fix(agent): close async chat persistence lifecycle

app/api/dependencies/agent.py

@@ -4,6 +4,7 @@ from fastapi import Depends
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.api.context import (
+    get_agent_chat_runtime,
     get_agent_chat_repository,
     get_agent_chat_transaction,
     get_async_session,
@@ -11,6 +12,7 @@ from app.api.context import (
 )
 from app.application.messaging.chat import (
     AgentChatService,
+    AgentChatPersistenceService,
     AsyncAgentChatRepository,
     AsyncUnitOfWork,
 )
@@ -26,6 +28,13 @@ def get_agent_chat_service(
     return AgentChatService(chat_repository, unit_of_work)
 
 
+def get_agent_chat_persistence(
+    runtime: HostRuntime = Depends(get_agent_chat_runtime),
+) -> AgentChatPersistenceService:
+    """从类型化 Agent 运行时获取有界会话写入端口。"""
+    return runtime.persistence
+
+
 def get_message_query_service(
     db: AsyncSession = Depends(get_async_session),
     runtime: HostRuntime = Depends(get_host_runtime),

app/api/deps.py

@@ -5,6 +5,7 @@
 """
 
 from app.api.dependencies.agent import (
+    get_agent_chat_persistence,
     get_agent_chat_service,
     get_message_query_service,
 )
@@ -52,6 +53,7 @@ from app.api.dependencies.workflow import (
 
 # 兼容聚合入口只显式列出既有 FastAPI 依赖，不向插件制造新的动态导出规则。
 __all__ = [
+    "get_agent_chat_persistence",
     "get_agent_chat_service",
     "get_auth_service",
     "get_current_active_manage_user",

app/api/endpoints/agent.py

@@ -47,17 +47,25 @@ from app.command import Command
 from app.runtime.config import global_vars
 from app.runtime.events import Event, EventManager
 from app.api.principal import ApiPrincipal
-from app.api.dependencies.agent import get_agent_chat_service
+from app.api.dependencies.agent import (
+    get_agent_chat_persistence,
+    get_agent_chat_service,
+)
 from app.api.dependencies.auth import get_current_active_user
 from app.application.messaging.chat import (
     AgentChatRecord,
+    AgentChatPersistenceService,
     AgentChatService,
     get_configured_agent_chat_service,
     get_configured_agent_chat_persistence,
 )
 from app.application.security.user import get_configured_user_id_lookup
 from app.application.configuration import get_api_runtime_config_snapshot
-from app.application.messaging.agent import attach_web_agent_edit_queue, detach_web_agent_edit_queue
+from app.application.messaging.agent import (
+    attach_web_agent_edit_queue,
+    create_web_agent_background_task,
+    detach_web_agent_edit_queue,
+)
 from app.application.messaging.agent import agent_interaction_manager
 from app.application.messaging.agent import (
     build_agent_choice_button_rows,
@@ -88,7 +96,6 @@ _WEB_AGENT_FILE_REGISTRY: dict[str, dict[str, Any]] = {}
 _WEB_AGENT_MESSAGE_QUEUES: dict[str, list[Queue[_SchemaMessage]]] = {}
 _WEB_AGENT_MESSAGE_LOCK = Lock()
 _WEB_AGENT_MESSAGE_LISTENER_REGISTERED = False
-_WEB_AGENT_BACKGROUND_TASKS: set[asyncio.Task] = set()
 
 
 class _WebAgentEventPublisher:
@@ -517,13 +524,16 @@ def _build_web_agent_session_id(user: ApiPrincipal, session_id: Optional[str]) -
 async def _build_web_agent_session_id_async(
     user: ApiPrincipal,
     session_id: Optional[str],
+    service: Optional[AgentChatService] = None,
 ) -> str:
     """异步解析 Web Agent 会话 ID，并复用异步会话查询端口。"""
     seed = str(session_id or "").strip() or uuid.uuid4().hex
     if seed.startswith(WEB_AGENT_SESSION_PREFIX):
         return seed
     try:
-        existing_chat = await get_configured_agent_chat_service().get(seed)
+        if service is None:
+            service = get_configured_agent_chat_service()
+        existing_chat = await service.get(seed)
         if existing_chat and AgentChatService.can_access(existing_chat, user):
             return seed
     except Exception as e:
@@ -658,36 +668,41 @@ async def _save_web_agent_display_snapshot(
     current_user: ApiPrincipal,
     messages: list[dict],
     client_session_id: Optional[str] = None,
+    service: Optional[AgentChatService] = None,
+    persistence: Optional[AgentChatPersistenceService] = None,
 ) -> None:
     """
     保存 WebAgent 当前展示消息快照。
     """
-    try:
-        existing_chat = await get_configured_agent_chat_service().get(session_id)
-        await get_configured_agent_chat_persistence().async_save_display_messages(
-            session_id=session_id,
-            user_id=(existing_chat.user_id if existing_chat else str(current_user.id)),
-            username=(existing_chat.username if existing_chat else current_user.name),
-            channel=(
-                existing_chat.channel
-                if existing_chat and existing_chat.channel
-                else NotificationChannel.WebAgent
-            ),
-            source=(
-                existing_chat.source
-                if existing_chat and existing_chat.source
-                else WEB_AGENT_SOURCE
-            ),
-            original_chat_id=existing_chat.original_chat_id if existing_chat else None,
-            client_session_id=(
-                existing_chat.client_session_id
-                if existing_chat and existing_chat.client_session_id
-                else client_session_id
-            ),
-            messages=messages,
-        )
-    except Exception as e:
-        logger.debug(f"保存WebAgent展示历史失败: {e}")
+    if service is None:
+        # 直接调用该内部 helper 时没有 FastAPI 依赖注入上下文。
+        service = get_configured_agent_chat_service()
+    existing_chat = await service.get(session_id)
+    if persistence is None:
+        # 直接调用该内部 helper 时没有 FastAPI 依赖注入上下文。
+        persistence = get_configured_agent_chat_persistence()
+    await persistence.async_save_display_messages(
+        session_id=session_id,
+        user_id=(existing_chat.user_id if existing_chat else str(current_user.id)),
+        username=(existing_chat.username if existing_chat else current_user.name),
+        channel=(
+            existing_chat.channel
+            if existing_chat and existing_chat.channel
+            else NotificationChannel.WebAgent
+        ),
+        source=(
+            existing_chat.source
+            if existing_chat and existing_chat.source
+            else WEB_AGENT_SOURCE
+        ),
+        original_chat_id=existing_chat.original_chat_id if existing_chat else None,
+        client_session_id=(
+            existing_chat.client_session_id
+            if existing_chat and existing_chat.client_session_id
+            else client_session_id
+        ),
+        messages=messages,
+    )
 
 
 def _build_web_agent_sse(
@@ -757,6 +772,7 @@ def _sanitize_web_agent_upload_name(
 async def _get_web_agent_upload_dir(
     user: ApiPrincipal,
     session_id: Optional[str],
+    service: Optional[AgentChatService] = None,
 ) -> Path:
     """
     计算当前 Web Agent 会话的临时附件目录。
@@ -765,7 +781,11 @@ async def _get_web_agent_upload_dir(
     :param session_id: 前端会话标识
     :return: 已创建的临时附件目录
     """
-    server_session_id = await _build_web_agent_session_id_async(user, session_id)
+    server_session_id = await _build_web_agent_session_id_async(
+        user,
+        session_id,
+        service,
+    )
     safe_session_id = server_session_id.replace(":", "_")
     upload_dir = get_api_runtime_config_snapshot().temp_path / "agent_uploads" / safe_session_id
     upload_dir.mkdir(parents=True, exist_ok=True)
@@ -1701,6 +1721,7 @@ async def upload_web_agent_file(
     file: UploadFile = File(...),
     session_id: Optional[str] = Form(None),
     current_user: ApiPrincipal = Depends(get_current_active_user),
+    service: AgentChatService = Depends(get_agent_chat_service),
 ) -> _SchemaResponse:
     """
     上传 Web 智能助手对话附件。
@@ -1712,7 +1733,7 @@ async def upload_web_agent_file(
     """
     mime_type = file.content_type or mimetypes.guess_type(file.filename or "")[0]
     safe_name = _sanitize_web_agent_upload_name(file.filename, mime_type)
-    upload_dir = await _get_web_agent_upload_dir(current_user, session_id)
+    upload_dir = await _get_web_agent_upload_dir(current_user, session_id, service)
     target_path = upload_dir / f"{uuid.uuid4().hex[:8]}_{safe_name}"
     size = await _save_web_agent_upload(file, target_path)
     attachment = _register_web_agent_file(
@@ -1846,6 +1867,7 @@ async def get_agent_chat_session(
         server_session_id = await _build_web_agent_session_id_async(
             current_user,
             session_id,
+            service,
         )
         if server_session_id != session_id:
             chat = await _get_accessible_agent_chat(
@@ -1884,6 +1906,7 @@ async def save_agent_chat_display(
     payload: _SchemaAgentChatDisplaySaveRequest,
     current_user: ApiPrincipal = Depends(get_current_active_user),
     service: AgentChatService = Depends(get_agent_chat_service),
+    persistence: AgentChatPersistenceService = Depends(get_agent_chat_persistence),
 ) -> _SchemaResponse:
     """
     保存前端聚合后的 Agent 展示消息。
@@ -1911,6 +1934,8 @@ async def save_agent_chat_display(
         current_user=current_user,
         messages=messages,
         client_session_id=existing_chat.client_session_id if existing_chat else session_id,
+        service=service,
+        persistence=persistence,
     )
     chat = await service.get_accessible(session_id, current_user)
     if not chat:
@@ -1964,6 +1989,7 @@ async def stop_web_agent_session_task(
     server_session_id = await _build_web_agent_session_id_async(
         current_user,
         session_id,
+        service,
     )
     chat = await _get_accessible_agent_chat(
         service,
@@ -1988,6 +2014,8 @@ async def _web_agent_stream_impl(
     payload: _SchemaAgentWebChatRequest,
     request: Request,
     current_user: ApiPrincipal = Depends(get_current_active_user),
+    service: Optional[AgentChatService] = None,
+    persistence: Optional[AgentChatPersistenceService] = None,
 ) -> StreamingResponse:
     """
     Web 智能助手流式对话。
@@ -1998,11 +2026,18 @@ async def _web_agent_stream_impl(
     :return: SSE 流式响应
     """
     prompt = payload.text.strip()
+    if not isinstance(service, AgentChatService):
+        # 直接调用公开函数时不经过 FastAPI 依赖解析；生产路由总是传入运行时服务。
+        service = get_configured_agent_chat_service()
+    if not isinstance(persistence, AgentChatPersistenceService):
+        # 直接调用公开函数时不经过 FastAPI 依赖解析；生产路由总是传入运行时端口。
+        persistence = get_configured_agent_chat_persistence()
     locale = LocaleHelper.get_locale_from_request(request)
     display_prompt = (payload.display_text or payload.text).strip()
     session_id = await _build_web_agent_session_id_async(
         current_user,
         payload.session_id,
+        service,
     )
     is_secret_confirmation_candidate = (
         prompt in {"确认", "取消"}
@@ -2110,13 +2145,13 @@ async def _web_agent_stream_impl(
                         current_user=current_user,
                         messages=display_messages,
                         client_session_id=payload.session_id or session_id,
+                        service=service,
+                        persistence=persistence,
                     )
                 except Exception as err:
                     logger.error(f"保存WebAgent传统消息快照失败: {str(err)}")
 
-            snapshot_task = asyncio.create_task(save_display_snapshot())
-            _WEB_AGENT_BACKGROUND_TASKS.add(snapshot_task)
-            snapshot_task.add_done_callback(_WEB_AGENT_BACKGROUND_TASKS.discard)
+            snapshot_task = create_web_agent_background_task(save_display_snapshot())
             await asyncio.sleep(0)
             for event in events:
                 event_payload = copy.deepcopy(event)
@@ -2268,16 +2303,19 @@ async def _web_agent_stream_impl(
                 # 终态先进入 SSE 队列，避免展示快照落库延迟前端结束动画。
                 event_publisher.publish(done_event)
                 if not is_secret_confirmation_control:
-                    await _save_web_agent_display_snapshot(
-                        session_id=session_id,
-                        current_user=current_user,
-                        messages=display_messages,
-                        client_session_id=payload.session_id or session_id,
-                    )
+                    try:
+                        await _save_web_agent_display_snapshot(
+                            session_id=session_id,
+                            current_user=current_user,
+                            messages=display_messages,
+                            client_session_id=payload.session_id or session_id,
+                            service=service,
+                            persistence=persistence,
+                        )
+                    except Exception as err:
+                        logger.error(f"保存WebAgent展示历史失败：{err}")
 
-        task = asyncio.create_task(run_agent())
-        _WEB_AGENT_BACKGROUND_TASKS.add(task)
-        task.add_done_callback(_WEB_AGENT_BACKGROUND_TASKS.discard)
+        task = create_web_agent_background_task(run_agent())
         disconnected = False
         terminal_sent = False
         try:
@@ -2349,6 +2387,14 @@ async def web_agent_stream(
     payload: _SchemaAgentWebChatRequest,
     request: Request,
     current_user: ApiPrincipal = Depends(get_current_active_user),
+    service: AgentChatService = Depends(get_agent_chat_service),
+    persistence: AgentChatPersistenceService = Depends(get_agent_chat_persistence),
 ) -> StreamingResponse:
     """Web 智能助手流式对话的稳定公开路由入口。"""
-    return await _web_agent_stream_impl(payload, request, current_user)
+    return await _web_agent_stream_impl(
+        payload,
+        request,
+        current_user,
+        service,
+        persistence,
+    )

app/application/messaging/agent.py

@@ -1,9 +1,10 @@
+import asyncio
 import uuid
 from dataclasses import dataclass, field
 from datetime import datetime, timedelta
 from queue import Queue
 from threading import Lock
-from typing import Callable, Dict, Iterable, List, Optional, Tuple, Union
+from typing import Awaitable, Callable, Dict, Iterable, List, Optional, Tuple, Union
 
 from app.schemas.types import NotificationChannel
 
@@ -177,6 +178,26 @@ _WEB_AGENT_EDIT_QUEUES: dict[str, list[Queue[dict]]] = {}
 _WEB_AGENT_EDIT_LOCK = Lock()
 _ChannelAdminResolver = Callable[[Optional[dict]], Iterable[Union[str, int]]]
 _CHANNEL_ADMIN_RESOLVERS: dict[str, _ChannelAdminResolver] = {}
+_WEB_AGENT_BACKGROUND_TASKS: set[asyncio.Task[object]] = set()
+
+
+def create_web_agent_background_task(
+    coroutine: Awaitable[object],
+) -> asyncio.Task[object]:
+    """登记 Web Agent 后台任务，使应用关闭时可以统一收口。"""
+    task = asyncio.create_task(coroutine)
+    _WEB_AGENT_BACKGROUND_TASKS.add(task)
+    task.add_done_callback(_WEB_AGENT_BACKGROUND_TASKS.discard)
+    return task
+
+
+async def shutdown_web_agent_background_tasks() -> None:
+    """取消并等待 Web Agent 后台任务，避免关闭数据库后仍提交快照。"""
+    tasks = tuple(_WEB_AGENT_BACKGROUND_TASKS)
+    for task in tasks:
+        task.cancel()
+    if tasks:
+        await asyncio.gather(*tasks, return_exceptions=True)
 
 
 def register_channel_admin_resolver(

app/application/messaging/chat.py

@@ -10,6 +10,7 @@ from weakref import WeakValueDictionary
 
 from app.application.database import (
     AsyncDatabaseExecutor,
+    DatabaseWorkerClosedError,
     DatabaseWorkerOverloadedError,
 )
 from app.schemas.agent import AgentChatSessionDetail, AgentChatSessionSummary
@@ -17,6 +18,7 @@ from app.runtime.observability import record_metric
 
 
 DEFAULT_AGENT_CHAT_WRITE_CAPACITY = 32
+DEFAULT_AGENT_CHAT_SESSION_CAPACITY = 4
 
 
 def has_custom_agent_chat_title(value: Optional[str]) -> bool:
@@ -143,7 +145,8 @@ class SyncAgentChatRepository(Protocol):
         ...
 
 
-SyncAgentChatRepositoryFactory = Callable[[], SyncAgentChatRepository]
+SyncAgentChatRepositoryFactory = Callable[[object], SyncAgentChatRepository]
+SyncAgentChatTransaction = Callable[[Callable[[object], object]], object]
 
 
 @dataclass(frozen=True, slots=True)
@@ -355,15 +358,24 @@ class AgentChatPersistenceService:
         self,
         repository: SyncAgentChatRepositoryFactory,
         async_executor: AsyncDatabaseExecutor,
+        sync_transaction: SyncAgentChatTransaction,
         capacity: int = DEFAULT_AGENT_CHAT_WRITE_CAPACITY,
+        session_capacity: int = DEFAULT_AGENT_CHAT_SESSION_CAPACITY,
     ) -> None:
-        """保存同步仓储工厂和异步执行端口。"""
+        """保存同步仓储工厂、事务端口和两级写入容量。"""
         if capacity < 1:
             raise ValueError("AgentChat 写入容量必须大于 0")
+        if session_capacity < 1:
+            raise ValueError("AgentChat 单会话写入容量必须大于 0")
         self._repository = repository
         self._async_executor = async_executor
+        self._sync_transaction = sync_transaction
         self._capacity = capacity
+        self._session_capacity = session_capacity
         self._pending_writes = 0
+        self._pending_by_session: dict[str, int] = {}
+        self._active_tasks: set[asyncio.Task[object]] = set()
+        self._closing = False
         # append_display_messages 属于读取旧快照后整列写回的复合操作；按会话串行化，
         # 才能在 worker 并发下保持首次建行和既有会话追加的完整性。弱引用避免长期运行
         # 中为一次性会话永久保留锁对象，不限制不同会话之间的 worker 并行度。
@@ -383,24 +395,52 @@ class AgentChatPersistenceService:
         operation: Callable[[SyncAgentChatRepository], object],
     ) -> None:
         """在线程 worker 内完成同步写入并丢弃仓储对象返回值。"""
-        # 会话锁前的等待也纳入固定总量，避免公开展示保存入口形成无界应用层队列。
-        if self._pending_writes >= self._capacity:
+        # 同时限制全局和单会话等待量，避免一个热点会话占满总 admission 后饿死其他会话。
+        if self._closing:
+            raise DatabaseWorkerClosedError("AgentChat 持久化服务当前不可接收任务")
+        session_pending = self._pending_by_session.get(session_id, 0)
+        if (
+            self._pending_writes >= self._capacity
+            or session_pending >= self._session_capacity
+        ):
             record_metric("agent.chat.persistence.rejected")
             raise DatabaseWorkerOverloadedError(
-                f"AgentChat 写入容量已用尽（上限 {self._capacity}）"
+                f"AgentChat 写入容量已用尽（全局上限 {self._capacity}，"
+                f"单会话上限 {self._session_capacity}）"
             )
         self._pending_writes += 1
-        record_metric("agent.chat.persistence.pending", self._pending_writes)
+        self._pending_by_session[session_id] = session_pending + 1
+        current = asyncio.current_task()
+        if current is not None:
+            self._active_tasks.add(current)
+        record_metric("agent.chat.persistence.pending", 1)
         try:
             async with self._session_lock(session_id):
                 def execute() -> None:
-                    """执行同步写入，不让 ORM 对象越过 worker 边界。"""
-                    operation(self._repository())
+                    """在单一同步事务中执行写入，不让 ORM 对象越过 worker 边界。"""
+                    self._sync_transaction(
+                        lambda session: operation(self._repository(session))
+                    )
 
                 await self._async_executor.run(execute)
         finally:
             self._pending_writes -= 1
-            record_metric("agent.chat.persistence.pending", self._pending_writes)
+            remaining = self._pending_by_session.get(session_id, 1) - 1
+            if remaining:
+                self._pending_by_session[session_id] = remaining
+            else:
+                self._pending_by_session.pop(session_id, None)
+            if current is not None:
+                self._active_tasks.discard(current)
+            record_metric("agent.chat.persistence.pending", -1)
+
+    async def shutdown(self) -> None:
+        """拒绝新写入并等待当前会话锁和 worker 操作取得终态。"""
+        self._closing = True
+        current = asyncio.current_task()
+        tasks = tuple(task for task in self._active_tasks if task is not current)
+        if tasks:
+            await asyncio.gather(*tasks, return_exceptions=True)
 
     async def async_append_display_messages(
         self,

app/db/oper/agentchat.py

@@ -269,7 +269,8 @@ class AgentChatOper(DbOper):
         )
         if not chat:
             return None
-        display_messages = self._normalize_messages(chat.display_messages)
+        # JSON 列不是 MutableList；必须复制旧列表，原地 extend 会让 SQLAlchemy 误认为字段未变化。
+        display_messages = list(self._normalize_messages(chat.display_messages))
         display_messages.extend(self._normalize_messages(messages))
         title = chat.title if self.has_custom_title(chat.title) else None
         return self.save_display_messages(

app/startup/context.py

@@ -6,6 +6,7 @@ from typing import Protocol
 
 from app.application.messaging.chat import (
     AsyncAgentChatRepository,
+    AgentChatPersistenceService,
     AsyncUnitOfWork,
 )
 from app.application.outbox import AsyncOutboxTransaction
@@ -113,6 +114,7 @@ class AgentChatRuntime:
     async_session: AsyncSessionProvider
     repository: AgentChatRepositoryFactory
     transaction: AsyncUnitOfWorkFactory
+    persistence: AgentChatPersistenceService
 
 
 @dataclass(frozen=True, slots=True)

app/startup/modules_initializer.py

@@ -67,7 +67,9 @@ from app.application.messaging.chat import (
     AgentChatService,
     configure_agent_chat_persistence,
     configure_agent_chat_service,
+    get_configured_agent_chat_persistence,
 )
+from app.application.messaging.agent import shutdown_web_agent_background_tasks
 from app.application.security.user import configure_user_lookups
 from app.application.security.auth import AuthService, configure_auth_service
 from app.application.security.passkeys import PasskeyService, configure_passkey_service
@@ -576,6 +578,11 @@ async def stop_modules():
     await run_step("消息服务", stop_message)
     await run_step("Redis缓存连接", lambda: RedisHelper().close())
     await run_step("异步Redis缓存连接", lambda: AsyncRedisHelper().close())
+    await run_step("Web Agent后台任务", shutdown_web_agent_background_tasks)
+    await run_step(
+        "Agent会话持久化",
+        lambda: get_configured_agent_chat_persistence().shutdown(),
+    )
     await run_step("数据库任务", stop_database_worker)
     if _database_worker is None:
         await run_step("数据库连接", close_database)
@@ -642,11 +649,18 @@ async def init_modules() -> HostRuntime:
         chain=lambda: build_chain_runtime_config(settings),
     )
     runtime_settings = _build_runtime_settings_service()
+    agent_chat_persistence = AgentChatPersistenceService(
+        repository=lambda session: AgentChatOper(session),
+        async_executor=database_worker,
+        sync_transaction=transaction_runner.sync,
+        capacity=database_worker.snapshot().capacity,
+    )
     host_runtime = HostRuntime(
         agent_chat=AgentChatRuntime(
             async_session=get_async_db,
             repository=AgentChatOper,
             transaction=SqlAlchemyAsyncUnitOfWork,
+            persistence=agent_chat_persistence,
         ),
         persistence=PersistenceRuntime(
             sync_session=get_db,
@@ -714,13 +728,7 @@ async def init_modules() -> HostRuntime:
     )
     configure_database_governance(build_database_governance())
     configure_agent_chat_service(AgentChatService(repository=AgentChatOper()))
-    configure_agent_chat_persistence(
-        AgentChatPersistenceService(
-            repository=AgentChatOper,
-            async_executor=database_worker,
-            capacity=database_worker.snapshot().capacity,
-        )
-    )
+    configure_agent_chat_persistence(agent_chat_persistence)
     configure_user_lookups(
         by_id=lambda user_id: UserOper().get_by_id(user_id),
         by_name=lambda username: UserOper().get_by_name(username),