feat: change default value for legacy put_inner_thoughts_in_kwargs (#6395)

* feat: change default value for legacy put_inner_thoughts_in_kwargs

* update default in validation layer

fern/openapi.json

@@ -28873,7 +28873,7 @@
             ],
             "title": "Put Inner Thoughts In Kwargs",
             "description": "Puts 'inner_thoughts' as a kwarg in the function call if this is set to True. This helps with function calling performance and also the generation of inner thoughts.",
-            "default": true
+            "default": false
           },
           "handle": {
             "anyOf": [

letta/schemas/llm_config.py

@@ -55,7 +55,7 @@ class LLMConfig(BaseModel):
     model_wrapper: Optional[str] = Field(None, description="The wrapper for the model.")
     context_window: int = Field(..., description="The context window size for the model.")
     put_inner_thoughts_in_kwargs: Optional[bool] = Field(
-        True,
+        False,
         description="Puts 'inner_thoughts' as a kwarg in the function call if this is set to True. This helps with function calling performance and also the generation of inner thoughts.",
     )
     handle: Optional[str] = Field(None, description="The handle for this config, in the format provider/model-name.")
@@ -163,11 +163,11 @@ class LLMConfig(BaseModel):
         if model is None:
             return values
 
-        # Define models where we want put_inner_thoughts_in_kwargs to be False
+        # Default put_inner_thoughts_in_kwargs to False for all models
-        avoid_put_inner_thoughts_in_kwargs = ["gpt-4"]
+        # Reasoning models (o1, o3, o4, claude-sonnet-4, etc.) will have this set to False below
-
+        # Non-reasoner models should also default to False to avoid unwanted reasoning token generation
         if values.get("put_inner_thoughts_in_kwargs") is None:
-            values["put_inner_thoughts_in_kwargs"] = False if model in avoid_put_inner_thoughts_in_kwargs else True
+            values["put_inner_thoughts_in_kwargs"] = False
 
         # For the o1/o3 series from OpenAI, set to False by default
         # We can set this flag to `true` if desired, which will enable "double-think"

tests/integration_test_send_message.py

@@ -2391,11 +2391,9 @@ def test_inner_thoughts_false_non_reasoner_models(
         pytest.skip(f"Skipping test for reasoning model {model_handle}")
 
     # Note: This test is for models without reasoning, so model_settings should already have reasoning disabled
-    # We don't need to modify anything
 
     last_message_page = client.agents.messages.list(agent_id=agent_state.id, limit=1)
     last_message = last_message_page.items[0] if last_message_page.items else None
-    model_settings["put_inner_thoughts_in_kwargs"] = False
     agent_state = client.agents.update(agent_id=agent_state.id, model=model_handle, model_settings=model_settings)
     response = client.agents.messages.create(
         agent_id=agent_state.id,