Closesqwibitai/nanoclaw#2325.
When the Claude Code SDK auto-compacts the conversation context, the
compaction summary tends to drop the agent's learned <message to="…">
wrapping discipline. The destinations table is still populated and the
system prompt still lists them, but the behavioral pattern degrades —
A2A sends and multi-channel routing silently revert to bare-text or
single-channel delivery for the rest of the session, until the next
/clear.
Three small changes wire a reminder back into the live query when this
fires:
- New `compacted` event on ProviderEvent. Distinct from `result` so it
doesn't mark the turn completed or get dispatched as a chat message
(which is also why "Context compacted (N tokens compacted)." stops
appearing as noise in user-facing chats — it was a side-effect of
reusing the result event path).
- ClaudeProvider yields `compacted` instead of `result` for the SDK's
compact_boundary system event.
- Poll-loop's event handler reacts by pushing a system-tagged reminder
back into the active query when there are >1 destinations. Single-
destination groups skip the push since they have a fallback that
works without wrapping.
Tests cover both branches (multi-destination → reminder fires;
single-destination → no reminder) using a CompactingProvider that
emits the new event mid-stream.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Before, every provider stored its opaque continuation id under the
single outbound.db key `sdk_session_id`. Flipping a session's
agent_provider (e.g. Codex → Claude) meant the new provider read the
old provider's id at wake, handed it to its own SDK, and got a
"No conversation found" error that cost the user one sacrificed
message before the stale-session recovery path cleared the id.
This reshapes session_state so continuations are keyed
`continuation:<provider>` instead. Consequences:
- Per-provider continuations coexist. Flipping Claude → Codex → Claude
resumes the Claude thread exactly where it left off, with the
intervening Codex thread also still on file.
- No provider ever reads another provider's id. Switching costs no
sacrificed message and emits no transient error.
- Legacy installs are migrated forward on first startup:
migrateLegacyContinuation() adopts any pre-existing `sdk_session_id`
row into the current provider's slot (best guess — it was whichever
provider ran last), then deletes the legacy row unconditionally so
it can't poison a future provider's read.
runPollLoop now takes providerName alongside the provider instance,
and threads it through processQuery to setContinuation on init.
Tests: 9 new tests covering set/get isolation across providers,
clear-specificity, legacy-adoption, legacy-always-deleted,
prefer-existing-slot-over-legacy, and idempotency of a second
migration call.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Container side:
- agent-runner switches to Bun. Drops better-sqlite3 (native compile gone),
drops tsc build step in-image AND the tsc-on-every-session-wake in the
entrypoint — bun runs src/index.ts directly. bun:sqlite replaces
better-sqlite3; cross-mount DB invariants (journal_mode=DELETE, busy_timeout)
preserved. Named params converted from @name to $name because bun:sqlite
does not auto-strip the prefix the way better-sqlite3 does.
- Tests ported from vitest to bun:test (only describe/it/expect/before/afterEach
used, API-compatible). vitest.config.ts excludes container/agent-runner/.
- bun.lock replaces pnpm-lock.yaml + pnpm-workspace.yaml under
container/agent-runner/. Host pnpm workspace does NOT include this tree.
Dockerfile improvements (independent of Bun but bundled while touching the file):
- tini as PID 1 for correct SIGTERM propagation (prevents half-written
outbound.db on shutdown).
- Extracted entrypoint.sh — readable and diffable vs the old inline printf.
- BuildKit cache mounts for apt + bun install + pnpm install.
- --no-install-recommends on apt, pinned CLAUDE_CODE_VERSION, AGENT_BROWSER,
VERCEL, BUN_VERSION.
- CJK fonts (~200MB) behind ARG INSTALL_CJK_FONTS=false; build.sh reads from
.env; setup/container.ts reads the same .env so /setup and manual rebuild
stay in sync.
- PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 in case any postinstall tries to pull a
redundant Chromium.
- /home/node 755 (was 777).
Host side:
- src/container-runner.ts dynamic spawn command collapses from
`pnpm exec tsc --outDir /tmp/dist … && node /tmp/dist/index.js` to
`exec bun run /app/src/index.ts` — cold start ~200-500ms faster per wake.
CI:
- oven-sh/setup-bun@v2 alongside Node/pnpm. Adds explicit container
typecheck (was documented in CLAUDE.md, not enforced) and `bun test` for
agent-runner tests.
Reshape AgentProvider so provider-specific assumptions stop leaking into
the generic layer. No change to what reaches sdkQuery() — same values,
different plumbing.
- QueryInput: opaque `continuation` replaces `sessionId` + `resumeAt`;
`systemContext.instructions` replaces ambiguous `systemPrompt`;
`mcpServers`, `env`, `additionalDirectories` move to `ProviderOptions`
at construction time.
- AgentProvider gains `isSessionInvalid(err)` and
`supportsNativeSlashCommands` so the poll-loop stops regex-matching
Claude error strings and gates passthrough slash commands per provider.
- ClaudeProvider owns `CLAUDE_CODE_AUTO_COMPACT_WINDOW` and the
stale-session regex internally.
- ProviderEvent.activity kept and documented as the liveness signal
(fires on every SDK message so the idle timer stays honest during
long tool runs); init carries `continuation` instead of `sessionId`.
- poll-loop drops mcpServers/env/systemPrompt from its config; admin
user id now passed explicitly.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The per-session destination map was being written as a sidecar JSON file
(/workspace/.nanoclaw-destinations.json) — inconsistent with the rest of
v2, where all host↔container IO goes through inbound.db / outbound.db.
Move it into a `destinations` table in INBOUND_SCHEMA. The host writes
it before every container wake AND on demand (e.g. after create_agent)
so the creator sees the new child destination mid-session without a
restart. The container queries the table live on every lookup — no
cache, no staleness window.
- src/db/schema.ts: add `destinations` table to INBOUND_SCHEMA.
- src/session-manager.ts: writeDestinationsFile → writeDestinations,
writes via DELETE + INSERT inside a transaction.
- src/delivery.ts: create_agent handler calls writeDestinations on the
creator's session after inserting the new destination rows.
- container/agent-runner/src/destinations.ts: queries inbound.db
directly in every findByName/getAllDestinations/findByRouting call.
No more cache. No setDestinationsForTest (obsolete). No fs import.
- container/agent-runner/src/index.ts and mcp-tools/index.ts: remove
loadDestinations() calls — no longer needed.
- Test helper initTestSessionDb creates the destinations table.
Integration test inserts a row directly instead of mocking the cache.
No backwards compatibility: sessions predating the schema update must
be recreated. This is fine on the v2 branch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces implicit routing context (NANOCLAW_PLATFORM_ID env vars) with
per-agent named destination maps. Agents reference channels and peer
agents by local names; the host re-validates every outbound route against
a new agent_destinations table that is both the routing map and the ACL.
Model changes:
- New migration 004 adds agent_destinations (agent_group_id, local_name,
target_type, target_id). Backfills from existing messaging_group_agents.
- Host writes /workspace/.nanoclaw-destinations.json before every container
wake so admin changes take effect on next start.
- Container loads map at startup, appends system-prompt addendum listing
available destinations and the <message to="name">…</message> syntax.
- Agent main output is parsed for <message to="..."> blocks; each block
becomes a messages_out row with routing resolved via the local map.
Untagged text and <internal>…</internal> are scratchpad (logged only).
- send_message MCP tool now takes `to` (destination name) instead of raw
routing fields. send_to_agent deleted (redundant — agents are just
destinations). send_file/edit_message/add_reaction route via map too.
- Inbound formatter adds from="name" attribute via reverse-lookup so the
agent sees a consistent namespace in both directions.
Permission enforcement:
- Host checks hasDestination() before every channel delivery AND every
agent-to-agent route. Unauthorized messages dropped and logged.
- routeAgentMessage simplified: ~15 lines, no JSON parse, content copied
verbatim (target formatter resolves the sender via its own local map).
- create_agent is admin-only, checked at both the container (tool not
registered for non-admins) and the host (re-check on receive). Inserts
bidirectional destination rows so parent↔child comms work immediately.
Includes path-traversal guard on folder name.
Self-modification cleanup:
- add_mcp_server now requires admin approval (previously had none).
- install_packages validates package names on BOTH sides (container tool
+ host receiver) with strict regex. Max 20 packages per request.
- All three self-mod tools are fire-and-forget: write request, return
immediately with "submitted" message. Admin approval triggers a chat
notification to the requesting agent — no tool-call polling, no 5-min
holds. On rebuild/mcp_server approval, the container is killed so the
next wake picks up new config/image.
- Approval delivery extracted into requestApproval() helper (the one
place where three call sites were literally identical).
Also folded in the phase-1 dynamic import cleanup (create_agent no longer
does `await import('./db/agent-groups.js')`) and removes NANOCLAW_PLATFORM_ID
/ CHANNEL_TYPE / THREAD_ID env-var routing entirely.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eliminates SQLite write contention across the host-container mount
boundary by splitting the single session.db into two files, each with
exactly one writer:
inbound.db — host writes (messages_in, delivered tracking)
outbound.db — container writes (messages_out, processing_ack)
Key changes:
- Host uses even seq numbers, container uses odd (collision-free)
- Container heartbeat via file touch instead of DB UPDATE
- Scheduling MCP tools now emit system actions via messages_out
(host applies them to inbound.db during delivery)
- Host sweep reads processing_ack + heartbeat file for stale detection
- OneCLI ensureAgent() call added (was missing from v2, caused
applyContainerConfig to reject unknown agent identifiers)
Verified: tsc clean, 327 tests pass, real e2e through Docker works.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Poll loop end-to-end with mock provider: message pickup, batch
processing, concurrent polling for late arrivals.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
