Adversarial Review

Adversarial Review

Multi-agent code review with Claude and GPT Codex in an adversarial debate loop.

Based on patterns from asimov-ralph and research on AI Debate.

Concept

Two AI agents (Claude and GPT Codex) independently review code, then critique each other's findings through multiple rounds of debate. This adversarial process helps:

- Find more issues: Different models catch different problems
- Eliminate false positives: Cross-validation filters out incorrect findings
- Build consensus: Disagreements are resolved through structured debate
- Improve confidence: Issues both agents agree on are high-confidence fixes

The 4-Phase Loop

Quick Start

Requirements

- claude CLI: npm install -g @anthropic-ai/claude-code
- codex CLI: npm install -g @openai/codex
- jq: brew install jq (macOS) or apt install jq (Linux)
- coreutils (macOS only, for timeout): brew install coreutils

Usage

Project Structure

Circuit Breaker

Prevents runaway loops by detecting:

- No progress: 3 iterations with no fixes made
- Persistent disagreement: 5+ iterations where agents can't agree
- Same issues: 3+ iterations finding the same unfixable issues

Customization

Custom Review Prompts

Environment Variables

How It Works

Agent Status Blocks

Each agent outputs a structured status block that gets parsed:

Exit Conditions

The loop exits when:
1. Both agents report NO_ISSUES in Phase 1
2. Synthesis completes with EXIT_SIGNAL: true
3. Max iterations reached
4. Circuit breaker opens (stagnation detected)

Artifacts

Each iteration produces:
- iter{N}_1_claude_review.md - Claude's initial review
- iter{N}_1_codex_review.md - Codex's initial review
- iter{N}_2_claude_on_codex.md - Claude's cross-review
- iter{N}_2_codex_on_claude.md - Codex's cross-review
- iter{N}_3_claude_meta.md -…