diff --git a/backend/app/chains/agents/base.py b/backend/app/chains/agents/base.py
index d46c71db..339678da 100644
--- a/backend/app/chains/agents/base.py
+++ b/backend/app/chains/agents/base.py
@@ -130,6 +130,18 @@ def _load_json_like(text: str) -> Any:
     raise ValueError("Empty output from LLM")
 
 
+def _thinking_extra_body(model: BaseChatModel, enable_thinking: bool) -> dict[str, Any] | None:
+    """Extra body toggling Qwen-style thinking for OpenAI-compatible endpoints.
+
+    `enable_thinking` is a DashScope extension; the official OpenAI API rejects
+    it as an unknown parameter, so nothing is bound for api.openai.com targets.
+    """
+    base = str(getattr(model, "openai_api_base", "") or "").rstrip("/")
+    if not base or "api.openai.com" in base:
+        return None
+    return {"enable_thinking": enable_thinking}
+
+
 class AgentBase(ABC, Generic[T]):
     """通用 Agent 基类：子类固化 prompt_template 与 output_model。"""
 
@@ -143,7 +155,9 @@ def __init__(
         agent_kwargs: dict[str, Any] | None = None,
     ) -> None:
         self._model = model
-        self._model.bind(extra_body={"enable_thinking": self.enable_thinking})
+        thinking_body = _thinking_extra_body(model, self.enable_thinking)
+        if thinking_body is not None:
+            self._model.bind(extra_body=thinking_body)
         self._structured_output_method = structured_output_method
         self._agent_kwargs = dict(agent_kwargs or {})
         self._structured_chain: Runnable | None = None
diff --git a/backend/app/services/llm/resolver.py b/backend/app/services/llm/resolver.py
index a0792ac0..e403f0cd 100644
--- a/backend/app/services/llm/resolver.py
+++ b/backend/app/services/llm/resolver.py
@@ -178,8 +178,16 @@ def _build_chat_openai_model(
         kwargs.setdefault("base_url", base_url)
 
     if not thinking:
-        extra_body = dict(kwargs.get("extra_body") or {})
-        extra_body["enable_thinking"] = False
-        kwargs["extra_body"] = extra_body
+        # DashScope-style OpenAI-compatible endpoints take the Qwen
+        # `enable_thinking` extension, but the official OpenAI API rejects it
+        # with 400 Unknown parameter; disable reasoning through the official
+        # `reasoning_effort` parameter there instead.
+        base = (base_url or "").rstrip("/")
+        if not base or "api.openai.com" in base:
+            kwargs.setdefault("reasoning_effort", "none")
+        else:
+            extra_body = dict(kwargs.get("extra_body") or {})
+            extra_body["enable_thinking"] = False
+            kwargs["extra_body"] = extra_body
 
     return ChatOpenAI(**kwargs)
diff --git a/backend/app/services/llm/runtime.py b/backend/app/services/llm/runtime.py
index 9bbb2116..3eb9367e 100644
--- a/backend/app/services/llm/runtime.py
+++ b/backend/app/services/llm/runtime.py
@@ -69,8 +69,16 @@ def build_default_text_llm_sync(
         kwargs.setdefault("base_url", base_url)
 
     if not thinking:
-        extra_body = dict(kwargs.get("extra_body") or {})
-        extra_body["enable_thinking"] = False
-        kwargs["extra_body"] = extra_body
+        # DashScope-style OpenAI-compatible endpoints take the Qwen
+        # `enable_thinking` extension, but the official OpenAI API rejects it
+        # with 400 Unknown parameter; disable reasoning through the official
+        # `reasoning_effort` parameter there instead.
+        base = (base_url or "").rstrip("/")
+        if not base or "api.openai.com" in base:
+            kwargs.setdefault("reasoning_effort", "none")
+        else:
+            extra_body = dict(kwargs.get("extra_body") or {})
+            extra_body["enable_thinking"] = False
+            kwargs["extra_body"] = extra_body
 
     return ChatOpenAI(**kwargs)