feat: add force_reasoning_effort litellm_params config

Add force_reasoning_effort as a per-deployment litellm_params field
that overrides reasoning_effort regardless of what the request sends.
Follows the same pattern as merge_reasoning_content_in_choices:
declared in GenericLiteLLMParams, captured in get_litellm_params(),
consumed downstream.

Override points for each entry path:
- /v1/chat/completions: main.py reads from litellm_params after
  get_litellm_params(), overrides optional_params["reasoning_effort"]
- /v1/messages → responses adapter: _build_responses_kwargs() pops
  from extra_kwargs, overrides reasoning.effort in responses format
- /v1/responses direct: responses/main.py pops from local_vars,
  overrides reasoning dict

Config example:
  litellm_params:
    model: github_copilot/gpt-5.4
    force_reasoning_effort: xhigh

Tested: all three endpoints with explicit low effort in request,
all overridden to xhigh with high reasoning_tokens in response.

litellm/litellm_core_utils/get_litellm_params.py

@@ -90,6 +90,7 @@ def get_litellm_params(
     async_call: Optional[bool] = None,
     ssl_verify: Optional[bool] = None,
     merge_reasoning_content_in_choices: Optional[bool] = None,
+    force_reasoning_effort: Optional[str] = None,
     use_litellm_proxy: Optional[bool] = None,
     api_version: Optional[str] = None,
     max_retries: Optional[int] = None,
@@ -149,6 +150,7 @@ def get_litellm_params(
         "async_call": async_call,
         "ssl_verify": ssl_verify,
         "merge_reasoning_content_in_choices": merge_reasoning_content_in_choices,
+        "force_reasoning_effort": force_reasoning_effort,
         "api_version": api_version,
         "max_retries": max_retries,
         "use_litellm_proxy": use_litellm_proxy,

litellm/llms/anthropic/experimental_pass_through/messages/handler.py

@@ -441,6 +441,18 @@ def anthropic_messages_handler(
             params=local_vars
         )
     )
+    force_reasoning_effort = litellm_params.force_reasoning_effort
+    if force_reasoning_effort:
+        output_config = anthropic_messages_optional_request_params.get("output_config") or {}
+        anthropic_messages_optional_request_params["output_config"] = {
+            **output_config,
+            "effort": force_reasoning_effort,
+        }
+        # Effort alone does not trigger reasoning — force-enable thinking too
+        thinking = anthropic_messages_optional_request_params.get("thinking")
+        if not isinstance(thinking, dict) or thinking.get("type") == "disabled":
+            anthropic_messages_optional_request_params["thinking"] = {"type": "adaptive"}
+
     return base_llm_http_handler.anthropic_messages_handler(
         model=model,
         messages=messages,

litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py

@@ -15,6 +15,7 @@ from litellm.types.llms.openai import ResponsesAPIResponse
 
 from .streaming_iterator import AnthropicResponsesStreamWrapper
 from .transformation import LiteLLMAnthropicToResponsesAPIAdapter
+from ..utils import is_reasoning_auto_summary_enabled
 
 _ADAPTER = LiteLLMAnthropicToResponsesAPIAdapter()
 
@@ -76,7 +77,8 @@ def _build_responses_kwargs(
         responses_kwargs["stream"] = True
 
     # Forward litellm-specific kwargs (api_key, api_base, logging obj, etc.)
-    excluded = {"anthropic_messages"}
+    excluded = {"anthropic_messages", "force_reasoning_effort"}
+    force_reasoning_effort = (extra_kwargs or {}).get("force_reasoning_effort")
     for key, value in (extra_kwargs or {}).items():
         if key == "litellm_logging_obj" and value is not None:
             from litellm.litellm_core_utils.litellm_logging import (
@@ -93,6 +95,17 @@ def _build_responses_kwargs(
         elif key not in excluded and key not in responses_kwargs and value is not None:
             responses_kwargs[key] = value
 
+    # Override reasoning effort if force_reasoning_effort is configured
+    if force_reasoning_effort:
+        reasoning = responses_kwargs.get("reasoning")
+        if isinstance(reasoning, dict):
+            responses_kwargs["reasoning"] = {**reasoning, "effort": force_reasoning_effort}
+        else:
+            responses_kwargs["reasoning"] = {
+                **({"summary": "detailed"} if is_reasoning_auto_summary_enabled() else {}),
+                "effort": force_reasoning_effort,
+            }
+
     return responses_kwargs
 
 

litellm/main.py

@@ -1577,6 +1577,7 @@ def completion(  # type: ignore # noqa: PLR0915
             merge_reasoning_content_in_choices=kwargs.get(
                 "merge_reasoning_content_in_choices", None
             ),
+            force_reasoning_effort=kwargs.get("force_reasoning_effort"),
             use_litellm_proxy=kwargs.get("use_litellm_proxy", False),
             api_version=api_version,
             azure_ad_token=kwargs.get("azure_ad_token"),
@@ -1592,6 +1593,11 @@ def completion(  # type: ignore # noqa: PLR0915
             tpm=kwargs.get("tpm"),
             rpm=kwargs.get("rpm"),
         )
+        # Override reasoning_effort if force_reasoning_effort is configured
+        force_reasoning_effort = litellm_params.get("force_reasoning_effort")
+        if force_reasoning_effort:
+            optional_params["reasoning_effort"] = force_reasoning_effort
+
         cast(LiteLLMLoggingObj, logging).update_environment_variables(
             model=model,
             user=user,

litellm/responses/main.py

@@ -862,6 +862,14 @@ def responses(
             if _mapped is not None:
                 reasoning = _mapped
                 local_vars["reasoning"] = _mapped
+        # Override reasoning effort if force_reasoning_effort is configured
+        force_reasoning_effort = local_vars.pop("force_reasoning_effort", None)
+        if force_reasoning_effort:
+            if isinstance(reasoning, dict):
+                reasoning = {**reasoning, "effort": force_reasoning_effort}
+            else:
+                reasoning = {"effort": force_reasoning_effort}
+            local_vars["reasoning"] = reasoning
         # Get ResponsesAPIOptionalRequestParams with only valid parameters
         response_api_optional_params: ResponsesAPIOptionalRequestParams = (
             ResponsesAPIRequestUtils.get_requested_response_api_optional_param(

litellm/types/router.py

@@ -203,6 +203,7 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
     use_litellm_proxy: Optional[bool] = False
     model_config = ConfigDict(extra="allow", arbitrary_types_allowed=True)
     merge_reasoning_content_in_choices: Optional[bool] = False
+    force_reasoning_effort: Optional[str] = None
     model_info: Optional[Dict] = None
     mock_response: Optional[Union[str, ModelResponse, Exception, Any]] = None