| name | ad-audit |
| description | Run this skill when the user explicitly invokes `/ad-audit` or names it ("run ad-audit", "audit against the rules"), or asks for the maximum quality gate before work reaches the team ("adversarial audit", "rules audit", "maximum gate", "verify before I post", "exhaustive review"). Heavier and more exhaustive than `ad-review` (light two-axis diff review); distinct from `ad-drift` (documentation-drift only). Auto-trigger note: `allow_implicit_invocation: true` is set, so audit-language can fire the skill; if a request is ambiguous, confirm scope before invoking.
Mechanical shape: ONE pass in the current session. The skill resolves the project's rule-set (repo binding docs plus optional curated machine and project rule layers), walks every rule-group as a checklist with structural separation, emits an explicit verdict for every rule (pass / violation / judgement-call / n-a — none skipped, so coverage is a matrix), grounds every finding on the actual code/output, and gates every teammate-visible claim on a real evidence artifact. Never emits an "approve" verdict. Rule gaps route to `/ad-level-up`.
|
| summary | Maximum-gate rules-anchored audit. Walks every rule-group as a checklist with structural separation, exhaustive per-rule verdicts (coverage matrix), evidence-gated, never approves. Codex reviews inline; ships an audit-group-reviewer subagent for user-initiated isolated + cross-model escalation. Heavier than ad-review. |
Codex skills run inline in the current session. Codex supports subagents, but spawning is explicit user-directed orchestration, not something this skill does silently. So the default audit is one pass with disciplined per-group checklist output; the optional escalation (true isolation per group, and the cross-model pass on critical groups) uses the bundled `audit-group-reviewer` subagent against the persisted audit-trail file.
Mechanical shape:
THIS SESSION:
1. Target + tree (what is under audit, and on which tree/SHA?).
2. Resolve the rule-set: repo binding docs (always) + curated store at
$AGENTIC_RULES_DIR or ~/.agentic/rules/ (optional) + project rules at
.agentic/rules/ (optional). The rule-set defines the groups and any
CRITICAL tag — never hardcode them.
3. Enumerate every group. For each: review it, or record explicit N/A + reason.
4. Write the assembled context to .agentic/reviews/<ISO>-audit-<scope>.md (audit trail).
5. Review each dispatched group in this session, as a checklist. Output one
section per group; give EVERY rule an explicit verdict.
6. Coverage matrix + verdict (never "approve") + rule-gap handoff to /ad-level-up.
- Do NOT emit an "approve" verdict. The audit is adversarial; nothing clears until every teammate-visible claim carries a reproducible artifact and every blocker is resolved or refuted with evidence.
- Do NOT hardcode rule content, group names, or a CRITICAL tag. They come from the resolved rule-set only.
- Do NOT skip a rule. Every rule in every enumerated group gets an explicit verdict, or the audit is invalid. A group whose subject is untouched is logged N/A-with-reason, never silently omitted.
- Do NOT clear a rule you could not ground on the actual code/output. Mark it an OPEN QUESTION.
- Do NOT accept a teammate-visible claim backed only by code-reading or inference. No run/observed artifact → BLOCKER.
- Do NOT skip writing the audit-trail file — it is the context packet for the user-spawned escalation.
- Do NOT begin any file I/O before printing the Step 0 announce line.
Step 0 — announce. Print the shape before any work:
- Resolve every blocker with the evidence artifact it named, then re-run `/ad-audit` on the fix.
- Route any rule gap to `/ad-level-up`.
- For a lighter routine diff review, use `/ad-review`. For documentation-vs-code drift only, use `/ad-drift`.
