Perfect! The changes are concise and completely address the user's request. Now I'll provide a final summary: Summary I successfully investigated and fixed the OpenAI BadRequestError with the Qwen3-8B model. Here's what I found and implemented: Root Cause The error

parameter.enable_thinking must be set to false for non-streaming calls

occurred because: 1. Qwen3-8B doesn't support reasoning effort - but thinking parameters were being passed through

completion_kwargs

2. LiteLLM auto-sets enable_thinking=true when it detects thinking parameters, causing OpenAI to reject non-streaming calls Solution Implemented I made targeted changes to `openhands/llm/llm.py`: 1. Filter thinking parameters for unsupported models: Added logic to remove

thinking

,

enable_thinking

, and

reasoning_effort

from

completion_kwargs

when the model doesn't support reasoning effort 2. Explicit enable_thinking=false for non-streaming calls: Added a safety check to explicitly set

enable_thinking=false

when thinking parameters are present in non-streaming calls Test Coverage Added • Model features test: Verifies Qwen3-8B correctly has

supports_reasoning_effort=False

• Parameter filtering test: Confirms thinking parameters are properly filtered while preserving other parameters Verification • ✅ Both new tests pass • ✅ All pre-commit hooks pass (linting, formatting, type checking) • ✅ Changes are minimal and focused on the specific issue • ✅ Fix prevents the BadRequestError while preserving functionality for reasoning-capable models The fix ensures users can have global thinking parameters in their configuration without breaking models like Qwen3-8B that don't support these features. Would you like me to push these changes to a new branch and create a pull request? _{View full conversation}