diff --git a/pipeline_service/llm_bridge.py b/pipeline_service/llm_bridge.py
index 769805a..e7ee630 100644
--- a/pipeline_service/llm_bridge.py
+++ b/pipeline_service/llm_bridge.py
@@ -105,11 +105,13 @@ def _no_llm_error(model_name=None, org_id=None) -> ValueError:
 
 
 async def llm_call(prompt: str, model: str = None, temperature: float = 0.7,
-                   org_id: str = None, user_id: str = None) -> str:
+                   org_id: str = None, user_id: str = None, purpose: str = '') -> str:
     """Call LLM and return text response.
 
     统一走模型治理推理 API（门禁链 + 双维度记账）。
     org_id 为空 = 系统级（'0'），与旧语义（不过滤机构）等价。
+    purpose='utility'：辅助任务（分类/选择/摘要），治理层按用途选模型链
+    （机构策略配置辅助模型优先），经 payload 的 _purpose 键透传到端点。
     """
     # 兼容旧优先级：harnessed_agent（若宿主加载了独立推理后端）
     try:
@@ -129,6 +131,8 @@ async def llm_call(prompt: str, model: str = None, temperature: float = 0.7,
         "messages": [{"role": "user", "content": prompt}],
         "temperature": temperature,
     }
+    if purpose:
+        payload["_purpose"] = purpose
     data = await _http_chat(payload, org_id or '0', user_id or '', model or '')
     try:
         return data["choices"][0]["message"]["content"]
@@ -143,9 +147,14 @@ async def call_llm(tenant_id: str, prompt: str, model: str = None, temperature:
 
 
 async def llm_call_msgs(messages: list, model: str = None, temperature: float = 0.7,
-                        org_id: str = None, user_id: str = None) -> str:
-    """Call LLM with full message array (system/user/assistant)."""
+                        org_id: str = None, user_id: str = None, purpose: str = '') -> str:
+    """Call LLM with full message array (system/user/assistant).
+
+    purpose='utility'：辅助任务（分类/选择/摘要），治理层按用途选模型链。
+    """
     payload = {"model": model or '', "messages": messages, "temperature": temperature}
+    if purpose:
+        payload["_purpose"] = purpose
     data = await _http_chat(payload, org_id or '0', user_id or '', model or '')
     try:
         return data["choices"][0]["message"]["content"]
@@ -156,17 +165,20 @@ async def llm_call_msgs(messages: list, model: str = None, temperature: float =
 
 async def llm_call_msgs_native(messages: list, tools: list = None, model: str = None,
                                temperature: float = 0.7, org_id: str = None,
-                               user_id: str = None) -> dict:
+                               user_id: str = None, purpose: str = '') -> dict:
     """Native function calling. 传入 tools JSON schema，返回 message dict。
 
     Returns:
         {"content": str, "tool_calls": [{"id","type","function":{"name","arguments"}}]}
     当模型返回 tool_calls 时，content 通常为空字符串。
+    purpose='utility'：辅助任务（分类/选择/摘要），治理层按用途选模型链。
     """
     payload = {"model": model or '', "messages": messages, "temperature": temperature}
     if tools:
         payload["tools"] = tools
         payload["tool_choice"] = "auto"
+    if purpose:
+        payload["_purpose"] = purpose
     data = await _http_chat(payload, org_id or '0', user_id or '', model or '')
     try:
         msg = data["choices"][0]["message"]