Noureddine RAMDI / open-ralph-wiggum: a self-correcting AI coding agent loop using git state feedback

Created Mon, 04 May 2026 10:23:01 +0000 Modified Sat, 23 May 2026 20:41:27 +0000

Th0rgal/open-ralph-wiggum

Open Ralph Wiggum takes a different approach to autonomous AI coding agents by turning the filesystem and git history into the agent’s working memory. Instead of passing explicit context between iterations, it repeatedly feeds the same prompt to an AI agent, letting the agent see its own previous work through file changes and git commits. This creates a self-correcting loop that continues until the agent signals task completion.

What open-ralph-wiggum does and how it works

Open Ralph Wiggum is a TypeScript CLI tool designed to orchestrate AI coding agents such as Claude Code, Codex, Copilot CLI, Cursor Agent, and OpenCode in an autonomous agentic loop. The core idea is the “Ralph Wiggum technique” where the same prompt is issued repeatedly to an agent until a completion promise is detected in its output.

Under the hood, the tool acts as a shell loop wrapper around the chosen AI coding agent CLI. It runs the agent subprocess, parses its output for specific completion or abort signals, and monitors iteration counts. Between iterations, it relies on git history and filesystem state changes as implicit feedback for the agent — effectively turning the repository’s state into the agent’s short-term memory.

The architecture is minimalistic: the CLI manages the loop and agent invocation, while the AI agents themselves remain separate executables. This design allows multi-agent support via a --agent flag to switch between different AI providers.

Additional features include task tracking mode that helps manage ongoing tasks, live status monitoring to observe the agent’s progress, and mid-loop context injection to add extra information dynamically.

The stack is TypeScript, with the CLI distributed as an npm or Bun global package. It requires the Bun runtime and at least one supported AI agent CLI installed separately.

What sets open-ralph-wiggum apart technically

The standout technical aspect is the use of the filesystem and git commits as the agent’s implicit memory. Since current AI coding agents often lack persistent memory across calls, this repo cleverly circumvents that limitation by using the git history and file changes as a feedback loop.

This means the agent “sees” its previous work by reading the updated files and the git commit log, enabling self-correction without explicit context passing in the prompt. It avoids complex state management or external memory stores.

The tradeoff here is that it depends heavily on git and the file system state, which may not suit all project types or workflows. Also, the loop is bounded by completion promise detection heuristics, which may not always be reliable.

The code is surprisingly clean and focused, mostly orchestrating subprocess calls, monitoring outputs, and managing iteration logic. The multi-agent CLI orchestration is flexible, allowing developers to experiment with different AI providers easily.

The task tracking and live status features add practical value, making it easier to follow what the agent is doing in real time, which is often a pain point in autonomous agent workflows.

Quick start

Installation

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
    • OpenCode — Open-source AI coding assistant
npm install -g @th0rgal/ralph-wiggum

Bun

bun add -g @th0rgal/ralph-wiggum

From source

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.

verdict

Open Ralph Wiggum is a neat experiment in autonomous AI coding loops that solves the persistent memory problem by offloading state to git and the filesystem. This makes it a practical tool for developers exploring multi-agent AI workflows or self-correcting automation where explicit memory passing is challenging.

That said, its reliance on git history means it fits best with projects already using git and comfortable with iterative file commits as communication. The completion detection mechanism may require tuning or might not cover all edge cases.

If you’re working with AI coding agents in CLI form and want to experiment with autonomous loops without building complex memory systems, this repo offers a clean, focused starting point. It’s also a good reference for understanding how to orchestrate multiple AI agents flexibly.

The requirement for Bun runtime and external AI agent CLIs means some setup overhead, but the global install via npm or Bun is straightforward. Overall, it’s worth exploring if you want to test self-correcting agent workflows or build on the agentic loop pattern.

The code is accessible and well-scoped, making it a solid base for further experimentation or customization in autonomous AI coding tasks.


→ GitHub Repo: Th0rgal/open-ralph-wiggum ⭐ 1,647 · TypeScript