When the agent outputs bare text without <message to="..."> blocks,
nothing gets delivered — silent failure. Now the poll-loop pushes a
one-shot correction back into the active query telling the agent to
re-send with proper wrapping. Capped at once per user turn to avoid
loops; resets when a new follow-up message arrives.
Also updates destination instructions to require explicit <internal>
wrapping for scratchpad instead of treating bare text as implicit
scratchpad.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The poll loop had a bare-text routing fallback in dispatchResultText: when
the agent produced text without <message to="..."> wrapping, it would auto-
route to the session's originating channel (via a frozen RoutingContext) or
to the single configured destination. This caused three problems:
1. Routing drift: RoutingContext was extracted once from the initial batch
and never refreshed. When the initial batch was a null-routed cron task
and a real chat arrived mid-query, replies were silently dropped to
scratchpad because the frozen routing had all-null fields.
2. Cross-channel thread bleed: sendToDestination applied a single
routing.threadId to every outbound message regardless of destination.
In agent-shared sessions (multiple channels sharing one session), one
channel's thread ID was stamped onto messages to a different channel.
3. Inconsistent formatting: task, webhook, and system messages had no
origin metadata in their formatted output, so the agent couldn't tell
which destination they came from — even when the underlying messages_in
rows carried routing fields.
Changes:
- Remove the bare-text routing fallbacks in dispatchResultText (both the
routing-based and single-destination shortcuts). All agent output must
be wrapped in <message to="name">...</message>. Bare text is scratchpad.
- Update buildDestinationsSection() to require explicit wrapping for all
groups, including single-destination. No more "no special wrapping
needed" shortcut.
- Resolve thread_id per-destination via resolveDestinationThread(), which
queries messages_in for the most recent message matching the target
channel+platform. Falls back to null (top-level channel message) when
no prior inbound exists for that destination.
- Extract originAttr() helper in formatter.ts and apply it to all message
types. Tasks now render as <task from="dest" time="...">, webhooks as
<webhook from="dest" source="..." event="...">, system responses as
<system_response from="dest" ...>. The agent always sees where a
message originated.
- Add a PreCompact shell hook (compact-instructions.ts) that outputs
custom compaction instructions, telling the compactor to preserve
recent message XML structure and routing metadata in the summary.
Wired via settings.json in the .claude-shared scaffold, with a
migration path (ensurePreCompactHook) for existing groups.
Relation to open PRs:
- #2277 (mergeRouting) becomes unnecessary — the routing fallback it
patches no longer exists. Can be closed.
- #2327 (post-compaction destination reminder) is complementary — it
handles the post-compaction push, this handles pre-compaction
instructions. Both can merge independently.
- #2328 (default routing instruction) is complementary — it adds "reply
to the from= destination" guidance to the multi-destination section.
Compatible with the unified instruction format here.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a multi-destination agent receives an inbound message, the model
had no explicit guidance about which destination to address by default
and would sometimes pick the wrong one — e.g. Casa replying to the
admin's group questions in Laura's DM instead of in the group itself.
The formatter already injects `from="<destname>"` on every inbound
<message> tag (formatter.ts:184), so the origin is right there in the
prompt — the system prompt just never told the agent to use it.
Added one line to buildDestinationsSection() that nudges the agent
toward replying via the same destination the message came from, with
an out for explicit cross-destination requests ("tell Laura that…").
Single-destination groups are unaffected (they take a separate
short-circuit path with a fallback that auto-replies to the origin).
Tests cover the multi-destination, single-destination, and
no-destination cases.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Move every agent-specific instruction out of the shared container/CLAUDE.md
so the base is genuinely universal. Persona/identity now comes from the
system-prompt addendum (buildSystemPromptAddendum now takes assistantName
and prepends "# You are {name}"). Per-module instructions live alongside
each MCP tool source:
container/agent-runner/src/mcp-tools/core.instructions.md
container/agent-runner/src/mcp-tools/scheduling.instructions.md
container/agent-runner/src/mcp-tools/self-mod.instructions.md
composeGroupClaudeMd() scans that directory and emits `module-<name>.md`
fragments as symlinks to /app/src/mcp-tools/<name>.instructions.md (valid
via the existing RO source mount). Skill fragments renamed to
`skill-<name>.md` for naming consistency with `module-*` and `mcp-*`.
Mount tightening so composer-managed files can't be clobbered by agent
writes: nested RO mounts for /workspace/agent/CLAUDE.md and
/workspace/agent/.claude-fragments/. CLAUDE.local.md (per-group memory)
stays RW as the only writable CLAUDE.md-family file.
.gitignore: ignore CLAUDE.local.md, .claude-shared.md, .claude-fragments/
everywhere, and simplify groups/ rules to ignore the whole tree (per-
installation state, not tracked).
Co-Authored-By: Claude Opus 4.7 (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>
When an agent has exactly one configured destination, wrapping output in
<message to="..."> blocks is unnecessary. Plain text goes to the sole
destination automatically. This preserves the simple "just reply" flow
for the common case of one user on one channel.
Applies in three places:
- System prompt addendum: single-destination case gets a simplified
explanation ("your messages are delivered to X, just write directly").
Multi-destination case keeps the <message to="..."> syntax docs.
- Main output parser: if zero <message> blocks are found and there is
exactly one destination, the entire cleaned text (with <internal>
stripped) is sent to that destination.
- send_message / send_file MCP tools: `to` parameter is now optional.
With one destination, omitted defaults to it. With multiple, omitting
returns an error listing the options.
Multi-destination behavior is unchanged — explicit <message to="..."> is
still required, and untagged text is still scratchpad.
groups/global/CLAUDE.md updated to describe both cases.
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>
