技能管家 · Skill Curator

Xiaomi MiMo（小米 MiMo）/ DeepSeek（深度求索）/ Qwen（通义千问）/ Gemini / OpenAI 多模型驱动的 AI 技能（skills）管理器。

一个 VS Code 扩展：自动扫描你本机所有 coding agent（Claude Code、Codex、GitHub Copilot、Gemini、OpenCode）下的 skills（技能），调用你配置的大模型做质量诊断（分级 + 跨技能重叠检测），并以只增不删的软链接方式把技能同步到所有 agent。可在多家模型之间一键切换：MiMo、DeepSeek、Qwen（通义千问）、Gemini、OpenAI。

• 🔍 一处总览所有 agent 的全部技能
• 🩺 大模型诊断：用 MiMo / DeepSeek / Qwen / Gemini / OpenAI 评估技能描述与触发词、检测重复重叠
• 🔁 跨 agent 同步技能（只增不删，先 dry-run 预览再确认）
• 🤝 一键把技能交给右侧的 Copilot / Claude Code 去优化

关键词 / keywords：技能管理、skill manager、MiMo、Xiaomi MiMo、小米 MiMo、小米大模型、DeepSeek、深度求索、Qwen、通义千问、Gemini、OpenAI、ChatGPT、Claude Code、Codex、GitHub Copilot、OpenCode、Kimi、大模型、LLM、AI agent、SKILL.md。

诊断用的是你自己配置的大模型 API（OpenAI 兼容端点），API key 只存在本机，不随扩展发布。

Skill Curator (English)

Inventory, LLM-diagnose, and add-only sync the skills installed for your coding agents — all from one VS Code sidebar. Built to manage a large, multi-agent skill library and then hand individual skills off to Copilot / Claude Code for editing.

Features

• Inventory — scans skills across Claude, Codex, Copilot, Gemini, and OpenCode (user dirs plus Claude/Codex plugin layers), grouped in a sidebar tree.
• LLM diagnosis — runs your configured OpenAI-compatible model over each skill and tags it attention / improve / good with a short note, plus a whole-library overlap pass that flags duplicate or trigger-conflicting skills.
• Static health checks — broken symlinks, missing/unparseable SKILL.md, name↔directory mismatch, cross-surface drift, stale plugin cache. These work without any LLM configured.
• Add-only sync — mirrors user-level skills into every enabled agent via symlinks. It shows a dry-run preview and asks for confirmation, never overwrites or deletes, and only ever creates missing links.
• Hand-off — open a skill's SKILL.md, or copy a diagnosis-laden prompt and send it to your Copilot / Claude Code chat to do the actual editing.

Requirements

• VS Code 1.90+
• For diagnosis: one API key for a built-in or custom OpenAI-compatible / Anthropic-compatible chat endpoint. Inventory and static health checks work without it.

Configuration

The API key is not a setting. Store it with the "Skill Curator: Set LLM API Key" command (kept in VS Code SecretStorage), or point skillCurator.llm.apiKeyEnv at an environment variable.

If skillCurator.llm.providers is empty, the settings panel shows a built-in catalog:

The extension auto-routes /anthropic endpoints to /v1/messages; OpenAI and Gemini use /chat/completions.

Example custom provider with selectable models:

{
  "skillCurator.llm.providers": [
    {
      "id": "deepseek-pro",
      "label": "DeepSeek Pro",
      "baseUrl": "https://api.deepseek.com/anthropic",
      "model": "deepseek-v4-pro",
      "models": ["deepseek-v4-pro", "deepseek-v4-flash"],
      "apiKeyEnv": "DEEPSEEK_API_KEY"
    }
  ],
  "skillCurator.llm.activeProvider": "deepseek-pro"
}

Commands

• Skill Curator: Refresh — re-scan skills
• Skill Curator: Diagnose All Skills — run LLM diagnosis + overlap
• Skill Curator: Diagnose This Skill — diagnose one (picks from a list)
• Skill Curator: Sync Skills to All Agents — dry-run preview, then add-only sync
• Skill Curator: Set LLM API Key — store the key in SecretStorage
• Skill Curator: Select Diagnosis Provider — switch the endpoint/key used for diagnosis
• Skill Curator: Select Diagnosis Model — switch the model under the active provider, including custom names

Privacy

Diagnosis sends each skill's name, description, trigger phrases, and a truncated body excerpt to the LLM endpoint you configure. Nothing is sent anywhere else. Sync only touches your local filesystem.

Development

npm install
npm run build      # esbuild bundle → dist/extension.js
npm test           # vitest (core / llm / sync)
npm run typecheck  # tsc --noEmit

Press F5 to launch the Extension Development Host.

License

MIT