Multilingual Triggers

Output language: Respond automatically in the user's conversation language.

C31-multi-review — 4-Agent Adversarial Review

Trigger

• User says: "review", "审查", "multi-review", "代码审查", "检查这个"
• C31-lfg Gate 5 auto-call (strict mode)
• PR submit pre-check (C31-git-worktree integration)

Input

• Target: file path, diff, or document to review
• Optional: --strict (Gate 5 mode, fail on any blocker)

Workflow

Phase 1: Scope & Target Read

1. Read the target file(s) or diff.
2. If diff not provided and in git repo: git diff HEAD.
3. Produce target digest (≤500 lines of relevant code + file list).

Phase 2: Spawn 4 Review Agents (Parallel)

Use sessions_spawn to dispatch 4 subagents simultaneously. Each receives:

• Full target digest
• Its persona prompt (below)
• Strict rule: find ≥1 real issue or exit (no hard-fabricated issues)
• Output schema (JSON)

{
  "agent": "correctness",
  "issues": [
    {
      "id": "C-01",
      "severity": "blocker|warning|nit",
      "line": 42,
      "file": "src/foo.ts",
      "description": "...",
      "suggestion": "..."
    }
  ]
}

Agents:

Each agent is isolated — no shared context, no knowledge of other agents' findings.

Phase 3: Conflict Detection

After all 4 agents return, spawn Conflict Agent (1 subagent):

Task: Compare the 4 issue lists and flag contradictions:

• Correctness says "add mutex" vs Simplicity says "remove mutex, unnecessary"
• Security says "sanitize all inputs" vs Maintainability says "too many guards, refactor instead"
• Severity disagreements on same line (blocker vs nit)

Conflict output schema:

{
  "conflicts": [
    {
      "id": "X-01",
      "agents": ["correctness", "simplicity"],
      "topic": "mutex necessity at line 88",
      "positions": {
        "correctness": "need lock for thread safety",
        "simplicity": "single-threaded context, no lock needed"
      },
      "resolution": "needs_human_decision"
    }
  ]
}

Phase 4: Synthesize Report

Merge 4 agent outputs + conflict list into unified markdown:

# C31 Multi-Review Report
**4 agents reviewed, X issues found, Y conflicts need human decision**

## Blockers (must fix)
| # | Agent | File:Line | Description | Suggestion |
|---|-------|-----------|-------------|------------|
| 1 | [security] | auth.ts:23 | SQL injection via raw query | Use parameterized queries |

## Warnings (should fix)
...

## Nits (optional)
...

## Conflicts Requiring Human Decision
| # | Agents | Topic | Positions |
|---|--------|-------|-----------|
| X-01 | correctness vs simplicity | Mutex at line 88 | "need lock" vs "unneeded" |

## Verdict
- **BLOCKED**: ≥1 blocker found → fix before merge
- **WARNED**: warnings but no blockers → address or acknowledge
- **APPROVED**: only nits or clean → ready

Phase 5: Integration Actions

With C31-git-worktree:

• Run review in isolated worktree
• If verdict = BLOCKED → abort merge to main, report issues
• If verdict = APPROVED → auto-merge worktree, cleanup

With C31-tdd:

• Maintainability Agent findings include test-coverage gap → spawn C31-tdd to fill
• Correctness Agent finds untested edge case → add test before merge

Gate 5 (strict mode):

• Any blocker = FAIL, stop pipeline, require human override
• Any unresolved conflict = FAIL, no auto-resolution

Agent Persona Prompts

Correctness Agent prompt (excerpt):

You are a Correctness Reviewer. Your job is to find logic errors, boundary condition failures, unhandled error paths, and type safety issues. You must find at least 1 real issue or explicitly state "No issues found" and exit. Do not invent issues to satisfy a quota. Return JSON only.

Security Agent prompt (excerpt):

You are a Security Reviewer. Check for injection vulnerabilities, missing permission guards, sensitive data exposure, and known dependency CVEs. Must find ≥1 real issue or exit. JSON only.

Maintainability Agent prompt (excerpt):

You are a Maintainability Reviewer. Evaluate readability, naming conventions, cyclomatic complexity, documentation gaps, and test coverage. Must find ≥1 real issue or exit. JSON only.

Simplicity Agent prompt (excerpt):

You are a Simplicity Reviewer. Look for over-engineering, unnecessary abstractions, premature generalization, and YAGNI violations. Ask: can this be done with 50% less code? Must find ≥1 real issue or exit. JSON only.

Rules

1. No fabricated issues — agents exit clean if no real findings.
2. Line refs mandatory — every issue must cite file:line.
3. Severity calibration — blocker = breaks production, warning = likely bug, nit = style.
4. Conflicts are features — disagreement between agents is data, not failure.
5. Human is tie-breaker — Conflict Agent never auto-resolves, only flags.
6. Under 400 lines — this skill file stays lightweight.