fix(plugin): serialize dynamic route projection (#6441)

app/adapters/web/plugin/routes.py

@@ -1,5 +1,8 @@
 """FastAPI 动态插件路由适配器。"""
 
+import asyncio
+from concurrent.futures import Future, TimeoutError as FutureTimeoutError
+from threading import Lock
 from typing import Any, Callable, Optional
 
 from fastapi import Depends, FastAPI
@@ -9,6 +12,8 @@ from fastapi.routing import APIRoute
 class FastAPIDynamicRouteRegistry:
     """在 FastAPI 上注册插件自由响应路由，并维护 OpenAPI 缓存。"""
 
+    _dispatch_admission_timeout = 5.0
+
     def __init__(
         self,
         app: FastAPI,
@@ -19,6 +24,7 @@ class FastAPIDynamicRouteRegistry:
         prefix: str,
         protected_routes: set[str],
         log: Any,
+        event_loop: Callable[[], asyncio.AbstractEventLoop | None] | None = None,
     ) -> None:
         """注入应用、插件投影、认证依赖和日志端口。"""
         self._app = app
@@ -29,9 +35,62 @@ class FastAPIDynamicRouteRegistry:
         self._prefix = prefix
         self._protected_routes = protected_routes
         self._logger = log
+        self._event_loop = event_loop
 
     def update(self, plugin_id: Optional[str], action: str) -> None:
-        """按插件生命周期新增或移除动态路由。"""
+        """在主事件循环中按插件生命周期新增或移除动态路由。"""
+        if self._event_loop is None:
+            self._update(plugin_id, action)
+            return
+        target_loop = self._event_loop()
+        if (
+            target_loop is None
+            or not target_loop.is_running()
+            or target_loop.is_closed()
+        ):
+            raise RuntimeError("主事件循环未运行，无法更新插件动态路由")
+        try:
+            current_loop = asyncio.get_running_loop()
+        except RuntimeError:
+            current_loop = None
+        if current_loop is target_loop:
+            self._update(plugin_id, action)
+            return
+
+        completed: Future[None] = Future()
+        dispatch_lock = Lock()
+        dispatch_started = False
+        dispatch_abandoned = False
+
+        def apply_update() -> None:
+            """在目标 loop 的单个回调中完成路由表与 OpenAPI 投影切换。"""
+            nonlocal dispatch_started
+            with dispatch_lock:
+                if dispatch_abandoned:
+                    return
+                dispatch_started = True
+            try:
+                self._update(plugin_id, action)
+            except BaseException as error:
+                completed.set_exception(error)
+            else:
+                completed.set_result(None)
+
+        target_loop.call_soon_threadsafe(apply_update)
+        try:
+            completed.result(timeout=self._dispatch_admission_timeout)
+        except FutureTimeoutError as error:
+            with dispatch_lock:
+                if not dispatch_started:
+                    dispatch_abandoned = True
+                    raise RuntimeError(
+                        "主事件循环未及时接收插件动态路由更新"
+                    ) from error
+            # 回调一旦开始便不可撤销，等待确定终态以免失败回滚后发生迟到写入。
+            completed.result()
+
+    def _update(self, plugin_id: Optional[str], action: str) -> None:
+        """执行不可中断的路由表与 OpenAPI 投影更新。"""
         if action not in {"add", "remove"}:
             raise ValueError("Action must be 'add' or 'remove'")
 

app/factory.py

@@ -24,6 +24,7 @@ from app.adapters.web.security.access import (
     verify_token,
 )
 from app.application.security.token import create_access_token, decode_access_token
+from app.runtime.config import global_vars
 from app.runtime.settings import RuntimeSettingsCompat
 
 settings = RuntimeSettingsCompat()
@@ -393,6 +394,7 @@ def create_app() -> FastAPI:
             "/redoc",
         },
         log=logger,
+        event_loop=lambda: global_vars.CURRENT_EVENT_LOOP,
     ))
 
     return _app

app/runtime/extensions/plugin_manager.py

@@ -71,6 +71,7 @@ LegacyPluginImportPreparer = Callable[..., None]
 PluginInstallReporter = Callable[..., None]
 SiteAuthLevelProvider = Callable[[], int]
 PluginCatalogFactory = Callable[["PluginManager"], Any]
+PluginRouteRefresher = Callable[[str], None]
 
 
 def _ignore_legacy_diagnostics(**_kwargs) -> None:
@@ -109,6 +110,11 @@ def _warn_if_plugin_enabled_gil(
     )
 
 
+def _unavailable_plugin_route_refresher(_plugin_id: str) -> None:
+    """在 HTTP 组合尚未装配时拒绝发布不完整的热重载投影。"""
+    raise RuntimeError("插件动态路由刷新器尚未由启动组合根装配")
+
+
 _legacy_diagnostics_configurator: LegacyDiagnosticsConfigurator = (
     _ignore_legacy_diagnostics
 )
@@ -119,6 +125,7 @@ _legacy_plugin_import_preparer: LegacyPluginImportPreparer = (
 _plugin_install_reporter: PluginInstallReporter = _ignore_legacy_diagnostics
 _site_auth_level_provider: SiteAuthLevelProvider = _unavailable_site_auth_level
 _plugin_catalog_factory: PluginCatalogFactory = _unavailable_plugin_catalog_factory
+_plugin_route_refresher: PluginRouteRefresher = _unavailable_plugin_route_refresher
 
 
 def configure_plugin_legacy_import_services(
@@ -158,6 +165,12 @@ def configure_plugin_catalog_factory(factory: PluginCatalogFactory) -> None:
     _plugin_catalog_factory = factory
 
 
+def configure_plugin_route_refresher(refresher: PluginRouteRefresher) -> None:
+    """由启动组合根注入热重载后的动态路由投影刷新器。"""
+    global _plugin_route_refresher
+    _plugin_route_refresher = refresher
+
+
 @observe_compat_facade("PluginManager")
 class PluginManager(ConfigReloadMixin, metaclass=Singleton):
     """插件管理器"""
@@ -656,7 +669,7 @@ class PluginManager(ConfigReloadMixin, metaclass=Singleton):
             monitor_suppressed=self.is_plugin_monitor_suppressed,
             local_candidate=self._get_local_plugin_candidate_from_path,
             sync_local=self._sync_local_plugin_if_installed,
-            reload_plugin=self.reload_plugin_tree,
+            reload_plugin=self._reload_plugin_tree_from_monitor,
             dependency_manifest_status=(
                 get_plugin_system().dependency_manifest_status
             ),
@@ -664,6 +677,18 @@ class PluginManager(ConfigReloadMixin, metaclass=Singleton):
             log=logger,
         ).run()
 
+    def _reload_plugin_tree_from_monitor(
+        self,
+        plugin_id: str,
+    ) -> PluginRuntimeStatus:
+        """重载源码树，并发布源插件及虚拟实例的动态路由投影。"""
+        with self.mutation("热重载插件路由"):
+            status = self.reload_plugin_tree(plugin_id)
+            reload_targets = self.get_plugin_reload_targets(plugin_id)
+            for reload_target in reload_targets:
+                _plugin_route_refresher(reload_target)
+            return status
+
     def _get_federated_plugin_change(
         self,
         event_path: Path,

app/scheduler.py

@@ -32,6 +32,7 @@ from app.application.agentdata import get_agent_task_port
 from app.application.database import get_database_governance
 from app.application.outbox import dispatch_pending_outbox
 from app.application.plugin.runtime import get_plugin_manager
+from app.application.plugin.routes import register_plugin_api
 from app.application.configuration import (
     SchedulerRuntimeConfig,
     get_configured_system_config,
@@ -152,6 +153,8 @@ class Scheduler(ConfigReloadMixin, metaclass=SingletonClass):
         self._auth_count = 0
         # 用户认证失败消息发送
         self._auth_message = False
+        # 插件已按认证结果重建，但动态路由尚未完成投影时保留重试状态。
+        self._auth_plugin_routes_pending = False
 
     async def on_config_changed(self) -> None:
         """
@@ -2054,6 +2057,9 @@ class Scheduler(ConfigReloadMixin, metaclass=SingletonClass):
         """
         config = get_scheduler_runtime_config()
         if SitesHelper().auth_level >= 2:
+            if self._auth_plugin_routes_pending:
+                register_plugin_api()
+                self._auth_plugin_routes_pending = False
             return
         # 最大重试次数
         __max_try__ = 30
@@ -2088,6 +2094,9 @@ class Scheduler(ConfigReloadMixin, metaclass=SingletonClass):
             # 认证通过后重新初始化插件
             get_plugin_manager().init_config()
             self.init_plugin_jobs()
+            self._auth_plugin_routes_pending = True
+            register_plugin_api()
+            self._auth_plugin_routes_pending = False
 
         else:
             self._auth_count += 1

app/startup/initializers/plugins.py

@@ -15,6 +15,7 @@ from app.runtime.extensions.plugin_manager import (
     configure_plugin_catalog_factory,
     configure_plugin_install_reporter,
     configure_plugin_legacy_import_services,
+    configure_plugin_route_refresher,
     configure_plugin_resource_import_preparer,
     configure_site_auth_level_provider,
 )
@@ -95,6 +96,7 @@ def configure_plugin_services() -> None:
         lambda: get_configured_system_config().get(SystemConfigKey.UserInstalledPlugins) or []
     )
     configure_plugin_catalog_factory(_build_plugin_catalog)
+    configure_plugin_route_refresher(register_plugin_api)
     configure_plugin_system(PluginSystemServices(
         market=market_client,
         package=PluginPackageManager(plugin_helper),

tests/fixtures/architecture/dependency-baseline.json

@@ -13,8 +13,8 @@
     "runtime_to_db": [],
     "workflow_to_db": []
   },
-  "edge_count": 6599,
+  "edge_count": 6601,
-  "edge_sha256": "33b44ea7d1f05a4c76fd22c4c1b798689ac8e580099c50141fbd91345a625714",
+  "edge_sha256": "4ebc1db325d75ac04418e89c199dc2b09eb95905e6e03a4101b7780d6b184c7c",
   "edges": [
     "app -> app.runtime",
     "app -> app.runtime.compat",
@@ -3937,6 +3937,7 @@
     "app.factory -> app.application.security",
     "app.factory -> app.application.security.token",
     "app.factory -> app.runtime",
+    "app.factory -> app.runtime.config",
     "app.factory -> app.runtime.correlation",
     "app.factory -> app.runtime.localization",
     "app.factory -> app.runtime.log",
@@ -5857,6 +5858,7 @@
     "app.scheduler -> app.application.messaging.message",
     "app.scheduler -> app.application.outbox",
     "app.scheduler -> app.application.plugin",
+    "app.scheduler -> app.application.plugin.routes",
     "app.scheduler -> app.application.plugin.runtime",
     "app.scheduler -> app.application.scheduling",
     "app.scheduler -> app.application.site",

tests/test_api_response.py

@@ -1,5 +1,10 @@
+import asyncio
+import threading
+import time
+from concurrent.futures import ThreadPoolExecutor
 from types import SimpleNamespace
 from typing import Any
+from unittest.mock import MagicMock
 
 import httpx
 import pytest
@@ -928,6 +933,144 @@ async def test_plugin_routes_ignore_included_router_wrappers():
     assert removed_response.status_code == 404
 
 
+async def test_plugin_route_updates_run_on_application_event_loop() -> None:
+    """线程侧插件变更必须回投主 loop 后再修改 FastAPI 路由表。"""
+    app = FastAPI()
+    main_thread = threading.get_ident()
+    mutation_threads: list[int] = []
+    original_add_api_route = app.router.add_api_route
+    original_setup = app.setup
+
+    def record_add_api_route(*args, **kwargs):
+        mutation_threads.append(threading.get_ident())
+        return original_add_api_route(*args, **kwargs)
+
+    def record_setup() -> None:
+        mutation_threads.append(threading.get_ident())
+        original_setup()
+
+    app.router.add_api_route = record_add_api_route
+    app.setup = record_setup
+    loop = asyncio.get_running_loop()
+    registry = FastAPIDynamicRouteRegistry(
+        app=app,
+        plugin_ids=lambda: ["DemoPlugin"],
+        plugin_apis=lambda _plugin_id: [{
+            "path": "/DemoPlugin/health",
+            "endpoint": lambda: {"ok": True},
+            "methods": ["GET"],
+            "allow_anonymous": True,
+        }],
+        verify_token=lambda: None,
+        verify_apikey=lambda: None,
+        prefix="/api/v1/plugin",
+        protected_routes=set(),
+        log=SimpleNamespace(debug=lambda *_args: None, error=lambda *_args: None),
+        event_loop=lambda: loop,
+    )
+
+    await asyncio.to_thread(registry.update, "DemoPlugin", "add")
+
+    assert mutation_threads
+    assert set(mutation_threads) == {main_thread}
+    assert any(
+        getattr(route, "path", None) == "/api/v1/plugin/DemoPlugin/health"
+        for route in app.routes
+    )
+
+
+def test_plugin_route_update_rejects_stopped_application_loop() -> None:
+    """生产 loop 已释放时不得退回调用线程修改路由。"""
+    registry = FastAPIDynamicRouteRegistry(
+        app=FastAPI(),
+        plugin_ids=lambda: [],
+        plugin_apis=lambda _plugin_id: [],
+        verify_token=lambda: None,
+        verify_apikey=lambda: None,
+        prefix="/api/v1/plugin",
+        protected_routes=set(),
+        log=SimpleNamespace(debug=lambda *_args: None, error=lambda *_args: None),
+        event_loop=lambda: None,
+    )
+
+    with pytest.raises(RuntimeError, match="主事件循环未运行"):
+        registry.update("DemoPlugin", "remove")
+
+
+async def test_plugin_route_update_abandons_late_application_loop_callback() -> None:
+    """超时前未开始的回调可以迟到执行，但不得再写入路由或污染事件循环。"""
+    app = FastAPI()
+    loop = asyncio.get_running_loop()
+    plugin_apis = MagicMock(return_value=[])
+    loop_errors: list[dict] = []
+    previous_handler = loop.get_exception_handler()
+    loop.set_exception_handler(lambda _loop, context: loop_errors.append(context))
+    registry = FastAPIDynamicRouteRegistry(
+        app=app,
+        plugin_ids=lambda: [],
+        plugin_apis=plugin_apis,
+        verify_token=lambda: None,
+        verify_apikey=lambda: None,
+        prefix="/api/v1/plugin",
+        protected_routes=set(),
+        log=SimpleNamespace(debug=lambda *_args: None, error=lambda *_args: None),
+        event_loop=lambda: loop,
+    )
+    registry._dispatch_admission_timeout = 0.01
+
+    try:
+        with ThreadPoolExecutor(max_workers=1) as executor:
+            update = executor.submit(registry.update, "DemoPlugin", "add")
+            time.sleep(0.05)
+            with pytest.raises(RuntimeError, match="未及时接收"):
+                update.result()
+        await asyncio.sleep(0)
+    finally:
+        loop.set_exception_handler(previous_handler)
+
+    plugin_apis.assert_not_called()
+    assert not any(
+        getattr(route, "path", None) == "/api/v1/plugin/DemoPlugin/health"
+        for route in app.routes
+    )
+    assert loop_errors == []
+
+
+async def test_started_plugin_route_update_waits_for_terminal_result() -> None:
+    """回调开始后即使超过 admission 预算，也不得先报失败再迟到写入。"""
+    app = FastAPI()
+    loop = asyncio.get_running_loop()
+
+    def delayed_plugin_apis(_plugin_id: str) -> list[dict]:
+        time.sleep(0.05)
+        return [{
+            "path": "/DemoPlugin/health",
+            "endpoint": lambda: {"ok": True},
+            "methods": ["GET"],
+            "allow_anonymous": True,
+        }]
+
+    registry = FastAPIDynamicRouteRegistry(
+        app=app,
+        plugin_ids=lambda: ["DemoPlugin"],
+        plugin_apis=delayed_plugin_apis,
+        verify_token=lambda: None,
+        verify_apikey=lambda: None,
+        prefix="/api/v1/plugin",
+        protected_routes=set(),
+        log=SimpleNamespace(debug=lambda *_args: None, error=lambda *_args: None),
+        event_loop=lambda: loop,
+    )
+    registry._dispatch_admission_timeout = 0.01
+
+    await asyncio.to_thread(registry.update, "DemoPlugin", "add")
+
+    assert sum(
+        getattr(route, "path", None) == "/api/v1/plugin/DemoPlugin/health"
+        for route in app.routes
+    ) == 1
+
+
 def test_response_router_uses_response_route_class():
     """统一路由器应默认创建统一响应路由。"""
     router = ResponseAPIRouter()

tests/test_plugin_monitor_lifecycle.py

@@ -19,6 +19,7 @@ from app.runtime.extensions.plugin.monitor import (
 )
 from app.runtime.extensions.plugin.admission import PluginMutationAdmission
 from app.runtime.extensions.plugin.system import reset_plugin_system
+from app.runtime.extensions import plugin_manager as plugin_manager_module
 from app.runtime.extensions.plugin_manager import PluginManager
 from app.schemas.plugin import PluginRuntimeStatus
 from app.startup.initializers import plugins as plugins_initializer
@@ -749,6 +750,35 @@ def test_plugin_monitor_skips_installing_plugin_until_package_write_finishes(tmp
     reload_plugin.assert_not_called()
 
 
+def test_monitor_reload_refreshes_source_and_instance_routes(monkeypatch) -> None:
+    """源码热重载后必须刷新源插件及其虚拟实例的动态路由投影。"""
+    admission = PluginMutationAdmission()
+    manager = SimpleNamespace(
+        mutation=admission.hold,
+        get_plugin_reload_targets=MagicMock(
+            return_value=["DemoPlugin", "DemoPlugin_1"]
+        ),
+        reload_plugin_tree=MagicMock(return_value=PluginRuntimeStatus.ACTIVE),
+    )
+    refresh_holds: list[bool] = []
+    refresh = MagicMock(
+        side_effect=lambda _plugin_id: refresh_holds.append(admission.is_held())
+    )
+    monkeypatch.setattr(plugin_manager_module, "_plugin_route_refresher", refresh)
+
+    status = PluginManager._reload_plugin_tree_from_monitor(manager, "DemoPlugin")
+
+    assert status is PluginRuntimeStatus.ACTIVE
+    assert admission.active_count == 0
+    assert refresh_holds == [True, True]
+    manager.get_plugin_reload_targets.assert_called_once_with("DemoPlugin")
+    manager.reload_plugin_tree.assert_called_once_with("DemoPlugin")
+    assert [item.args for item in refresh.call_args_list] == [
+        ("DemoPlugin",),
+        ("DemoPlugin_1",),
+    ]
+
+
 def test_plugin_monitor_suppression_is_reference_counted(monkeypatch) -> None:
     """同一插件的重叠写入必须等最后一个事务退出后才解除监控抑制。"""
     _reset_plugin_manager()

tests/test_scheduler_cache_expiry.py

@@ -2,6 +2,8 @@ import asyncio
 import threading
 from unittest.mock import AsyncMock, Mock
 
+import pytest
+
 from app import scheduler as scheduler_module
 from app.scheduler import Scheduler
 from app.application.configuration import SchedulerRuntimeConfig
@@ -133,3 +135,90 @@ def test_clear_cache_is_manual_only(monkeypatch):
     assert "clear_cache" in scheduler._jobs
     assert scheduler._jobs["clear_cache"]["manual"] is True
     assert background_scheduler.started is True
+
+
+def test_user_auth_refreshes_plugin_routes_after_runtime_reinitialization(monkeypatch):
+    """自动认证重建插件实例后必须同步刷新动态路由投影。"""
+    scheduler = object.__new__(Scheduler)
+    scheduler._auth_count = 1
+    scheduler._auth_message = False
+    scheduler._auth_plugin_routes_pending = False
+    plugin_manager = Mock()
+    plugin_jobs = Mock()
+    refresh_routes = Mock()
+    message_chain = Mock()
+    monkeypatch.setattr(
+        scheduler_module,
+        "get_scheduler_runtime_config",
+        lambda: Mock(site_link="https://example.invalid"),
+    )
+    monkeypatch.setattr(
+        scheduler_module,
+        "SitesHelper",
+        lambda: Mock(auth_level=0, check_user=Mock(return_value=(True, "demo"))),
+    )
+    monkeypatch.setattr(
+        scheduler_module,
+        "get_configured_system_config",
+        lambda: Mock(get=Mock(return_value=None)),
+    )
+    monkeypatch.setattr(scheduler_module, "SchedulerChain", lambda: message_chain)
+    monkeypatch.setattr(
+        scheduler_module,
+        "get_plugin_manager",
+        lambda: plugin_manager,
+    )
+    monkeypatch.setattr(scheduler, "init_plugin_jobs", plugin_jobs)
+    monkeypatch.setattr(scheduler_module, "register_plugin_api", refresh_routes)
+
+    scheduler.user_auth()
+
+    plugin_manager.init_config.assert_called_once_with()
+    plugin_jobs.assert_called_once_with()
+    refresh_routes.assert_called_once_with()
+    assert scheduler._auth_plugin_routes_pending is False
+
+
+def test_user_auth_retries_pending_plugin_route_projection(monkeypatch):
+    """认证已成功但路由投影失败时，后续认证任务只重试未完成的投影。"""
+    scheduler = object.__new__(Scheduler)
+    scheduler._auth_count = 1
+    scheduler._auth_message = False
+    scheduler._auth_plugin_routes_pending = False
+    plugin_manager = Mock()
+    plugin_jobs = Mock()
+    refresh_routes = Mock(side_effect=[RuntimeError("loop unavailable"), None])
+    message_chain = Mock()
+    sites = Mock(auth_level=0, check_user=Mock(return_value=(True, "demo")))
+    monkeypatch.setattr(
+        scheduler_module,
+        "get_scheduler_runtime_config",
+        lambda: Mock(site_link="https://example.invalid"),
+    )
+    monkeypatch.setattr(scheduler_module, "SitesHelper", lambda: sites)
+    monkeypatch.setattr(
+        scheduler_module,
+        "get_configured_system_config",
+        lambda: Mock(get=Mock(return_value=None)),
+    )
+    monkeypatch.setattr(scheduler_module, "SchedulerChain", lambda: message_chain)
+    monkeypatch.setattr(
+        scheduler_module,
+        "get_plugin_manager",
+        lambda: plugin_manager,
+    )
+    monkeypatch.setattr(scheduler, "init_plugin_jobs", plugin_jobs)
+    monkeypatch.setattr(scheduler_module, "register_plugin_api", refresh_routes)
+
+    with pytest.raises(RuntimeError, match="loop unavailable"):
+        scheduler.user_auth()
+
+    assert scheduler._auth_plugin_routes_pending is True
+    sites.auth_level = 2
+    scheduler.user_auth()
+
+    assert scheduler._auth_plugin_routes_pending is False
+    assert refresh_routes.call_count == 2
+    plugin_manager.init_config.assert_called_once_with()
+    plugin_jobs.assert_called_once_with()
+    sites.check_user.assert_called_once_with()