| name | recall |
| description | Answer from the bureau canon with citations and trust tiers — read dossiers as memory, honor each page's status, and never present an unverified claim as fact. Use when running bureau:query, or when the user asks what the project knows / what was decided / what the canon says about something. |
| argument-hint | "<question>" [--workspace <name>] |
Recall — read the canon, tier-aware
The cabinets are repo memory. Recall is the read side of the trust gate: it answers from the
compiled canon (not raw retrieval), cites what it used, and carries each claim's tier into
the answer so an unverified claim can never masquerade as fact.
Steps
- Locate the workspace (
bureau.json; default bureau). If none, tell the user to run
bureau:init first and stop.
- Find the pages that bear on it. Search the cabinet drawers for pages on the question — by
title, by drawer topic, and by following
[[links]] between pages. EXCLUDE logbook/,
board/, lint/, and every _-prefixed entry (those are history, output, or state, not
canon).
- Answer only from those pages. Synthesize a direct answer. Do NOT add knowledge the canon
does not contain; if you must reason beyond it, mark that clearly as inference.
- Cite with tiers. For each claim used, name the page, its
status: (canonical /
verified / proposed / stale / contested), and the [[session …]] provenance.
- Never state non-
canonical as fact. canonical is fact; verified is "checked, not
approved"; proposed/stale/contested are claims to confirm — flag them in the answer.
- Name gaps. If the canon does not cover the question, say so; do not fabricate. Suggest
the next step (file a session, run
bureau:lint).
- Offer to keep it. If the synthesis reaches a conclusion worth remembering, offer to
capture it (
bureau:note) so it enters capture → compile → review, never straight to canon.
Rules
- Compiled canon only. Read dossiers; never re-derive an answer from raw sources or
the logbook prose — that is what compile is for.
- Tier travels. Every cited claim carries its
status:; the answer's confidence is capped
by its weakest load-bearing tier.
- No invention. A gap is reported as a gap, not filled with a guess.
- Read-only. Recall never edits cabinets; capturing a new conclusion goes through the gate.
Examples
Context: A repo with a bureau canon; the user asks a settled question.
user: "bureau:query \"how long do auth tokens last?\""
assistant: "From **Token TTL** (`canonical`, sourced to session a1b2c3): auth tokens last 24h. That's an approved fact."
A canonical page answers as fact, with the page, its tier, and its provenance cited.
Context: the page that answers it is unreviewed.
user: "What did we decide about the queue retry policy?"
assistant: "The canon has **Queue retry** but it's `proposed` (unreviewed, from session d4e5f6) — so treat this as unverified: it says 3 retries with backoff. Want me to put it through bureau:review before you rely on it?"
A non-canonical claim is answered WITH its tier and an explicit 'not fact yet' flag — never presented as settled.
Scope note
This skill covers ONLY reading the canon to answer a question. It does not capture sessions
(capture / bureau:file-session), not distil the logbook (compile), not approve
claims (review), and not render the gazette (bureau:inspect). It is invoked by the
bureau:query command, and auto-triggers when the user asks what the project knows.
