review-loop（审查循环）

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 /review-loop, determine:

• Target: plan or work? Check conversation context, recent activity, user's words.
• Reviewers: user-specified, or your judgment. Available: code-reviewer, architect-reviewer, security-auditor, technical-editor, accessibility-tester, design-reviewer, codex-specialist. 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 /plan-review 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 with Co-Authored-By trailer per AGENTS.md.
• 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}) section to the plan file.
• Work: gh issue create --label review-finding --title "R{N}: <summary>". Cap 10/round, 20 total across all rounds. If gh unavailable, append to .branch-context.md.

c–f. Panel governance — invoke /assemble-panel for RETAIN, EXPAND, CONVERGE, and ESCALATE_RECURRING. The policy algebra and reviewer maps live there — do not duplicate them here.

Codex/Cursor: read .claude/skills/assemble-panel/SKILL.md directly and apply the policy algebra inline.

If /assemble-panel is unavailable, fall back to panel [technical-editor, code-reviewer] with gate=P2, cap=3. Keep all reviewers with above-gate findings, exit at round 3 or when all clean.

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

• /plan-review unavailable → dispatch technical-editor + architect-reviewer directly.
• gh not authenticated → degrade to .branch-context.md or printed list.
• Reviewer returns DROP → exit immediately, print summary.
• No reviewers available → use code-reviewer as minimum viable panel.