| name | reviewer |
| description | Multi-lens review orchestrator for code, specs, architecture, config, docs, and infrastructure. Hybrid execution: direct technical review using sub-reviewer references, or delegate to independent agents/workers for multi-perspective, audit, or specialized reviews. Sub-lenses: code-quality, design-rigor, adversarial, security, edge-case, editorial. Activate on "review", "check", "audit", "challenge", "QA", or any explicit review request. |
Reviewer
Hybrid review: direct technical lenses or delegated multi-perspective / audit. Lazy-load sub-reviewers — never load all at once.
Gate 1 — Author bias
If you created or edited the artifact → MUST delegate all reviews. Self-review with sub-reviewer files is not independent judgment.
Trivial exception: <10 lines, no structural impact → skip review rather than fake self-review.
Gate 2 — Self-grounded verification
Any in-context review must follow rules/self-grounded-verification.md: criteria + disconfirming test before judging the artifact; then PASS/FAIL/UNVERIFIED with evidence.
Direct vs delegate (summary)
| Situation | Approach |
|---|
| You are author | Delegate always |
| Audit / independent judgment | Delegate (strongest independent reviewer available) |
| Multi-stakeholder personas | Delegate separately per persona |
| Pure technical, you are not author | Direct + sub-reviewer refs |
| No independent delegated reviewers + you are author | Do not self-review; ask human / external tools |
Full decision tree, delegated-reviewer detection, synthesis: references/delegation/framework.md.
Persona prompts: references/delegation/personas.md.
Expanded direct/delegation steps: references/direct-and-delegation-detail.md.
Fallback + review modes + quick start: references/fallback-modes-quickstart.md.
Worked scenarios: references/examples/delegation-scenarios.md (load only if needed).
Direct mode — routing (load only matching lenses)
| Artifact | Load (in order) |
|---|
| Code / PR | code-quality → design-rigor → adversarial → security → edge-case-hunter |
| Specs / PRD / TRD | adversarial → editorial |
| Architecture | code-quality → design-rigor → adversarial → security |
| Docs / prose | editorial |
| Config / infra | security → edge-case-hunter |
| Bug fix / incident | design-rigor → code-quality → adversarial → edge-case-hunter |
| API contracts | code-quality → design-rigor → adversarial → security → edge-case-hunter |
| Skills / commands | adversarial → editorial |
Paths: references/sub-reviewers/<name>.md.
- Quick: first lens only. Deep: all listed. Aggregate by severity with file:line.
| Lens | Axis |
|---|
| adversarial | assumptions, decisions |
| code-quality | structure, smells, AI laziness |
| design-rigor | designed vs grown; root cause |
| security | threats, secrets, auth |
| edge-case-hunter | boundaries, null paths |
| editorial | prose clarity |
Delegation mode (short)
- Pick audit / multi-persona / specialized pattern (
framework.md).
- Isolate personas (
personas.md) — one agent per hat.
- Synthesize: CRITICAL → HIGH → MEDIUM → LOW; call out conflicts and trade-offs.
Modes of use
- Explicit — user asked for review.
- Self-review before present — delegate (author bias).
- Proactive — flag structural/security/assumption issues mid-conversation when not in hot-fix/throwaway mode.
