fix: keep reasoning_effort for mode: responses bridge deployments

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/main.py

@@ -5398,6 +5398,16 @@ def completion(
         if dynamic_api_key is not None:
             api_key = dynamic_api_key
         # check if user passed in any of the OpenAI optional params
+        allowed_openai_params: Final[list[str] | None] = cast(
+            list[str] | None,
+            (
+                [*(kwargs.get("allowed_openai_params") or []), "reasoning_effort"]
+                if responses_api_model_info.get("mode") == "responses"
+                and not skip_responses_api_bridge
+                and "reasoning_effort" not in (kwargs.get("allowed_openai_params") or [])
+                else kwargs.get("allowed_openai_params")
+            ),
+        )
         optional_param_args: Final = {
             "functions": functions,
             "function_call": function_call,
@@ -5442,7 +5452,7 @@ def completion(
             "service_tier": service_tier,
             "store": store,
             "prompt_cache_key": prompt_cache_key,
-            "allowed_openai_params": kwargs.get("allowed_openai_params"),
+            "allowed_openai_params": allowed_openai_params,
             "base_model": base_model,
         }
         optional_params = get_optional_params(**optional_param_args, **non_default_params)

tests/test_litellm/test_main.py

@@ -1367,6 +1367,40 @@ def test_gpt_5_4_responses_bridge_preserves_reasoning_summary_dict(
     }
 
 
+@pytest.mark.parametrize("reasoning_effort", ["high", {"effort": "high"}])
+@patch("litellm.completion_extras.responses_api_bridge.completion")
+def test_responses_bridge_preserves_reasoning_effort_with_drop_params(
+    mock_responses_completion, reasoning_effort, restore_model_registry
+):
+    mock_responses_completion.return_value = MagicMock()
+    model = "perplexity/test-responses-bridge"
+    litellm.register_model(
+        {
+            model: {
+                "litellm_provider": "perplexity",
+                "mode": "responses",
+                "supports_reasoning": True,
+                "input_cost_per_token": 0.0,
+                "output_cost_per_token": 0.0,
+            }
+        },
+        persist_across_reloads=False,
+    )
+
+    with patch.object(litellm, "supports_reasoning", return_value=False):
+        litellm.completion(
+            model=model,
+            messages=[{"role": "user", "content": "hello"}],
+            reasoning_effort=reasoning_effort,
+            drop_params=True,
+            api_key="fake-key",
+            api_base="https://api.perplexity.ai",
+        )
+
+    optional_params = mock_responses_completion.call_args.kwargs["optional_params"]
+    assert optional_params["reasoning_effort"] == reasoning_effort
+
+
 @pytest.mark.parametrize(
     "model, model_info, expected_model_param, expected_base_model_param",
     [