hunjing/backend/app/main.py

"""FastAPI 应用入口。

既导出 ASGI 应用实例 ``app``（供 ``uvicorn app.main:app`` 使用），又在作为脚本执行
（``python -m app.main``）时启动 uvicorn。日志系统在模块级初始化，确保无论从哪种
入口启动，uvicorn 横幅、第三方库与应用日志都统一走 ``app.core.logging`` 出口。
"""

from collections.abc import AsyncGenerator, Awaitable, Callable
from contextlib import asynccontextmanager
from time import perf_counter

import httpx
from fastapi import FastAPI, Request, Response
from fastapi.middleware.cors import CORSMiddleware
from sqlalchemy.exc import OperationalError

from app.api.router import api_router
from app.application.tool_runtime.runtime import ToolRuntime
from app.core.config import settings
from app.core.errors import register_exception_handlers, trace_id_var
from app.core.ids import new_trace_id
from app.core.logging import (
    clear_log_context,
    log_access,
    log_with_context,
    set_request_id,
    setup_logging,
)
from app.db.init_data import seed_default_data
from app.db.session import AsyncSessionLocal, close_engine
from app.integrations.sandbox.client import SandboxClient
from app.modules.safety.confirmation import InMemoryConfirmationStore
from app.schemas.common import ErrorResponse
from app.services.dispatcher import Dispatcher
from app.services.event_store import DbEventStore
from app.services.safety import load_governor_from_db

# 模块级初始化日志系统：import 本模块即完成 loguru 配置并接管标准 logging，
# 覆盖 ``python -m app.main`` 与 ``uvicorn app.main:app`` 两种启动入口。
setup_logging()
log_with_context(
    "应用日志已初始化",
    log_dir=str(settings.resolved_log_dir),
    log_level=settings.log_level,
    app_env=settings.app_env,
)


@asynccontextmanager
async def lifespan(_app: FastAPI) -> AsyncGenerator[None, None]:
    try:
        async with AsyncSessionLocal() as session:
            await seed_default_data(session)
    except OperationalError:
        log_with_context("数据库尚未完成迁移，跳过默认数据初始化。", level="WARNING")

    # SQLite WAL 模式：用 sqlite3 直接在文件层面设置（必须跑在事务外才能生效），
    # 后续所有 aiosqlite 连接自动继承 WAL 模式；允许读写并发，
    # 避免叩心枢后台循环与请求处理同时写时锁冲突
    if settings.resolved_database_url.startswith("sqlite"):
        try:
            import sqlite3

            _db_path = settings.resolved_database_url.replace("sqlite+aiosqlite:///", "")
            _raw_conn = sqlite3.connect(_db_path)
            _raw_conn.execute("PRAGMA journal_mode=WAL")
            _raw_conn.execute("PRAGMA busy_timeout=5000")
            _raw_conn.close()
            log_with_context("SQLite WAL 模式已启用，busy_timeout=5000ms")
        except Exception:  # noqa: BLE001
            log_with_context("SQLite WAL 模式设置失败，已跳过", level="WARNING")

    # 运行时注入：事件存储始终就绪（DbEventStore 持久化到 events 表，重启不丢）；
    # 工具运行时依赖沙箱配置，未配置（缺 SANDBOX_AUTH_SECRET 等）则降级 None（工具 API 503）。
    # 事件流与历史查询不依赖沙箱，始终可用。
    _app.state.event_store = DbEventStore(AsyncSessionLocal)

    # Kernel（rewrite/1.0 M0）：极小硬核 + 事件日志 + 前台席位；神识客户端按配置装配，
    # 未配置神识地址时为进程内回声桩（骨架形态，见 app.shenshi_client）。
    from app.kernel import Kernel, KernelEventLog
    from app.shenshi_client import build_shenshi_client

    _app.state.kernel = Kernel(
        shenshi=build_shenshi_client(settings.shenshi_base_url, settings.shenshi_service_token),
        session_factory=AsyncSessionLocal,
        event_log=KernelEventLog(_app.state.event_store),
        timeout_seconds=float(settings.run_timeout_seconds),
    )

    # 镜生域事件写入绑定一炁（灵胚/启镜/命台/死亡事件落 events 表）
    from app.domains.jingsheng import bind_event_store

    bind_event_store(_app.state.event_store)

    # ── MCP 进程管理器（stdio 子进程生命周期管理） ──────────────
    from app.integrations.mcp.process_manager import (
        ProcessManager,
        set_process_manager,
    )

    process_manager = ProcessManager()
    _app.state.mcp_process_manager = process_manager
    set_process_manager(process_manager)
    log_with_context("MCP 进程管理器就绪")

    # 一炁进程内订阅分发器：订阅 EventStore 全量事件，按 glob pattern 分发给注册的 handler。
    # 各枢核心就绪后在 start() 前 register；lifespan 管理 start/stop。
    _app.state.dispatcher = Dispatcher(_app.state.event_store)
    await _app.state.dispatcher.start()
    _app.state.confirmation_store = InMemoryConfirmationStore()
    if settings.use_db_confirmation_store:
        from app.modules.safety.confirmation import DbConfirmationStore

        _app.state.confirmation_store = DbConfirmationStore(AsyncSessionLocal)
        log_with_context("confirmation_store_set_to_db")
    sandbox_http: httpx.AsyncClient | None = None
    tools_http: httpx.AsyncClient | None = None
    if settings.enable_tool_runtime:
        try:
            if settings.sandbox_auth_enabled and not settings.sandbox_auth_secret:
                raise ValueError("SANDBOX_AUTH_ENABLED=true 但 SANDBOX_AUTH_SECRET 未配置")
            sandbox_http = httpx.AsyncClient(trust_env=False)
            # executor 工具（fetch_url）的长生命周期 HTTP 客户端，
            # 与 sandbox client 同生命周期，由 lifespan 统一 aclose。
            tools_http = httpx.AsyncClient(trust_env=False)
            sandbox = SandboxClient(sandbox_http, settings)
            governor = await load_governor_from_db()
            _app.state.tool_runtime = ToolRuntime(
                sandbox,
                governor,
                _app.state.event_store,
                http_client=tools_http,
            )
            log_with_context("百工工具运行时就绪")
            # 启动加载 MCP server 工具（按 transport 类型注册，失败不阻断启动）
            try:
                from app.services.mcp import load_registered_mcp_servers

                await load_registered_mcp_servers(
                    _app.state.tool_runtime,
                    process_manager=_app.state.mcp_process_manager,
                )
            except Exception:  # noqa: BLE001
                log_with_context("MCP 启动加载失败，已跳过", level="WARNING")
        except Exception as exc:  # noqa: BLE001 — 任意配置/构造失败均降级，不阻断启动
            _app.state.tool_runtime = None
            log_with_context(f"百工工具运行时未就绪，工具 API 将返回 503：{exc}", level="WARNING")
    else:
        _app.state.tool_runtime = None
        log_with_context("ENABLE_TOOL_RUNTIME=false，工具运行时未启用", level="INFO")

    yield

    # Dispatcher 先停（取消 pump/consumer task + unsubscribe all_queue），再关 http/engine
    if getattr(_app.state, "dispatcher", None) is not None:
        await _app.state.dispatcher.stop()
    if sandbox_http is not None:
        await sandbox_http.aclose()
    if tools_http is not None:
        await tools_http.aclose()
    # MCP 进程管理器关闭所有 stdio 子进程
    if getattr(_app.state, "mcp_process_manager", None) is not None:
        await _app.state.mcp_process_manager.shutdown_all()
        log_with_context("MCP stdio 子进程已全部关闭")
    await close_engine()


def create_app() -> FastAPI:
    # OpenAPI servers 仅作展示：backend_host 为通配地址（0.0.0.0 / ::）时回退 127.0.0.1，
    # 避免在 Swagger 文档里暴露不可访问的地址或写死的占位域名。
    server_host = (
        settings.backend_host if settings.backend_host not in ("0.0.0.0", "::") else "127.0.0.1"
    )
    app = FastAPI(
        title=settings.app_name,
        debug=settings.debug,
        lifespan=lifespan,
        servers=[
            {"url": f"http://{server_host}:{settings.backend_port}", "description": "本地开发"},
        ],
        responses={
            400: {"model": ErrorResponse, "description": "请求无效或业务前置条件不满足。"},
            404: {"model": ErrorResponse, "description": "资源不存在或不属于当前人物。"},
            422: {"model": ErrorResponse, "description": "请求字段校验失败。"},
            502: {"model": ErrorResponse, "description": "模型或外部工具调用失败。"},
        },
    )
    app.add_middleware(
        CORSMiddleware,
        allow_origins=settings.cors_origins,
        allow_credentials=True,
        allow_methods=["*"],
        allow_headers=["*"],
    )

    @app.middleware("http")
    async def trace_middleware(
        request: Request, call_next: Callable[[Request], Awaitable[Response]]
    ) -> Response:
        trace_id = request.headers.get("X-Trace-Id") or new_trace_id()
        token = trace_id_var.set(trace_id)
        set_request_id(trace_id)
        started_at = perf_counter()
        try:
            response = await call_next(request)
            response.headers["X-Trace-Id"] = trace_id
            duration_ms = round((perf_counter() - started_at) * 1000, 2)
            log_access(
                "HTTP 请求完成",
                method=request.method,
                path=request.url.path,
                status=response.status_code,
                duration_ms=duration_ms,
            )
            return response
        finally:
            trace_id_var.reset(token)
            clear_log_context()

    register_exception_handlers(app)
    app.include_router(api_router, prefix=settings.api_prefix)
    return app


app = create_app()


if __name__ == "__main__":
    import uvicorn

    uvicorn.run(
        "app.main:app",
        host=settings.backend_host,
        port=settings.backend_port,
        reload=settings.is_development,
        workers=1 if settings.is_development else 4,
        log_config=None,  # 禁用 uvicorn 默认 dictConfig，统一由 InterceptHandler 接管
    )