�� AI Agent 学习笔记(Week3下)- SKILL是什么 & Function Calling / MCP / CLI / Skill 串讲

# skill.md
name: weather查询
description: 查询城市天气
trigger: 当用户询问天气时使用
steps:
  1. 解析城市名
  2. 调用 weather_api 获取数据
  3. 格式化回复

飞书消息 → Gateway → agent-runner.runAgentTurnWithFallback()

源码位置：agent-runner.runtime-*.js:2944

const agentTool = {{
    name: safeToolName,           // 工具名
    label: tool.title,           // 显示名
    description: tool.description, // 简短描述（~100 token）
    parameters: tool.inputSchema,
    execute: async (_toolCallId, input) => {{
        const result = await params.runtime.callTool(...);
        return toAgentToolResult({{...}});
    }}
}};

response = await llm.chat.completions.create({{
    messages: [...systemPrompt, ...conversation],
    tools: toolDefinitions,  // MCP tools + built-in tools
    tool_choice: "auto"
}});

if (response.choices[0].message.tool_calls) {{
    const toolCall = response.choices[0].message.tool_calls[0];
    // {{ id: "call_xxx", name: "mcp_github__list_issues", arguments: {{...}} }}

async callTool(serverName, toolName, input) {{
    const session = getMcpSession(serverName);
    return await session.client.callTool({{
        name: toolName,
        arguments: input
    }});
}}

execute: async (_toolCallId, args, signal, onUpdate) => {{
    const child = spawn(preparedSpawn.command, preparedSpawn.args, {{
        detached: process.platform !== "win32",
        shell: false,
    }});
    // 处理输出流
}}

second_response = await llm.chat.completions.create({{
    messages: [
        {{"role": "user", "content": "用户问题"}},
        assistant_msg_with_tool_call,
        {{"role": "tool", "tool_call_id": tool_call.id, "content": result}}
    ]
}});

const child = spawn(command, args, {{
    detached: process.platform !== "win32",
    shell: false,
    windowsHide: process.platform === "win32"
}});

const transport = new StreamableHTTPClientTransport({{ url }});
const client = new Client({{ name, version }}, {{ timeout: 30000 }});
await client.connect(transport);

function createLazyFacadeProxyValue(params) {{
    const resolve = createLazyFacadeValueLoader(params.load);
    return new Proxy(params.target, {{
        get(_target, property, receiver) {{
            return Reflect.get(resolve(), property, receiver);
        }}
    }});
}}

1. 飞书 → Gateway → agent-runner
         ↓
2. 组装 System Prompt（含 github skill 简短描述）
         ↓
3. LLM 判断需要 github 工具 → 返回 tool_calls
         ↓
4. bundle-mcp 调用 MCP Server
   → JSON-RPC: {{"method": "tools/call", ...}}
         ↓
5. MCP Server（gh CLI）执行
         ↓
6. 结果返回 Agent Runner
         ↓
7. LLM 生成最终回复
         ↓
8. Gateway → 飞书用户