fix(llm): remove default reasoning_effort; fix Gemini special case (#11567)

openhands/core/config/llm_config.py

@@ -179,10 +179,7 @@ class LLMConfig(BaseModel):
         if self.openrouter_app_name:
             os.environ['OR_APP_NAME'] = self.openrouter_app_name
 
-        # Set reasoning_effort to 'high' by default for non-Gemini models
-        # Gemini models use optimized thinking budget when reasoning_effort is None
-        if self.reasoning_effort is None and 'gemini-2.5-pro' not in self.model:
-            self.reasoning_effort = 'high'
+        # Do not set a default reasoning_effort. Leave as None unless user-configured.
 
         # Set an API version by default for Azure models
         # Required for newer models.

openhands/llm/async_llm.py

@@ -62,8 +62,11 @@ class AsyncLLM(LLM):
             elif 'messages' in kwargs:
                 messages = kwargs['messages']
 
-            # Set reasoning effort for models that support it
-            if get_features(self.config.model).supports_reasoning_effort:
+            # Set reasoning effort for models that support it, only if explicitly provided
+            if (
+                get_features(self.config.model).supports_reasoning_effort
+                and self.config.reasoning_effort is not None
+            ):
                 kwargs['reasoning_effort'] = self.config.reasoning_effort
 
             # ensure we work with a list of messages

openhands/llm/llm.py

@@ -155,7 +155,8 @@ class LLM(RetryMixin, DebugMixin):
                 # don't send reasoning_effort to specific Claude Sonnet/Haiku 4.5 variants
                 kwargs.pop('reasoning_effort', None)
             else:
-                kwargs['reasoning_effort'] = self.config.reasoning_effort
+                if self.config.reasoning_effort is not None:
+                    kwargs['reasoning_effort'] = self.config.reasoning_effort
             kwargs.pop(
                 'temperature'
             )  # temperature is not supported for reasoning models

openhands/llm/streaming_llm.py

@@ -64,8 +64,11 @@ class StreamingLLM(AsyncLLM):
                     'The messages list is empty. At least one message is required.'
                 )
 
-            # Set reasoning effort for models that support it
-            if get_features(self.config.model).supports_reasoning_effort:
+            # Set reasoning effort for models that support it, only if explicitly provided
+            if (
+                get_features(self.config.model).supports_reasoning_effort
+                and self.config.reasoning_effort is not None
+            ):
                 kwargs['reasoning_effort'] = self.config.reasoning_effort
 
             self.log_prompt(messages)

tests/unit/llm/test_llm.py

@@ -1143,13 +1143,19 @@ def test_gemini_model_keeps_none_reasoning_effort():
     assert config.reasoning_effort is None
 
 
-def test_non_gemini_model_gets_high_reasoning_effort():
-    """Test that non-Gemini models get reasoning_effort='high' by default."""
-    config = LLMConfig(model='gpt-4o', api_key='test_key')
-    # Non-Gemini models should get reasoning_effort='high'
+def test_non_gemini_model_explicit_reasoning_effort():
+    """Test that non-Gemini models get reasoning_effort ONLY if explicitly set."""
+    config = LLMConfig(model='gpt-4o', api_key='test_key', reasoning_effort='high')
     assert config.reasoning_effort == 'high'
 
 
+def test_non_gemini_model_default_reasoning_effort_none():
+    """Test that non-Gemini models do NOT get reasoning_effort by default after PR."""
+    config = LLMConfig(model='gpt-4o', api_key='test_key')
+    # Should be None by default after your change
+    assert config.reasoning_effort is None
+
+
 def test_explicit_reasoning_effort_preserved():
     """Test that explicitly set reasoning_effort is preserved."""
     config = LLMConfig(