I investigated the reported Qwen3.8-27B/vLLM issue with Agent Canvas. Using the supplied image and flags, direct tool calls worked. With OpenHands SDK 1.58.1, the default

reasoning_effort="high"

failed because the checkpoint accepts low/medium/xhigh. Setting it to medium made the agent loop work with streaming on and off. vLLM’s Rust frontend obscures the cause with `raise_exception is unknown`; its Python frontend reports the unsupported reasoning-effort value directly. Do we want to address this through custom-endpoint defaults, model capability validation, or configuration guidance? I have reproduction scripts and traces. I wasn't sure if there is a way to keep this from happening again or if this is just too narrow of an edge case.

of what are you speaking? which issues?

Sorry, agent copy and paste. I got an email support request saying they had an issue when trying to use Qwen/VLLM/Agent canvas and tracked it down to this reasoning effort parameter.

ok can you open an issue in the sdk (I suppose) with all the details? Thanks 🙂

Yup, can do