RFC-0003 Plugin

• Feature Name: plugin
• Start Date: 2026-10-29
• RFC PR: #14
• Updated by: #15, #17, #21

A Plugin is a pluggable component of PenguinHarness that can bundle server modules, Skills, Hooks, and Workbenches. Unlike Benchmarks and Evaluations, a Plugin must be a standard NPM package.

Through this pluggable Plugin architecture, PenguinHarness can adapt to an unlimited variety of RSI algorithms simultaneously. Typically, we implement each RSI algorithm as a standalone Plugin. We built the open-source PenguinHarness Plugins repository to curate all community Plugins.

Design Principle: Decoupling Algorithm from Data

A Plugin implementation should be decoupled from specific datasets: in principle, an RSI algorithm Plugin should contain only the method itself (such as experience distillation, reflection loops, or prompt and skill evolution procedures) and must not be hardcoded or bound to any particular Benchmark dataset. By keeping the algorithm decoupled from data, the same RSI Plugin can be applied directly to any Benchmark conforming to RFC-0001 Benchmark.

Directory Structure

Each Plugin is stored under <plugin-id>/. The file structure of the repository is organized as follows:

penguin-harness-plugins/
└── <plugin-id>/
    ├── skills/
    │   └── <skill-id>/
    │       └── SKILL.md
    ├── hooks/
    │   └── stop.mjs
    ├── workbenches/
    │   └── <workbench-id>/
    │       └── package.json
    ├── icon.svg
    └── package.json

• skills/<skill-id>/SKILL.md: Skill definitions contributed by the plugin (optional; may contain one or more Skills).
• hooks/*.mjs: Lifecycle hook scripts contributed by the plugin (optional; e.g., stop.mjs).
• workbenches/<workbench-id>/: Custom UI workbench packages contributed by the plugin (optional).
• icon.svg: SVG icon file for the plugin (optional).
• package.json: Standard NPM package manifest, augmented with a penguin field that provides display and quick-start metadata in PenguinHarness.

Example: opro-rsi-toolkit/package.json

Adding a Plugin

To add a new Plugin to the repository, organize its files as follows:

1. Create a <plugin-id> directory and organize skills/, hooks/, workbenches/, and an optional icon.svg according to the plugin's capabilities, keeping the algorithm logic decoupled from specific datasets.
2. Create a package.json file inside <plugin-id>/ containing standard NPM fields together with the penguin metadata block:

{
  "name": "@penguinharness/<plugin-id>",
  "version": "0.1.0",
  "description": "English description of the RSI plugin.",
  "license": "Apache-2.0",
  "repository": {
    "type": "git",
    "url": "git+https://github.com/Prism-Shadow/penguin-harness-plugins.git",
    "directory": "<plugin-id>"
  },
  "files": [
    "icon.svg",
    "skills",
    "hooks",
    "workbenches"
  ],
  "publishConfig": {
    "access": "public"
  },
  "penguin": {
    "title": "Plugin Title",
    "title_zh": "插件标题",
    "description": "Detailed English description of the plugin.",
    "description_zh": "插件的详细中文描述。",
    "short_description": "One-line English summary for the plugin card.",
    "short_description_zh": "插件卡片展示的单行中文简介。",
    "category": "rsi",
    "preinstall": false,
    "icon": "icon.svg",
    "references": [
      "https://arxiv.org/abs/...",
      "https://github.com/..."
    ],
    "quick_start": {
      "prompt": "Use the `<skill-id>` Skill to ...",
      "prompt_zh": "使用 `<skill-id>` 技能……",
      "skills": [
        "<skill-id>"
      ]
    }
  }
}

package.json Specification

Standard NPM fields:

• name — NPM package name (e.g., "@penguinharness/<plugin-id>" or "<plugin-id>"). The unscoped name is <plugin-id> and may contain only letters, digits, and hyphens -.
• version — Plugin version in standard x.y.z SemVer format (e.g., "0.1.0").
• description — English description of the NPM package.
• license — Open-source license identifier (e.g., "Apache-2.0" or "MIT").
• repository — Repository metadata object (type, url, directory).
• files — Array of files and directories included when publishing the NPM package (e.g., ["icon.svg", "skills", "hooks", "workbenches"]).
• publishConfig — NPM publishing configuration (typically { "access": "public" }).

penguin extension fields (required):

• penguin.title — English display name shown in PenguinHarness.
• penguin.description — Full English description shown in the plugin detail view.
• penguin.short_description — One-line English summary shown on the plugin card.
• penguin.category — Plugin category; set to "rsi" for RSI algorithm plugins.
• penguin.preinstall — Whether the plugin is preinstalled by default; set to false for community plugins.
• penguin.quick_start — Quick-start configuration object whose role is to provide a ready-to-run invocation example for the algorithm: when a user clicks Quick Start on the plugin in PenguinHarness, it opens a new chat draft with the required Skills pre-selected and an example invocation prompt pre-filled:
  • prompt — Pre-filled English quick-start invocation prompt (required).
  • skills — Array of Skill IDs pre-selected for the quick-start example (required, array of strings).

penguin extension fields (optional):

• penguin.title_zh — Chinese display name shown in the Chinese UI.
• penguin.description_zh — Full Chinese description shown in the plugin detail view.
• penguin.short_description_zh — One-line Chinese summary shown on the plugin card.
• penguin.icon — Relative path to the SVG icon inside the package (e.g., "icon.svg").
• penguin.references — List of original reference links (array of strings) recording the RSI algorithm's original paper URL, upstream code repository, or project page.
• penguin.quick_start.prompt_zh — Pre-filled Chinese quick-start invocation prompt in the Chinese UI.