Peer Review

Peer Review Skill

Purpose

Enable cross-agent code reviews where Gemini and Codex review each other's work to catch blind spots, with Claude Code (Lead Engineer) verifying all findings before acting on them.

Why Peer Review Matters

Different models have different blind spots:

• Fresh eyes catch issues the original author missed
• Cross-domain review (frontend ↔ backend) reveals integration issues
• Verification step prevents false positives and wasted work

Workflow

Step 1: Identify Review Opportunity

When to trigger peer review:

• After significant implementation (new feature, refactor)
• Before merging complex changes
• When code touches critical systems (auth, payments, data model)
• When user explicitly requests review

Examples:

• Gemini implemented frontend → Codex reviews it
• Codex implemented backend → Gemini reviews it

Step 2: Assign Review

As Claude Code (Lead Engineer), assign the review:

"[Agent name], please review [other agent]'s work in [files/directories].

Use the Code Review Checklist from CLAUDE.md and provide findings with:
- Severity levels (CRITICAL, HIGH, MEDIUM, LOW)
- File:line references
- Specific fix recommendations

Focus on:
[Specific areas to review, e.g., "auth checks", "bandwidth optimization", "error handling"]"

Step 3: Reviewer Conducts Review

The reviewing agent should:

1. Read the code (not trust documentation)

2. Apply Code Review Checklist from CLAUDE.md:

   • Security & Auth
   • Performance & Bandwidth
   • TypeScript & Type Safety
   • Code Quality
   • React/Frontend (if applicable)
   • Production Readiness

3. Provide structured findings:

### ✅ Looks Good
- Authentication properly implemented
- TypeScript types well-defined
- Follows bandwidth optimization patterns

### ⚠️ Issues Found
- **CRITICAL** [convex/mutations.ts:45](convex/mutations.ts:45) - Missing auth check in updateRFP
  - Fix: Add `const identity = await ctx.auth.getUserIdentity()` guard
- **HIGH** [services/api.ts:120](services/api.ts:120) - Using .collect() without limit
  - Fix: Replace with `.take(100)` or use stats aggregation
- **MEDIUM** [components/Card.tsx:30](components/Card.tsx:30) - Expensive filter in render
  - Fix: Move to useMemo

### 📊 Summary
- Files reviewed: 5
- Critical issues: 1
- High priority: 1
- Medium priority: 1
- Low priority: 0

Step 4: Verify Findings (Lead Engineer)

⚠️ CRITICAL: You (Claude Code) must verify each finding.

The reviewer has less context than you. Don't accept findings at face value.

For each finding:

1. Read the relevant code - Does this issue actually exist?
2. Check if it's a real problem - Or is it a misunderstanding?
3. Assess severity - Is the severity level correct?

Mark each finding as:

• ✅ Valid - Issue exists and should be fixed
• ❌ Invalid - Not an issue, explain why
• 🔄 Debatable - Ask original implementer for context

Step 5: Provide Consolidated Feedback

After verification, provide feedback to the original implementer:

## Peer Review Results

[Agent name] reviewed your work. Here are the validated findings:

### Valid Issues to Address

1. **CRITICAL** [file.ts:line] - [Issue]
   - Fix: [Recommendation]
   - Rationale: [Why this matters]

2. **HIGH** [file.ts:line] - [Issue]
   - Fix: [Recommendation]

### Findings Reviewed (Not Issues)

- "Should use Promise.all instead of await loop" - Sequential is correct here due to rate limiting
- "Missing error handling" - Already handled by Convex runtime

### Next Steps

Please address the valid findings and re-submit.

Example Exchange

User: "Codex just added the SAM.gov ingestion feature"

Claude Code: "Gemini, please review Codex's work in convex/ingestion/samGov.ts.

Focus on:
- Error handling for API failures
- Rate limiting implementation
- TypeScript types for API responses
- Bandwidth optimization patterns"

[Gemini reviews and returns 5 findings]

Claude Code: [Reads the actual code and verifies each finding]

Findings verification:
1. "Missing error handling for API timeout" ✅ Valid
2. "Should use Promise.all instead of await loop" ❌ Invalid - Sequential is correct due to rate limiting
3. "Hardcoded API key" ✅ Valid, CRITICAL
4. "Type `any` used for response" ✅ Valid, HIGH
5. "No retry logic" 🔄 Debatable - Let me ask Codex

Claude Code to Codex: "Peer review found 3 valid issues:

1. **CRITICAL** [samGov.ts:45] - API key is hardcoded
   - Fix: Move to environment variable or Convex secrets

2. **HIGH** [samGov.ts:120] - Response typed as `any`
   - Fix: Create proper interface for SAM.gov API response

3. **DEBATABLE** [samGov.ts:88] - No retry logic for API failures
   - Question: Should we implement retry, or is fail-fast acceptable here?

Please address items 1-2 and clarify item 3."

Red Flags to Watch For

Invalid Findings (Common False Positives)

• "This could be simplified" without clear benefit
• "Should use pattern X" when existing pattern is project standard
• "Missing optimization" for code that isn't a bottleneck
• "Needs more error handling" when current handling is sufficient

Valid Findings (Don't Dismiss These)

• Security issues (auth bypass, injection vulnerabilities)
• Data loss risks (missing validation, race conditions)
• Bandwidth violations (unbounded queries, missing limits)
• Type safety violations (any types, missing null checks)

Anti-Patterns to Avoid

❌ Accepting all findings without verification

• Results in wasted work on non-issues

❌ Dismissing findings defensively

• Defeats the purpose of fresh eyes

❌ Turning review into architecture debate

• Focus on code quality, not design preferences

✅ Verify, prioritize, consolidate

• Check if issue exists → assess severity → provide clear feedback

Invocation

The user invokes this skill by typing /peer-review when they want to coordinate a cross-agent review.

Usage:

/peer-review [files-or-directories] [reviewer-agent]

Examples:

/peer-review convex/ingestion/samGov.ts Gemini
/peer-review components/AdminView.tsx Codex
/peer-review convex/ Gemini

If no reviewer specified, Claude Code will assign the appropriate agent based on domain (frontend vs backend).

Notes

• Always verify findings - The reviewer might not have full context
• Be specific in feedback - File:line references, not vague comments
• Prioritize by severity - CRITICAL must be fixed, LOW is optional
• This is collaborative, not combative - Goal is better code, not proving someone wrong