Eino Skills - Claude Skills 集成方案

基于 Anthropic Agent Skills 和 deepagents-cli 的设计，为 cloudwego/eino 框架实现 Skills 支持。

核心概念

什么是 Skills？

Skills 是包含 SKILL.md 文件的文件夹，提供： - 渐进式披露 (Progressive Disclosure)：只在需要时加载完整指令 - Token 效率：启动时仅加载元数据（name + description） - 认知负担降低：Agent 使用少量原子工具 + 按需加载的技能指令

SKILL.md 结构

---
name: skill-name
description: Brief description of what this skill does and when to use it
---

# Skill Name

## Instructions
[具体操作指令]

## Examples
[使用示例]

架构设计

graph TD
    User([User]) -->|Input| Middleware[Skills Middleware]

    subgraph "Eino App Runtime"
        Middleware -->|1. Inject System Prompt| Agent[ReAct Agent]

        subgraph "Skill Core"
            Registry[Skill Registry]
            Loader[Skill Loader]
        end

        subgraph "Tools Layer"
            List[list_skills]
            View[view_skill]
            Terminal[run_terminal_command]
        end

        Agent <-->|2. Tool Calls| List & View & Terminal
        List & View -->|Query| Registry
        Registry -->|Manage| Loader
    end

    subgraph "External Environment"
        FileSystem[Skill Directory

~/.claude/skills]
        OS[Operating System]
    end

    Loader -->|Load SKILL.md| FileSystem
    Terminal -->|Execute Scripts| OS
    OS -.->|Runs| FileSystem

执行流程

sequenceDiagram
    participant User
    participant Agent
    participant Tools
    participant System

    Note over Agent: System Prompt 包含 Skill 列表摘要

    User->>Agent: "帮我提交代码"

    rect rgb(240, 248, 255)
    Note right of Agent: 1. Discovery (发现)
    Agent->>Tools: list_skills()
    Tools-->>Agent: 返回可用 Skills 列表
    end

    rect rgb(255, 250, 240)
    Note right of Agent: 2. Loading (加载)
    Agent->>Tools: view_skill("git-commit")
    Tools-->>Agent: 返回详细通过 SKILL.md 定义的指令
    end

    rect rgb(240, 255, 240)
    Note right of Agent: 3. Execution (执行)
    Note right of Agent: 遵循指令: Analysis -> Execution
    Agent->>System: run_terminal_command("python3 ... --analyze")
    System-->>Agent: Output / Error

    opt Error Recovery (错误自愈)
        Agent->>System: Retry with absolute path / Fix context
    end

    Agent->>System: run_terminal_command("python3 ... -m 'feat: ...'")
    System-->>Agent: Success
    end

    Agent->>User: "代码提交完成！"

项目结构

eino-skills/
├── README.md
├── go.mod
├── pkg/
│   ├── skill/
│   │   ├── types.go            # Skill 类型定义
│   │   ├── loader.go           # Skills 加载器
│   │   ├── parser.go           # SKILL.md 解析器
│   │   └── registry.go         # Skills 注册中心
│   ├── tools/
│   │   ├── skills.go           # 工具包入口
│   │   ├── view_skill.go       # view_skill Tool
│   │   ├── list_skills.go      # list_skills Tool
│   │   └── run_terminal_command.go # 终端执行工具
│   └── middleware/
│       └── skills.go           # Skills 中间件
├── cmd/
│   ├── example/
│   │   └── main.go             # 完整 Agent 示例
│   └── eino-skills-cli/
│       └── main.go             # CLI 管理工具

快速开始

package main

import (
    "context"
    "fmt"
    "os"

    // Eino 核心组件
    "github.com/cloudwego/eino-ext/components/model/claude"
    "github.com/cloudwego/eino/compose"
    "github.com/cloudwego/eino/flow/agent"
    "github.com/cloudwego/eino/flow/agent/react"
    "github.com/cloudwego/eino/schema"

    // Skills 相关包
    skillsmw "github.com/dyike/eino-skills/pkg/middleware"
    skillpkg "github.com/dyike/eino-skills/pkg/skill"
    skilltools "github.com/dyike/eino-skills/pkg/tools"
)

func main() {
    ctx := context.Background()

    // 1. 初始化 Skills (加载器 & 注册表)
    loader := skillpkg.NewLoader(
        skillpkg.WithGlobalSkillsDir("~/.claude/skills"), // 指向实际的 skills 目录
    )

    registry := skillpkg.NewRegistry(loader)
    if err := registry.Initialize(ctx); err != nil {
        panic(err)
    }

    // 2. 创建 Skills 中间件
    skillsMiddleware := skillsmw.NewSkillsMiddleware(registry)

    // 3. 准备 Tools (基础 Skill 工具 + 终端执行能力的工具)
    tools := skilltools.NewSkillTools(registry) // 包含 list_skills, view_skill
    cwd, _ := os.Getwd()
    tools = append(tools, skilltools.NewRunTerminalCommandTool(cwd))

    // 4. 配置 Chat Model (以 Claude 为例)
    baseURL := "http://127.0.0.1:8045" // 你的 Proxy 地址
    chatModel, err := claude.NewChatModel(ctx, &claude.Config{
        Model:     "claude-3-5-sonnet-20240620",
        APIKey:    os.Getenv("ANTHROPIC_API_KEY"), // 建议从环境变量获取
        BaseURL:   &baseURL,
        MaxTokens: 4096,
    })
    if err != nil {
        panic(err)
    }

    // 5. 注入 System Prompt (包含 Skills 使用规范)
    basePrompt := `You are a helpful AI assistant with access to specialized skills.

1. Use 'list_skills' to see what's available.
2. Use 'view_skill' to load detailed instructions.
3. Follow the skill's workflow strictly step-by-step.
4. Execute commands directly (no pre-checks like 'ls').
5. If a script fails, try to fix paths or fall back to native commands.`

    systemPrompt := skillsMiddleware.InjectPrompt(basePrompt)

    // 6. 创建 Agent
    myAgent, err := react.NewAgent(ctx, &react.AgentConfig{
        ToolCallingModel: chatModel,
        ToolsConfig: compose.ToolsNodeConfig{
            Tools: tools,
        },
        MaxStep: 50, // 增加步数限制以支持多步骤 Skill
    })
    if err != nil {
        panic(err)
    }

    // 7. 运行 Agent
    // 实际使用建议使用 Stream 模式
    resp, err := myAgent.Generate(ctx, []*schema.Message{
        {Role: schema.System, Content: systemPrompt},
        {Role: schema.User, Content: "帮我写一个 git commit message"},
    })

    if err != nil {
        fmt.Println("Error:", err)
        return
    }

    fmt.Println(resp.Content)
}

Feature Roadmap