newtw66

core/orchestrator.py

@@ -37,6 +37,18 @@ _OUTBOUND_BLOCK_MARKERS = (
     '[{"name":',
 )
 
+# 历史记录格式检测模式（AI 转述历史时容易泄露）
+_HISTORY_LEAK_PATTERNS = [
+    r'\[\d{4}-\d{2}-\d{2}[^\]]*\]\s*(客户|客服)[：:]',  # [2026-03-07 12:00:00] 客户：
+    r'\[\d{2}:\d{2}:\d{2}\]\s*(客户|客服|我)[：:]',       # [12:00:00] 客户：
+    r'(根据|查看|查询|翻看)(历史|聊天|对话)(记录|内容)',   # 根据历史记录
+    r'历史(记录|对话|消息)(显示|表明|中)',                # 历史记录显示
+    r'之前的(聊天|对话|记录)(中|里|显示)',               # 之前的聊天中
+    r'共\d+条(历史|对话)?消息',                          # 共30条历史消息
+    r'订单号[:：]\s*\d{10,}',                            # 订单号:xxxxxxxxxx
+    r'(状态|金额|数量)[:：].*(状态|金额|数量)[:：]',     # 状态:xxx 金额:xxx 连续出现
+]
+
 class SystemOrchestrator:
     """
     全系统总编排：具备转接冷却、防抖合并、多消息去重、以及精准日志。
@@ -84,6 +96,11 @@ class SystemOrchestrator:
         if any(marker in cleaned for marker in _OUTBOUND_BLOCK_MARKERS):
             logger.warning("[Orchestrator] 拦截到内部内容外发，替换为安全兜底回复")
             return "我在帮你看记录，稍等哈"
+        # 检查历史记录泄露模式
+        for pattern in _HISTORY_LEAK_PATTERNS:
+            if re.search(pattern, cleaned):
+                logger.warning(f"[Orchestrator] 检测到历史记录泄露模式: {pattern[:30]}...")
+                return "我在帮你看记录，稍等哈"
         return cleaned
 
     async def on_raw_message_received(self, platform: str, raw_data: dict):
@@ -243,11 +260,17 @@ class SystemOrchestrator:
 
     async def _debounced_process(self, session_key: str, user_id: str, platform: str):
         try:
+            # 记录开始时间（防抖前）
+            process_start = time.time()
+            
             await asyncio.sleep(self._debounce_seconds)
             async with self._get_user_lock(session_key):
                 messages = self._pending_messages.pop(session_key, [])
                 if not messages: return
                 
+                debounce_elapsed = time.time() - process_start
+                logger.info(f"[计时] user={user_id} 防抖等待完成: {debounce_elapsed:.1f}s")
+                
                 # A. 合并与元数据修复（去重：同一防抖窗口内完全相同的内容只保留一条）
                 seen_contents = set()
                 unique_parts = []
@@ -278,9 +301,12 @@ class SystemOrchestrator:
                 )
 
                 # B. 持久化
+                db_start = time.time()
                 db_content = combined_content
                 if all_image_urls: db_content = f"【系统：已收到{len(all_image_urls)}张图】\n{combined_content}"
                 await repo.save_chat(platform, user_id, db_content, "in", acc_id=acc_id, image_urls=all_image_urls)
+                db_elapsed = time.time() - db_start
+                logger.info(f"[计时] user={user_id} 消息入库: {db_elapsed:.2f}s")
                 
                 # B2. 后台图片分析（不阻塞主流程，用于数据标定）
                 if all_image_urls:
@@ -302,9 +328,17 @@ class SystemOrchestrator:
                     )
                 else:
                     # D. 正常流程：调用AI思考
+                    history_start = time.time()
                     history = await repo.get_chat_history(user_id, limit=10, acc_id=acc_id)
                     if history and history[-1].get('content') == db_content: history = history[:-1]
+                    history_elapsed = time.time() - history_start
+                    logger.info(f"[计时] user={user_id} 查询历史: {history_elapsed:.2f}s (共{len(history)}条)")
+                    
+                    ai_start = time.time()
                     std_res = await self.brain.think_and_reply(final_msg, history=history)
+                    ai_elapsed = time.time() - ai_start
+                    total_elapsed = time.time() - process_start
+                    logger.info(f"[计时] user={user_id} AI思考: {ai_elapsed:.1f}s | 总耗时: {total_elapsed:.1f}s")
 
                 # E. 发送并记录时间
                 if std_res.should_reply: