One writes. One reviews. You architect.
AI can generate code. But it cannot distrust itself.
A single agent writes code AND decides if it's done — like grading your own exam. Different agents fail in characteristically different ways: Claude Code skips error handling when context grows long; Codex over-engineers abstractions but catches edge cases Claude misses. Pairing them means each catches what the other misses.
Ralph writes → Lisa reviews → Consensus → Next step
↑ |
└────────────────────────────────────────┘
- Ralph (Claude Code): Lead developer — researches, plans, codes, tests
- Lisa (Codex): Code reviewer — reviews diffs, checks edge cases
- You: Tech lead — architecture, scope, tiebreaking
An agent never reviews its own output.
npm i -g ralph-lisa-loop
cd your-project
ralph-lisa init
ralph-lisa start "implement login feature"See the User Guide for the full walkthrough.
Used to fork AionUI (~3k stars, Electron + React app) into an independent production product:
| Metric | Value |
|---|---|
| Project | AionUI fork → Margay |
| Commits | 30 |
| Manual code | 0 lines |
| Review rounds | 40 |
| Status | In production use as internal AI assistant |
- User Guide — Setup, workflows, configuration
- Command Reference — All CLI commands
- FAQ — Common questions and troubleshooting
- Changelog — Version history
- Design Philosophy — Why dual-agent collaboration works
Node.js >= 18, Claude Code, Codex CLI. Auto mode also requires tmux. See the User Guide for details.
Part of the TigerHill project family.
The iterative loop concept builds on Geoffrey Huntley's Ralph Wiggum technique. Ralph-Lisa Loop adds structured dual-agent review discipline on top — enforcing role separation between generation and critique.