diff --git a/frontend b/frontend
index 1053f9c..5b1e145 160000
--- a/frontend
+++ b/frontend
@@ -1 +1 @@
-Subproject commit 1053f9c3fbb0857e4cc0a8acefc01d523a9f3c52
+Subproject commit 5b1e145ff77dce497dedeed6453b14f50a99987b
diff --git a/src/api/routers/configs.py b/src/api/routers/configs.py
index f249793..ac239aa 100644
--- a/src/api/routers/configs.py
+++ b/src/api/routers/configs.py
@@ -87,9 +87,22 @@ def test_llm_config() -> OkResponse[str]:
 
     try:
         client = LLMClient(runtime_cfg)
+        # 暴露实际路由参数，便于排查「连不上 / 报错」时定位是 model 前缀还是 base_url 问题
+        routed_kwargs = client._litellm_kwargs(temperature=0.7, max_tokens=None)
+        diagnosis = {
+            "route": {
+                "model": routed_kwargs.get("model"),
+                "custom_llm_provider": routed_kwargs.get("custom_llm_provider"),
+                "api_base": routed_kwargs.get("api_base"),
+                "is_custom": (runtime_cfg.type or "").strip().lower() == "custom",
+                "llm_format": runtime_cfg.llm_format,
+            },
+        }
         resp = client.call([{"role": "user", "content": "hello"}])
-        raw_json = json.dumps(_to_jsonable(resp.raw), ensure_ascii=False)
-        return OkResponse[str](data=raw_json)
+        diagnosis["reply"] = resp.content
+        diagnosis["usage"] = resp.usage
+        diagnosis["ok"] = True
+        return OkResponse[str](data=json.dumps(diagnosis, ensure_ascii=False))
     except Exception as ex:
         return OkResponse[str](
             success=False,
diff --git a/src/config.py b/src/config.py
index 2737866..4bd008d 100644
--- a/src/config.py
+++ b/src/config.py
@@ -138,6 +138,11 @@ class LLMConfig:
     base_url: Optional[str] = None
     api_version: Optional[str] = None
     azure_endpoint: Optional[str] = None
+    # API 协议格式：仅对「自定义 Provider」生效，用于路由到对应的上游端点。
+    #   "openai"     -> OpenAI 兼容（/v1/chat/completions）
+    #   "anthropic"  -> Anthropic 格式（/v1/messages）
+    # 为空时按 base_url 启发式推断（含 /anthropic 视作 anthropic，否则 openai）。
+    llm_format: Optional[str] = None
 
 
 @dataclass
diff --git a/src/llm/client.py b/src/llm/client.py
index fd43412..d6c5722 100644
--- a/src/llm/client.py
+++ b/src/llm/client.py
@@ -109,23 +109,52 @@ def __init__(self, config: LLMConfig):
         self.config = config
 
     def _litellm_kwargs(self, temperature: float = 0.7, max_tokens: Optional[int] = None, **kwargs) -> Dict:
-        """从 config 构建 litellm completion 参数，不依赖环境变量"""
-        model = self.config.model
-        if "/" not in model and self.config.provider:
-            model = f"{self.config.provider}/{model}"
+        """从 config 构建 litellm completion 参数，不依赖环境变量。
+
+        路由策略（单点收口）：
+        - 内置 provider（deepseek / anthropic / openai / azure ...）：使用 litellm 原生
+          `provider/model` 形式，由 litellm 自行解析 base_url / 鉴权。
+        - 自定义 provider：统一通过 ``custom_llm_provider`` 指定上游协议格式，
+          传 **裸模型名**（不带 provider 前缀）+ ``api_base``，避免把任意 provider ID
+          拼成 ``<custom-id>/<model>`` 这种 litellm 无法识别的前缀。
+          协议格式由 ``llm_format`` 决定（openai / anthropic），缺省时按 base_url 启发式。
+        """
+        model = (self.config.model or "").strip()
+        provider = (self.config.provider or "").strip()
+        is_custom = (self.config.type or "").strip().lower() == "custom"
+
         kw: Dict = {
-            "model": model,
             "temperature": temperature,
             "max_tokens": max_tokens,
         }
         if self.config.api_key:
             kw["api_key"] = self.config.api_key
-        # base_url：自建或自定义端点；Azure 时也可用 azure_endpoint
-        api_base = self.config.base_url or getattr(self.config, "azure_endpoint", None)
-        if api_base and self.config.type != "builtin":
-            kw["api_base"] = api_base
         if getattr(self.config, "api_version", None):
             kw["api_version"] = self.config.api_version
+
+        api_base = self.config.base_url or getattr(self.config, "azure_endpoint", None)
+
+        if is_custom:
+            # 自定义 Provider：不要拼 provider 前缀，用 custom_llm_provider 路由协议格式
+            fmt = (self.config.llm_format or "").strip().lower()
+            if not fmt:
+                # 启发式：base_url 含 /anthropic 视作 Anthropic 格式端点，否则 OpenAI 兼容
+                fmt = "anthropic" if "/anthropic" in (api_base or "").lower() else "openai"
+            fmt = fmt if fmt in ("openai", "anthropic") else "openai"
+            kw["model"] = model
+            kw["custom_llm_provider"] = fmt
+            if api_base:
+                kw["api_base"] = api_base
+        else:
+            # 内置 provider：litellm 原生 provider/model 形式
+            litellm_model = model
+            if "/" not in litellm_model and provider:
+                litellm_model = f"{provider}/{litellm_model}"
+            kw["model"] = litellm_model
+            # 自建/兼容端点：builtin 模式下仍允许覆盖 base_url（如自托管的 openai 网关）
+            if api_base:
+                kw["api_base"] = api_base
+
         kw.update(kwargs)
         return {k: v for k, v in kw.items() if v is not None}
 
diff --git a/src/schemas/config.py b/src/schemas/config.py
index 543b25c..c57ed22 100644
--- a/src/schemas/config.py
+++ b/src/schemas/config.py
@@ -29,6 +29,8 @@ class LLMConfigUpdate(BaseModel):
     LLM_model: Optional[str] = None
     LLM_baseurl: Optional[str] = None
     type: Optional[str] = None
+    # 自定义 Provider 的 API 协议格式：openai | anthropic
+    LLM_format: Optional[str] = None
 
 
 class CodeAgentConfigUpdate(BaseModel):
diff --git a/src/services/config_service.py b/src/services/config_service.py
index f10efbe..23b6d22 100644
--- a/src/services/config_service.py
+++ b/src/services/config_service.py
@@ -113,6 +113,7 @@ def get_llm_runtime_config() -> Optional[LLMConfig]:
         model=model,
         type=tp,
         base_url=cfg.get("LLM_baseurl") or None,
+        llm_format=(cfg.get("LLM_format") or "").strip().lower() or None,
     )