Doc Sync

doc-sync — keep docs and code converged

This repo treats doc lag as an unfinished bug (see AGENTS.md § Doc-sync discipline). This skill is the executable procedure; the canonical mapping table (which change owes which doc) lives in AGENTS.md — read it there, do not copy it here.

This file is canonical at .agents/skills/doc-sync/SKILL.md (.claude/skills/doc-sync is a symlink). The skill itself iterates: when a run shows an instruction here is wrong or misleading, fix it in the same pass, with the run's evidence in the commit message.

Mode 1 — per-change sync (run at the end of every change)

1. Determine the change surface. git diff --stat origin/main...HEAD (or the staged diff for uncommitted work).

2. Map each changed path against the AGENTS.md doc-sync table. Produce the list of docs owed by this change. The high-traffic rows:

   • src/db/schema.ts → docs/generated/db-schema.md (hand-maintained snapshot — no generator script exists despite the generated/ dir name; update the table row and any enum lists by hand; prod DB migrates via prod:up → db:push:prod).
   • Routes / CLI subcommands / daemon protocol → ARCHITECTURE.md §II codemap + §IV contracts.
   • Module purpose / boundary / invariant changed → ARCHITECTURE.md §II–IV.
   • Feature completed or modified → FEATURES.md checkbox (checkbox + a short note; long verification narratives belong in the PR, not the checklist).
   • TODO / known drift left behind → docs/tech-debt-tracker.md new entry (next free ID — check the archive too so IDs are never reused).
   • src/daemon/** shipped in the bundle → bump packages/daemon/package.json, cut a GitHub Release, and add the version's CHANGELOG.md entry. Merged ≠ shipped.

3. Status single-source rule. Security/authorization item states (F*/C*/IDOR-B*) live only in docs/authorization.md §6. Other docs (PLANS, tech-debt I44) may point there but must never mirror per-item states — mirrored lists are how C10 stayed "remaining" for weeks after it was fixed. Detect mechanically, don't trust prose: grep -rnE '\b(C1[0-2]|C[1-9]|B[1-6]|F[0-9]|IDOR)\b' --include='*.md' docs/ *.md | grep -vE 'authorization\.md|docs/exec-plans/|docs/superpowers/' (historical plan/spec records legitimately narrate the states they shipped — the rule polices live docs). A hit naming an item's state (open/fixed/remaining) is a violation; a bare pointer to authorization.md §6 is fine.

4. Staleness check on every doc you touched. For each edited doc: file paths it names exist; tech-debt I<n> / D<n> references resolve (tracker or archive); counts/enums match code (grep, wc -l — better: drop volatile numbers entirely, per ARCHITECTURE.md's header rule).

5. Hygiene gates.

   • No personal/machine refs in what this change adds (core belief #9). Grep the added diff lines, not whole files: git diff origin/main -- <files> | grep '^+' | grep -E '~/\.claude|/Users/'. A personal absolute path (/Users/<name>/…) is a violation. Two known non-violations that this grep still surfaces: a generic runtime path that is product behavior (e.g. ~/.claude/skills as Claude CLI's global skills dir), and a tech-debt entry quoting a personal path as the evidence of the debt it records. Judge, don't blind-fail — but never add a new personal path outside those two shapes.
   • Touched src/daemon/prompt.ts? Grep for provider-specific tool names (Read, cat, grep, vision hints) → expect zero hits (code-quality red line).

6. Fail loud. In the PR/summary, list: docs updated · docs checked-and-clean · drift found but deferred (with its new tech-debt entry ID).

Mode 2 — periodic full audit (on request / doc-gardening)

Cross-check the status documents against code and each other; they drift fastest:

1. FEATURES.md — sample every [ ] unchecked line: is the feature actually still missing? (grep the endpoint / CLI verb / component). Unchecked-but-shipped is the most common rot.
2. docs/PLANS.md — every referenced plan file exists; Active items are still active; completed plans moved to docs/exec-plans/completed/ with their status line updated (a moved plan still saying "merge pending " is half-finished rot).
3. docs/tech-debt-tracker.md — no duplicate IDs (incl. vs the archive); ⬜/🟡 entries spot-checked against code; newly-✅ entries: move the full row to the archive. The tracker keeps no stub — it holds open items only; cross-references resolve by grepping both files.
4. docs/authorization.md §6 vs any doc that mentions security items — pointers only, no mirrored states (run the Mode 1 step-3 grep).
5. ARCHITECTURE.md §II — new substantive modules (>50 lines) present in the codemap. Recipe: git log --since=<last audit> --name-only --diff-filter=A --pretty=format: -- 'src/*' 'web/src/*', then check each surviving non-test file's basename appears in the codemap. Named files exist; CHANGELOG.md's newest version heading (the [Unreleased] section doesn't count) matches packages/daemon/package.json.
6. Size / density audit — docs must stay maps, not manuals. Line counts hide bloat (long table rows); rank by bytes: git ls-files '*.md' | xargs wc -c | sort -rn | head. For each fat doc, check it against its own stated form rule (FEATURES' "checkbox + short note" header, ARCHITECTURE's "write only what doesn't change often" header) — self-rule violations are the strongest trim mandate. Cut change-history narrative ("the former X was removed…", "before this fix…") and verification evidence — git log and PRs own those. Frozen plan/spec archives of shipped work are deletable (git history retains them); before deleting, grep for inbound references incl. from src/.
7. README.zh-CN.md parity — diff section structure + bullet counts against README.md; the zh mirror silently misses EN feature edits (drift is one-directional). Small deltas: translate in the same pass. A backlog: open a tech-debt entry.

Report findings with file:line evidence; fix mechanically-safe drift in the same pass, open tech-debt entries for anything needing a decision.