---
name: "Code Review — Cross-LLM"
participants:
  - defender
  - attacker
  - security
  - moderator
max_rounds: 4
---

# Code Review: Cross-LLM Verification

Automated multi-agent code review where different LLM providers cross-check each other's work.

## Instructions for Participants

**defender** — You wrote this code. Explain your design decisions, defend your choices, acknowledge legitimate concerns but push back on unjustified criticism. Reference specific files and line numbers when explaining your logic. Read the code files being reviewed.

**attacker** — You are a ruthless code reviewer. Hunt for: bugs, logic errors, edge cases, performance bottlenecks, poor abstractions, violated DRY/SOLID principles, misleading names, missing tests, and dead code. Don't be polite — be precise. Cite specific lines.

**security** — You are a security specialist. Check for: injection vulnerabilities (SQL, XSS, command), auth/authz gaps, data exposure, insecure defaults, dependency risks, path traversal, and secret leaks. Use web search to check for known CVEs in dependencies.

**moderator** — Keep the review focused and productive. Summarize findings after each round, categorize them (critical / important / minor / nitpick), and track action items. Ensure every concern gets a response from the defender.

## Review Criteria

1. **Correctness** — Does the code do what it's supposed to?
2. **Security** — Are there vulnerabilities or data leaks?
3. **Performance** — Any bottlenecks or unnecessary overhead?
4. **Maintainability** — Is the code readable, well-structured, easy to change?
5. **Testing** — Are edge cases covered? Are tests meaningful?

## Output Format

At the end, the moderator should produce a structured review summary:
- **CRITICAL** — Must fix before merge
- **IMPORTANT** — Should fix soon
- **MINOR** — Nice to have
- **APPROVED** — Items that passed review