SDK sends unsupported reasoning_effort to MiniMax M3

Summary

Switching an OpenHands conversation/profile to MiniMax M3 can fail on the next LLM call because the SDK sends the default reasoning_effort=high parameter to minimax/MiniMax-M3.

LiteLLM rejects that request with:

litellm.UnsupportedParamsError: minimax does not support parameters: ['reasoning_effort'], for model=MiniMax-M3

Reproduction

1. Start a conversation with a model that can use the switch-model/profile tool, for example deepseek-v4-flash.

2. Ask it to switch to the minimax-m3 profile and continue the task, e.g.:

   hey switch to minimax-m3 then tell me a fun fact about today
   

3. The switch succeeds, e.g. the UI shows:

   Switched to profile openhands_minimax-m3
   

4. The next model call errors because reasoning_effort is sent to MiniMax M3.

This is reproducible independently of the Model Router; the switch-model/profile tool is enough to expose it.

Expected behavior

MiniMax M3 is a valid model/profile target. Switching to it should produce a valid request shape, and the SDK should not send provider-invalid default parameters.

Likely cause

The SDK currently treats generic LiteLLM/model metadata supports_reasoning: true as equivalent to support for the OpenAI-compatible reasoning_effort parameter. MiniMax M3 appears to support reasoning/thinking features, but LiteLLM's MiniMax adapter does not accept the reasoning_effort parameter; it supports other thinking-related parameters instead.

The SDK should distinguish:

• supports_reasoning: model has some reasoning/thinking capability
• supports_reasoning_effort: safe to send the exact OpenAI-style reasoning_effort parameter

Suggested fix

Only send reasoning_effort when the exact parameter is known to be supported, such as via explicit capability_overrides.supports_reasoning_effort, model_info.supports_reasoning_effort, model_info.supported_openai_params containing reasoning_effort, LiteLLM get_supported_openai_params(...) containing reasoning_effort, or an SDK explicit allowlist.

Generic model_info.supports_reasoning alone should not cause reasoning_effort to be sent.

This issue was created by an AI agent (OpenHands) on behalf of Juan Michelini.

Search the SDK's model capability handling and request construction for supports_reasoning, reasoning_effort, capability_overrides, and supported_openai_params; the issue does not name specific files or tests. Reproduce the profile switch to minimax-m3, then verify that MiniMax M3 requests no longer include reasoning_effort while models with explicit support still can use it.