diff --git a/pipeline_service/llm_bridge.py b/pipeline_service/llm_bridge.py
index 7f6695c..527e0e1 100644
--- a/pipeline_service/llm_bridge.py
+++ b/pipeline_service/llm_bridge.py
@@ -197,7 +197,10 @@ async def llm_call_msgs(messages: list, model: str = None, temperature: float =
         payload["_timeout"] = int(timeout)
     if session_id:
         payload["_session_id"] = session_id
+    # 客户端等待须覆盖端点侧预算 + 余量（同 llm_infer 模式），否则客户端先断连、
+    # 报模糊 TimeoutError 而非端点的结构化错误
     data = await _http_chat(payload, org_id or '0', user_id or '', model or '',
+                            timeout=(int(timeout) + 60 if timeout else 0),
                             project_id=project_id or '')
     try:
         return data["choices"][0]["message"]["content"]
@@ -225,6 +228,14 @@ async def llm_call_msgs_native(messages: list, tools: list = None, model: str =
         payload["tool_choice"] = "auto"
     if purpose:
         payload["_purpose"] = purpose
+    if timeout:
+        # 端点侧上游预算必须随调用方透传（2026-09-15 pbls design 三连超时根因）：
+        # 原来 native 只把 timeout 用作客户端等待，payload 不带 _timeout → 端点按
+        # 供应商端点配置（百炼 timeout=120）掐断上游，而 design 收尾大请求天然
+        # 72~115s，一半概率 120s 被杀 → 内部 3 重试全灭 → 任务 failed。
+        # 端点预算 = 客户端等待 - 60s 余量：让端点先掐并返回结构化错误（可分类
+        # 重试），而不是客户端 aiohttp 断连报模糊 TimeoutError。
+        payload["_timeout"] = max(int(timeout) - 60, 60)
     if session_id:
         payload["_session_id"] = session_id
     data = await _http_chat(payload, org_id or '0', user_id or '', model or '',