更新：

backend/rpa_engine/douyin_im/service.py

@@ -1,5 +1,7 @@
 import asyncio
 import logging
+import os
+import random
 import time
 from typing import Awaitable, Callable, Optional
 
@@ -374,6 +376,22 @@ class DouyinImService:
         )
         await self._ws_client.start()
 
+        # 轮询错峰：多账号同时托管时，若所有账号按同一节奏轮询，请求会在同一
+        # 时刻叠峰。这里给每个账号随机相位偏移 + 每轮 ±20% 抖动，把请求摊平。
+        # WS 可用时轮询只是兜底，可以适当放缓（KEFU_IM_POLL_INTERVAL_SECONDS 可调）。
+        try:
+            poll_interval = float(os.getenv("KEFU_IM_POLL_INTERVAL_SECONDS", "") or 15)
+        except ValueError:
+            poll_interval = 15.0
+        poll_interval = max(5.0, poll_interval)
+        if has_ws:
+            poll_interval = max(poll_interval, 30.0)
+
+        # 首轮轮询前的随机延迟（相位偏移），批量启动时错开各账号的首波请求
+        await asyncio.sleep(random.uniform(0.5, min(10.0, poll_interval)))
+        if not self._running:
+            return
+
         try:
             await self._poll_conversations()
         except Exception as e:
@@ -387,10 +405,12 @@ class DouyinImService:
             )
 
         loop_count = 0
+        next_poll = time.monotonic() + poll_interval * random.uniform(0.8, 1.2)
         while self._running:
             loop_count += 1
             try:
-                if loop_count % 3 == 1:
+                if time.monotonic() >= next_poll:
+                    next_poll = time.monotonic() + poll_interval * random.uniform(0.8, 1.2)
                     await self._poll_conversations()
                 if loop_count % 6 == 1:
                     logger.info(f"IM direct tick #{loop_count} account={self.account_id}")