lint

Lint — semantic consistency of the canon

the press's health lane is mechanical: it catches dangling links, orphans, stale dates, and typed contradicts: edges. It cannot read prose. Lint is the judgment layer — an LLM sweep that reads the dossiers and finds the inconsistencies that survive a structural check.

Run it on a cadence or before a milestone, not on every edit — it costs tokens and time. It is the right-hand column of the consistency table: judgment, not mechanism.

What it finds

1. Contradiction — two dossiers assert claims that cannot both be true (e.g. page A says auth tokens last 24h, page B says 1h).
2. Superseded — a page's settled (verified/canonical) claim is obsoleted by a newer dossier on the same topic; the canon still states the old truth. (Lint reads the cabinets, not the logbook — superseded evidence is cabinet-vs-cabinet.)
3. Gap — a concept referenced across pages but never defined on its own page.
4. Drift — one concept named differently across pages (vocabulary drift: "cabinet" vs "drawer" vs "page" for the same thing).

Rigor — verify before recording (the audit→fix→verify discipline)

A false contradiction is worse than a missed one — it erodes trust in the canon. So every candidate finding is adversarially verified before it is recorded:

1. Find. Read the corpus (chunk by drawer for a large workspace; fan out one reader per drawer when subagents are available). Collect candidate findings of each type.
2. Refute. For each candidate, try to disprove it — are the two claims actually about the same scope/time/entity, or is the conflict only apparent (different context, superseded already, profile-specific)? Default to discarding a candidate the refutation survives.
3. Record only survivors. A finding is written only if the refutation fails.

This mirrors cc-suite:audit-fix: dimensions → find → adversarially verify → report.

Profile lenses

Load bureau.json.profiles and weight the sweep:

• software → decision reversals, API/behavior claims, version/config values.
• story → canon facts, timeline order, character traits and relationships.

Output

Lint writes one rendered findings page: <workspace>/lint/findings.md (title Lint findings, the lint/ drawer renders as its own board section). It is overwritten each run — it states the CURRENT findings, not history (history lives in the logbook). Each finding lists its type, severity, the involved [[pages]] (body links, so they show as backlinks), and a one-line suggested resolution.

With --apply, lint also writes conservative, reversible in-place markers so the press's health lane surfaces the hard cases:

• a verified contradiction → set both pages status: contested and add a reciprocal single-line contradicts: [[Other page]] edge to each (2+ pages → one comma list contradicts: [[A]], [[B]], deduped, preserving any existing edge; never a multi-line YAML list, which the press ignores). the press renders the contradiction;
• a verified superseded claim → set the page status: stale. Gaps and drift are report-only (creating pages or renaming a concept is a human/compile decision, never an automatic edit).

Steps

1. Locate the workspace (bureau.json; default bureau). If none, tell the user to run bureau:init first and stop.
2. Read the cabinets. Load every cabinet drawer, EXCLUDING logbook/, the rendered board/, lint/ (lint's own findings — never lint the findings), and every _-prefixed file/dir (state ledgers, not canon). If there are no dossiers yet, report "no cabinets to lint" and stop.
3. Find + refute. Run the find→refute→record loop above for all four finding types, weighted by the active profiles.
4. Write the report. Write lint/findings.md with the surviving findings; if none survive, write a page stating the canon is consistent.
5. Apply markers (only when --apply): set contested/stale status and contradicts: edges for verified contradictions and superseded claims. Never edit a page's prose claims — only its status and the contradicts edge.
6. Structural check. Run bureau:inspect; report the contradictions the press now renders alongside the lint report.
7. Report. Summarize counts by type and severity, name each contested page, and point the user at lint/findings.md.

Rules

1. Read-mostly. Without --apply, lint writes only the report page; it never edits the canon's claims.
2. No prose edits. Even with --apply, lint changes only status: and contradicts: — resolving a claim is a human act.
3. Verified findings only. A recorded finding survived adversarial refutation.
4. Idempotent report. Re-running overwrites lint/findings.md with the current state.

Examples

Context: Before a milestone, the user wants the canon checked for inconsistencies. user: "bureau:lint" assistant: "Swept the cabinets. One survivor after refutation: **Token TTL** says 24h, **Session TTL** says 1h — a real contradiction (same artifact, same scope). I wrote it to lint/findings.md with both pages linked. The 'cabinet' vs 'drawer' wording I considered drift was refuted — they name different things — so I did not record it." Only findings that survive adversarial refutation are recorded; a false finding would erode trust in the canon. Context: The user wants lint to also mark the hard cases so the gazette surfaces them. user: "bureau:lint --apply" assistant: "Same findings, plus markers: I set both TTL pages to status: contested with reciprocal contradicts: edges, so the press's health lane now reports the contradiction. I changed only status and edges — never the prose claims; resolving which TTL is true is your call." --apply writes conservative, reversible markers only; prose claims are never edited.

Scope note

This skill covers ONLY the semantic-consistency sweep of dossiers. It does not capture sessions (capture / bureau:file-session), does not distil the logbook into cabinets (compile / bureau:compile), and does not render the gazette (bureau:inspect). It reads the cabinets that compile produced and is invoked by the bureau:lint command.