修复第三方报告追问事务和上下文

app/services/conversations.py

@@ -308,6 +308,11 @@ class ConversationService:
             yield {"type": "error", "message": "LLM 未返回有效内容"}
             return
 
+        if persist_reply is not None:
+            assistant_message = await persist_reply(reply, semantics)
+            yield {"type": "done", "data": {"assistantMessageId": assistant_message.id}}
+            return
+
         # Persist the complete assistant text before optional card enrichment.
         # A slow or failed RAG/score lookup must not lose the conversation turn.
         assistant_message = ConversationMessage(
@@ -378,6 +383,8 @@ class ConversationService:
         conversation: Conversation,
         current_text: str,
         semantics: dict[str, Any],
+        *,
+        external_context: str | None = None,
     ) -> Sequence[dict[str, str]]:
         messages = reply_system_messages(semantics)
         messages.append(
@@ -390,6 +397,8 @@ class ConversationService:
         structured_context = ConversationService._structured_context(conversation, summary=is_summary)
         if structured_context:
             messages.append({"role": "system", "content": structured_context})
+        if external_context:
+            messages.append({"role": "system", "content": external_context})
         if is_summary:
             messages.append(
                 {