Support for reasoning=ReasoningEffort(effort="none") to explicitly disable reasoning for non-openAI models

[okdolly]

Please read this first

Currently, when using models that differ from OpenAI's reasoning format (like Qwen3 etc.) through custom providers like LiteLLM, there's no way to explicitly pass reasoning_effort="none" to the underlying API. Setting ModelSettings(reasoning=None) is treated as a default/unset value rather than an explicit instruction to disable reasoning.

See Qwen API doc:

Feature Request

Could you add "none" to the enum?
https://github.com/openai/openai-python/blob/main/src/openai/types/shared_params/reasoning_effort.py

class Reasoning(BaseModel):
    effort: Optional[ReasoningEffort] = None

ReasoningEffort: TypeAlias = Optional[Literal["minimal", "low", "medium", "high"]]

Describe the question

When integrating non-OpenAI models through LiteLLM or other providers, some models require an explicit reasoning_effort="none" parameter to indicate that reasoning should be disabled. However:

1. ModelSettings(reasoning=None) doesn't pass any reasoning parameter to the provider
2. Reasoning(effort="none") is not a valid option (only "minimal", "low", "medium", "high" are allowed)
3. Using extra_body or extra_args to pass reasoning_effort can cause conflicts with LiteLLM's own parameter handling

Current Behavior

# This doesn't pass reasoning_effort to the provider at all
ModelSettings(reasoning=None)  # Treated as unset/default

# This raises a validation error
ModelSettings(reasoning=Reasoning(effort="none"))  # "none" is not a valid effort level

# This causes duplicate parameter errors in LiteLLM
ModelSettings(extra_body={"reasoning_effort": "none"})  # Conflicts with LiteLLM's reasoning_effort parameter

Expected Behavior

# Should explicitly pass reasoning_effort="none" to the provider
ModelSettings(reasoning=Reasoning(effort="none"))

Debug information

• Agents SDK version: (e.g. v0.0.3)
• Python version (e.g. Python 3.10)

Repro steps

Ideally provide a minimal python script that can be run to reproduce the issue.

Expected behavior

A clear and concise description of what you expected to happen.

[seratch]

ModelSettings.reasoning.effort accepts only the ones supported by OpenAI, so we won't add none to it. That said, adding an extra option to support LiteLLM models may be worth considering.