feat：优化流式输出，增加临时多轮对话记录

app/prompt/intent.py

@@ -9,7 +9,7 @@ def build_intent_question() -> ChoiceDecision:
         instructions=(
             "Classify the user's primary intent. "
             "Classify the latest user answer in `clarification_history`, or `prompt` when there is no history. "
-            "Use earlier turns and the selected option to resolve references; the latest answer overrides earlier conflicting intent. "
+            "Use earlier turns, `history` (previous turns of this conversation) and the selected option to resolve references; the latest answer overrides earlier conflicting intent. "
             "Once the requested action or response is clear, earlier ambiguity does not require another confirmation. "
             "Choose modify_screen only when the user explicitly requests an actual edit. "
             "Choose qa for clear requests for a textual answer or general conversation, including information, explanation, diagnosis, advice, and assistant identity or capabilities. "

app/prompt/qa.py

@@ -5,7 +5,8 @@ QA_SYSTEM_PROMPT = """## 任务
 也可回答助手身份、能力等一般交流请求；介绍自己时以大屏助手的职责和已知能力为依据，不编造模型身份或未提供的能力
 
 ## 回答依据
-结合当前问题、大屏快照和 clarification_history，以用户后续明确补充或纠正为准
+结合当前问题、大屏快照、history（此前各轮的用户输入与助手摘要）和 clarification_history，以用户后续明确补充或纠正为准
+理解“刚才那个”“再大一点”等指代时参考 history，但当前大屏快照才是最新事实
 仅在与请求相关时使用大屏快照；一般交流无需强行关联大屏或要求用户补充大屏信息
 区分快照事实与待验证的推测，证据不足时说明缺少的信息，不编造故障原因
 

app/prompt/requirement.py

@@ -49,10 +49,14 @@ REQUIREMENT_INPUT_TEMPLATE = """历史会话：
 
 
 def build_requirement_input(
-    context: InputContext, intent: Intent, *, history_text: str = "未提供"
+    context: InputContext, intent: Intent, *, history_text: str | None = None
 ) -> str:
-    """整理本次模型输入，历史会话由调用方加载后传入"""
+    """整理本次模型输入，历史会话默认取输入上下文中由服务端写入的记录"""
     data = context.model_dump(mode="json", exclude_unset=True)
+    # 会话历史
+    history = data.pop("history", [])
+    if history_text is None:
+        history_text = json.dumps(history, ensure_ascii=False) if history else "未提供"
     content = REQUIREMENT_INPUT_TEMPLATE.format(
         history_text=history_text,
         clarification_text=json.dumps(