fix(litellm): scope the unset-effort responses bridge to constraint-enforcing endpoints

Chat-only OpenAI-compatible backends registered under the openai
provider with custom api_base and gpt-5.4+ model names served
tools-without-reasoning fine and have no /responses route, so the
unset-effort arm added for real OpenAI would have silently rerouted
previously working deployments. The arm now fires only when api_base is
unset (default OpenAI endpoint) or the provider is azure; an explicit
reasoning_effort keeps its pre-existing bridging behavior on any
api_base. Flagged lines also modernized to PEP 604

litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py

@@ -531,12 +531,9 @@ class LiteLLMResponseObjectHandler:
 
 
 def _should_convert_tool_call_to_json_mode(
-    tool_calls: Optional[
-        Union[
-            List[Union[ChatCompletionMessageToolCall, ChatCompletionMessageCustomToolCall]],
-            List[DatabricksTool],
-        ]
-    ] = None,
+    tool_calls: (
+        list[ChatCompletionMessageToolCall | ChatCompletionMessageCustomToolCall] | list[DatabricksTool] | None
+    ) = None,
     convert_tool_call_to_json_mode: Optional[bool] = None,
 ) -> bool:
     """

litellm/llms/openai/chat/gpt_transformation.py

@@ -533,9 +533,7 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig):
         for choice in choices:
             ## HANDLE JSON MODE - anthropic returns single function call]
             tool_calls = choice["message"].get("tool_calls", None)
-            new_tool_calls: Optional[
-                List[Union[ChatCompletionMessageToolCall, ChatCompletionMessageCustomToolCall]]
-            ] = None
+            new_tool_calls: list[ChatCompletionMessageToolCall | ChatCompletionMessageCustomToolCall] | None = None
             message_content = choice["message"].get("content", None)
             if tool_calls is not None:
                 _openai_tool_calls = []

litellm/main.py

@@ -986,6 +986,7 @@ def responses_api_bridge_check(
     tools: Optional[List[Any]] = None,
     reasoning_effort: Optional[Any] = None,
     reasoning_summary: Optional[Any] = None,
+    api_base: str | None = None,
 ) -> Tuple[dict, str]:
     model_info: Dict[str, Any] = {}
 
@@ -1030,6 +1031,12 @@ def responses_api_bridge_check(
     #   ``"none"`` keeps the request chat-servable. Custom (grammar) tools are served
     #   natively by Chat Completions with reasoning on, so custom-only requests stay on
     #   chat and keep their native custom tool_call response shape.
+    # - The UNSET-effort arm only fires against endpoints known to enforce that
+    #   constraint (the default OpenAI endpoint, or Azure OpenAI where api_base is
+    #   always set): chat-only OpenAI-compatible backends registered under the openai
+    #   provider with a custom api_base and gpt-5.4+ model names serve tools without
+    #   reasoning fine and have no /responses route, so they keep pre-existing
+    #   behavior (bridge only on an explicit reasoning_effort).
     # - Older GPT-5 names (e.g. ``gpt-5``, ``gpt-5.1``): bridge only when a reasoning
     #   summary alias is present with ``reasoning_effort`` (tools alone stay on chat).
     has_function_tool = any(
@@ -1040,6 +1047,7 @@ def responses_api_bridge_check(
         reasoning_active = reasoning_effort.get("effort") != "none" or reasoning_effort.get("summary") is not None
     else:
         reasoning_active = reasoning_effort != "none"
+    on_constraint_enforcing_endpoint = custom_llm_provider == "azure" or api_base is None
     if (
         custom_llm_provider in ("openai", "azure")
         and model_info.get("mode") != "responses"
@@ -1047,7 +1055,12 @@ def responses_api_bridge_check(
         and not OpenAIGPT5Config.is_model_gpt_5_search_model(model)
         and (
             (reasoning_effort is not None and reasoning_summary is not None)
-            or (OpenAIGPT5Config.is_model_gpt_5_4_plus_model(model) and has_function_tool and reasoning_active)
+            or (
+                OpenAIGPT5Config.is_model_gpt_5_4_plus_model(model)
+                and has_function_tool
+                and reasoning_active
+                and (reasoning_effort is not None or on_constraint_enforcing_endpoint)
+            )
         )
     ):
         model_info["mode"] = "responses"
@@ -5173,6 +5186,7 @@ def completion(  # type: ignore
             model=model,
             custom_llm_provider=custom_llm_provider,
             web_search_options=web_search_options,
+            api_base=api_base,
         )
 
         if not _should_allow_input_examples(custom_llm_provider=custom_llm_provider, model=model):
@@ -5412,6 +5426,7 @@ def completion(  # type: ignore
                 tools=tools,
                 reasoning_effort=reasoning_effort,
                 reasoning_summary=_reasoning_summary_for_bridge,
+                api_base=api_base,
             )
 
         # Use base_model (the true underlying model) for Azure model-type

tests/test_litellm/test_main.py

@@ -998,6 +998,64 @@ def test_responses_api_bridge_check_dict_effort_none_with_summary_routes_to_resp
     assert model_info.get("mode") == "responses"
 
 
+def test_responses_api_bridge_check_custom_api_base_with_unset_effort_stays_chat():
+    """
+    Chat-only OpenAI-compatible backends registered under the openai provider with a
+    custom api_base and gpt-5.4+ model names serve tools-without-reasoning fine and
+    have no /responses route; the unset-effort arm must not reroute them.
+    """
+    from litellm.main import responses_api_bridge_check
+
+    with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
+        mock_get_model_info.return_value = {"max_tokens": 128000}
+        model_info, model = responses_api_bridge_check(
+            model="gpt-5.6",
+            custom_llm_provider="openai",
+            tools=[{"type": "function", "function": {"name": "get_capital"}}],
+            reasoning_effort=None,
+            api_base="http://vllm.internal:8000/v1",
+        )
+
+    assert model == "gpt-5.6"
+    assert model_info.get("mode") != "responses"
+
+
+def test_responses_api_bridge_check_custom_api_base_with_explicit_effort_still_routes():
+    """Explicit reasoning_effort keeps its pre-existing bridging behavior on any api_base."""
+    from litellm.main import responses_api_bridge_check
+
+    with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
+        mock_get_model_info.return_value = {"max_tokens": 128000}
+        model_info, model = responses_api_bridge_check(
+            model="gpt-5.6",
+            custom_llm_provider="openai",
+            tools=[{"type": "function", "function": {"name": "get_capital"}}],
+            reasoning_effort="high",
+            api_base="http://vllm.internal:8000/v1",
+        )
+
+    assert model == "gpt-5.6"
+    assert model_info.get("mode") == "responses"
+
+
+def test_responses_api_bridge_check_azure_with_api_base_and_unset_effort_routes():
+    """Azure OpenAI always sets api_base and does enforce the constraint; keep bridging."""
+    from litellm.main import responses_api_bridge_check
+
+    with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
+        mock_get_model_info.return_value = {"max_tokens": 128000}
+        model_info, model = responses_api_bridge_check(
+            model="gpt-5.4",
+            custom_llm_provider="azure",
+            tools=[{"type": "function", "function": {"name": "get_capital"}}],
+            reasoning_effort=None,
+            api_base="https://myresource.openai.azure.com",
+        )
+
+    assert model == "gpt-5.4"
+    assert model_info.get("mode") == "responses"
+
+
 def test_responses_api_bridge_check_older_gpt_5_tools_without_reasoning_stays_chat():
     """Pre-5.4 GPT-5 names keep the old boundary: tools alone never bridge."""
     from litellm.main import responses_api_bridge_check