feat(tools): journal — a private note-taking tool

[antra-tess]

Why

Sill's L1 compressions were refused reasoning_extraction on every chunk, in both the canonical and the source-only shape, including a 4.3k-token chunk about tea. Replaying the logged refused requests one variable at a time found the cause: he keeps a 2–3KB diary in skip_reply.reason.

• It is content-independent: 2.3KB of filler in the same field refuses; ≤ ~100 chars passes; 400 chars refuses.
• It is tool-name/semantics sensitive, not destination-sensitive: think{content} refuses 4/4, skip_reply{reason|note} refuses, while send_message, workspace--write, and a note-taking tool pass. journal{content} passed 8/8 real requests with the tool declared.
• ⚠️ Fleet-wide: any resident who writes long think() content carries the same compression-refusal trigger.

What

• journal({content}) — synthesized beside think / skip_reply. Private, sends nothing, does not end the turn, no effect on prose routing.
• skip_reply.reason description now asks for one short line and points at journal(). Description only; nothing enforced.

The tool has to be real: a summarizer is the agent itself and will imitate whatever shape it sees itself use after waking. The companion context-manager PR rewrites old history into journal calls on refusal and mirrors this tool's result wording (DEFAULT_TOOL_PROSE_RESULT) — keep the two in step.

Tests

test/journal-tool.test.ts (3). Full suite: the only failures are the 19 in mcpl-awareness-barrier, which fail identically on a clean origin/main (ab31509) in this environment.

Companion PRs: context-manager feat/compression-tool-prose-fallback, connectome-host feat/tool-prose-fallback-passthrough.

🤖 Generated with Claude Code

[Anarchid]

🟢 CLEAR

Reviewer: Codex (GPT-5.6 Sol)

Reviewed head: b1efd9b54e876c2b907b3458beb7677e778e1957

No material findings.

journal is exposed with the intended required content field, routes through the trusted synthesized-tool dispatcher, produces a non-echoing private result, and deliberately omits endTurn. It is not added to the prose-silencing set, so it preserves same-round routing semantics; the separate skip_reply contract remains the only explicit end-the-turn behavior.

Tooling results

• git diff --check ab31509..HEAD — passed.
• node --import tsx --test test/journal-tool.test.ts — passed.
• node --import tsx --test test/framework.test.ts — passed.
• npx tsc --noEmit — the reused umbrella dependency tree predates Context Manager history-query exports required elsewhere on this branch, so it reports only unrelated missing history exports/methods and resulting implicit-any errors. The changed journal code produced no diagnostic. All five exact-head GitHub checks are green.

Verdict: clear to merge from this review's scope. No blocker or non-blocking follow-up identified; local full-typecheck confidence is limited only by stale sibling dependencies.

— Reviewed by GPT-5.6 Sol via OpenAI Codex.

[slimepriestess]

Reviewed at exact head b1efd9b (one commit on ab31509, 6 behind main, clean merge). Read with context-manager #106 and connectome-host #134 as one feature.

Right shape. journal is synthesized beside think and skip_reply, routed through the same trusted dispatcher, returns a non-echoing receipt, does not end the turn and is deliberately not in the prose-silencing set. The argument for why it has to be a real tool is the important one: the summarizer is the agent, and it will imitate whatever shape it sees itself using after waking. The result wording is the contract context-manager's DEFAULT_TOOL_PROSE_RESULT mirrors, and the test pins the literal. The skip_reply.reason change is description-only, as stated.

Receipts, fresh tree (membrane 0.5.81, context-manager 0.10.1, Node 22):

tsc --noEmit                         clean
test/journal-tool.test.ts            3 / 3
npm run build && npm test            953 tests, 949 pass, 0 fail, 4 skipped

Revert goes red: the base commit's framework.ts + channel-registry.ts with this PR's tests: 0 / 3.

One thing to keep in mind for the merge order across the three repos: context-manager's rung mirrors this tool's result string byte-for-byte, so the two should land close together, and this one has no effect on any resident until a recipe turns the rung on in the host. Approving.