ansarullahanasz360/code-review-swarm

Code Review Swarm

This skill defines the review protocol used by verify-pr-proof.

Sharding

• Max 10 changed files per reviewer.
• Max 1,200 diff lines per reviewer.
• Prefer subsystem shards: app UI, API routes, database/types, tests, agents/jobs, packages.
• Assign high-risk large files alone.
• If subagents are not authorized, apply the same shard model serially.

Reviewer Instructions

Each reviewer is read-only. They must inspect the assigned diff and nearby code, then return:

{
  "verdict": "pass | needs-fix | blocker",
  "findings": [
    {
      "file": "path",
      "line": 10,
      "severity": "critical | high | medium | low",
      "issue": "Concrete failure mode",
      "fixRecommendation": "Specific fix",
      "confidence": 0.8
    }
  ],
  "test_gaps": ["Missing test or scenario"],
  "browser_scenarios": ["Route/flow to verify"],
  "confidence": 0.8
}

Review Rubric

Read ../../references/review-rubric.md. Read ../../references/codex-review.md if codex review, GitHub PR review comments, or automated Codex review output is available.

Prioritize:

• correctness and regressions
• auth/security/privacy
• data integrity and migrations
• external API contracts
• async/retry/idempotency/cancellation
• user-facing browser behavior
• missing tests that affect merge safety

Do not report pure style opinions.

Synthesis

The master agent must:

• deduplicate findings
• reject speculative findings with weak evidence
• turn accepted findings into a fix plan
• map browser scenarios into Agent Browser flows
• preserve unresolved blockers in the final report

External Review Sources

• codex review may be used as one advisory reviewer if available.
• Existing GitHub review comments must be inspected before duplicating review work.
• Normalize all review sources into the reviewer output schema.
• The master agent must verify external findings locally before accepting them.