跳到主要内容
    ↑↓ 选择↵ 打开esc 关闭
    中文English
    charfeng1

    Ralph Loop

    v1.0.10其他
    opencode-ralph-loop

    Minimal Ralph Loop plugin for opencode - auto-continues until task completion

    GitHub 星标

    105

    近 30 天 +1

    月装机量

    994

    近 7 天 187

    综合评分SCORE

    52.1

    生态多维模型

    最近提交

    2 个月前

    2026-05-25

    快速安装与配置

    opencode.json

    写入当前项目的 opencode.json,只对这个仓库生效。

    opencode.json

    {
      "$schema": "https://opencode.ai/config.json",
      "plugin": ["opencode-ralph-loop@1.0.10"]
    }

    opencode 启动时会通过内嵌运行时自动加载 npm 依赖并缓存至本地目录,无需手动在全局环境执行安装。

    CI npm version

    Minimal Ralph Loop plugin for opencode - auto-continues until task completion.

    Inspired by Anthropic's Ralph Wiggum technique for iterative, self-referential AI development loops.

    Why this plugin?

    oh-my-opencode is a fantastic, feature-rich plugin that includes Ralph Loop along with many other powerful capabilities like the Sisyphus orchestrator, background agents, and more.

    However, we personally found the full suite a bit heavy for our workflow. We also noticed others in the community expressing interest in specific features without needing the complete package. So we extracted just the Ralph Loop functionality into this standalone, lightweight plugin.

    If you want the full-featured experience, definitely check out oh-my-opencode. If you just want auto-continuation loops with minimal overhead, this plugin is for you.

    Installation

    Add to your ~/.config/opencode/opencode.json:

    {
      "plugin": ["opencode-ralph-loop"]
    }
    

    Restart opencode. That's it!

    On first run, the plugin will automatically install skills and commands to your ~/.config/opencode/ directory.

    Usage

    Start a loop

    /ralph-loop "Build a REST API with authentication"
    

    The AI will work on your task and automatically continue until completion.

    Cancel a loop

    /cancel-ralph
    

    Get help

    /help
    

    How it works

    1. /ralph-loop creates a state file at .opencode/ralph-loop.local.md
    2. When the AI goes idle, the plugin checks if <promise>DONE</promise> was output
    3. If not found, it injects "Continue from where you left off"
    4. Loop continues until DONE is found or max iterations (100) reached
    5. State file is deleted when complete

    Completion Promise

    When the AI finishes a task, it outputs:

    <promise>DONE</promise>
    

    Important: The AI should ONLY output this when the task is COMPLETELY and VERIFIABLY finished. False promises are not allowed.

    State File

    The loop state is stored in your project directory:

    .opencode/ralph-loop.local.md
    

    Format (markdown with YAML frontmatter):

    ---
    active: true
    iteration: 3
    maxIterations: 100
    sessionId: ses_abc123
    ---
    
    Your original task prompt
    

    Add .opencode/ralph-loop.local.md to your .gitignore.

    Features

    • Plug-and-play: Just add to config and restart - no manual setup
    • Auto-setup: Skills and commands are automatically installed on first run
    • Minimal: ~300 lines, no bloat
    • Project-relative: State file in .opencode/, not global
    • Completion detection: Scans session messages for DONE promise
    • Progressive context: Skills provide context only when needed
    • Commands: /ralph-loop, /cancel-ralph, and /help

    Architecture

    Following Anthropic's Claude Code plugin pattern:

    opencode-ralph-loop/
    ├── src/
    │   └── index.ts        # Main plugin with event hooks and tools
    ├── skills/
    │   ├── ralph-loop/     # Progressive context for starting loops
    │   ├── cancel-ralph/   # Context for cancellation
    │   └── help/           # Plugin documentation
    ├── commands/
    │   ├── ralph-loop.md   # Slash command for starting
    │   ├── cancel-ralph.md # Slash command for cancelling
    │   └── help.md         # Slash command for help
    └── package.json
    

    Contributing

    PRs welcome. See CONTRIBUTING.md.

    Releases

    Auto-published from CI with npm provenance via OIDC trusted publishing — each tarball is cryptographically linked to the exact commit, no long-lived tokens involved.

    Credits

    License

    MIT