fix(agent): serialize session persistence writes

app/application/messaging/chat.py

@@ -2,9 +2,11 @@
 
 from __future__ import annotations
 
+import asyncio
 from dataclasses import dataclass
 from collections.abc import Callable
 from typing import Any, Optional, Protocol
+from weakref import WeakValueDictionary
 
 from app.application.database import AsyncDatabaseExecutor
 from app.schemas.agent import AgentChatSessionDetail, AgentChatSessionSummary
@@ -350,17 +352,31 @@ class AgentChatPersistenceService:
         """保存同步仓储工厂和异步执行端口。"""
         self._repository = repository
         self._async_executor = async_executor
+        # 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。
+        self._session_locks: WeakValueDictionary[str, asyncio.Lock] = WeakValueDictionary()
+
+    def _session_lock(self, session_id: str) -> asyncio.Lock:
+        """返回当前进程内指定会话的写锁。"""
+        lock = self._session_locks.get(session_id)
+        if lock is None:
+            lock = asyncio.Lock()
+            self._session_locks[session_id] = lock
+        return lock
 
     async def _run_write(
         self,
+        session_id: str,
         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。"""
-        def execute() -> None:
+        async with self._session_lock(session_id):
-            """执行同步写入，不让 ORM 对象越过 worker 边界。"""
+            def execute() -> None:
-            operation(self._repository())
+                """执行同步写入，不让 ORM 对象越过 worker 边界。"""
+                operation(self._repository())
 
-        await self._async_executor.run(execute)
+            await self._async_executor.run(execute)
 
     async def async_append_display_messages(
         self,
@@ -376,6 +392,7 @@ class AgentChatPersistenceService:
     ) -> None:
         """异步追加展示消息，等待同步事务取得确定终态。"""
         await self._run_write(
+            session_id,
             lambda repository: repository.append_display_messages(
                 session_id=session_id,
                 user_id=user_id,
@@ -403,6 +420,7 @@ class AgentChatPersistenceService:
     ) -> None:
         """异步保存展示消息快照，实际写入由有界 worker 承接。"""
         await self._run_write(
+            session_id,
             lambda repository: repository.save_display_messages(
                 session_id=session_id,
                 user_id=user_id,
@@ -425,6 +443,7 @@ class AgentChatPersistenceService:
     ) -> None:
         """异步保存可恢复的原始消息。"""
         await self._run_write(
+            session_id,
             lambda repository: repository.save_agent_messages(
                 session_id=session_id,
                 user_id=user_id,
@@ -446,6 +465,7 @@ class AgentChatPersistenceService:
     ) -> None:
         """异步写入首次生成的会话标题。"""
         await self._run_write(
+            session_id,
             lambda repository: repository.update_title_if_empty(
                 session_id=session_id,
                 user_id=user_id,

tests/test_agent_chat_persistence.py

@@ -7,9 +7,12 @@ import threading
 from uuid import uuid4
 
 import pytest
+from sqlalchemy import delete, select
 
 from app.application.messaging.chat import AgentChatPersistenceService, AgentChatService
+from app.db.models.agentchat import AgentChat
 from app.db.oper.agentchat import AgentChatOper
+from app.db.session import SessionFactory, async_session_scope
 from app.db.worker import DatabaseWorker
 
 
@@ -153,3 +156,67 @@ async def test_agent_chat_persistence_uses_real_worker_and_sqlite_transaction()
             user_id="worker-user",
         )
         await worker.shutdown()
+
+
+@pytest.mark.asyncio
+async def test_agent_chat_persistence_serializes_same_session_writes() -> None:
+    """同一会话的首次创建和既有快照追加都必须串行。"""
+    worker = DatabaseWorker(max_workers=4, capacity=16)
+    await worker.start()
+    session_id = f"worker-race-{uuid4().hex}"
+    existing_session_id = f"worker-race-existing-{uuid4().hex}"
+    persistence = AgentChatPersistenceService(
+        repository=AgentChatOper,
+        async_executor=worker,
+    )
+
+    async def append(content: str) -> None:
+        await persistence.async_append_display_messages(
+            session_id=session_id,
+            user_id="worker-race-user",
+            messages=[{"role": "user", "content": content}],
+        )
+
+    async def append_existing(content: str) -> None:
+        await persistence.async_append_display_messages(
+            session_id=existing_session_id,
+            user_id="worker-race-user",
+            messages=[{"role": "user", "content": content}],
+        )
+
+    try:
+        await asyncio.gather(*(append(f"message-{index}") for index in range(4)))
+        await persistence.async_save_display_messages(
+            session_id=existing_session_id,
+            user_id="worker-race-user",
+            messages=[{"role": "user", "content": "seed"}],
+        )
+        await asyncio.gather(
+            *(append_existing(f"existing-{index}") for index in range(4))
+        )
+        async with async_session_scope() as session:
+            result = await session.execute(
+                select(AgentChat).where(
+                    AgentChat.session_id.in_((session_id, existing_session_id))
+                )
+            )
+            rows = list(result.scalars().all())
+        assert len(rows) == 2
+        row_by_session = {row.session_id: row for row in rows}
+        assert {
+            message["content"]
+            for message in row_by_session[session_id].display_messages
+        } == {f"message-{index}" for index in range(4)}
+        assert {
+            message["content"]
+            for message in row_by_session[existing_session_id].display_messages
+        } == {"seed"} | {f"existing-{index}" for index in range(4)}
+    finally:
+        with SessionFactory() as session:
+            session.execute(
+                delete(AgentChat).where(
+                    AgentChat.session_id.in_((session_id, existing_session_id))
+                )
+            )
+            session.commit()
+        await worker.shutdown()