audit

docaudit:audit — change-driven documentation audit

Report-only orchestrator. Reads the per-project adapter ${CLAUDE_PROJECT_DIR}/.claude/doc-audit.json. If that file is absent, tell the user to run /docaudit:init (Plan 3) or that this repo has no adapter yet — do NOT invent project facts.

SD="${CLAUDE_SKILL_DIR}" ; CFG="${CLAUDE_PROJECT_DIR}/.claude/doc-audit.json". Also bind ANCHOR_PATH="$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1])).get("anchorPath",""))' "$CFG")" for the run lifecycle and Phase 5. --full forces whole-corpus mode (ignores the anchor diff scope).

Open the run before every Phase-0 probe — this is the audit's first action after confirming the config exists. Bind RUN_BASE="$CLAUDE_PROJECT_DIR/.claude/state/docaudit-run". If --break-lock was supplied, run only python3 "$SD/scripts/open-run.py" --run-base "$RUN_BASE" --repo-root "$CLAUDE_PROJECT_DIR" --anchor-path "$ANCHOR_PATH" --break-lock, report its JSON (including holder), and exit; this path does not acquire a new lock or run any phase. It is an emergency operation that intentionally breaks report serialization; while the gate holds its flock, including the complete gate+report interval, it MUST be refused with reason:"gate-running" and must never be bypassed. Otherwise run Before acquiring a run lock, run AUDIT_SCOPE_CHECK="$(python3 "$SD/scripts/import-audit-scope.py" --repo-root "$CLAUDE_PROJECT_DIR" --config "$CFG" --check --json)". Bind PRECHECK_CONFIG_SHA, AUDIT_SCOPE_PATH, and AUDIT_SCOPE_STATE from that same JSON's configSha, scopePath, and state fields; do not read the config again between this check and open. If AUDIT_SCOPE_STATE is drift or path-mismatch, or errors[] is non-empty, stop without calling the lock-acquiring open-run.py; for path-mismatch, show both scopePath and recordedScopePath, otherwise show diff.missing / diff.extra (or errors[]), and tell the user to run /docaudit:init --import-audit-scope to restore the generated map. If it is not-imported, show only 💡 audit-scope.json は未導入です。 and continue. absent and in-sync are silent. python3 "$SD/scripts/open-run.py" --run-base "$RUN_BASE" --repo-root "$CLAUDE_PROJECT_DIR" --anchor-path "$ANCHOR_PATH" --expect-config-sha "$PRECHECK_CONFIG_SHA" --skill-version 0.22.0 [--accept-config], adding --accept-config only when the skill received it. Assign the complete stdout JSON, unchanged, to EVIDENCE; bind RUNID and RUN_DIR from its runid and runDir fields. Do not create RUN_DIR yourself. If stdout includes previousReportStatus with pending, failed, or written-durability-unknown, report that unresolved or uncertain prior report state to the user before Phase 0; do not silently discard it. Continue the newly locked run normally. Exit 4 with {locked:true,holder} means a prior run owns the lock: show the holder and stop with “先行 run が lock を保持しています。死んでいるなら /docaudit:audit --break-lock を実行してください。” Exit 6 means an earlier run detected an unapproved config change: stop, ask the user to inspect git diff .claude/doc-audit.json, and re-run with --accept-config only after approving that difference. Neither exit path owns a lock. Exit 2 with stderr containing unrecognized arguments: --skill-version or does not match plugin engine version means the plugin and this skill have different versions: start a new session and rerun /docaudit:audit. This exit path also owns no lock.

After every successful open and at the beginning of every later turn or phase before its first config consumer, re-derive CONFIG_SHA="$(python3 -c 'import json,sys; print(json.loads(sys.argv[1])["config"])' "$EVIDENCE")". CONFIG_SHA is never restored from a checkpoint as an independent value.

EVIDENCE is the sole transport for evidence hashes. open-run.py seeds it; every later evidence producer (write-evidence.py, plan-dispatch.py, start-run.py, and seal-run.py) receives --evidence "$EVIDENCE" and its complete stdout JSON MUST replace EVIDENCE verbatim. Never parse hashes into separate carried variables, merge JSON by hand, add sentinel fields by hand, or reconstruct an earlier value. Operational values such as MODE, tool availability, and status lines may be bound normally, but the only carried evidence state is RUNID plus the one EVIDENCE string. The gate receives it only as --expect-json "$EVIDENCE". probe-record.py also receives --evidence "$EVIDENCE" for run-dir validation only; it is not an evidence producer and its stdout MUST NOT replace EVIDENCE.