ralph-code

An autonomous AI agent loop that turns a description into working code. Powered by Claude Code and Codex. Inspired by Ralph.

• Describe → Plan → Execute — tell it what to build, it generates tasks and works through them
• Commit per task — each completed task is a clean git commit
• Context between tasks — git diff + structured progress log passed to every iteration
• Auto retry — failed tasks retry up to 3 times before stopping
• Mix agents — use Claude for planning and Codex for execution, or any combo
• Customizable prompts — edit the prompt templates to tune agent behavior

Warning: ralph-code runs agents in fully autonomous mode (--dangerously-skip-permissions for Claude, --full-auto for Codex). Always review generated tasks before executing.

Install

npm install -g ralph-code

Requirements

• Node.js 18+
• Claude Code or Codex CLI installed and authenticated

Quick Start

cd ~/my-project
ralph-code

  ╭───────────────────────────╮
  │  ralph-code               │
  │  v0.1.0                   │
  ╰───────────────────────────╯

  plan: claude/opus · execution: claude/sonnet

> /run "create a todo app"
  ✓ Created tasks.md with 4 task(s)
  ✓ Tasks are ready. Review tasks.md and edit if needed before continuing.
  Start execution? (y/n) y

  [1/4]  ·  4 pending, 0 done
  ▸ Set up project structure
  ✓ Set up project structure

  [2/4]  ·  3 pending, 1 done
  ▸ Build todo list UI
  ...

How It Works

/run "description"
───────────────────
  .ralph/tasks.md exists?
    no  → plan agent generates tasks
    yes ↓

  Parse tasks → find first [pending]
          │
          ▼
  Build prompt (task + git diff + progress + rules)
          │
          ▼
  Execution agent: implement → test → commit
       → mark [done] → log to task-progress.md
          │
          ▼
  Failed? → retry (up to maxRetries)
  Done?   → next task or exit

Each task receives:

• Current task description and all tasks for context
• git diff from the previous task's commit
• Structured progress log from completed tasks

Config

On first run, a .ralph/ directory is created in your project. Add .ralph/ to your .gitignore.

.ralph/
  config.json        Configuration
  tasks.md           Generated task list
  task-progress.md   Progress log (context between tasks)

.ralph/config.json:

{
  "model": {
    "plan": "claude/opus",
    "execution": "claude/sonnet"
  },
  "tasks": "tasks.md",
  "maxIterations": 100,
  "maxRetries": 3,
  "timeout": 300
}

Field	Description
model.plan	Agent/model for task generation
model.execution	Agent/model for task execution
tasks	Task file name inside .ralph/
maxIterations	Max tasks to process per /run
maxRetries	Retry attempts per failed task
timeout	Seconds before killing an agent invocation

Model selection

Models use agent/model format:

Agent	Planning	Execution
Claude	claude/opus	claude/sonnet, claude/haiku
Codex	codex/gpt-5.3-codex	codex/gpt-5.3-codex, codex/gpt-5-codex-mini

Mix agents freely:

{
  "model": {
    "plan": "claude/opus",
    "execution": "codex/gpt-5.3-codex"
  }
}

Commands

Command	Description
/run	Resume execution from the next pending task
/run "desc"	Generate a new plan, or replan if tasks already exist
/config	Show current config
/help	Show available commands
/exit	Quit
Esc	Pause execution

Task Format

Tasks use ## headers with a [pending] or [done] tag:

# Project: my-app

## [pending] Set up project structure
Initialize the repo with the base framework.

## [pending] Add user authentication
Implement login/signup with email and password.

## [done] Create README
Already completed.

Tasks are processed top-to-bottom. The first [pending] task is picked each iteration.