mikeyobrien/ralph-orchestrator
An improved implementation of the Ralph Wiggum technique for autonomous AI agent orchestration
What it solves
Ralph Orchestrator 提供一个自主任务完成的框架,通过让 AI 代理保持在持续迭代循环中直至达到目标,防止常见的“一次性”失败——AI 助手提供部分或错误答案后就停止;相反,它会迭代并验证工作,直到任务完全完成。
How it works
Ralph 实现了“Ralph Wiggum technique”,使用“帽子系统”,让专门化的角色(如 research、debug、review)通过事件协同执行多步骤任务。它加入了“回压门”——自动检查(如 lint、类型检查、测试),会拒绝不完整或损坏的工作,迫使代理回环修复。系统支持多种 AI 后端(包括 Claude Code、Gemini CLI、Copilot CLI),并可作为 MCP 服务器集成。
Who it’s for
想要使用 AI 代理自动化复杂软件工程任务且不需要持续人工监督的开发者,以及希望通过 Telegram 将人类环路指导集成进来的用户。
Highlights
- Multi-Backend Support: Works with various AI coding assistants like Claude Code, Gemini CLI, and Codex.
- Hat System: Uses specialized roles (e.g.,
code-assist,debug,review) to structure task execution. - Backpressure Gates: Ensures quality by rejecting work that fails tests or linting.
- Human-in-the-Loop: Integration with Telegram allows humans to provide proactive guidance or answer agent questions mid-loop.
- Web Dashboard: An alpha-stage dashboard for monitoring and managing orchestration loops.
- MCP Server Mode: Can be run as a Model Context Protocol server for compatible clients.