bureau

Turn AI sessions into a maintained, inspectable canon.

Every AI session is an unrecorded meeting; natural-language docs drift, go stale, and contradict each other. bureau turns sessions into memory you can trust:

• Write, gated. Capture each session to an append-only logbook → compile into consistency-checked cabinet pages (the SSOT, with provenance) → review, the human gate that promotes a claim to canonical. AI-written claims are never fact until you approve them — memory works like version control, not a notepad the AI scribbles in.
• Read, tier-aware. query answers from the canon, citing each claim's trust tier and refusing to state an unverified one as fact. And BUREAU.md — the instructions init writes at your repo root and imports from CLAUDE.md — makes every AI session honor those tiers, so the gate governs all work, not just bureau commands.
• Inspect. A navigable offline gazette (the board), built by the bundled press inside this plugin (nothing else to install).

This is the Karpathy LLM-wiki pattern (LLM as compiler, not retriever) plus session provenance, a review gate, and an always-on BUREAU.md instruction — so the canon is current, traceable, and trusted.

New here? Start with the User Guide — a 60-second quickstart and a worked example.

Trust tiers — why memory here is safe

The cabinets double as repo memory, and no AI claim is recalled as fact until a human approves it. Every page carries a status: the AI must honor on recall:

status:	trust	who writes it
proposed	AI claim, unchecked	capture / compile
verified	checked against the repo	compile (automatic)
canonical	a human approved it	bureau:review only
stale	a verified source changed	staleness re-check
contested	two claims disagree	lint

AI writes only proposed/verified; the proposed → review → canonical gate is the double-check. Facts-about-artifacts auto-verify; judgments route to the human.

One self-contained plugin

	role
bureau (this plugin)	the engine: capture · compile · review · lint
press (press/, bundled inside bureau)	builds the gazette + runs the structural checks (deterministic)
bureau/ (in your repo)	your data: cabinet drawers + the logbook/ drawer

the press is a self-contained Node bundle vendored into the plugin (press/bin/gazette.mjs, no node_modules). bureau:inspect runs it directly — there is no separate install. The bundle is regenerated from the upstream renderer source by scripts/build-gazette.mjs.

Workspace layout

canon/             ← the content dir (default name); top-level folders are nav sections
  decisions/       ← a cabinet drawer (ADRs)
  architecture/    ← cabinet drawer (software profile)
  characters/      ← cabinet drawer (story profile)
  logbook/         ← append-only history — RENDERS as its own section
  _config.json     ← gazette meta
  bureau.json      ← profiles, board dir, autoCompile
bureau/crew/       ← bureau's control dir (the crew) — reserved, never rendered
gazette/           ← the rendered gazette (derived, gitignored, outside the workspace)

The canonical drawers (collectively, "cabinets") are the SSOT; logbook/ is the append-only history. Every cabinet claim links back to the minute that introduced it.

Commands

Command	Does
bureau:init	scaffold the workspace, write BUREAU.md + import it from CLAUDE.md, wire the press
bureau:note	take a live note into the running minute (run at decision points)
bureau:file-session	file the rich minute for the current session
bureau:compile	distil minutes into dossiers (with provenance)
bureau:review	the human gate — promote vetted claims to canonical, reject the rest
bureau:lint	semantic consistency sweep across the cabinets
bureau:query	answer a question from the canon — cited, tier-aware, never stating an unverified claim as fact
bureau:status	what's uncompiled / pending review / stale / contested
bureau:inspect	build + open the gazette (gazette)
bureau:crew	enable or author specialized agents (a "crew") that work the canon

Write (gated): note/file-session → compile → review. Read (tier-aware): query, plus BUREAU.md — written by init and imported from CLAUDE.md — which makes every AI session honor the trust tiers, so the gate governs all work, not just bureau commands.

Two hooks run automatically: SessionEnd writes a mechanical logbook stub (no session is ever lost); SessionStart-after-compaction re-grounds the agent from the logbook so decisions survive a context compaction.

Status

The full loop is implemented: capture (note/file-session + hooks) → compile → review → lint, read via query/status under the BUREAU.md gate, built by the bundled press. See dev-docs/plan.md.

Requirements

• Node.js ≥ 18 on PATH — the SessionEnd capture hook and the bundled press both run node. If Node is absent the hook simply no-ops (it never blocks session end); you can still capture with bureau:file-session.
• Nothing else — the press is bundled (press/), so there is no separate renderer to install.

Install

claude plugin install bureau@xiaolai --scope project

Part of the xiaolai marketplace.

License

MIT