I really love how you can send instructions to the agent while its runnings, and it will adjust on the fly I wonder if there's a way to make that more obvious, I've been using openhands for many months and only found out 3 weeks ago 😅

is this available in CLI Mode as well? Sometimes I see that the agent tell me "we need to edit" some file, and I want to interrupt it and just let me tell about the issue, not fix it without providing a draft

I think in CLI mode you type ctrl-p

^^ We can Pause, and tell it whatever we want. If it started to edit, it will finish that exact action before it pauses.

it is not obvious if ctrl-p is available at this step I want add additional instructions AND proceed with yes or no

Ah, at that step you can say N

Ah got it, thanks!!

@quick-window-91705 maybe the problem is that it’s a “stop” button in the UI?

We can also send instructions in the UI while the agent is running, without having to pause it Its a weird mental model at first but super convenient I dont have to press a button and wait for the agent to pause/stop

I find it is super convenient for me as well. At first I didn't know if the agent receives and acknowledges my messages while running, but then I noticed it did adjust the course

@nutritious-balloon-65098 Currently the “send message” button turns in to a “stop” button in the UI when the agent is running. A preferable interface would be to have a “send” button that is always active, and a “stop” button that is always visible, but disabled when the agent is running. Implement this, run the app (make sure to use port 12000), and share screenshots of both conditions.

I think the Stop button should be enabled while the agent is running, it stops the agent. Otherwise I’m not sure what its purpose is? thinky spin

haha, that was my mistake, I’ll fix it

I'm on it! Graham Neubig can track my progress here.

how about expanding the helpfulness feedback flow but for every tool call, thought, action, and add the ability to "annotate" or comment, in real time? I'd like the ability to tell OpenHands "what you just did was totally dumb, stop it" as well as "Clever, keep to this going forward" real-time steering? Not clear what should happen if I provide this kind of feedback before a condensation step(?)