AstroResearch/docs/architecture/agent/overview.md

# Agent 架构概览

AstroResearch 内置了一个基于 **ReAct** (Thought → Action → Observation) 范式的科研智能体引擎 (`src/agent/`)，参考 Claude Code 的分层设计。以下对各子系统的架构、数据流和内部逻辑进行完整说明。

### 整体架构

```mermaid
graph TD
    subgraph API["API 层"]
        SSE["SSE /api/chat/agent"]
        Sessions["Session CRUD"]
        Metrics["GET /api/chat/metrics"]
        Audit["GET /api/chat/sessions/:id/audit"]
        AskUser["问答 /api/chat/questions + /api/chat/answer"]
    end

    subgraph Runtime["AgentRuntime — ReAct 引擎"]
        RunTurn["run_turn() 主入口"]
        SP["SystemPrompt 组装器"]
        CtxBuild["Context Builder 上下文构建"]
        ReAct["ReAct 主循环"]
        Streaming["streaming.rs 流式处理"]
        Executor["executor.rs 并行执行"]
        Finalize["finalize.rs 会话收尾"]
        TokenBudget["token_budget.rs"]
        CircuitBreaker["circuit_breaker.rs"]
    end

    subgraph Tools["工具系统 (tools/)"]
        AgentTool["AgentTool trait"]
        Registry["ToolRegistry"]
        FS["filesystem/ (6 工具)"]
        Astro["astro/ (7 工具)"]
        AskUserT["ask_user"]
        SubAgentT["subagent/delegate_research"]
        TeamT["team/ (4 工具)"]
        BG["background (2 工具)"]
    end

    subgraph CrossCutting["横切关注点"]
        Hooks["HookRegistry (9 事件)"]
        Skills["SkillRegistry (两层加载)"]
        Memory["MemoryManager (项目记忆)"]
        Permission["PermissionChecker"]
        FileCache["FileStateCache (Read 去重)"]
    end

    subgraph DB["持久化"]
        AgentSessions["agent_sessions"]
        AgentMessages["agent_messages"]
        AgentTasks["agent_tasks"]
        AgentAudit["agent_audit_log"]
    end

    SSE --> RunTurn
    RunTurn --> SP
    RunTurn --> CtxBuild --> DB
    RunTurn --> ReAct
    ReAct --> Streaming --> Tools
    ReAct --> Executor --> Tools
    ReAct --> TokenBudget
    ReAct --> CircuitBreaker
    ReAct --> Finalize

    Hooks -.-> ReAct
    Hooks -.-> Tools
    Skills -.-> Tools
    Memory -.-> SP
    Permission -.-> Tools
    FileCache -.-> Tools
```

---

### ReAct 运行循环 (`runtime/`)

主循环由 `AgentRuntime::run_turn()` 驱动，分为 4 个阶段：

#### 完整生命周期

```mermaid
sequenceDiagram
    participant FE as 前端 SSE
    participant RT as AgentRuntime
    participant DB as SQLite
    participant LLM as LLM API
    participant Tools as ToolRegistry

    FE->>RT: POST /api/chat/agent { question, session_id? }

    Note over RT: Phase 1 — 会话管理
    RT->>DB: create_or_resume_session()
    alt 新会话
        DB-->>RT: session_id = uuid, turn_index = 0
    else 恢复会话
        DB-->>RT: 验证存在 + 计算 turn_index
    end
    RT->>RT: 触发 OnSessionStart hook
    RT-->>FE: SSE session { session_id, title }

    Note over RT: Phase 2 — 上下文构建
    RT->>RT: build_initial_context()
    RT->>RT: ① 组装 SystemPrompt (静态 section + 记忆注入)
    RT->>DB: ② 加载历史消息 load_history_for_llm()
    RT->>DB: ③ 恢复未完成任务 (agent_tasks)
    RT->>RT: ④ 检查压缩/清理上下文
    RT->>DB: ⑤ 保存用户消息
    RT->>RT: ⑥ 运行 PreToolUse hooks 过滤

    Note over RT: Phase 3 — ReAct 循环
    loop 每步迭代 (step ≤ max_steps)
        RT->>LLM: chat_stream(messages + tool_defs)
        LLM-->>RT: ReasoningDelta / TextDelta / ToolCallsComplete
        RT-->>FE: SSE thought / text_delta / tool_call

        alt 无工具调用 → 最终答案
            RT->>DB: 保存 assistant 消息
            RT-->>FE: SSE text_delta → usage → done
            Note over RT: break 循环
        else 有工具调用
            RT->>RT: 检查 token 预算 + 熔断器
            RT->>RT: validate_and_prepare() — 去重 + 过滤
            RT->>Tools: execute_parallel() — 并行执行
            Tools-->>RT: (tool_call_id, name, output)
            RT-->>FE: SSE tool_result { tool_call_id, name, output }
            RT->>DB: 保存 tool 消息 + 审计日志
            RT->>RT: 运行 PostToolUse hooks
            RT->>RT: 检测压缩需求 (auto_compact)
        end
        RT->>RT: 检测循环终止条件
    end

    Note over RT: Phase 4 — 会话收尾
    RT->>DB: 更新 turn_count + updated_at
    RT->>DB: calculate_and_persist_metrics()
    RT->>RT: 运行 OnSessionStop hook
    RT-->>FE: SSE done
```

#### 并行工具执行模型

```mermaid
sequenceDiagram
    participant ReAct as ReAct 循环
    participant Val as validate_and_prepare
    participant Exec as execute_parallel
    participant T1 as Tool A
    participant T2 as Tool B
    participant FE as 前端 SSE

    ReAct->>Val: LLM 返回 [tool_call_a, tool_call_b]
    Val->>Val: 去重检测 + 权限验证
    Val-->>ReAct: prepared_calls[] + has_duplicate 标志
    ReAct->>FE: 发送 tool_call SSE (逐一)
    ReAct->>Exec: 启动 execute_parallel()
    par 并行执行
        Exec->>T1: tool_a.execute(args_a)
        Exec->>T2: tool_b.execute(args_b)
    end
    T1-->>Exec: ToolOutput { content, is_error }
    Exec-->>FE: SSE tool_result (立即推送)
    T2-->>Exec: ToolOutput { content, is_error }
    Exec-->>FE: SSE tool_result (立即推送)
    Exec-->>ReAct: Vec<(tool_call_id, name, args, output, cancelled)>
    ReAct->>ReAct: PostToolUse hooks + 审计日志 + 持久化
```

- **并发上限**：由 `max_concurrent_tools` 环境变量控制，默认不限制
- **Sibling Abort**：仅 `causes_sibling_abort() = true` 的工具（如 `download_paper`）能在出错时中断兄弟任务
- **InterruptBehavior**：`Block` 工具（如 `ask_user`）不可被用户取消；`Cancel` 工具可在取消信号时中断
- **超时控制**：每个工具有独立超时，默认 120s（`AGENT_TOOL_TIMEOUT_SECS`）

---