Review Loop
Bounded review-fix-re-review loop. You are the router — parse intent, select reviewers, run the loop.
1. Parse Intent
From natural language after
, determine:
- Target: plan or work? Check conversation context, recent activity, user's words.
- Reviewers: user-specified, or your judgment. Available: , , , , , , . Match to artifact type.
- Round cap: default 3. User can say "thorough" (5) or "quick" (1).
- Severity gate: default P2 (fix P0-P2, record P3+). User can say "strict" (P1) or "lenient" (P3).
If ambiguous, ask: "Reviewing the plan or the work?"
2. Dispatch
- Plan: invoke via Skill tool. Pass round cap. Its internal P0 loop counts as R1.
- Work: dispatch reviewer(s) via Agent tool. Request severity-tagged findings (P0-P3).
Untagged findings: treat as P1.
3. Loop: Fix, Record, Re-review
For each round (R1..R{cap}):
a. Triage findings by severity gate (P0 is highest severity, P3 is lowest):
- Fix (higher severity than gate — lower P-number, i.e. P0 through P{gate}): fix immediately. Commit:
fix: address R{N} review findings
- Record (lower severity than gate — higher P-number, i.e. P{gate+1}+): record, do not fix.
b. Record below-gate findings:
- Plans: append
### Deferred Findings (R{N})
- Work:
gh issue create --label review-finding --title "R{N}: <summary>"
c. Reviewer retention — zero trust in fixes. A reviewer stays on the panel until they individually return no above-gate findings. Only then are they done.
- R1: full panel
- R2+: only reviewers who had above-gate findings in the prior round re-review
- Reviewers with no above-gate findings are done — they already approved implicitly
d. Exit when: all reviewers have individually confirmed clean (converged), OR cap reached, OR DROP verdict.
e. Recurring findings: if same finding appears across 2+ rounds, escalate severity by 1 level. If an escalated finding reaches P0 and remains unfixed at cap, stop and hand off to human.
4. Summary
Review Loop Summary
───────────────────────────────────────
Target: <plan name or file description>
Rounds: <N of cap>
Gate: P<N>
│ Round │ P0 │ P1 │ P2 │ P3+ │ Fixed │ Recorded │
│───────│────│────│────│─────│───────│──────────│
Verdict: APPROVE | ESCALATE
APPROVE = all above-gate findings resolved. ESCALATE = unresolved above-gate findings remain — hand off to human.
Guardrails
- Never modify protected files (see AGENTS.md).
- Hard max: 5 rounds even if user requests more. Note: "thorough" (5) is already the maximum.
- Do not auto-merge or auto-approve. The loop produces a verdict, not an action.
- Plain text output only — cross-surface safe.
Failure Modes
- unavailable → dispatch + directly.
- not authenticated → degrade to or printed list.
- Reviewer returns DROP → exit immediately, print summary.
- No reviewers available → use as minimum viable panel.