Audit Docs

A read-first audit answering "do the docs still match reality?" Produces a findings report; it does not silently rewrite docs.

Turn contract — verify before ending the turn

✓ The AUDIT DOCS fixed-format report was printed, ending in `Decision: PASS | FAIL`
✓ No doc was rewritten without explicit --fix / user go-ahead
✓ The closing `→ Next:` block is printed as the ABSOLUTE last output

About to end the turn with any box unchecked? The turn is NOT done — complete the missing box first (weak models drop end-of-document duties; this list is first on purpose).

When to use

• Before a release or milestone, after merging several features/fixes, or whenever the doc set might have drifted from the code and issues.

Step 0 — Discover the project (always first)

Per the agent guide's Workflow conventions + documentation map, then read what THIS skill needs: the roadmap, the fix index + template, and the feature folder layout — the map tells you which links and invariants to check.

Checks

Run these and collect findings (cite paths/lines/issue numbers each):

1. Roadmap ↔ feature folders. Every docs/features/<NN>-<slug>/ is in the roadmap, and every roadmap entry has a folder (or is explicitly "scheduled").
2. Feature dependencies. SPEC Depends on / Branch fields are valid; no dependency cycles; ordering is consistent with the roadmap.
3. Fix index hygiene. Every entry maps to an open issue and an unmerged branch; flag entries whose issue is closed or whose PR merged (should have been removed). Flag open fix branches missing from the index.
4. Documentation-map links resolve. Every file the map references exists; flag "scheduled, not yet authored" items so they aren't mistaken for drift.
5. Broken intra-doc links. Relative links/anchors point at real files/sections.
6. Issue references. Acceptance/known-issues lines referencing #N aren't pointing at long-closed issues without note.
7. Language & naming conventions. Artifacts in the project's docs language (this repo: English); file/dir naming matches conventions (e.g. kebab-case TS files, PascalCase components).
8. Invariant tags. If the project uses invariant/decision IDs (e.g. INV-/D-/KI-), spot-check that referenced IDs exist where claimed.
9. PR-link integrity on done rows. Every roadmap row (and fix-index entry) with status done carries a linked PR reference — done · [#<pr>](<pr-url>). For each bare done, find its PR in the forge (gh pr list --state all --head feat/<NN>-<slug> or by title) and propose the row update; a done with no PR found at all is HIGH severity (the unit may never have been closed out).

Workflow discipline (checks 10–14) — the executor skills enforce these at write time; this audit verifies they actually held. Each check is mechanical: run the command shown, don't infer.

10. Phase naming — keyed on the carrier, like check 11.
    • lane-era — bun scripts/phase-lint.mjs <unit-doc> (owner skills/phase-contract/SKILL.md) must print verdict PASS, and bun scripts/ceremony.mjs docs check-phase-labels --slug <slug> must answer valid: true. A S# hit in a lane-era doc is not a phase label: since #308 S# is the requirement id.
    • legacy carriers only — grep -rnE '\bS[0-9]+\b|\bStep [0-9]' docs/features/*/{PLAN,TASKS,progress}.md must return nothing: plans use P1, P2, … ("phases") only. Any hit: LOW (rename), plus check the executor argument still resolves. Today's 22 hits are all literal Step 0 — Discover the project citations; the grep is the legacy carrier's, so a lane-era S# is never read through it.
11. Per-phase doc discipline — keyed on the unit's own carrier. Pick the carrier mechanically, never by guess: TASKS.md in the unit folder means legacy carriage; its absence means the unit doc's ## Tasks/## Log.
    • lane-era — bun scripts/ceremony.mjs docs check-tasks-ticks --slug <slug> and bun scripts/ceremony.mjs docs check-progress-entry --slug <slug> must both report valid: true: every - [x] P<n> has its P<n> line in ## Log, and no task lacks one. Cite those two fields, never check-all (its commit-format/docs-language fields are checks 14 and 7). A doc with no ## Tasks section answers n/a — no ## Tasks section — legacy carrier — so the vacuous case routes to the legacy branch, never to a silent pass.
    • legacy — the original assertions on that unit's own files: completed phases ticked in TASKS.md, one progress.md entry per completed phase, the literal close-out tasks in the final phase.
    • Severity — a done unit whose ledger disagrees with its tasks: HIGH. The same on an in-progress unit: LOW (it may still be settling). Only high/low exist in this report's legend.
    • Close-out is audit-pr's contract (receipt + ordered gates): cite its verdict, never re-derive the retired literal tasks here.
12. Branch & PR discipline vs the forge. For every done unit: its PR exists, targets the default branch, has a non-empty body, and carries Closes #<n> when the unit is issue-born (SPEC references an issue). Also scan recent default-branch history (git log --first-parent) for feature/fix-scoped changes committed directly without a PR: HIGH.
13. Generated-docs provenance (only when the documentation map declares a Docs site block; otherwise state n/a). Scan the declared content dir for orphan pages each: (a) its source-unit exists in the roadmap or fix index — no match is an orphan (low: propose deletion or re-attribution); (b) the unit's PR merged after the page's updated date with commits touching the page's subject paths — that page is stale (LOW: propose ). Cite page path + unit per finding.
14. Commit format & dependency discipline. Sample the unit branches' commits: <type>(<scope>): <summary> conventional format (violations: LOW). Every in-progress/done row's Depends on: closure was merged — a unit built on unmerged deps is HIGH unless decisions.md records a user-forced override (--force), which downgrades it to LOW (documented risk).
15. README currency against what shipped. Run each README against the tree: rg -n '<a retired name or stale claim>' README.md packages/*/README.md and every hit that names a surface the tree no longer carries (a retired skill in an install/uninstall example, a command count or package count the tree contradicts, a "none yet" beside shipped subcommands) is HIGH — the claim misleads a reader who acts on it. Its mirror: a shipped, consumer-facing surface with no README entry at all is HIGH (the document-or-don't-coin residue read at README scope — product-audit owns the product-wide sweep, this check is the docs-tree half). Superseded prose left standing beside its replacement is LOW. Owners: the rule is AGENTS.md §Working rules, the step is skills/unit-lane/references/DOCS.md.
16. Project-charter coverage (issue #435). Applies when the agent guide declares the charter — a ## Project charter section or a documentation-map row for the criteria catalogue; an older install that declares neither is stated n/a (absent structure, never a finding). Mechanical, the CAPABILITIES pattern — run the commands, never infer:
    • the declared map row's file exists — a row naming docs/workflow/VERIFICATION_CRITERIA.md with no such file is HIGH (the map points at nothing);
    • the charter section is present when either the section or the row is declared — a missing ## Project charter section is HIGH (identity answers with no tier-0 home push agents back to guessing);

    • no unfilled charter slot survives — a bullet whose entire value is a <…> placeholder is the raw-template failure. The scan is bounded to the ## Project charter section, because the guide's other honest placeholders — the Hard-rules stubs init-workspace leaves unfilled by design — are not the charter's finding, and a guide-wide scan reports them forever:

      awk '/^## Project charter/{f=1; next} f && /^## /{f=0} f && /^- \*\*[^*]+:\*\* `?<[^>]+>`?$/ {print FILENAME":"FNR":"$0}' AGENTS.md   → prints nothing

      (the template's own file is the positive case — its 11 slots print with path:line; a legal unresolved — <question> value never matches; an identity bullet — the five the template names — is HIGH, a domain bullet LOW);

    • the catalogue carries no raw placeholder cell — same scan as the capability seed: rg -n '\| *\?<[^>]+>`? *|' docs/workflow/VERIFICATION_CRITERIA.md must print nothing, and every criterion row's provenance cell is one of the closed set (workflow-generic/project-declared/stamped assumption (dated)`) — a row naming none is LOW (it cannot be consumed as declared);

Adapt the list to what the project has; skip checks for absent structures and say so.

Process

1. Discover, then run the checks with grep, file reads, and the forge CLI (per Workflow conventions; examples use gh).

2. Produce the findings report — return exactly (fixed output contract):

   AUDIT DOCS — scope: <docs tree / roadmap / fix index / issues checked>
   
   | # | Check (1-16) | Finding | Sev | Evidence | Proposed fix |
   |---|-------------|---------|-----|----------|--------------|
   | 1 | <which>     | <what>  | high|low | <path:line / #issue> | <smallest action> |
   
   Checks run: <n>/16 (skipped: <which + why — absent structures only>)
   Summary: <1-2 sentences>
   Decision: PASS | FAIL   (FAIL if any high-severity finding is open)

   Sev: high = misleading or broken; low = cosmetic. Only high and low are defined in this legend

3. Fix only on request. With explicit --fix (or user go-ahead), apply the low-risk corrections (remove a merged fix-index row, fix a dead link, register a missing roadmap entry, add a verified PR link to a bare done row). Leave judgment calls to the user.

Guardrails

• Read-first; never bulk-rewrite docs unprompted.
• Distinguish genuine drift from intentionally "scheduled/deferred" items — don't report deliberate tradeoffs as errors.
• Keep changes surgical and within docs; no code or behavior changes.

Portability (agents other than Claude Code)

The workflow is the contract; Claude Code features are conveniences. On an agent that lacks one, apply the fallback — never skip the step the feature enables:

• No slash-command menu — where this skill says /<skill>, open that skill's SKILL.md (wherever your agent installed the skills) and follow it literally, in a fresh conversation: hand-offs assume a clean context.
• No per-skill model:/effort: — no skill pins a tier; pick tiers yourself: these are mostly mechanical cross-document checks — a mid-tier model is enough; escalate to your strongest only for a deep audit.

Relationship to other skills

• Complements unit-lane (which creates the docs this audits) and triage-issue (which keeps the fix index honest).
• Run standalone anytime; no required predecessor.

Done when

• A severity-ranked findings report exists, and any approved low-risk fixes are applied — with genuine deferrals left untouched and labeled as such.

• The closing → Next: block is printed:

  → Next: apply the approved low-risk fixes, then re-run /audit-docs to confirm clean
    · real drift (not cosmetic) → /triage-issue   · a concrete defect → /unit-lane --fix <n>
    · already clean → nothing to do