What happens when an iteration fails?

If validation fails, Ralph logs the failure details and keeps the story in a pending state, allowing the supervisor to analyze the error and attempt a fix in the next iteration.

How does Ralph prevent context drift in AI sessions?

By starting each implementation iteration with a fresh context window, Ralph avoids the accumulation of irrelevant history, focusing only on the specific task and relevant project metadata.

What is the Ralph Supervisor pattern?

It is an architecture where Claude acts as a supervisor to plan and review work, while a specialized executor model (Codex) performs the actual coding in clean, isolated sessions.

Does Ralph work with my existing test suite?

Yes, Ralph uses a configurable validation command, such as 'npm test' or 'pytest', to ensure that all code meets your project's specific quality standards before committing.

How is progress tracked across multiple sessions?

Progress is tracked through a prd.json file for stories and a progress.txt file that captures semantic learnings, ensuring the agent 'remembers' important implementation details across steps.

Ralph Supervisor Coding Loop

Name: Ralph Supervisor Coding Loop
Author: joelhooks

byjoelhooks

•

566

•

生産性とワークフロー

Implements an autonomous multi-agent coding loop where Claude supervises while Codex executes implementation tasks with validation gates.

Ralph Supervisor is a sophisticated multi-agent coordination pattern designed to prevent context drift and ensure high-quality output in AI-driven development. By separating roles into a Supervisor (Claude) and an Executor (Codex), it manages complex engineering tasks through a series of isolated, fresh-context iterations. The skill utilizes a git-backed persistence layer, mandatory validation gates (like unit tests and type checks), and a structured PRD system to track progress, making it an essential tool for developers building complex features autonomously.

主な機能

01Fresh context per iteration to eliminate model drift and context window bloat

02566 GitHub stars

03Progress carryover mechanism to flow learnings between implementation sessions

04Mandatory validation gates ensuring all tests pass before moving to the next task

05Git-backed persistence that automatically commits successful code implementations

06Granular story tracking via a structured JSON PRD for precise project management

ユースケース

01Automating long-running feature implementations that require sequential logic and multiple steps

02Managing complex refactoring projects that need isolated execution and strict verification

03Implementing Test-Driven Development (TDD) workflows where AI must satisfy specific test criteria

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add joelhooks/swarm-tools ralph-supervisor

For use in Claude.ai and ChatGPT

Download Skill