Review

Be thorough. Check for:

• Logic errors, off-by-one, null/undefined paths
• Security: injection, XSS, auth bypass, secrets in code
• Performance: O(n^2) loops, unnecessary allocations, missing caching
• Correctness: invariant violations, race conditions, error handling gaps
• Design: SOLID violations, missing abstractions, tight coupling

Step 3: Gemini Review

Run Gemini on the same diff using this command:

git diff HEAD~1 | gemini -m gemini-2.5-pro -p "$(cat ${CLAUDE_PLUGIN_DIR}/templates/review-prompt-gemini.md)"

If gemini CLI is not available, skip this step and note it.

Parse Gemini's output for findings in the same format.

Step 4: Codex Review

Run Codex review:

codex review --uncommitted 2>&1

If codex CLI is not available, skip this step and note it.

Parse Codex's output for findings.

Step 5: Cross-Model Agreement Matrix

For each finding from any model, check if other models found the same or similar issue. Build an agreement matrix:

| # | Finding | Claude | Gemini | Codex | Severity |
|---|---------|--------|--------|-------|----------|
| 1 | Missing null check in foo() | YES | YES | - | high |
| 2 | SQL injection in query() | YES | YES | YES | critical |
| 3 | Unused import | - | YES | - | low |

Findings agreed upon by 2+ models are higher confidence.

Step 6: Output Report

Output the final report in this exact format:

## Code Review Report

**Date**: <date>
**Scope**: <what was reviewed>
**Models**: Claude, Gemini, Codex (or subset if some unavailable)

### Findings

| # | Severity | Category | File | Description | Models |
|---|----------|----------|------|-------------|--------|
| 1 | critical | ... | ... | ... | C,G,X |
| 2 | high | ... | ... | ... | C,G |

### Agreement Matrix

<cross-model agreement table>

### Summary

- **Critical**: N findings
- **High**: N findings
- **Medium**: N findings
- **Total**: N findings
- **Cross-model agreement**: N/M findings confirmed by 2+ models

### Recommendations (prioritized)

1. [CRITICAL] ...
2. [HIGH] ...

Step 7: Save Results

Save the structured report to ${CLAUDE_PLUGIN_DIR}/data/latest-review.json as:

{
  "date": "...",
  "scope": "...",
  "models_used": ["claude", "gemini", "codex"],
  "findings": [...],
  "agreement_matrix": [...],
  "summary": { "critical": N, "high": N, "medium": N, "low": N, "total": N }
}