I wanted to do some more testing around openhands and check out the latest version so I ran

uv tool upgrade openhands --python 3.12

and then

openhands serve

and thought I would try to ask a quick basic question to see how well openhands could take care of the request. I asked "I want to view this site in openhands. Do whatever changes are required to view this site and run it and let me know so I can see it." Seems like half the things openhands is attempting to do are responding with "Your last response did not include a function call or a message. Please use a tool to proceed with the task." and it ended up starting the nodejs/react project on port 3000 and could not close the app or run it on the correct port for openhands to see it. It also keeps telling me that the app is running on port 41981 and I'm on the container and checking and nothing is running on that port and it never stopped running on 3000. I'm using lm studio and

qwen/qwen3.5-35b-a3b

on a M3 max with 64gb memory. Curious what the possible issues could be and what I should do. Any advice is appreciated.

What is the context window set to in lm studio? We commonly see these errors with a short context window, so I'd suggest you use at least 32k

I just slid the slider to 86567 and using it for now.

Got it, what was it before?

No before. That was the first setting I used and the only setting I've tried. I just loaded this model for the test.

Hm, that should be big enough

For the port part, you often need to mention to use an externally accessible port or refer it to the work_hosts skill. In future OH upgrade to SDK 1.16.1 it will more likely pick correct port on its own as skill will be fully visible in context then (there is a skills bug in current version)

I'm ok with the llm not figuring out the port, I'm more confused on why it kept saying it checked port 3000 and it wasn't running on that port and was running on port 41981. Current Status: • Port 41981: ✅ Running with login page (username/password fields visible) • Port 3000: ❌ Not running (connection refused) The browser is currently showing the app on port 41981 with a login interface. When I'm on the container and I curl localhost:41981 I get a failed to connect but a correct output on localhost:3000. So curious why it seems to be so disconnected from the actual state and how it was attempting to verify what ports are running, as it's not verifying correctly.

I've had weaker models do same thing a few times. You might ask it to screenshot the page with browser tool and you can look at it in browser tab to confirm that way as well. That might convince it that it's confused once it sees no snapshot.

I'm trying to do some research to see if I could offload enough work to cover the costs of building an AI server. I've been pricing out building a server around the RTX Pro 6000 w/96gb ram and wanted to prove the concept, even if slow or using OpenRouter to test models I can't host today but could host with a more powerful server, to verify I can offload tasks I currently leverage Kiro to do today. Hopefully faster. Any suggestions on a better model I should be using?

Awesome, glad you figured it out! I should try it too