lint

/lint — Structural + corpus-level checks over <paths.wiki>/

Also reachable via /hi <natural language> (e.g., /hi lint the wiki, /hi wiki audit, /hi wiki orphans). See harness/intents.toml [intents.lint] for the row's example phrases. Both paths execute this same procedure.

Deterministic Python pass. The LLM never hand-checks structure — scripts/lint.py is the single source of truth, mirroring the scripts/trust.py pattern.

Scope: Three passes. (0) Harness portability, $OV ingestion hygiene, and privacy checks. (1) Structural: everything under <paths.wiki>/. (2) Staleness: L2 working-layer directories (<paths.agent_findings>/, <paths.wip>/, <paths.gtd>/, <paths.preprints>/, <paths.reflections>/, <paths.research>/). Structural lint enforces the wiki schema; staleness lint surfaces L2 notes that need attention (archival, compaction, or promotion to L4).

What gets checked:

Check	Severity	Source
Per-note parse errors: schema items 1-10, malformed citation metadata, or dangling targets (parse-error)	ERROR	scripts/trust.py parser + resolver
Disputed claim or fully expired evidence (claim-attention)	WARN	scripts/wiki_review.py; reasons are in autoevo-applied-* receipts. Route to the user.
Duplicate titles across wiki entries (breaks citation resolution)	ERROR	scripts/lint.py
Slug ↔ title alignment (filename stem matches H1 title)	WARN	scripts/lint.py
Orphan entry: no inbound citation from another wiki entry	WARN	scripts/lint.py graph topology
No outbound citation to another wiki entry	INFO	scripts/lint.py graph topology
Shared anchor without a citation edge between entries	INFO	scripts/lint.py graph topology
url: or gist: anchor missing readwise: field (readwise-missing)	WARN	scripts/lint.py — save to Readwise with anchor-evidence tag and backfill the document ID; fix via uv run scripts/snapshot_anchors.py --apply --note "<paths.wiki>/<Title>.md"
Technical term in claim body not in vocabulary allowlist and not matching any wiki entry title (unfounded-term)	INFO	scripts/lint.py — add term to scripts/wiki_vocabulary.txt if common knowledge, or add a wiki entry, or add a parenthetical definition inline
Localized shadow missing for a configured language (shadow-missing)	WARN	scripts/lint.py — regenerate the shadow from the English entry.
Localized shadow older than English source (shadow-stale)	WARN	scripts/lint.py — re-translate the localized shadow to match the updated English source
Public configuration schema and Claude/Codex harness alignment (registry-schema, registry-read, registry-validator, source/reference/edge findings)	ERROR/WARN/INFO	harness/registry.schema.json + scripts/harness_lint.py
$OV ingestion hygiene (missing READMEs, raw-without-digest, archive↔working-tier overlap, root-level orphans, empty .md files, suspicious top-level dirs, vault layout, duplicate Reflect titles)	INFO (advisory)	scripts/zk_audit.py — see protocols/drive-zk-ingestion.md § Post-ingestion verification

Not checked: cross-note @anchor date consistency. Per protocols/wiki-schema.md, valid_at is the day the marker was added to its home note, so the same source being anchored from two notes on different days is the normal case.

Exit code: 0 if no ERROR-level findings, 1 otherwise. WARN and INFO never fail the run.

Process

Phase 0: Harness health

Bash: python3 scripts/harness_lint.py --json

Read counts and findings (severity, code, where, message). Any ERROR blocks the run until fixed; WARN and INFO are advisory. This pass owns the AGENTS.md size and bold-marker checks; do not repeat them manually.

registry-schema identifies the source file and failing field. A registry-validator setup error requires installing the pinned dependencies with uv sync --locked before retrying; lint itself never installs them.

Phase 0b: $OV ingestion hygiene audit

Bash: uv run scripts/zk_audit.py --json

Read categories and total; category rows identify where and detail.

Advisory only: never blocks the run. Exit code 0 unless $OV is missing (exit 2). Surface a one-line summary per non-empty category. Detailed listings are read on demand via uv run scripts/zk_audit.py (no --json). Source of truth: protocols/drive-zk-ingestion.md § Post-ingestion verification.

Phase 0c: Privacy leak scan

Bash: uv run scripts/privacy_check.py --json

--json mode emits a document on every run regardless of exit code. Route on its action field: "proceed" → pass (WARN first on any coverage_warnings); "soft_skip" → note "privacy gate skipped (<reason>)" and continue; "abort" → ERROR, block and present each hits entry verbatim. No JSON / exit ≥ 2 without JSON → real script error: surface stderr, soft-skip.

Any non-empty hits array is an ERROR: each entry is a private identifier (filename stem, wikilink target, slug from profile/private_slugs.txt, or exact term from profile/private_terms.txt) that appears in a public-bound pathname, worktree file, or staged blob. Present each hit verbatim with its file, source, and line number. Remediation:

• Replace the private title with a generic placeholder (e.g., Sample Wiki Entry, Topic A).
• Add private names, places, program labels, and preference phrases that cannot be inferred from vault titles to gitignored profile/private_terms.txt, one exact term per line.
• Or, if the exposure is deliberate (e.g., the title is fully public and appears as an illustrative example), add the stem to scripts/privacy_allowlist.txt and document the rationale in the commit message.

The check is a blocking quality gate for any commit that touches tracked files when the gate ran meaningfully (no skip flag). Do not proceed to structural lint if Phase 0c returns hits. An absent exact-term sidecar is a coverage warning, not proof of a leak. Keep the semantic privacy round enabled and populate the gitignored sidecar before claiming exact identity or preference coverage.

Phase 1a: Structural lint

Bash: uv run scripts/lint.py --json

Read wiki_dir, counts, and findings using the Phase 0 finding fields.

Phase 1b: Staleness lint

Bash: python3 scripts/staleness.py --json

Read thresholds, counts, and notes (path, staleness, category).

Staleness findings are always advisory (no ERROR level). They surface L2 notes that have gone cold, using the formula days_since_modified / (1 + log(1 + reference_count)). Notes referenced from wiki entries or recent reflections decay slower.

Phase 2: Present

Group findings by severity. If the corpus is clean, say so and stop.

For each ERROR-level finding: show the code, file path, and message verbatim. Do not rephrase the message — scripts/trust.py and scripts/lint.py emit precise line numbers that the user needs.

For WARN-level findings: show them but mark them as non-blocking.

Summarize INFO findings (e.g., "4 entries lack outbound citations") unless the user asks for the full list.

Staleness section (from Phase 1b): present after the structural findings, under a separate heading. Group by category:

• stale notes: list paths, suggest archiving to <paths.archive>/
• dormant notes: list paths, suggest review or compaction
• promote candidates: list paths; nightly Autoevo grows wiki entries from them
• If all notes are active, say so in one line and move on.

Phase 3: Offer fixes

For each fixable category, ask the user before acting:

Finding code	Fix	How
slug-mismatch	Rename file or edit H1	Ask the user which side to change. Never rename without confirmation — downstream @cite targets key off the title.
parse-error (e.g., missing valid_at, duplicate claim IDs, invalid ranges or dangling citations)	Edit the wiki entry	Route to the user; do not auto-edit wiki entries.
duplicate-title	Edit one of the H1 titles	Ask the user which note keeps the title.
orphan-entry	Add inline citation references inside related claims	Suggest supported claim links using protocols/wiki-schema.md; shared-anchor findings can identify candidates.
shared-anchor-no-cite	Add a supported citation reference	Show the shared anchor and propose the citing claim and target. A shared source alone does not establish dependence.

Phase 4: Rerun (if fixes applied)

If any user-driven fixes were made, suggest rerunning /lint manually to confirm the state.

Rules

1. Never hand-check structure. Always call scripts/lint.py. If the script is missing or errors, fix the script — do not re-derive the checks in the LLM.
2. Ask before fixing. All fixes route back to the user; scripts/lint.py is advisory-only.
3. ERROR findings block further wiki work. If a wiki entry has a parse error, do not cite it from new entries; its claims need work until it is fixed.
4. WARN findings are advisory. The system keeps working; the user decides whether to act.