Works with Claude Code, OpenAI Codex, Copilot CLI, Cursor Agent, Qwen Code, and OpenCode — switch agents with --agent.
Based on the Ralph Wiggum technique by Geoffrey Huntley

Supported Agents • What is Ralph? • Installation • Quick Start • Commands

Tired of agents breaking your local environment?
🏝️ sandboxed.sh gives each task an isolated Linux workspace. Self-hosted. Git-backed.

💬 Join the community: relens.ai/community

Open Ralph Wiggum works with multiple AI coding agents. Switch between them using the --agent flag:

# Use Claude Code
ralph "Build a REST API" --agent claude-code --max-iterations 10

# Use OpenAI Codex
ralph "Create a CLI tool" --agent codex --max-iterations 10

# Use Copilot CLI
ralph "Refactor the auth module" --agent copilot --max-iterations 10

# Use Cursor Agent
ralph "Add unit tests" --agent cursor-agent --max-iterations 10

# Use Qwen Code
ralph "Add unit tests" --agent qwen-code --max-iterations 10

# Use OpenCode (default)
ralph "Fix the failing tests" --max-iterations 10

Open Ralph Wiggum implements the Ralph Wiggum technique — an autonomous agentic loop where an AI coding agent (Claude Code, Codex, Copilot CLI, Cursor Agent, Qwen Code, or OpenCode) receives the same prompt repeatedly until it completes a task. Each iteration, the AI sees its previous work in files and git history, enabling self-correction and incremental progress.

This is a CLI tool that wraps any supported AI coding agent in a persistent development loop. No plugins required — just install and run.

# The essence of the Ralph loop:
while true; do
  claude-code "Build feature X. Output <promise>DONE</promise> when complete."  # or codex, copilot, cursor-agent, qwen, opencode
done

Why this works: The AI doesn't talk to itself between iterations. It sees the same prompt each time, but the codebase has changed from previous iterations. This creates a powerful feedback loop where the agent iteratively improves its work until all tests pass.

Switch between AI coding agents without changing your workflow:

• Claude Code (--agent claude-code) — Anthropic's powerful coding agent
• Codex (--agent codex) — OpenAI's code-specialized model
• Copilot CLI (--agent copilot) — GitHub's agentic coding tool
• Cursor Agent (--agent cursor-agent) — Cursor's headless AI coding agent
• Qwen Code (--agent qwen-code) — Alibaba's Qwen Code headless CLI agent
• OpenCode (--agent opencode) — Open-source default option

• Multi-Agent Support — Use Claude Code, Codex, Copilot CLI, Cursor Agent, Qwen Code, or OpenCode with the same workflow
• Self-Correcting Loops — Agent sees its previous work and fixes its own mistakes
• Autonomous Execution — Set it running and come back to finished code
• Task Tracking — Built-in task management with --tasks mode
• Live Monitoring — Check progress with --status from another terminal
• Mid-Loop Hints — Inject guidance with --add-context without stopping

Prerequisites:

• Bun runtime
• At least one AI coding agent CLI:
  • Claude Code — Anthropic's Claude Code CLI
  • Codex — OpenAI's Codex CLI
  • Copilot CLI — GitHub's Copilot CLI
  • Cursor Agent — Cursor's headless Agent CLI
  • Qwen Code — Alibaba's Qwen Code CLI
  • OpenCode — Open-source AI coding assistant

npm install -g @th0rgal/ralph-wiggum

bun add -g @th0rgal/ralph-wiggum

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

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

This installs the ralph CLI command globally.

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

# Build something real
ralph "Build a REST API for todos with CRUD operations and tests. \
  Run tests after each change. Output <promise>COMPLETE</promise> when all tests pass." \
  --max-iterations 20

# Use Claude Code instead of OpenCode
ralph "Create a small CLI and document usage. Output <promise>COMPLETE</promise> when done." \
  --agent claude-code --model claude-sonnet-4 --max-iterations 5

# Use Codex instead of OpenCode
ralph "Create a small CLI and document usage. Output <promise>COMPLETE</promise> when done." \
  --agent codex --model gpt-5-codex --max-iterations 5

# Use Codex goal mode through OMX: Ralph owns cross-iteration retries, /goal owns one iteration
RALPH_CODEX_GOAL=1 RALPH_CODEX_BACKEND=omx \
ralph "Complete the task described in .harness/goal.md. Output <promise>COMPLETE</promise> when everything passes." \
  --agent codex --max-iterations 5

# Use Copilot CLI
ralph "Create a small CLI and document usage. Output <promise>COMPLETE</promise> when done." \
  --agent copilot --max-iterations 5

# Use Qwen Code
ralph "Create a small CLI and document usage. Output <promise>COMPLETE</promise> when done." \
  --agent qwen-code --max-iterations 5

# Complex project with Tasks Mode
ralph "Build a full-stack web application with user auth and database" \
  --tasks --max-iterations 50

Configure agent binaries with these environment variables:

Note for Windows users: Ralph automatically resolves .cmd extensions for npm-installed CLIs. If you encounter "command not found" errors, you can use these environment variables to specify the full path to the executable.

ralph "<prompt>" [options]

Options:
  --agent AGENT            AI agent to use: opencode (default), claude-code, codex, copilot, cursor-agent, qwen-code
  --codex-goal             Run Codex iterations in goal mode; final Codex/OMX prompt starts with /goal
  --codex-backend BACKEND  Backend for --codex-goal: codex or omx (default: detect)
  --codex-goal-native      Force a native /goal attempt even when backend support is unconfirmed
  --min-iterations N       Minimum iterations before completion allowed (default: 1)
  --max-iterations N       Stop after N iterations (default: unlimited)
  --completion-promise T   Text that signals completion (default: COMPLETE)
  --abort-promise TEXT     Phrase that signals early abort (e.g., precondition failed)
  --tasks, -t              Enable Tasks Mode for structured task tracking
  --task-min-iterations N   Require each top-level task to receive N Ralph iterations before completion (default: 1)
  --task-promise T         Text that signals task completion (default: READY_FOR_NEXT_TASK)
  --model MODEL            Model to use (agent-specific)
  --rotation LIST          Agent/model rotation for each iteration (comma-separated)
  --prompt-file, --file, -f  Read prompt content from a file
  --prompt-template PATH   Use custom prompt template (see Custom Prompts)
  --no-stream              Buffer agent output and print at the end
  --verbose-tools          Print every tool line (disable compact tool summary)
  --questions              Enable interactive question handling (default: enabled)
  --no-questions           Disable interactive question handling (agent will loop on questions)
  --no-plugins             Disable non-auth OpenCode plugins for this run (opencode only)
  --no-commit              Don't auto-commit after iterations
  --allow-all              Auto-approve all tool permissions (default: on)
  --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
  --help                   Show help

--codex-goal is an opt-in mode for --agent codex. Open Ralph still owns the outer loop: max iterations, process restarts, promise detection, .ralph/ralph-history.json, git/file-system state, and optional auto-commits. Inside each Ralph iteration, the Codex backend receives a final prompt whose first token is /goal, so Codex goal mode can own the single-session push for that iteration.

ralph \
  "Complete the task in .harness/goal.md. Run .harness/checks.sh. Output <promise>COMPLETE</promise> when everything passes." \
  --agent codex \
  --codex-goal \
  --codex-backend omx \
  --max-iterations 5

Equivalent environment form:

RALPH_CODEX_GOAL=1 RALPH_CODEX_BACKEND=omx ralph \
  "Complete the task in .harness/goal.md. Run .harness/checks.sh. Output <promise>COMPLETE</promise> when everything passes." \
  --agent codex \
  --max-iterations 5

Goal-mode iterations write a small durable audit ledger to .ralph/codex-goal-ledger.jsonl. This is intentionally file-system state, not Codex thread state, so later Ralph iterations can start fresh while retaining repo-native progress evidence. If native /goal cannot be pre-confirmed, Ralph prints an explicit warning; simulated fallback is used only when native Codex goal support is not available for the selected backend.

Tasks Mode allows you to break complex projects into smaller, manageable tasks. Ralph works on one task at a time and tracks progress in a markdown file.

# Enable Tasks Mode
ralph "Build a complete web application" --tasks --max-iterations 20

# Custom task completion signal
ralph "Multi-feature project" --tasks --task-promise "TASK_DONE"

# Require each top-level task to receive three Ralph iterations before task/final completion
ralph "Multi-feature project" --tasks --task-min-iterations 3 --max-iterations 20

# List current tasks
ralph --list-tasks

# Add a new task
ralph --add-task "Implement user authentication"

# Remove task by index
ralph --remove-task 3

# Show status (tasks shown automatically when tasks mode is active)
ralph --status

1. Task File: Tasks are stored in .ralph/ralph-tasks.md
2. One Task Per Iteration: Ralph focuses on a single task to reduce confusion
3. Automatic Progression: When a task completes (<promise>READY_FOR_NEXT_TASK</promise>), Ralph moves to the next eligible task
4. Task Minimum Iterations: --task-min-iterations N requires every top-level task to receive N Ralph outer-loop iterations before task or final completion is accepted. This is separate from global --min-iterations, which gates the whole run.
5. Task Ledger Fallback: With task minimums enabled, if the selected top-level task is already [x] with all subtasks [x] after its required iterations, Ralph advances from the task ledger even if the agent forgets to emit READY_FOR_NEXT_TASK
6. Persistent State: Tasks survive loop restarts; per-task attempt counts are stored in .ralph/ralph-task-runs.json when task minimums are enabled
7. Focused Context: Smaller contexts per iteration reduce costs and improve reliability

Task status indicators:

• [ ] - Not started
• [/] - In progress
• [x] - Complete

Example task file:

# Ralph Tasks

- [ ] Set up project structure
- [x] Initialize git repository
- [/] Implement user authentication
  - [ ] Create login page
  - [ ] Add JWT handling
- [ ] Build dashboard UI

You can fully customize the prompt sent to the agent using --prompt-template. This is useful for integrating with custom workflows or tools.

ralph "Build a REST API" --prompt-template ./my-template.md

Available variables:

Example template (my-template.md):

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

## Task
{{prompt}}

## Instructions
1. Check beads for current status
2. Decide what to do next
3. When the epic in beads is complete, output:
   <promise>{{completion_promise}}</promise>

{{context}}

# Check status of active loop (run from another terminal)
ralph --status

# Add context/hints for the next iteration
ralph --add-context "Focus on fixing the auth module first"

# Clear pending context
ralph --clear-context

The --status command shows:

• Active loop info: Current iteration, elapsed time, prompt
• Pending context: Any hints queued for next iteration
• Current tasks: Automatically shown when tasks mode is active (or use --tasks)
• Iteration history: Last 5 iterations with tools used, duration
• Struggle indicators: Warnings if agent is stuck (no progress, repeated errors)

╔══════════════════════════════════════════════════════════════════╗
║                    Ralph Wiggum Status                           ║
╚══════════════════════════════════════════════════════════════════╝

🔄 ACTIVE LOOP
   Iteration:    3 / 10
   Elapsed:      5m 23s
   Promise:      COMPLETE
   Prompt:       Build a REST API...

📊 HISTORY (3 iterations)
   Total time:   5m 23s

   Recent iterations:
   🔄 #1: 2m 10s | Bash:5 Write:3 Read:2
   🔄 #2: 1m 45s | Edit:4 Bash:3 Read:2
   🔄 #3: 1m 28s | Bash:2 Edit:1

⚠️  STRUGGLE INDICATORS:
   - No file changes in 3 iterations
   💡 Consider using: ralph --add-context "your hint here"

Guide a struggling agent without stopping the loop:

# In another terminal while loop is running
ralph --add-context "The bug is in utils/parser.ts line 42"
ralph --add-context "Try using the singleton pattern for config"

Context is automatically consumed after one iteration.

This package is CLI-only. If OpenCode tries to load a ralph-wiggum or open-ralph-wiggum plugin, remove it from your OpenCode plugin list (opencode.json), or run:

ralph "Your task" --no-plugins

If you see ProviderModelNotFoundError or "Provider returned error", you need to configure a default model:

For OpenCode:

1. Edit ~/.config/opencode/opencode.json:

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

2. Or use the --model flag: ralph "task" --model provider/model

For other agents: Use the --model flag to specify the model explicitly.

Ralph automatically tries .cmd extensions on Windows. If you still have issues:

1. Set the full path using environment variables:

   $env:RALPH_OPENCODE_BINARY = "C:\path\to\opencode.cmd"

2. Or add the CLI to your PATH

Install Bun: https://bun.sh

❌ Bad:

Build a todo API

✅ Good:

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

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

❌ Bad:

Make the code better

✅ Good:

Refactor auth.ts to:
1. Extract validation into separate functions
2. Add error handling for network failures
3. Ensure all existing tests still pass

Output <promise>DONE</promise> when refactored and tests pass.

# Safety net for runaway loops
ralph "Your task" --max-iterations 20

Ralph treats prompt files as plain text, so any format works. For best results, use a concise PRD with:

• 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: include <promise>COMPLETE</promise> (or match your --completion-promise)

Example (Markdown):

# PRD: Add Export Button

## Goal
Let users export reports as CSV from the dashboard.

## Scope
- In: export current report view
- Out: background exports, scheduling

## Requirements
1. Add "Export CSV" button to dashboard header.
2. CSV includes columns: date, revenue, sessions.
3. Works for reports up to 10k rows.

## Constraints
- Keep current UI styling.
- Use existing CSV utility in utils/csv.ts.

## Acceptance Criteria
- Clicking button downloads a valid CSV.
- CSV opens cleanly in Excel/Sheets.
- All existing tests pass.

## Completion Promise
<promise>COMPLETE</promise>

For larger projects, a structured JSON feature list works better than prose. Based on Anthropic's research on effective agent harnesses, JSON format reduces the chance of agents inappropriately modifying test definitions.

Create a features.json file:

{
  "features": [
    {
      "category": "functional",
      "description": "Export button downloads CSV with current report data",
      "steps": [
        "Navigate to dashboard",
        "Click 'Export CSV' button",
        "Verify CSV file downloads",
        "Open CSV and verify columns: date, revenue, sessions",
        "Verify data matches displayed report"
      ],
      "passes": false
    },
    {
      "category": "functional",
      "description": "Export handles large reports up to 10k rows",
      "steps": [
        "Load report with 10,000 rows",
        "Click 'Export CSV' button",
        "Verify export completes without timeout",
        "Verify all rows present in CSV"
      ],
      "passes": false
    },
    {
      "category": "ui",
      "description": "Export button matches existing dashboard styling",
      "steps": [
        "Navigate to dashboard",
        "Verify button uses existing button component",
        "Verify button placement in header area"
      ],
      "passes": false
    }
  ]
}

Then reference it in your prompt:

Read features.json for the feature list. Work through each feature one at a time.
After verifying a feature works end-to-end, update its "passes" field to true.
Do NOT modify the description or steps - only change the passes boolean.
Output <promise>COMPLETE</promise> when all features pass.

Why JSON? Agents are less likely to inappropriately modify JSON test definitions compared to Markdown. The structured format keeps agents focused on implementation rather than redefining success criteria.

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

┌─────────────────────────────────────────────────────────────┐
│                                                             │
│   ┌──────────┐    same prompt    ┌──────────┐              │
│   │          │ ───────────────▶  │          │              │
│   │  ralph   │                   │ AI Agent │              │
│   │   CLI    │ ◀─────────────── │          │              │
│   │          │   output + files  │          │              │
│   └──────────┘                   └──────────┘              │
│        │                              │                     │
│        │ check for                    │ modify              │
│        │ <promise>                    │ files               │
│        ▼                              ▼                     │
│   ┌──────────┐                   ┌──────────┐              │
│   │ Complete │                   │   Git    │              │
│   │   or     │                   │  Repo    │              │
│   │  Retry   │                   │ (state)  │              │
│   └──────────┘                   └──────────┘              │
│                                                             │
└─────────────────────────────────────────────────────────────┘

1. Ralph sends your prompt to the selected agent
2. The agent works on the task, modifies files
3. Ralph checks output for completion promise
4. If not found, repeat with same prompt
5. AI sees previous work in files
6. Loop until success or max iterations

ralph-wiggum/
├── bin/ralph.js                  # CLI entrypoint (npm wrapper)
├── ralph.ts                      # Main loop implementation
├── package.json                  # Package config
├── install.sh / install.ps1     # Installation scripts
└── uninstall.sh / uninstall.ps1 # Uninstallation scripts

During operation, Ralph stores state in .ralph/:

• ralph-loop.state.json - Active loop state
• ralph-history.json - Iteration history and metrics
• ralph-context.md - Pending context for next iteration
• ralph-tasks.md - Task list for Tasks Mode (created when --tasks is used)
• ralph-questions.json - Pending user answers to agent questions

npm uninstall -g @th0rgal/ralph-wiggum

npm uninstall -g @th0rgal/ralph-wiggum

Claude Code is Anthropic's official CLI for Claude. Use it with Open Ralph Wiggum for powerful autonomous coding:

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

Codex is OpenAI's code-specialized agent. Perfect for code generation and refactoring tasks:

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

OpenCode is an open-source AI coding assistant. It's the default agent:

ralph "Fix all TypeScript errors" --max-iterations 10

Copilot CLI is GitHub's agentic coding tool (public preview). It requires a GitHub Copilot subscription and authentication via GH_TOKEN, GITHUB_TOKEN, or prior copilot /login.

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
• --allow-all (default) maps to --allow-all + --no-ask-user in Copilot CLI
• --no-plugins has no effect with Copilot CLI
• Authentication: set GH_TOKEN / GITHUB_TOKEN env var, or run copilot /login first

Qwen Code is Alibaba's headless AI coding CLI agent based on the Qwen model family. It uses the same stream-JSON output format as Claude 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:

• The CLI binary is qwen; override path with RALPH_QWEN_CODE_BINARY env var
• --allow-all (default) maps to --yolo in Qwen Code CLI (auto-approves all actions)
• --no-plugins has no effect with Qwen Code
• Headless mode uses --output-format stream-json --include-partial-messages
• See headless docs for details

Cursor Agent is Cursor's headless CLI agent. It works with any model available through your Cursor subscription.

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 (no Cursor window), set the CURSOR_API_KEY environment variable
• The binary is cursor-agent; override with RALPH_CURSOR_AGENT_BINARY env var

Agent rotation lets you cycle through different agent/model combinations across iterations. This is useful for leveraging the strengths of different models or comparing their performance on a task.

Each rotation entry uses the agent:model format:

--rotation "agent1:model1,agent2:model2,agent3:model3"

Valid agents: opencode, claude-code, codex, copilot, cursor-agent, qwen-code

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

# Cycle through three different configurations
ralph "Refactor the auth module" \
  --rotation "opencode:claude-sonnet-4,claude-code:claude-sonnet-4,codex:gpt-5-codex" \
  --max-iterations 15

# Include Copilot in the rotation
ralph "Build a REST API" \
  --rotation "opencode:claude-sonnet-4,copilot:claude-sonnet-4" \
  --max-iterations 10

When --rotation is used, the --agent and --model flags are ignored. The rotation list takes precedence for agent/model selection.

The rotation cycles back to the first entry after reaching the end:

• Iteration 1 → Entry 1
• Iteration 2 → Entry 2
• Iteration 3 → Entry 1 (wraps around for a 2-entry rotation)
• ...and so on

Invalid rotation entries produce clear error messages:

Invalid agent name:

Error: Invalid agent 'invalid' in rotation entry 'invalid:model'. Valid agents: opencode, claude-code, codex, copilot, cursor-agent, qwen-code

Malformed entry (missing colon):

Error: Invalid rotation entry 'opencode-model'. Expected format: agent:model

Empty values:

Error: Invalid rotation entry 'opencode:'. Both agent and model are required.

When using --status with an active rotation, the output shows all rotation entries and marks the current one:

🔄 ACTIVE LOOP
   Iteration:    3 / 10
   Prompt:       Build a REST API...

   Rotation (position 1/2):
   1. opencode:claude-sonnet-4  **ACTIVE**
   2. claude-code:claude-sonnet-4

The --status command shows which agent and model was used for each iteration:

📊 HISTORY (3 iterations)
   Total time:   5m 23s

   Recent iterations:
   #1  2m 10s  opencode / claude-sonnet-4  Bash(5) Write(3) Read(2)
   #2  1m 45s  claude-code / claude-sonnet-4  Edit(4) Bash(3) Read(2)
   #3  1m 28s  opencode / claude-sonnet-4  Bash(2) Edit(1)

• Original Ralph Wiggum technique by Geoffrey Huntley
• Ralph Orchestrator

Check out 🏝️ sandboxed.sh — a dashboard for orchestrating AI agents with workspace management, real-time monitoring, and multi-agent workflows.

MIT