Review Loop Skill 安全审计

185186## Files to review187{paste file contents or list file paths with relevant excerpts}188```

135136### Step 2: Spawn Reviewer Subagent137138Use the Agent tool to dispatch a reviewer. The reviewer must:139- Be a **separate subagent** (fresh context, no anchoring to your reasoning)140- Receive only the **work product** (files, diffs) — not your thought process141- Score 1-10 with **specific, actionable feedback**142- Use a balanced/standard model by default (or an advanced reasoning model for complex/specialized tasks)143

266You:2671. Read the auth module2682. Spawn reviewer → Score 5, feedback: SQL injection, no rate limiting2693. Fix: parameterize queries, add rate limiter2704. Spawn reviewer → Score 8, approved2715. Done

210211### Step 5: Revise212213Address the reviewer's feedback. Fix Critical and Important issues. Minor issues are optional. Then go back to Step 2.214215**Important:** Each revision should be targeted. Don't rewrite everything — fix what the reviewer flagged. Maintain a mental list of ALL prior feedback to avoid regressions.216

265266You:2671. Read the auth module2682. Spawn reviewer → Score 5, feedback: SQL injection, no rate limiting2693. Fix: parameterize queries, add rate limiter2704. Spawn reviewer → Score 8, approved2715. Done272```

202203### Step 4: Check Stop Conditions204205In this order:2061. If loops completed < min_loops → **continue** (always)2072. If score >= quality_gate → **stop, go to final polish**2083. If loops completed >= max_loops → **stop, go to final polish**2094. Otherwise → **revise and loop**210

218219Once the loop exits (quality gate met or max loops hit):220- Address any remaining minor issues if trivial221- Verify the final output is coherent (no artifacts from revision cycles)222- Report final score and loop count to user223

262. The agent does the work (or reads existing work)273. A separate critic subagent scores it 1-10 with specific feedback284. The agent revises and repeats until score >= 8 (default quality gate)295. A loop summary is delivered with the final output30

54|---------|---------|55| Min loops | 2 |56| Max loops | 4 |57| Quality gate | 8/10 |58| Worker model | (your current model) |59| Reviewer model | (your current model or fast/balanced alternative) |60

8283**Default behavior:** Since you (the main agent) ARE the worker, spawn a reviewer subagent that matches or exceeds your current capability based on the task:84- Most tasks → Standard/Balanced reviewer85- Specialized/hard tasks → Advanced/Reasoning reviewer86- Quick checks → Fast/Lightweight reviewer (only if you are also acting as a lightweight worker)87

320321## Cost and Speed322323- Each loop = 1 reviewer subagent call324- Budget roughly 1–2x the time of a single implementation pass for a full 3-loop cycle325- This is cheap compared to shipping buggy code, vague specs, or triggering a late-stage review cycle326- Use your default/standard model for most reviews; only upgrade to advanced reasoning models for specialized domains (security audits, distributed systems, quant finance)327

3233## When to Use3435- User says "use review-loop", "polish this", "iterate on this", "/review-loop"36- Complex implementations where quality matters more than speed37- Design docs, specs, or technical writing38- Code that needs to be robust (security, data pipelines, financial logic)39- When user wants adversarial critique baked into the process

5051## Defaults5253| Setting | Default |54|---------|---------|55| Min loops | 2 |56| Max loops | 4 |57| Quality gate | 8/10 |58| Worker model | (your current model) |59| Reviewer model | (your current model or fast/balanced alternative) |60