iAs/docs/TOOL_PLATFORM_PLAN_IMPL.md

实现细节

2024-06-18 当前状态

---

关键文件清单

src/
├── daemon.rs          → 主控 (LLM/Tool/Send Consumer, Registry, execute_builtin)
├── main.rs            → CLI (cmd_tool_dynamic, cmd_tools)
├── cli.rs             → clap 命令定义
├── logger.rs          → 三路日志初始化
├── tools/
│   ├── mod.rs         → 模块声明 (6 个)
│   ├── spec.rs        → ToolSpec + ParamSpec
│   ├── parser.rs      → parse_dir() YAML 解析 (4 tests)
│   ├── registry.rs    → Registry (load/list/dispatch)
│   ├── executor.rs    → execute_loop (spawn→stdout→http/db/result)
│   └── approval.rs    → ApprovalManager (6位确认码)
├── llm/
│   ├── types.rs       → Message, ToolCall, StreamChunk, ConversationConfig
│   ├── conversation.rs → Conversation (run_tool_loop, summarize)
│   ├── deepseek.rs    → DeepSeek provider (tool_call 流式解析)
│   └── provider.rs    → LlmProvider trait
└── context/
    ├── types.rs       → ChatSession, add_tool_result
    └── builder.rs     → build_context (system prompt + 摘要 + 消息)

---

工具调用全链路

用户消息
  → LLM Consumer: 创建 Conversation, 注册 2 元工具
  → LLM: 调 query_capabilities (tool_call)
  → ToolExecutor → PipelineMessage::ToolCall → 消息队列
  → ToolConsumer → execute_builtin → query_capabilities → 返回工具列表
  → ToolResult → LLM Consumer → resume

  → LLM: 调 call_capability(name="weather_now", prompt='{"location":"北京"}')
  → ToolConsumer → execute_builtin → call_capability 解包
    → registry.dispatch("weather_now", {"location":"北京"})
    → executor::execute_loop
      → spawn weather --command now
      → stdout: {"type":"http","url":"...","desc":"..."}
      → 审批(如需要) → reqwest::get → response
      → spawn weather --command now (带 response)
      → stdout: {"type":"result","content":"北京晴，25°C"}
    → ToolResult
  → LLM Consumer → resume → 生成回复
  → Send Consumer → 微信

---

executor::execute_loop 状态机

spawn tool
  → stdout JSON parse
    ├─ {"type":"result","content":"..."} → 返回 content
    ├─ {"type":"http","method":"GET","url":"..."}
    │   → 审批(high risk) → HTTP → 存入 response
    │   → re-spawn (带 response 字段)
    └─ {"type":"db","operation":"read_memories",...}
        → 审批 → MemoryStore 操作 → 存入 db_response
        → re-spawn (带 db_response 字段)

---

消息队列路由

MessageKind::UserMessage | ToolResult | ScheduledTask → LLM Consumer
MessageKind::ToolCall | ApprovalRequest             → Tool Consumer
MessageKind::LLMReply                              → Send Consumer

---

System Prompt

你是 iAs 智能助手，通过微信与用户交流。回复简洁自然。

你可以通过 function calling 调用两个元工具：
1. query_capabilities — 查询所有可用能力工具
2. call_capability — 调用能力工具，参数 {"name":"工具名","prompt":"{\"参数\":\"值\"}"}

流程：收到消息 → query_capabilities → call_capability → 回复