tune xiaoniu reply brevity and flow thresholds

plugins/ai_auto_response/main.py

@@ -248,6 +248,8 @@ class AIAutoResponsePlugin(MessagePluginInterface):
             )
             return False, "empty_response"
 
+        response = self._finalize_reply(response, reply_mode)
+
         await bot.send_text_message(room_id, response, sender)
         self.last_reply_at[room_id] = time.time()
         self.flow_manager.note_bot_reply(room_id)
@@ -314,7 +316,7 @@ class AIAutoResponsePlugin(MessagePluginInterface):
         return (current_ts - last_room_reply) >= room_cd
 
     def _build_user_prompt(self, context: Dict, memory_hints: Dict) -> str:
-        recent_text = "\n".join(context.get("recent_messages", [])[-20:]) or "暂无"
+        recent_text = "\n".join(context.get("recent_messages", [])[-8:]) or "暂无"
         reply_mode = context.get("reply_mode", "social_short")
         length_rule = self._build_length_rule(reply_mode)
         return (
@@ -333,6 +335,11 @@ class AIAutoResponsePlugin(MessagePluginInterface):
             f"4. 如果信息不足，不要硬编。\n"
             f"5. 输出最终可直接发到群里的内容，不要解释你的思路。\n"
             f"6. {length_rule}\n"
+            f"7. 优先直接回应“当前发言”本身，不要被较早上下文带跑。\n"
+            f"8. 成员记忆和向量召回只有在与当前问题直接相关时才允许使用，否则忽略。\n"
+            f"9. 如果你不确定自己是否理解对了，就宁可不展开，只回很短。\n"
+            f"10. 把这次回复当作真人聊天里的第一反应，先只给第一层结论，不要主动补第二层解释。\n"
+            f"11. 如果一句话已经够了，就立刻停，不要为了完整而补充。\n"
         )
 
     @staticmethod
@@ -343,16 +350,42 @@ class AIAutoResponsePlugin(MessagePluginInterface):
         response = re.sub(r"\n{3,}", "\n\n", response)
         return response[:500].strip()
 
+    def _finalize_reply(self, response: str, reply_mode: str) -> str:
+        text = (response or "").strip()
+        if not text:
+            return ""
+        text = re.sub(r"\s+", " ", text)
+        text = text.replace("\n", " ").strip()
+
+        if reply_mode == "social_short":
+            text = self._take_first_sentence(text, 12)
+        elif reply_mode == "qa_fast":
+            text = self._take_first_sentence(text, 28)
+        elif reply_mode == "qa_with_context":
+            text = self._take_first_sentence(text, 36)
+        else:
+            text = self._take_first_sentence(text, 24)
+        return text.strip()
+
     @staticmethod
     def _build_length_rule(reply_mode: str) -> str:
         if reply_mode == "social_short":
-            return "默认只回一句短话，最好控制在2到12个字，除非非常不自然。"
+            return "默认只回一句短话，最好控制在2到8个字，除非非常不自然。"
         if reply_mode == "qa_fast":
-            return "尽量只回1句话，必要时最多2句，先给结论，不要展开成长教程。"
+            return "尽量只回1句话，总长度优先控制在28字内，先给结论，不要主动补解释。"
         if reply_mode == "qa_with_context":
-            return "优先控制在1到2句，除非对方明显在等详细步骤。"
+            return "优先控制在1句话，必要时最多2句，总长度优先控制在36字内，只给第一层答案。"
         return "尽量短，像群友临时接一句，不要长篇大论。"
 
+    @staticmethod
+    def _take_first_sentence(text: str, limit: int) -> str:
+        parts = re.split(r"(?<=[。！？!?；;])", text)
+        first = parts[0].strip() if parts and parts[0].strip() else text.strip()
+        if len(first) <= limit:
+            return first
+        clipped = first[:limit].rstrip("，,、；;：:")
+        return clipped
+
 
     def _sync_member_memory(self, room_id: str, sender: str, sender_name: str, member_context: Dict) -> None:
         if not member_context: