diff --git a/pipeline_service/llm_bridge.py b/pipeline_service/llm_bridge.py
index 37124c7..f333856 100644
--- a/pipeline_service/llm_bridge.py
+++ b/pipeline_service/llm_bridge.py
@@ -110,7 +110,17 @@ async def _post_chat_completion(url: str, headers: dict, payload: dict) -> dict:
                             await asyncio.sleep(2 * (attempt + 1))
                             continue
                         raise err
-                    return await resp.json()
+                    data = await resp.json()
+                    # 网关偶发返回 200 但内容无 choices（错误 JSON），视为可重试的瞬时异常；
+                    # 不处理会导致上层 llm_call_msgs_native 的 data["choices"] 抛 KeyError。
+                    if not isinstance(data, dict) or "choices" not in data:
+                        err = ValueError("LLM 响应缺 choices: %s" % json.dumps(data, ensure_ascii=False)[:300])
+                        if attempt < _LLM_MAX_ATTEMPTS - 1:
+                            last_exc = err
+                            await asyncio.sleep(2 * (attempt + 1))
+                            continue
+                        raise err
+                    return data
         except (asyncio.TimeoutError, aiohttp.ClientError) as e:
             last_exc = e
             if attempt < _LLM_MAX_ATTEMPTS - 1: