| name | review |
| tier | capability |
| runtime | on-demand |
| trigger | mention |
| description | On-demand audit: security / coverage / tech-debt / full Gate / PR code review (dispatch specialist lenses + optional cross-model review). For Phase/Gate milestones, after a security incident, pre-merge PR review. Trigger: audit / security scan / full check / Gate check / review --security / review --gate / tech-debt inventory / security audit / review PR / code review / 审计 / 安全扫描 / 全量检查 / Gate检查 / 技术债盘点 / 代码审查 / PR审查. Skip: routine verification (/verify) / commit (/commit). |
| metadata | {"source":"oh-my-dag","version":"5.0.0","methodology":"on-demand audit + specialist lenses + optional cross-model review"} |
| disable-model-invocation | true |
/review — On-Demand Audit
On-demand full scan, for Phase completion / pre-Gate / post-security-incident. Routine changes go /verify → /commit.
Review model: review is read-only diagnosis (specialist lenses) + an optional cross-model adversarial pass + the agent's own judgment (the verdict — taste is not hostage to findings). The real check is your build/typecheck/test commands.
Modes
--security → scan trust boundaries (sig-verify/auth/injection/fail-open) across request/event/integration entrypoints
--coverage → test coverage + missing-test detection
--debt → typecheck + unfixed known issues + TODO/FIXME scan
--zod → input-validation coverage audit of untrusted entrypoints (route/callback/inbound)
--pr → PR-level review (specialist lenses in parallel + plan completion + scope drift)
--gate → the above in parallel, pre-Gate aggregate check
--release-gate → final delivery review (aggregate scan)
--with-codex → modifier (combine with --gate/--release-gate/--pr): adds a cross-model review pass in parallel in the same message
--with-codex — Cross-Model Phase Gate
--with-codex is a modifier flag. On top of the specialist lenses, it adds a cross-model review in parallel in the same message — independence from a second model family that the primary model cannot self-check.
| Gate | Max rounds (hard cap) | BLOCK fallback |
|---|
--gate (phase end) | 1 per phase end (fix folds into the next phase, no second round) | agent judgment |
--release-gate | 1 + max 1 fix = 2 | the developer |
--pr | 1 | agent judgment |
anti-slop: a cross-model finding ≠ ground truth. Must reject niche/theoretical/telemetry/tests-for-tests/style/doc-completeness; must accept P0/P1 reproducible bug (file:line+steps) + contract violation + ship-blocker + mechanism-level boundary failure.
Finding → fix routing: each finding is routed to the matching specialist lens (read-only diagnosis), then the fix is written on the main line. Type errors / typos (≤3 lines) are fixed directly without dispatching a lens.
Trigger
/review --<mode> or /review --gate
Security Mode (--security)
Dispatch the relevant specialist lenses (read-only), or scan directly:
- Scan scope: request handlers + websocket/event entrypoints + external integrations + SQL/migrations
- Checks: trust-boundary sig-verify (HMAC/timing-safe compare) / auth gap / injection / fail-open (
catch { return null } swallowing errors) / GET read-only / no existence leakage / input validation / row-level deny-all
- Output: P0/P1/P2 + verdict (PASS/BLOCK)
Coverage Mode (--coverage)
- Run tests with coverage
- Compare against
git diff changed files: added/changed files with no corresponding test → warning
- Identify gaps: 0-test danger zones
Debt Mode (--debt)
- Typecheck — type debt
- Read your known-issues log — unfixed P0/P1
- Grep TODO/FIXME/HACK + the next-state file's secondary tracks (known backlog)
Zod Mode (--zod)
Scan untrusted entrypoints (request bodies / channel callbacks / inbound endpoints):
- PROTECTED: has schema validation | UNPROTECTED: accepts external input but no validation
- Output: coverage table + unprotected entrypoints + suggested schemas
Gate Mode (--gate)
Phase completion / pre-Gate. In parallel:
| Lens/Check | maps to |
|---|
| specialist lens (routed by diff to data-layer / pipeline / interface / agent-protocol) | methodology review |
| typecheck + tests | types + tests |
| cross-model review (if --with-codex) | cross-model |
Synthesis: All green → GATE PASS | Any P1 → NEEDS-CHANGE | Any P0 → GATE BLOCK
Specialist prompt injection
Read the review checklist / required-reads kernel, then review the scope.
Your lens: {data-layer | pipeline | interface | agent-protocol | stack}
Review scope: git diff $(git merge-base HEAD main)..HEAD
Binary PASS/FAIL + stack-specific checks.
PR Review Mode (--pr)
PR-level code review.
Step 1: Branch Diff
git diff main...HEAD --stat && git log main..HEAD --oneline
No diff → "Nothing to review", terminate.
Step 2: Plan Completion Audit
Read the next-state file's active plan + commitments → compare against diff: DONE / PARTIAL / NOT DONE / EXTRA (scope creep).
Step 3: Critical Pass
Scan diff for high-risk patterns: SQL injection (IN instead of = ANY), races (mutation without compare-and-swap), trust-boundary writes bypassing validation, auth gap, enum completeness (switch without default). Each finding carries confidence 1-10 (1-3 suppress / 4-6 mark low / 7-10 show).
Step 4: Specialist Dispatch (parallel, read-only)
Routed by diff (parallel, same message):
| lens | condition | checks |
|---|
| data-layer | touches SQL / schema | schema safety + row-level security + indexes + immutability |
| pipeline | touches retrieval / memory | data flow + validated writes + naming |
| interface | touches routes / websocket | idempotency + sig-verify + error shape + type contract |
| agent-protocol | touches agent/skill/hook / DAG | agent/skill/hook + node contract |
| stack | touches dispatcher / entrypoint | loop blocking + compare-and-swap + lifecycle |
Each lens sees only the diff + context, outputs findings (criterion + verdict + confidence + evidence + fix).
Step 5: Synthesis
Dedup (multiple lenses reporting the same → merge and raise confidence) → grade P0/P1/P2 → 2+ lenses reporting the same P1 → escalate to P0.
Output
## PR Review — {branch} → main
### Plan Completion / ### Scope Check (CLEAN/DRIFT/MISSING)
### Findings ({N}: {P0} P0, {P1} P1, {P2} P2)
| # | Severity | File:Line | Finding | Confidence | Source |
### Verdict: APPROVE / NEEDS-CHANGE / BLOCK
Release Gate Mode (--release-gate)
Final delivery review. /handoff recommends it on significant changes.
- Read the release-gate checklist
- Dispatch the aggregate-review lens (read-only), inject the template
- Scan this session's changes, output the delivery checklist + VERDICT → a dated report under
docs/reports/
/handoff recommendation triggers
| Condition | Recommend |
|---|
| This session changed ≥5 files + cross-layer | /review --release-gate |
| This session has a new migration | /review --release-gate |
| This session changed schema / safeguards / orchestration | /review --release-gate |
| None of the above | Silently skip |
Fix-First Protocol (common to all modes)
- AUTO-FIX: mechanical issues (naming/formatting/import/dead code/typo) → fix directly
- ASK: needs judgment (architecture/performance tradeoff/security decision) → AskUserQuestion
Execution: classify → batch AUTO-FIX (one line each [AUTO-FIXED] file:line Problem → done) → remaining ASK ≤3 asked individually / >3 batched.
Scope Drift Detection (automatic for --gate / --release-gate)
Read the next-state file's active plan (stated) vs git diff --stat (delivered) → SCOPE CREEP (out-of-scope files) / MISSING (promised but absent) → Scope Check: [CLEAN / DRIFT / MISSING] (INFORMATIONAL, non-blocking).
Doc Staleness Check (automatic at the end)
diff's impact on docs/: code changed but docs not updated → flag (INFORMATIONAL).
Constraints
- Code may be changed under Fix-First — AUTO-FIX fixes directly, ASK needs confirmation
- Each mode usable independently; --gate runs all in parallel; --release-gate is delivery review
- Frequency: Phase completion / pre-Gate / post-security-incident (not routine)
- Output a dated
audit-report-{date}.md or release-gate-{date}.md under docs/reports/
