/review-docs

Quality gate for documentation. Catches stale stats, dead links, and drift between code and docs.

When to use: Verify doc accuracy after /update-docs, or as a standalone doc audit. Use /review-full instead if: your diff is mixed (code + docs) — it'll detect scope and dispatch. Called by: /ship Phase D (after /update-docs), /review-full (docs-only diffs).

Why opus?

Docs are a trust surface. Sonnet misses cross-file inconsistencies (e.g., README says "30 modules" but modules/ has 32 files). Opus catches these.

When to run

Phase D pipeline:
  /update-docs  → stage doc changes
  /review-docs  → verify them    ← you are here
  (if fail)     → fix in-place, re-stage, re-review (max 2 rounds)
  (if pass)     → continue to Phase S (ship)

Always run after /update-docs — even if it reports "came back clean." Verify the claim.

Bundling behavior

When called from /ship:

• Read-only review of staged doc changes
• If issues found: fix in-place, re-stage, re-review
• Do NOT commit — /ship commits per phase

Step 1: Gather doc changes

git diff --cached --stat -- '*.md'
git diff --cached -- '*.md'

If no staged doc changes AND /update-docs reported "clean" → report "Phase D clean — no doc changes to review" and pass.

Step 2: Review focus areas

Step 3: Automated checks

Run any doc-drift tools your project has:

# Example — replace with your project's tools
bun run tools/doc-drift.ts        # dead file refs + stale dates
bun run tools/stats-sync.ts       # stat accuracy (check-only mode)

Any non-zero output = automatic fail.

If you don't have tools, manually spot-check (assumes repo-root-relative links; relative paths like ../foo.md will need adjustment):

# Find dead doc links (root-relative only)
grep -roE '\[.*\]\([^)]+\.md\)' --include='*.md' . | \
  awk -F'[()]' '{print $2}' | sort -u | \
  while read f; do [ -f "$f" ] || echo "DEAD: $f"; done

Step 4: Scoring

Binary pass/fail. No numeric score.

Why binary? Either docs are accurate or they're not. "Mostly accurate" docs are a trap — users trust them, then hit the inaccurate part.

Step 5: Report

## Review Docs — Report

Files reviewed: X docs changed
Tiers covered: 1 (core) / 2 (arch) / 3 (journey) / 4 (config)

| Check | Status |
|-------|--------|
| Stat accuracy | pass / fail |
| File path validity | pass / fail |
| Code-doc sync | pass / fail |
| Frontmatter | pass / fail |
| Completeness | pass / fail |

Issues: (if any, with file path + description)

Result: PASS / FAIL
Ready for Phase S: Yes / No

Guidelines

• Binary pass/fail — no numeric score
• Scope to doc diff only — don't review impl code (that's Phase R's job)
• Fix in-place — if issues found, fix immediately, re-stage, re-review
• Reviewer is read-only — agent reports, never edits in review pass
• Max 2 fix rounds — if still failing after 2 rounds, flag to user

Exit conditions

• PASS → return "Ready for Phase S", OR
• FAIL after 2 fix rounds → return remaining issues, stop pipeline