open-ralph-wiggum

Open Ralph Wiggum

Open Ralph Wiggum (ralph) wraps any supported AI coding agent in an autonomous loop: it sends the same prompt on every iteration, and the agent self-corrects by observing the state of the repo. The loop ends when the agent outputs a configurable completion promise (e.g. <promise>COMPLETE</promise>).

Supported agents: Claude Code, OpenAI Codex, GitHub Copilot CLI, Cursor Agent, Qwen Code, OpenCode (default).

---

Installation

Prerequisites

• Bun runtime
• At least one of these AI coding agent CLIs installed and authenticated:
  • claude — Claude Code
  • codex — OpenAI Codex CLI
  • copilot — GitHub Copilot CLI
  • cursor-agent — Cursor Agent CLI
  • qwen — Qwen Code CLI
  • opencode — OpenCode

npm (recommended)

npm install -g @th0rgal/ralph-wiggum

Bun

bun add -g @th0rgal/ralph-wiggum

From source (Linux/macOS)

git clone https://github.com/Th0rgal/open-ralph-wiggum
cd open-ralph-wiggum
./install.sh

From source (Windows)

git clone https://github.com/Th0rgal/open-ralph-wiggum
cd open-ralph-wiggum
.\install.ps1

After installation, the ralph command is available globally.

---

Quick Start

Always include a completion promise in your prompt — this is how ralph knows the task is done.

OpenCode (default)

ralph "Create a hello.txt file with 'Hello World'. Output <promise>DONE</promise> when complete." \
  --max-iterations 5

Claude Code

ralph "Build a REST API with tests. Output <promise>COMPLETE</promise> when all tests pass." \
  --agent claude-code --model claude-sonnet-4 --max-iterations 20

Codex

ralph "Refactor auth module, ensure all tests pass. Output <promise>COMPLETE</promise> when done." \
  --agent codex --model gpt-5-codex --max-iterations 20

Copilot CLI

ralph "Implement login feature. Output <promise>COMPLETE</promise> when done." \
  --agent copilot --max-iterations 15

Requires GitHub Copilot subscription and prior authentication (copilot /login or GH_TOKEN env var).

Cursor Agent

ralph "Add integration tests for the API. Output <promise>COMPLETE</promise> when done." \
  --agent cursor-agent --max-iterations 15

Requires Cursor Agent CLI installed via curl https://cursor.com/install -fsSL | bash. For headless environments, set CURSOR_API_KEY.

Qwen Code

ralph "Refactor the database layer. Output <promise>COMPLETE</promise> when done." \
  --agent qwen-code --max-iterations 15

Requires Qwen Code installed via npm install -g @qwen-code/qwen-code.

---

Checking Available Agents and Models

Before running ralph, verify which agents are installed and what models they support.

Check available models per agent

OpenCode — lists all configured providers and models:

opencode models

Configure a default in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anthropic/claude-sonnet-4-5"
}

Claude Code — check version and available models:

claude --version

Common models: claude-opus-4, claude-sonnet-4, claude-haiku-4

Codex — check version and available models:

codex --version

Common models: gpt-5-codex, o4-mini

Copilot CLI — uses GitHub Copilot subscription; verify auth:

copilot /status   # shows login state and available models
# If not logged in:
copilot /login
# Or set env var:
export GH_TOKEN=your_token

Qwen Code — check version and install if needed:

qwen --version
# If not installed:
npm install -g @qwen-code/qwen-code

Quick environment check (Linux/macOS)

for bin in opencode claude codex copilot cursor-agent qwen; do
  if command -v "$bin" &>/dev/null; then echo "✅ $bin: $(which $bin)"; else echo "❌ $bin: not found"; fi
done && \
  [[ -n "$GH_TOKEN" ]] && echo "✅ GH_TOKEN set (Copilot CLI)" || echo "ℹ️  GH_TOKEN not set (needed only for Copilot CLI)"

---

Agent Selection

Use environment variables to point to a custom binary path if the CLI is not on $PATH.

---

Key Options

--agent AGENT            Agent to use (opencode|claude-code|codex|copilot|cursor-agent|qwen-code)
--model MODEL            Model name (agent-specific, e.g. claude-sonnet-4, gpt-5-codex)
--max-iterations N       Stop after N iterations (always set this as a safety net)
--min-iterations N       Require at least N iterations before allowing completion (default: 1)
--completion-promise T   Text that signals task completion (default: COMPLETE)
--abort-promise TEXT     Text that signals early abort/precondition failure
--tasks / -t             Enable Tasks Mode (structured multi-task tracking)
--prompt-file / -f PATH  Read prompt from a file instead of CLI argument
--prompt-template PATH   Use a custom Mustache-style prompt template
--no-commit              Skip git auto-commit after each iteration
--no-plugins             Disable OpenCode plugins (useful to avoid plugin conflicts)
--allow-all              Auto-approve all tool permission prompts (default: on)
--status                 Show live loop status from another terminal
--add-context TEXT       Inject a hint for the next iteration without stopping the loop
--clear-context          Remove pending context
--list-tasks             List current tasks (Tasks Mode)
--add-task TEXT          Add a task (Tasks Mode)
--remove-task N          Remove task by index (Tasks Mode)
--rotation LIST          Cycle through agent/model pairs each iteration (comma-separated agent:model)
--verbose-tools          Print every tool line (disable compact tool summary)
--last-activity-timeout DURATION  Kill and restart iteration after inactivity (e.g., 30m, 1h)
--no-questions           Disable interactive question handling (agent will loop on questions)
--task-promise T        Text that signals task completion (default: READY_FOR_NEXT_TASK)
--no-stream             Buffer agent output and print at the end
--no-allow-all          Require interactive permission prompts
--config PATH           Use custom agent config file
--init-config [PATH]    Write default agent config to PATH and exit
--questions             Enable interactive question handling (default: enabled)

---

IDE Integration

Ralph is a terminal CLI tool that runs inside any IDE's integrated terminal.

VS Code / Cursor

1. Open the integrated terminal (Ctrl+`` or View → Terminal).

2. Run ralph from your project root using the agent of your choice:

   OpenCode:

   ralph "Your task. Output <promise>COMPLETE</promise> when done." --max-iterations 20
   

   Claude Code:

   ralph "Your task. Output <promise>COMPLETE</promise> when done." \
     --agent claude-code --model claude-sonnet-4 --max-iterations 20
   

   Codex:

   ralph "Your task. Output <promise>COMPLETE</promise> when done." \
     --agent codex --model gpt-5-codex --max-iterations 20
   

   Copilot CLI:

   ralph "Your task. Output <promise>COMPLETE</promise> when done." \
     --agent copilot --max-iterations 20
   

   Note: --agent copilot uses the standalone Copilot CLI, not the VS Code extension. Both can be active at the same time.

   Cursor Agent:

   ralph "Your task. Output <promise>COMPLETE</promise> when done." \
     --agent cursor-agent --max-iterations 20
   

   Qwen Code:

   ralph "Your task. Output <promise>COMPLETE</promise> when done." \
     --agent qwen-code --max-iterations 20
   

3. Open a second terminal tab to monitor while the loop runs:

   ralph --status
   

4. Inject hints mid-loop from the second terminal:

   ralph --add-context "Focus on fixing the auth module first"
   

JetBrains IDEs (IntelliJ, WebStorm, PyCharm, etc.)

1. Open the integrated terminal (Alt+F12).
2. Run the same agent-specific commands as above.
3. Use Run Configurations → Shell Script to save common ralph invocations per agent as reusable run configurations.

Neovim / Vim

Run ralph in a split terminal. Examples per agent:

OpenCode:

:split | terminal ralph "Your task. Output <promise>COMPLETE</promise> when done." --max-iterations 20

Claude Code:

:split | terminal ralph "Your task. Output <promise>COMPLETE</promise> when done." --agent claude-code --model claude-sonnet-4 --max-iterations 20

Codex:

:split | terminal ralph "Your task. Output <promise>COMPLETE</promise> when done." --agent codex --model gpt-5-codex --max-iterations 20

Copilot CLI:

:split | terminal ralph "Your task. Output <promise>COMPLETE</promise> when done." --agent copilot --max-iterations 20

Cursor Agent:

:split | terminal ralph "Your task. Output <promise>COMPLETE</promise> when done." --agent cursor-agent --max-iterations 20

Qwen Code:

:split | terminal ralph "Your task. Output <promise>COMPLETE</promise> when done." --agent qwen-code --max-iterations 20

Or use a plugin like toggleterm.nvim for a persistent terminal.

Any IDE — Prompt File Workflow

For complex prompts, save them as a file to avoid shell escaping issues and make prompts versionable.

OpenCode:

ralph --prompt-file ./task.md --max-iterations 30

Claude Code:

ralph --prompt-file ./task.md --agent claude-code --model claude-sonnet-4 --max-iterations 30

Codex:

ralph --prompt-file ./task.md --agent codex --model gpt-5-codex --max-iterations 30

Copilot CLI:

ralph --prompt-file ./task.md --agent copilot --max-iterations 30

Cursor Agent:

ralph --prompt-file ./task.md --agent cursor-agent --max-iterations 30

Qwen Code:

ralph --prompt-file ./task.md --agent qwen-code --max-iterations 30

---

Tasks Mode

Break large projects into a tracked task list:

# Start a loop in Tasks Mode
ralph "Build a full-stack app" --tasks --max-iterations 50

# Manage tasks while the loop is idle (or before starting)
ralph --add-task "Set up database schema"
ralph --add-task "Implement REST API"
ralph --list-tasks
ralph --remove-task 2

Tasks are stored in .ralph/ralph-tasks.md. Each task uses one loop iteration, signaled by <promise>READY_FOR_NEXT_TASK</promise>.

---

Monitoring a Running Loop

From a second terminal in the same project directory:

ralph --status      # Shows iteration progress, history, struggle indicators
ralph --add-context "The bug is in utils/parser.ts line 42"  # Guide the agent
ralph --clear-context  # Remove queued hint

The status dashboard shows iteration count, time elapsed, tool usage per iteration, and struggle warnings (e.g., no file changes in N iterations).

The --status output includes:

• Active loop info: iteration, elapsed time, prompt, rotation position (if using --rotation)
• Pending context: hints queued for next iteration
• Iteration history: last 5 iterations with agent/model, tool usage, duration
• Struggle indicators: warnings if no file changes in N iterations

---

Writing Effective Prompts

Bad prompt (no verifiable criteria):

Build a todo API

Good prompt (verifiable, with completion promise):

Build a REST API for todos with:
- CRUD endpoints (GET, POST, PUT, DELETE)
- Input validation
- Tests for each endpoint

Run tests after each change.
Output <promise>COMPLETE</promise> when all tests pass.

Rules of thumb:

• Include explicit success criteria (tests passing, linter clean, files present)
• Always include a completion promise tag that the agent must output
• Set --max-iterations as a safety net (20–50 is a common range)
• For complex projects, use --tasks or a --prompt-file

---

Custom Prompt Templates

Create a Markdown template with Mustache-style variables:

# Iteration {{iteration}} / {{max_iterations}}

## Task

{{prompt}}

## Instructions

Check git history to see what was tried. Fix what failed.
Output <promise>{{completion_promise}}</promise> when done.

{{context}}

Available variables: {{iteration}}, {{max_iterations}}, {{min_iterations}}, {{prompt}}, {{completion_promise}}, {{abort_promise}}, {{task_promise}}, {{context}}, {{tasks}}.

Use with:

ralph "Your task" --prompt-template ./my-template.md

---

Agent Rotation

Cycle through different agent/model combinations across iterations:

# Alternate between OpenCode and Claude Code
ralph "Build a REST API" \
  --rotation "opencode:claude-sonnet-4,claude-code:claude-sonnet-4" \
  --max-iterations 10

# Three-way rotation
ralph "Refactor the auth module" \
  --rotation "opencode:claude-sonnet-4,claude-code:claude-sonnet-4,codex:gpt-5-codex" \
  --max-iterations 15

Valid agents are opencode, claude-code, codex, copilot, cursor-agent, and qwen-code.

Format: agent:model entries separated by commas. When --rotation is set, --agent and --model are ignored. The list cycles (iteration 3 of a 2-entry rotation goes back to entry 1).

---

Agent-Specific Notes

OpenCode (default)

• Default model can be set in ~/.config/opencode/opencode.json:

  {
    "$schema": "https://opencode.ai/config.json",
    "model": "your-provider/model-name"
  }
  

• Use --no-plugins if OpenCode tries to load a ralph-wiggum plugin.

Claude Code

ralph "Refactor the auth module and ensure tests pass" \
  --agent claude-code --model claude-sonnet-4 --max-iterations 15

OpenAI Codex

ralph "Generate unit tests for all utility functions" \
  --agent codex --model gpt-5-codex --max-iterations 10

Copilot CLI

Requires a GitHub Copilot subscription. Authenticate before running:

copilot /login   # or set GH_TOKEN / GITHUB_TOKEN env var

Install:

npm install -g @github/copilot
# or
brew install copilot-cli

Usage:

ralph "Refactor the auth module and add tests" \
  --agent copilot --max-iterations 15

# With a specific model
ralph "Build a REST API" \
  --agent copilot --model claude-opus-4.6 --max-iterations 10

Notes:

• Default model is Claude Sonnet 4.5; override with --model
• --no-plugins has no effect with Copilot CLI
• --allow-all (default) maps to --allow-all + --no-ask-user in Copilot CLI

Qwen Code

Install:

npm install -g @qwen-code/qwen-code

Usage:

ralph "Refactor the database layer" \
  --agent qwen-code --max-iterations 10

# With a specific model
ralph "Add integration tests for the API" \
  --agent qwen-code --model qwen-coder-plus --max-iterations 15

Notes:

• --allow-all (default) maps to --yolo in Qwen Code CLI
• --no-plugins has no effect with Qwen Code
• Headless mode uses --output-format stream-json --include-partial-messages
• Binary is qwen; override with RALPH_QWEN_CODE_BINARY

Cursor Agent

Install:

curl https://cursor.com/install -fsSL | bash

Usage:

ralph "Add integration tests for the API" \
  --agent cursor-agent --max-iterations 10

# With a specific model
ralph "Refactor the database layer" \
  --agent cursor-agent --model sonnet-4 --max-iterations 15

Notes:

• --allow-all (default) maps to --force in Cursor Agent CLI
• --no-plugins has no effect with Cursor Agent
• For headless environments without a Cursor window, set CURSOR_API_KEY env var
• Binary is cursor-agent; override with RALPH_CURSOR_AGENT_BINARY

---

---

When to Use Ralph

Good for:

• Tasks with automatic verification (tests, linters, type checking)
• Well-defined tasks with clear completion criteria
• Greenfield projects where you can walk away
• Iterative refinement (getting tests to pass)

Not good for:

• Tasks requiring human judgment
• One-shot operations
• Unclear success criteria
• Production debugging

---

Recommended PRD Format

For complex tasks, pass a prompt file with --prompt-file. Use this structure:

• Goal: one sentence summary of the desired outcome
• Scope: what is in/out
• Requirements: numbered, testable items
• Constraints: tech stack, performance, security, compatibility
• Acceptance criteria: explicit success checks
• Completion promise: <promise>COMPLETE</promise>

For larger projects, a JSON feature list reduces the chance of agents modifying test definitions:

{
  "features": [
    {
      "category": "functional",
      "description": "Feature description",
      "steps": ["Step 1", "Step 2"],
      "passes": false
    }
  ]
}

Reference it in your prompt: Read features.json. Work through each feature. Update "passes" to true when verified. Output <promise>COMPLETE</promise> when all pass.

---

Troubleshooting

---

Uninstall

npm uninstall -g @th0rgal/ralph-wiggum