feat：优化流式输出，增加临时多轮对话记录

app/services/conversation.py

@@ -0,0 +1,71 @@
+"""会话存储：创建时即占位，用于校验 ID 由后端签发；当前为进程内存实现"""
+
+from datetime import UTC, datetime, timedelta
+from typing import Protocol
+from uuid import uuid4
+
+from app.config import settings
+from app.schemas.conversation import Turn
+
+
+class ConversationStore(Protocol):
+    """后续接入 MongoDB 时实现同一组方法，业务代码不需要改动"""
+
+    async def create(self) -> str: ...
+
+    async def exists(self, conversation_id: str) -> bool: ...
+
+    async def load_turns(
+        self, conversation_id: str, limit: int | None
+    ) -> list[Turn]: ...
+
+    async def save_turn(self, turn: Turn) -> None: ...
+
+
+def _now() -> datetime:
+    return datetime.now(UTC)
+
+
+class InMemoryConversationStore:
+    def __init__(self, empty_ttl: int | None = None):
+        ttl = empty_ttl if empty_ttl is not None else settings.conversation_empty_ttl
+        self._empty_ttl = timedelta(seconds=ttl)
+        # 会话 ID -> 创建时间；轮次按 thread_id 去重并保持首次写入顺序
+        self._created: dict[str, datetime] = {}
+        self._turns: dict[str, dict[str, Turn]] = {}
+
+    def _purge_expired(self) -> None:
+        # 只清理没有任何轮次且超过保留时长的空会话，有对话的会话不受影响
+        deadline = _now() - self._empty_ttl
+        for conversation_id in [
+            key
+            for key, created in self._created.items()
+            if created < deadline and not self._turns.get(key)
+        ]:
+            del self._created[conversation_id]
+            self._turns.pop(conversation_id, None)
+
+    async def create(self) -> str:
+        self._purge_expired()
+        conversation_id = str(uuid4())
+        self._created[conversation_id] = _now()
+        return conversation_id
+
+    async def exists(self, conversation_id: str) -> bool:
+        self._purge_expired()
+        return conversation_id in self._created
+
+    async def load_turns(self, conversation_id: str, limit: int | None) -> list[Turn]:
+        turns = list(self._turns.get(conversation_id, {}).values())
+        turns.sort(key=lambda turn: turn.created_at)
+        if limit is None:
+            return turns
+        # 只取最近 limit 轮，limit 为 0 时不带历史
+        return turns[-limit:] if limit else []
+
+    async def save_turn(self, turn: Turn) -> None:
+        # 同一任务再次写入（澄清恢复后）保留首次创建时间，只更新内容
+        existing = self._turns.setdefault(turn.conversation_id, {}).get(turn.thread_id)
+        if existing is not None:
+            turn = turn.model_copy(update={"created_at": existing.created_at})
+        self._turns[turn.conversation_id][turn.thread_id] = turn