
+  return `nsa-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+}
+
+export interface RequestSenderApprovalInput {
+  messagingGroupId: string;
+  agentGroupId: string;
+  senderIdentity: string; // namespaced user id (channel_type:handle)
+  senderName: string | null;
+  event: InboundEvent;
+}
+
+export async function requestSenderApproval(input: RequestSenderApprovalInput): Promise<void> {
+  const { messagingGroupId, agentGroupId, senderIdentity, senderName, event } = input;
+
+  // In-flight dedup: don't spam the admin if the same unknown sender
+  // retries while a card is already pending.
+  if (hasInFlightSenderApproval(messagingGroupId, senderIdentity)) {
+    log.debug('Unknown-sender approval already in flight — dropping retry', {
+      messagingGroupId,
+      senderIdentity,
+    });
+    return;
+  }
+
+  const approvers = pickApprover(agentGroupId);
+  if (approvers.length === 0) {
+    log.warn('Unknown-sender approval skipped — no owner or admin configured', {
+      messagingGroupId,
+      agentGroupId,
+      senderIdentity,
+    });
+    return;
+  }
+
+  const originMg = getMessagingGroup(messagingGroupId);
+  const originChannelType = originMg?.channel_type ?? '';
+  const target = await pickApprovalDelivery(approvers, originChannelType);
+  if (!target) {
+    log.warn('Unknown-sender approval skipped — no DM channel for any approver', {
+      messagingGroupId,
+      agentGroupId,
+      senderIdentity,
+    });
+    return;
+  }
+
+  const approvalId = generateId();
+  const senderDisplay = senderName && senderName.length > 0 ? senderName : senderIdentity;
+  const originName = originMg?.name ?? `a ${originChannelType} channel`;
+
+  const title = '👤 New sender';
+  const question = `${senderDisplay} wants to talk to your agent in ${originName}. Allow?`;
+  const options = normalizeOptions(APPROVAL_OPTIONS);
+
+  createPendingSenderApproval({
+    id: approvalId,
+    messaging_group_id: messagingGroupId,
+    agent_group_id: agentGroupId,
+    sender_identity: senderIdentity,
+    sender_name: senderName,
+    original_message: JSON.stringify(event),
+    approver_user_id: target.userId,
+    created_at: new Date().toISOString(),
+    title,
+    options_json: JSON.stringify(options),
+  });
+
+  const adapter = getDeliveryAdapter();
+  if (!adapter) {
+    // Without a delivery adapter, the card can't be sent. Log + leave the
+    // row in place so the admin can see it via DB or manual tooling; the
+    // dedup gate will suppress further cards until it's cleared.
+    log.error('Unknown-sender approval row created but no delivery adapter is wired', {
+      approvalId,
+    });
+    return;
+  }
+
+  try {
+    await adapter.deliver(
+      target.messagingGroup.channel_type,
+      target.messagingGroup.platform_id,
+      null,
+      'chat-sdk',
+      JSON.stringify({
+        type: 'ask_question',
+        questionId: approvalId,
+        title,
+        question,
+        options,
+      }),
+    );
+    log.info('Unknown-sender approval card delivered', {
+      approvalId,
+      senderIdentity,
+      approver: target.userId,
+      messagingGroupId,
+      agentGroupId,
+    });
+  } catch (err) {
+    log.error('Unknown-sender approval card delivery failed', {
+      approvalId,
+      err,
+    });
+  }
+}
+
+/**
+ * Option value the admin clicked that means "allow" — shared with the
+ * response handler so the two sides can't drift.
+ */
+export const APPROVE_VALUE = 'approve';
+export const REJECT_VALUE = 'reject';
diff --git a/src/modules/permissions/user-dm.ts b/src/modules/permissions/user-dm.ts
index ef9566a..a5274d1 100644
--- a/src/modules/permissions/user-dm.ts
+++ b/src/modules/permissions/user-dm.ts
@@ -136,11 +136,14 @@ async function resolveDmPlatformId(channelType: string, handle: string): Promise
 function parseUserId(user: User): { channelType: string; handle: string } | { channelType: null; handle: null } {
   const idx = user.id.indexOf(':');
   if (idx < 0) return { channelType: null, handle: null };
-  const channelType = user.id.slice(0, idx);
+  const prefix = user.id.slice(0, idx);
   const handle = user.id.slice(idx + 1);
-  if (!channelType || !handle) return { channelType: null, handle: null };
-  // The `kind` on users mirrors the channel_type prefix in our current
-  // scheme. Pull it from `user.kind` if we ever decouple them later, but
-  // today the id prefix is authoritative.
-  return { channelType, handle };
+  if (!prefix || !handle) return { channelType: null, handle: null };
+  // Teams user IDs use a `29:` prefix, not `teams:`. When the id prefix
+  // isn't a registered adapter, fall back to user.kind and treat the full
+  // id as the handle.
+  if (!getChannelAdapter(prefix) && user.kind && getChannelAdapter(user.kind)) {
+    return { channelType: user.kind, handle: user.id };
+  }
+  return { channelType: prefix, handle };
 }
diff --git a/src/modules/scheduling/recurrence.test.ts b/src/modules/scheduling/recurrence.test.ts
new file mode 100644
index 0000000..358e6b4
--- /dev/null
+++ b/src/modules/scheduling/recurrence.test.ts
@@ -0,0 +1,98 @@
+/**
+ * Tests for `handleRecurrence` — specifically the timezone-aware cron
+ * interpretation ported from v1 (src/v1/task-scheduler.ts).
+ *
+ * Core invariant: cron expressions are interpreted in the user's TIMEZONE,
+ * not UTC. Without this, `"0 9 * * *"` fires at 09:00 UTC instead of 09:00
+ * user-local — a recurring scheduling bug users can't diagnose.
+ */
+import fs from 'fs';
+import path from 'path';
+import { afterEach, describe, expect, it } from 'vitest';
+
+import { ensureSchema, openInboundDb } from '../../db/session-db.js';
+import { insertTask } from './db.js';
+import { handleRecurrence } from './recurrence.js';
+import type { Session } from '../../types.js';
+
+const TEST_DIR = '/tmp/nanoclaw-recurrence-test';
+const DB_PATH = path.join(TEST_DIR, 'inbound.db');
+
+function freshDb() {
+  if (fs.existsSync(TEST_DIR)) fs.rmSync(TEST_DIR, { recursive: true });
+  fs.mkdirSync(TEST_DIR, { recursive: true });
+  ensureSchema(DB_PATH, 'inbound');
+  return openInboundDb(DB_PATH);
+}
+
+function fakeSession(): Session {
+  return {
+    id: 'sess-test',
+    agent_group_id: 'ag-test',
+    messaging_group_id: 'mg-test',
+    thread_id: null,
+    status: 'active',
+    created_at: new Date().toISOString(),
+    last_active: new Date().toISOString(),
+    container_status: 'stopped',
+  } as Session;
+}
+
+afterEach(() => {
+  if (fs.existsSync(TEST_DIR)) fs.rmSync(TEST_DIR, { recursive: true });
+});
+
+describe('handleRecurrence', () => {
+  it('clones a completed recurring task with a next-run in the future', async () => {
+    const db = freshDb();
+    insertTask(db, {
+      id: 'task-1',
+      processAfter: '2020-01-01T00:00:00.000Z',
+      recurrence: '0 9 * * *', // every day at 09:00 (user TZ)
+      platformId: null,
+      channelType: null,
+      threadId: null,
+      content: JSON.stringify({ prompt: 'daily digest' }),
+    });
+    db.prepare(`UPDATE messages_in SET status='completed' WHERE id='task-1'`).run();
+
+    await handleRecurrence(db, fakeSession());
+
+    const rows = db
+      .prepare(`SELECT id, status, process_after, recurrence, series_id FROM messages_in ORDER BY seq`)
+      .all() as Array<{
+      id: string;
+      status: string;
+      process_after: string;
+      recurrence: string | null;
+      series_id: string;
+    }>;
+    expect(rows).toHaveLength(2);
+    const original = rows.find((r) => r.id === 'task-1')!;
+    const follow = rows.find((r) => r.id !== 'task-1')!;
+    expect(original.recurrence).toBeNull();
+    expect(follow.status).toBe('pending');
+    expect(follow.recurrence).toBe('0 9 * * *');
+    expect(follow.series_id).toBe('task-1');
+    expect(new Date(follow.process_after).getTime()).toBeGreaterThan(Date.now());
+  });
+
+  it('does not clone rows whose recurrence is already cleared', async () => {
+    const db = freshDb();
+    insertTask(db, {
+      id: 'task-1',
+      processAfter: '2020-01-01T00:00:00.000Z',
+      recurrence: null,
+      platformId: null,
+      channelType: null,
+      threadId: null,
+      content: JSON.stringify({ prompt: 'one-off' }),
+    });
+    db.prepare(`UPDATE messages_in SET status='completed' WHERE id='task-1'`).run();
+
+    await handleRecurrence(db, fakeSession());
+
+    const count = (db.prepare(`SELECT COUNT(*) AS c FROM messages_in`).get() as { c: number }).c;
+    expect(count).toBe(1);
+  });
+});
diff --git a/src/modules/scheduling/recurrence.ts b/src/modules/scheduling/recurrence.ts
index a8a2e5c..d521f95 100644
--- a/src/modules/scheduling/recurrence.ts
+++ b/src/modules/scheduling/recurrence.ts
@@ -13,6 +13,7 @@
  */
 import type Database from 'better-sqlite3';
 
+import { TIMEZONE } from '../../config.js';
 import { log } from '../../log.js';
 import type { Session } from '../../types.js';
 import { clearRecurrence, getCompletedRecurring, insertRecurrence } from './db.js';
@@ -23,7 +24,11 @@ export async function handleRecurrence(inDb: Database.Database, session: Session
   for (const msg of recurring) {
     try {
       const { CronExpressionParser } = await import('cron-parser');
-      const interval = CronExpressionParser.parse(msg.recurrence);
+      // Interpret the cron expression in the user's timezone. v1 did this
+      // (src/v1/task-scheduler.ts:20-49); without it, a task written "0 9 * * *"
+      // by an agent running in a user's local TZ fires at 09:00 UTC instead of
+      // 09:00 user-local.
+      const interval = CronExpressionParser.parse(msg.recurrence, { tz: TIMEZONE });
       const nextRun = interval.next().toISOString();
       const prefix = msg.kind === 'task' ? 'task' : 'msg';
       const newId = `${prefix}-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
diff --git a/src/modules/self-mod/agent.md b/src/modules/self-mod/agent.md
index 33bca9b..9d67a4a 100644
--- a/src/modules/self-mod/agent.md
+++ b/src/modules/self-mod/agent.md
@@ -1,29 +1,28 @@
 # Self-modification
 
-You can install additional OS or npm packages, rebuild your container image,
-or add new MCP servers — but only with admin approval.
+You can install additional OS or npm packages or add new MCP servers — but
+only with admin approval.
 
 ## Tools
 
 - `install_packages({ apt?: string[], npm?: string[], reason?: string })` —
-  adds the listed packages to your container config and rebuilds the image
-  after admin approval. Package names are validated strictly (`[a-z0-9._+-]`
-  for apt, standard npm naming with optional scope). Max 20 packages per
-  request.
-
-- `request_rebuild({ reason?: string })` — rebuilds your container image
-  without config changes. Useful if the image has drifted from config.
+  adds the listed packages to your container config, rebuilds the image,
+  and restarts your container, all in a single admin approval step.
+  Package names are validated strictly (`[a-z0-9._+-]` for apt, standard
+  npm naming with optional scope). Max 20 packages per request.
 
 - `add_mcp_server({ name, command, args?, env? })` — adds a new MCP server
-  to your container config. The container restarts on next message so the
-  new server is available.
+  to your container config and restarts the container so the new server
+  is wired up on the next message. No image rebuild is required (bun runs
+  TS directly).
 
 ## Flow
 
 You call one of these tools → the host asks an admin via DM → admin approves
-or rejects. On approve, the config is applied and the container is killed;
-the host respawns it on the next message. You'll get a system chat message
-confirming the outcome (either "Packages installed..." or a failure reason).
+or rejects. On approve, the config is applied, the image is rebuilt if
+needed, and the container is killed; the host respawns it on the next
+message. You'll get a system chat message confirming the outcome (either
+"Packages installed..." or a failure reason).
 
 On reject you'll see "Your X request was rejected by admin."
 
diff --git a/src/modules/self-mod/apply.ts b/src/modules/self-mod/apply.ts
index da33fd0..5291937 100644
--- a/src/modules/self-mod/apply.ts
+++ b/src/modules/self-mod/apply.ts
@@ -5,6 +5,11 @@
  * pending_approvals row whose action matches. Each handler mutates the
  * container config, rebuilds/kills the container as needed, and lets the
  * host sweep respawn it on the new image on the next message.
+ *
+ * install_packages: rebuild image + kill container (apt/npm global installs
+ *   must be baked into the image layer).
+ * add_mcp_server: kill container only — bun runs TS directly, so a pure
+ *   MCP wiring change needs nothing more than a process restart.
  */
 import { updateContainerConfig } from '../../container-config.js';
 import { buildAgentGroupImage, killContainer } from '../../container-runner.js';
@@ -24,7 +29,10 @@ export const applyInstallPackages: ApprovalHandler = async ({ session, payload,
     if (payload.npm) cfg.packages.npm.push(...(payload.npm as string[]));
   });
 
-  const pkgs = [...((payload.apt as string[] | undefined) || []), ...((payload.npm as string[] | undefined) || [])].join(', ');
+  const pkgs = [
+    ...((payload.apt as string[] | undefined) || []),
+    ...((payload.npm as string[] | undefined) || []),
+  ].join(', ');
   log.info('Package install approved', { agentGroupId: session.agent_group_id, userId });
   try {
     await buildAgentGroupImage(session.agent_group_id);
@@ -51,24 +59,12 @@ export const applyInstallPackages: ApprovalHandler = async ({ session, payload,
     log.info('Container rebuild completed (bundled with install)', { agentGroupId: session.agent_group_id });
   } catch (e) {
     notify(
-      `Packages added to config (${pkgs}) but rebuild failed: ${e instanceof Error ? e.message : String(e)}. Call request_rebuild to retry.`,
+      `Packages added to config (${pkgs}) but rebuild failed: ${e instanceof Error ? e.message : String(e)}. Tell the user — an admin will need to retry the install_packages request or inspect the build logs.`,
     );
     log.error('Bundled rebuild failed after install approval', { agentGroupId: session.agent_group_id, err: e });
   }
 };
 
-export const applyRequestRebuild: ApprovalHandler = async ({ session, userId, notify }) => {
-  try {
-    await buildAgentGroupImage(session.agent_group_id);
-    killContainer(session.id, 'rebuild applied');
-    notify('Container image rebuilt. Your container will restart with the new image on the next message.');
-    log.info('Container rebuild approved and completed', { agentGroupId: session.agent_group_id, userId });
-  } catch (e) {
-    notify(`Rebuild failed: ${e instanceof Error ? e.message : String(e)}`);
-    log.error('Container rebuild failed', { agentGroupId: session.agent_group_id, err: e });
-  }
-};
-
 export const applyAddMcpServer: ApprovalHandler = async ({ session, payload, userId, notify }) => {
   const agentGroup = getAgentGroup(session.agent_group_id);
   if (!agentGroup) {
diff --git a/src/modules/self-mod/index.ts b/src/modules/self-mod/index.ts
index aedf1dc..e1f49e2 100644
--- a/src/modules/self-mod/index.ts
+++ b/src/modules/self-mod/index.ts
@@ -3,26 +3,28 @@
  *
  * Optional tier. Depends on the approvals default module for the request/
  * handler plumbing. On install the module registers:
- *   - Three delivery actions (install_packages, request_rebuild, add_mcp_server)
- *     that validate input and queue an approval via requestApproval().
- *   - Three matching approval handlers that run on approve: mutate the
- *     container config, rebuild the image, kill the container so the next
- *     wake picks up the change.
+ *   - Two delivery actions (install_packages, add_mcp_server) that validate
+ *     input and queue an approval via requestApproval().
+ *   - Two matching approval handlers that run on approve and perform the
+ *     complete follow-up:
+ *       install_packages → update container.json, rebuild image, kill
+ *         container (next wake respawns on the new image), schedule a
+ *         verify-and-report follow-up prompt.
+ *       add_mcp_server → update container.json, kill container. No image
+ *         rebuild — bun runs TS directly, so the new MCP server is wired
+ *         by the next container start.
  *
- * Without this module: the three MCP tools in the container still write
- * outbound system messages with these actions, but delivery logs
- * "Unknown system action" and drops them. Admin never sees a card; nothing
- * changes.
+ * Without this module: the MCP tools in the container still write outbound
+ * system messages with these actions, but delivery logs "Unknown system
+ * action" and drops them. Admin never sees a card; nothing changes.
  */
 import { registerDeliveryAction } from '../../delivery.js';
 import { registerApprovalHandler } from '../approvals/index.js';
-import { applyAddMcpServer, applyInstallPackages, applyRequestRebuild } from './apply.js';
-import { handleAddMcpServer, handleInstallPackages, handleRequestRebuild } from './request.js';
+import { applyAddMcpServer, applyInstallPackages } from './apply.js';
+import { handleAddMcpServer, handleInstallPackages } from './request.js';
 
 registerDeliveryAction('install_packages', handleInstallPackages);
-registerDeliveryAction('request_rebuild', handleRequestRebuild);
 registerDeliveryAction('add_mcp_server', handleAddMcpServer);
 
 registerApprovalHandler('install_packages', applyInstallPackages);
-registerApprovalHandler('request_rebuild', applyRequestRebuild);
 registerApprovalHandler('add_mcp_server', applyAddMcpServer);
diff --git a/src/modules/self-mod/project.md b/src/modules/self-mod/project.md
index bb6a0ec..556bcfe 100644
--- a/src/modules/self-mod/project.md
+++ b/src/modules/self-mod/project.md
@@ -1,20 +1,25 @@
 # Self-mod module
 
 Optional-tier module that gives agents admin-gated self-modification:
-installing OS/npm packages, rebuilding the container image, and registering
-new MCP servers. All three paths go through the approvals module's request
-primitive — no unapproved changes ever land.
+installing OS/npm packages and registering new MCP servers. Both paths go
+through the approvals module's request primitive — no unapproved changes
+ever land. The rebuild+restart (or restart-only) follow-up is bundled into
+the approval handler itself — there is no separate "request rebuild" step.
 
 ## What this module adds
 
-- Three delivery actions (`install_packages`, `request_rebuild`, `add_mcp_server`)
-  that the container's self-mod MCP tools write into outbound.db. On the host,
-  each handler validates input and queues an approval via
+- Two delivery actions (`install_packages`, `add_mcp_server`) that the
+  container's self-mod MCP tools write into outbound.db. On the host, each
+  handler validates input and queues an approval via
   `approvals.requestApproval()`.
-- Three matching approval handlers that run on approve: mutate the container
-  config via `updateContainerConfig`, rebuild the image via
-  `buildAgentGroupImage`, and kill the container so the host sweep respawns
-  it on the new image.
+- Two matching approval handlers that run on approve:
+  - `install_packages` → update `container.json`, rebuild the image via
+    `buildAgentGroupImage`, and kill the container so the host sweep
+    respawns it on the new image. Also schedules a verify-and-report
+    follow-up prompt ~5 s after kill.
+  - `add_mcp_server` → update `container.json` and kill the container.
+    No image rebuild — bun runs TS directly, so the new MCP wiring is
+    picked up on the next container start.
 
 ## Dependency
 
diff --git a/src/modules/self-mod/request.ts b/src/modules/self-mod/request.ts
index d965616..6cd7f05 100644
--- a/src/modules/self-mod/request.ts
+++ b/src/modules/self-mod/request.ts
@@ -1,10 +1,12 @@
 /**
  * Delivery-action handlers for agent-initiated self-modification requests.
  *
- * Three actions the container can write into messages_out (via the self-mod
- * MCP tools): install_packages, request_rebuild, add_mcp_server. Each one
- * validates input and queues an approval request. The admin's approval
- * triggers the matching approval handler in ./apply.ts.
+ * Two actions the container can write into messages_out (via the self-mod
+ * MCP tools): install_packages, add_mcp_server. Each one validates input
+ * and queues an approval request. The admin's approval triggers the
+ * matching approval handler in ./apply.ts, which also performs the
+ * required follow-up (rebuild+restart for install_packages, restart-only
+ * for add_mcp_server).
  *
  * Host-side sanitization for install_packages is defense-in-depth — the MCP
  * tool validates first. Both layers matter: the DB row carries the payload
@@ -61,23 +63,6 @@ export async function handleInstallPackages(content: Record<string, unknown>, se
   });
 }
 
-export async function handleRequestRebuild(content: Record<string, unknown>, session: Session): Promise<void> {
-  const agentGroup = getAgentGroup(session.agent_group_id);
-  if (!agentGroup) {
-    notifyAgent(session, 'request_rebuild failed: agent group not found.');
-    return;
-  }
-  const reason = (content.reason as string) || '';
-  await requestApproval({
-    session,
-    agentName: agentGroup.name,
-    action: 'request_rebuild',
-    payload: { reason },
-    title: 'Rebuild Request',
-    question: `Agent "${agentGroup.name}" is attempting to rebuild container.${reason ? `\nReason: ${reason}` : ''}`,
-  });
-}
-
 export async function handleAddMcpServer(content: Record<string, unknown>, session: Session): Promise<void> {
   const agentGroup = getAgentGroup(session.agent_group_id);
   if (!agentGroup) {
diff --git a/src/platform-id.ts b/src/platform-id.ts
new file mode 100644
index 0000000..1c49325
--- /dev/null
+++ b/src/platform-id.ts
@@ -0,0 +1,23 @@
+/**
+ * Determine whether a platform ID needs a channel-type prefix.
+ *
+ * Chat SDK adapters (Telegram, Discord, Slack, Teams, etc.) namespace their
+ * platform IDs with a channel prefix: "telegram:123456", "discord:guild:chan".
+ * The router stores channel_type and platform_id in separate columns, but
+ * Chat SDK adapters send the prefixed form as the platform_id — so any code
+ * that writes messaging_groups rows must produce the same shape the adapter
+ * will later emit as event.platformId, or router lookups miss and messages
+ * get silently dropped.
+ *
+ * Native adapters (Signal, WhatsApp, iMessage) use their own ID formats and
+ * send them as-is — no channel prefix. WhatsApp/iMessage emit JIDs/emails
+ * containing '@'. Signal emits raw phone numbers ('+15551234567') for DMs
+ * and 'group:<id>' for group chats. Prefixing any of these would cause a
+ * mismatch with what the adapter later emits.
+ */
+export function namespacedPlatformId(channel: string, raw: string): string {
+  if (raw.startsWith(`${channel}:`)) return raw;
+  if (raw.includes('@')) return raw;
+  if (raw.startsWith('+') || raw.startsWith('group:')) return raw;
+  return `${channel}:${raw}`;
+}
diff --git a/src/router.ts b/src/router.ts
index 8971f7f..3cf0192 100644
--- a/src/router.ts
+++ b/src/router.ts
@@ -18,31 +18,27 @@
  * for policy refusals.
  */
 import { getChannelAdapter } from './channels/channel-registry.js';
+import { gateCommand } from './command-gate.js';
+import { getAgentGroup } from './db/agent-groups.js';
 import { recordDroppedMessage } from './db/dropped-messages.js';
-import { getMessagingGroupByPlatform, createMessagingGroup, getMessagingGroupAgents } from './db/messaging-groups.js';
+import {
+  createMessagingGroup,
+  getMessagingGroupAgents,
+  getMessagingGroupWithAgentCount,
+} from './db/messaging-groups.js';
+import { findSessionForAgent } from './db/sessions.js';
 import { startTypingRefresh } from './modules/typing/index.js';
 import { log } from './log.js';
-import { resolveSession, writeSessionMessage } from './session-manager.js';
+import { resolveSession, writeSessionMessage, writeOutboundDirect } from './session-manager.js';
 import { wakeContainer } from './container-runner.js';
 import { getSession } from './db/sessions.js';
-import type { MessagingGroup, MessagingGroupAgent } from './types.js';
+import type { AgentGroup, MessagingGroup, MessagingGroupAgent } from './types.js';
+import type { InboundEvent } from './channels/adapter.js';
 
 function generateId(): string {
   return `msg-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
 }
 
-export interface InboundEvent {
-  channelType: string;
-  platformId: string;
-  threadId: string | null;
-  message: {
-    id: string;
-    kind: 'chat' | 'chat-sdk';
-    content: string; // JSON blob
-    timestamp: string;
-  };
-}
-
 /**
  * Sender-resolver hook. Runs before agent resolution.
  *
@@ -89,6 +85,50 @@ export function setAccessGate(fn: AccessGateFn): void {
   accessGate = fn;
 }
 
+/**
+ * Per-wiring sender-scope hook. Runs alongside the access gate for each
+ * agent that would otherwise engage — lets the permissions module enforce
+ * `sender_scope='known'` on wirings that are stricter than the messaging
+ * group's `unknown_sender_policy`. When the hook isn't registered (module
+ * not installed), sender_scope is a no-op.
+ */
+export type SenderScopeGateFn = (
+  event: InboundEvent,
+  userId: string | null,
+  mg: MessagingGroup,
+  agent: MessagingGroupAgent,
+) => AccessGateResult;
+
+let senderScopeGate: SenderScopeGateFn | null = null;
+
+export function setSenderScopeGate(fn: SenderScopeGateFn): void {
+  if (senderScopeGate) {
+    log.warn('Sender-scope gate overwritten');
+  }
+  senderScopeGate = fn;
+}
+
+/**
+ * Channel-registration hook. Runs when the router sees a mention/DM on a
+ * messaging group that has no wirings AND hasn't been denied. The hook is
+ * expected to escalate to an owner (card, etc.) and arrange for future
+ * replay via routeInbound after approval. Fire-and-forget from the
+ * router's perspective.
+ *
+ * Registered by the permissions module. Without the module the router
+ * silently records the drop with reason='no_agent_wired' and moves on.
+ */
+export type ChannelRequestGateFn = (mg: MessagingGroup, event: InboundEvent) => Promise<void>;
+
+let channelRequestGate: ChannelRequestGateFn | null = null;
+
+export function setChannelRequestGate(fn: ChannelRequestGateFn): void {
+  if (channelRequestGate) {
+    log.warn('Channel-request gate overwritten');
+  }
+  channelRequestGate = fn;
+}
+
 function safeParseContent(raw: string): { text?: string; sender?: string; senderId?: string } {
   try {
     return JSON.parse(raw);
@@ -109,18 +149,30 @@ export async function routeInbound(event: InboundEvent): Promise<void> {
     event = { ...event, threadId: null };
   }
 
-  // 1. Resolve messaging group
-  let mg = getMessagingGroupByPlatform(event.channelType, event.platformId);
+  const isMention = event.message.isMention === true;
 
-  if (!mg) {
+  // 1. Combined lookup: messaging_group row + count of wired agents in a
+  //    single query. Cheap short-circuit for the common "unwired channel"
+  //    case — one DB read and we're out, no auto-create, no sender
+  //    resolution, no log spam.
+  const found = getMessagingGroupWithAgentCount(event.channelType, event.platformId);
+
+  let mg: MessagingGroup;
+  let agentCount: number;
+  if (!found) {
+    // No messaging_groups row. Auto-create only when the message warrants
+    // attention (the bot was addressed — @mention or DM). Plain chatter in
+    // channels we merely sit in stays silent — no row, no DB writes.
+    if (!isMention) return;
     const mgId = `mg-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
     mg = {
       id: mgId,
       channel_type: event.channelType,
       platform_id: event.platformId,
       name: null,
-      is_group: 0,
-      unknown_sender_policy: 'strict',
+      is_group: event.message.isGroup ? 1 : 0,
+      unknown_sender_policy: 'request_approval',
+      denied_at: null,
       created_at: new Date().toISOString(),
     };
     createMessagingGroup(mg);
@@ -129,6 +181,51 @@ export async function routeInbound(event: InboundEvent): Promise<void> {
       channelType: event.channelType,
       platformId: event.platformId,
     });
+    agentCount = 0;
+  } else {
+    mg = found.mg;
+    agentCount = found.agentCount;
+  }
+
+  // 1b. No wirings — either silent drop (plain chatter / denied channel) or
+  //     escalate to owner for channel-registration approval.
+  if (agentCount === 0) {
+    if (!isMention) return;
+    if (mg.denied_at) {
+      log.debug('Message dropped — channel was denied by owner', {
+        messagingGroupId: mg.id,
+        deniedAt: mg.denied_at,
+      });
+      return;
+    }
+
+    const parsed = safeParseContent(event.message.content);
+    recordDroppedMessage({
+      channel_type: event.channelType,
+      platform_id: event.platformId,
+      user_id: null,
+      sender_name: parsed.sender ?? null,
+      reason: 'no_agent_wired',
+      messaging_group_id: mg.id,
+      agent_group_id: null,
+    });
+
+    if (channelRequestGate) {
+      // Fire-and-forget escalation. The gate is expected to build a card,
+      // persist pending_channel_approvals, and replay the event via
+      // routeInbound after approval. Errors are logged internally — the
+      // user's message still stays dropped here either way.
+      void channelRequestGate(mg, event).catch((err) =>
+        log.error('Channel-request gate threw', { messagingGroupId: mg.id, err }),
+      );
+    } else {
+      log.warn('MESSAGE DROPPED — no agent groups wired and no channel-request gate registered', {
+        messagingGroupId: mg.id,
+        channelType: event.channelType,
+        platformId: event.platformId,
+      });
+    }
+    return;
   }
 
   // 2. Sender resolution (permissions module upserts the users row as a
@@ -136,113 +233,235 @@ export async function routeInbound(event: InboundEvent): Promise<void> {
   //    Without the module, userId is null — downstream tolerates it.
   const userId: string | null = senderResolver ? senderResolver(event) : null;
 
-  // 3. Resolve agent groups wired to this messaging group. Structural
-  //    drops record to dropped_messages for audit.
+  // 3. Fetch wired agents in full (we already know the count is > 0; now
+  //    we need their actual rows for fan-out).
   const agents = getMessagingGroupAgents(mg.id);
-  if (agents.length === 0) {
-    log.warn('MESSAGE DROPPED — no agent groups wired to this channel. Run setup register step to configure.', {
-      messagingGroupId: mg.id,
-      channelType: event.channelType,
-      platformId: event.platformId,
-    });
-    const parsed = safeParseContent(event.message.content);
-    recordDroppedMessage({
-      channel_type: event.channelType,
-      platform_id: event.platformId,
-      user_id: userId,
-      sender_name: parsed.sender ?? null,
-      reason: 'no_agent_wired',
-      messaging_group_id: mg.id,
-      agent_group_id: null,
-    });
-    return;
-  }
 
-  const match = pickAgent(agents, event);
-  if (!match) {
-    log.warn('MESSAGE DROPPED — no agent matched trigger rules', {
-      messagingGroupId: mg.id,
-      channelType: event.channelType,
-    });
-    const parsed = safeParseContent(event.message.content);
-    recordDroppedMessage({
-      channel_type: event.channelType,
-      platform_id: event.platformId,
-      user_id: userId,
-      sender_name: parsed.sender ?? null,
-      reason: 'no_trigger_match',
-      messaging_group_id: mg.id,
-      agent_group_id: null,
-    });
-    return;
-  }
+  // 4. Fan-out: evaluate each wired agent independently against engage_mode,
+  //    sender_scope, and access gate. An agent that engages gets its own
+  //    session and container wake. An agent that declines but has
+  //    ignored_message_policy='accumulate' still gets the message stored in
+  //    its session (trigger=0) so the context is available when it does
+  //    engage later. Drop policy = skip silently.
+  //
+  //    Subscribe (for mention-sticky wirings on threaded platforms) fires
+  //    once per message from this loop — the first engaging mention-sticky
+  //    wiring triggers adapter.subscribe(...); subsequent wirings don't
+  //    re-subscribe (chat.subscribe is idempotent anyway, but the flag
+  //    avoids the extra await).
+  const parsed = safeParseContent(event.message.content);
+  const messageText = parsed.text ?? '';
 
-  // 4. Access gate (if the permissions module is loaded). Otherwise
-  //    allow-all.
-  if (accessGate) {
-    const result = accessGate(event, userId, mg, match.agent_group_id);
-    if (!result.allowed) {
-      log.info('MESSAGE DROPPED — access gate refused', {
-        messagingGroupId: mg.id,
-        agentGroupId: match.agent_group_id,
-        userId,
-        reason: result.reason,
+  let engagedCount = 0;
+  let accumulatedCount = 0;
+  let subscribed = false;
+
+  for (const agent of agents) {
+    const agentGroup = getAgentGroup(agent.agent_group_id);
+    if (!agentGroup) continue;
+
+    const engages = evaluateEngage(agent, messageText, isMention, mg, event.threadId);
+
+    const accessOk = engages && (!accessGate || accessGate(event, userId, mg, agent.agent_group_id).allowed);
+    const scopeOk = engages && (!senderScopeGate || senderScopeGate(event, userId, mg, agent).allowed);
+
+    if (engages && accessOk && scopeOk) {
+      await deliverToAgent(agent, agentGroup, mg, event, userId, adapter?.supportsThreads === true, true);
+      engagedCount++;
+
+      // Mention-sticky: ask the adapter to subscribe the thread so the
+      // platform's subscribed-message path carries follow-ups without
+      // requiring another @mention. Threaded-adapter only; DMs and
+      // non-threaded platforms skip.
+      if (
+        !subscribed &&
+        agent.engage_mode === 'mention-sticky' &&
+        adapter?.supportsThreads &&
+        adapter.subscribe &&
+        event.threadId !== null &&
+        mg.is_group !== 0
+      ) {
+        subscribed = true;
+        // Fire-and-forget — subscribe is platform-side bookkeeping and
+        // shouldn't block message routing. Errors are logged inside the
+        // adapter (or by the promise rejection handler below).
+        void adapter.subscribe(event.platformId, event.threadId).catch((err) => {
+          log.warn('adapter.subscribe failed', { channelType: event.channelType, threadId: event.threadId, err });
+        });
+      }
+    } else if (agent.ignored_message_policy === 'accumulate') {
+      await deliverToAgent(agent, agentGroup, mg, event, userId, adapter?.supportsThreads === true, false);
+      accumulatedCount++;
+    } else {
+      log.debug('Message not engaged for agent (drop policy)', {
+        agentGroupId: agent.agent_group_id,
+        engage_mode: agent.engage_mode,
+        engages,
+        accessOk,
+        scopeOk,
       });
-      return;
     }
   }
 
-  // 5. Resolve or create session.
-  //
-  // Adapter thread policy overrides the wiring's session_mode: if the adapter
-  // is threaded, each thread gets its own session regardless of what the
-  // wiring says. Agent-shared is preserved because it expresses a
-  // cross-channel intent the adapter can't know about.
-  //
-  // Exception: DMs (is_group=0). Sub-threads within a DM are a UX affordance,
-  // not a conversation boundary — treat the whole DM as one session and let
-  // threadId flow through to delivery so replies land in the right sub-thread.
-  let effectiveSessionMode = match.session_mode;
-  if (adapter && adapter.supportsThreads && effectiveSessionMode !== 'agent-shared' && mg.is_group !== 0) {
-    effectiveSessionMode = 'per-thread';
-  }
-  const { session, created } = resolveSession(match.agent_group_id, mg.id, event.threadId, effectiveSessionMode);
-
-  // 6. Write message to session DB
-  writeSessionMessage(session.agent_group_id, session.id, {
-    id: event.message.id || generateId(),
-    kind: event.message.kind,
-    timestamp: event.message.timestamp,
-    platformId: event.platformId,
-    channelType: event.channelType,
-    threadId: event.threadId,
-    content: event.message.content,
-  });
-
-  log.info('Message routed', {
-    sessionId: session.id,
-    agentGroup: match.agent_group_id,
-    kind: event.message.kind,
-    userId,
-    created,
-  });
-
-  // 7. Show typing indicator while the agent processes.
-  startTypingRefresh(session.id, session.agent_group_id, event.channelType, event.platformId, event.threadId);
-
-  // 8. Wake container
-  const freshSession = getSession(session.id);
-  if (freshSession) {
-    await wakeContainer(freshSession);
+  if (engagedCount + accumulatedCount === 0) {
+    recordDroppedMessage({
+      channel_type: event.channelType,
+      platform_id: event.platformId,
+      user_id: userId,
+      sender_name: parsed.sender ?? null,
+      reason: 'no_agent_engaged',
+      messaging_group_id: mg.id,
+      agent_group_id: null,
+    });
   }
 }
 
 /**
- * Pick the matching agent for an inbound event.
- * Currently: highest priority agent. Future: trigger rule matching.
+ * Decide whether a given wired agent should engage on this message.
+ *
+ *   'pattern'        — regex test on text; '.' = always
+ *   'mention'        — bot must be mentioned on the platform. Resolved by
+ *                      the adapter (SDK-level) and forwarded as
+ *                      `event.message.isMention`. Agent display name
+ *                      (`agent_group.name`) is irrelevant — users address
+ *                      the bot via its platform username (@botname on
+ *                      Telegram, user-id mention on Slack/Discord), not
+ *                      via the agent's NanoClaw-side display name. If a
+ *                      user wants to disambiguate between multiple agents
+ *                      wired to one chat, use engage_mode='pattern' with
+ *                      the disambiguator as the regex.
+ *   'mention-sticky' — platform mention OR an active per-thread session
+ *                      already exists for this (agent, mg, thread). The
+ *                      session existence IS our subscription state; once
+ *                      a thread has engaged us once, follow-ups arrive
+ *                      with no mention and should still fire.
  */
-function pickAgent(agents: MessagingGroupAgent[], _event: InboundEvent): MessagingGroupAgent | null {
-  // Agents are already ordered by priority DESC from the DB query
-  // TODO: apply trigger_rules matching (pattern, mentionOnly, etc.)
-  return agents[0] ?? null;
+function evaluateEngage(
+  agent: MessagingGroupAgent,
+  text: string,
+  isMention: boolean,
+  mg: MessagingGroup,
+  threadId: string | null,
+): boolean {
+  switch (agent.engage_mode) {
+    case 'pattern': {
+      const pat = agent.engage_pattern ?? '.';
+      if (pat === '.') return true;
+      try {
+        return new RegExp(pat).test(text);
+      } catch {
+        // Bad regex: fail open so admin sees the agent responding + can fix.
+        return true;
+      }
+    }
+    case 'mention':
+      return isMention;
+    case 'mention-sticky': {
+      if (isMention) return true;
+      // Sticky follow-up: session already exists for this (agent, mg, thread)
+      // — the thread was activated before, keep firing.
+      if (mg.is_group === 0) return false; // DMs never use mention-sticky sensibly
+      const existing = findSessionForAgent(agent.agent_group_id, mg.id, threadId);
+      return existing !== undefined;
+    }
+    default:
+      return false;
+  }
+}
+
+async function deliverToAgent(
+  agent: MessagingGroupAgent,
+  agentGroup: AgentGroup,
+  mg: MessagingGroup,
+  event: InboundEvent,
+  userId: string | null,
+  adapterSupportsThreads: boolean,
+  wake: boolean,
+): Promise<void> {
+  // Apply the adapter thread policy: threaded adapter in a group chat →
+  // per-thread session regardless of wiring. agent-shared preserved (it's
+  // a cross-channel directive the adapter doesn't know about). DMs collapse
+  // sub-threads to one session (is_group=0 short-circuit).
+  let effectiveSessionMode = agent.session_mode;
+  if (adapterSupportsThreads && effectiveSessionMode !== 'agent-shared' && mg.is_group !== 0) {
+    effectiveSessionMode = 'per-thread';
+  }
+
+  const { session, created } = resolveSession(agent.agent_group_id, mg.id, event.threadId, effectiveSessionMode);
+
+  // The inbound row's (channel_type, platform_id, thread_id) is the address
+  // the agent's reply will be delivered to. Normally it mirrors the source
+  // (stamped from the event). When the caller supplied `replyTo` (CLI admin
+  // transport acting on operator intent), the reply is redirected there.
+  const deliveryAddr = event.replyTo ?? {
+    channelType: event.channelType,
+    platformId: event.platformId,
+    threadId: event.threadId,
+  };
+
+  // Command gate: classify slash commands before they reach the container.
+  // Filtered commands are dropped silently. Denied admin commands get a
+  // permission-denied response written directly to messages_out.
+  if (event.message.kind === 'chat' || event.message.kind === 'chat-sdk') {
+    const gate = gateCommand(event.message.content, userId, agent.agent_group_id);
+    if (gate.action === 'filter') {
+      log.debug('Filtered command dropped by gate', { agentGroupId: agent.agent_group_id });
+      return;
+    }
+    if (gate.action === 'deny') {
+      writeOutboundDirect(session.agent_group_id, session.id, {
+        id: `deny-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
+        kind: 'chat',
+        platformId: deliveryAddr.platformId,
+        channelType: deliveryAddr.channelType,
+        threadId: deliveryAddr.threadId,
+        content: JSON.stringify({ text: `Permission denied: ${gate.command} requires admin access.` }),
+      });
+      log.info('Admin command denied by gate', { command: gate.command, userId, agentGroupId: agent.agent_group_id });
+      return;
+    }
+  }
+
+  writeSessionMessage(session.agent_group_id, session.id, {
+    id: messageIdForAgent(event.message.id, agent.agent_group_id),
+    kind: event.message.kind,
+    timestamp: event.message.timestamp,
+    platformId: deliveryAddr.platformId,
+    channelType: deliveryAddr.channelType,
+    threadId: deliveryAddr.threadId,
+    content: event.message.content,
+    trigger: wake ? 1 : 0,
+  });
+
+  log.info('Message routed', {
+    sessionId: session.id,
+    agentGroup: agent.agent_group_id,
+    engage_mode: agent.engage_mode,
+    kind: event.message.kind,
+    userId,
+    wake,
+    created,
+    agentGroupName: agentGroup.name,
+  });
+
+  if (wake) {
+    // Typing indicator + wake are only for the engaged branch; accumulated
+    // messages sit silently until a real trigger fires.
+    startTypingRefresh(session.id, session.agent_group_id, event.channelType, event.platformId, event.threadId);
+    const freshSession = getSession(session.id);
+    if (freshSession) {
+      await wakeContainer(freshSession);
+    }
+  }
+}
+
+/**
+ * When fanning out, the same inbound message lands in multiple per-agent
+ * session DBs. messages_in.id is PRIMARY KEY, so reuse of the raw id would
+ * collide across sessions (or, more subtly, within one session if re-routed
+ * after a retry). Namespace by agent_group_id to keep ids unique per session.
+ */
+function messageIdForAgent(baseId: string | undefined, agentGroupId: string): string {
+  const id = baseId && baseId.length > 0 ? baseId : generateId();
+  return `${id}:${agentGroupId}`;
 }
diff --git a/src/session-manager.ts b/src/session-manager.ts
index 7aaef24..38eaa0d 100644
--- a/src/session-manager.ts
+++ b/src/session-manager.ts
@@ -17,7 +17,14 @@ import path from 'path';
 import type { OutboundFile } from './channels/adapter.js';
 import { DATA_DIR } from './config.js';
 import { getMessagingGroup } from './db/messaging-groups.js';
-import { createSession, findSession, findSessionByAgentGroup, getSession, updateSession } from './db/sessions.js';
+import {
+  createSession,
+  findSession,
+  findSessionByAgentGroup,
+  findSessionForAgent,
+  getSession,
+  updateSession,
+} from './db/sessions.js';
 import {
   ensureSchema,
   openInboundDb as openInboundDbRaw,
@@ -89,7 +96,9 @@ export function resolveSession(
     }
   } else if (messagingGroupId) {
     const lookupThreadId = sessionMode === 'shared' ? null : threadId;
-    const existing = findSession(messagingGroupId, lookupThreadId);
+    // Scope lookup by agent_group_id so fan-out to multiple agents in the
+    // same chat doesn't accidentally deliver to the wrong agent's session.
+    const existing = findSessionForAgent(agentGroupId, messagingGroupId, lookupThreadId);
     if (existing) {
       return { session: existing, created: false };
     }
@@ -187,6 +196,13 @@ export function writeSessionMessage(
     content: string;
     processAfter?: string | null;
     recurrence?: string | null;
+    /**
+     * 1 = this message should wake the agent (the default); 0 = accumulate
+     * as context only, don't wake. Host's countDueMessages gates on this
+     * column; the container still reads all prior messages as context when
+     * a trigger-1 message does arrive.
+     */
+    trigger?: 0 | 1;
   },
 ): void {
   // Extract base64 attachment data, save to inbox, replace with file paths
@@ -204,6 +220,7 @@ export function writeSessionMessage(
       content,
       processAfter: message.processAfter ?? null,
       recurrence: message.recurrence ?? null,
+      trigger: message.trigger ?? 1,
     });
   } finally {
     db.close();
@@ -262,6 +279,34 @@ export function openOutboundDb(agentGroupId: string, sessionId: string): Databas
   return openOutboundDbRaw(outboundDbPath(agentGroupId, sessionId));
 }
 
+/**
+ * Write a message directly to a session's outbound DB so the host delivery
+ * loop picks it up. Used by the command gate to send denial responses
+ * without waking a container.
+ */
+export function writeOutboundDirect(
+  agentGroupId: string,
+  sessionId: string,
+  message: {
+    id: string;
+    kind: string;
+    platformId: string | null;
+    channelType: string | null;
+    threadId: string | null;
+    content: string;
+  },
+): void {
+  const db = openOutboundDb(agentGroupId, sessionId);
+  try {
+    db.prepare(
+      `INSERT OR IGNORE INTO messages_out (id, seq, timestamp, kind, platform_id, channel_type, thread_id, content)
+       VALUES (?, (SELECT COALESCE(MAX(seq), 0) + 2 FROM messages_out), datetime('now'), ?, ?, ?, ?, ?)`,
+    ).run(message.id, message.kind, message.platformId, message.channelType, message.threadId, message.content);
+  } finally {
+    db.close();
+  }
+}
+
 /**
  * @deprecated Use openInboundDb / openOutboundDb instead.
  */
diff --git a/src/types.ts b/src/types.ts
index ad14441..b3e2470 100644
--- a/src/types.ts
+++ b/src/types.ts
@@ -17,6 +17,16 @@ export interface MessagingGroup {
   name: string | null;
   is_group: number; // 0 | 1
   unknown_sender_policy: UnknownSenderPolicy;
+  /**
+   * When set, the owner explicitly denied registering this channel — the
+   * router drops silently and does not re-escalate. Cleared by any explicit
+   * wiring mutation (admin command). See migration 012.
+   *
+   * Optional on the TS type so pre-migration-012 callers that build
+   * MessagingGroup objects in code (fixtures, etc.) don't need to update;
+   * the column itself defaults to NULL in SQLite.
+   */
+  denied_at?: string | null;
   created_at: string;
 }
 
@@ -67,12 +77,23 @@ export interface UserDm {
   resolved_at: string;
 }
 
+export type EngageMode = 'pattern' | 'mention' | 'mention-sticky';
+export type SenderScope = 'all' | 'known';
+export type IgnoredMessagePolicy = 'drop' | 'accumulate';
+
 export interface MessagingGroupAgent {
   id: string;
   messaging_group_id: string;
   agent_group_id: string;
-  trigger_rules: string | null; // JSON: { pattern, mentionOnly, excludeSenders, includeSenders }
-  response_scope: 'all' | 'triggered' | 'allowlisted';
+  engage_mode: EngageMode;
+  /**
+   * Regex source string used when engage_mode='pattern'. `'.'` is the sentinel
+   * for "match every message" (the "always" flavor). Ignored for 'mention' /
+   * 'mention-sticky' modes.
+   */
+  engage_pattern: string | null;
+  sender_scope: SenderScope;
+  ignored_message_policy: IgnoredMessagePolicy;
   session_mode: 'shared' | 'per-thread' | 'agent-shared';
   priority: number;
   created_at: string;