Hi all, I made some important adjustments that is already in the latest

main

1. I realized that the system message uses the default from openhands. The system prompt has been changed to a modified version of LocAgent’s (referring to their prompt from the appendix) 2. The VLLMEngine in SkyRL does not seem to utilize the reasoning parser, so reasoning_content is always null. Also, by default Qwen3's chat template actually detects thinking brackets and removes them. But in our implementation this does not occur. I don’t have a clean solution for this yet, so what I did was just subclasses

Agent

from openhands and manually parse the message in the

step

function. This works so now the training more reflects Qwen’s suggested usage. If you are have a training run already underway, restarting the training from scratch is fine. But we could keep it running for comparison as well.