diff --git a/deploy/profile-scripts/llm_client.py b/deploy/profile-scripts/llm_client.py
index f86c932b..85fb3495 100644
--- a/deploy/profile-scripts/llm_client.py
+++ b/deploy/profile-scripts/llm_client.py
@@ -216,7 +216,7 @@ def gateway_alive(timeout=5):
 
 
 def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
-             retries=1, backoff=20, system=None):
+             retries=1, backoff=20, system=None, concurrent=False):
     """调用 LLM：OCG 直连优先，hermes gateway 兜底。带重试和结构化日志。
 
     Args:
@@ -227,6 +227,8 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
         retries: 每个通道的失败重试次数
         backoff: 重试间隔（秒）
         system: 可选 system message
+        concurrent: 并发模式（2026-08-13）：True 时 OCG router 用 round-robin 锁 key 轮询，
+                    避免并发压同一 key（header X-OCG-Mode: concurrent）
 
     Returns:
         {ok, content, error, model, elapsed, attempts, channel}
@@ -251,11 +253,15 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
     # key池直连保留为 fallback（router 不可达时回退）。
     channels = []
     # OCG 路由代理第一（2026-08-13 老莫：重评配置到 OCG 路由代理，支持pro升级+多key池）
-    channels.append(("ocg_router", OCG_ROUTER, {
+    _ocg_headers = {
         "Content-Type": "application/json",
         "Authorization": OCG_ROUTER_AUTH,
         "User-Agent": "curl/8.5.0",
-    }))
+    }
+    # 2026-08-13 并发模式：concurrent=True 时加 header，ocg_router round-robin 锁 key 轮询
+    if concurrent:
+        _ocg_headers["X-OCG-Mode"] = "concurrent"
+    channels.append(("ocg_router", OCG_ROUTER, _ocg_headers))
     # SenseNova 快通道（并发应急，不支持 pro 升级）
     if SN_HEADERS:
         channels.append(("sensenova", SN_URL, SN_HEADERS))