This message was deleted.
# proj-agent-canvas
d
This message was deleted.
w
The main flow is • it warms up for ~30s, then starts working • it comes up with a plan ◦ you click a button to approve the plan ◦ or it auto-approves after 5m ◦ I love this part--in particular it lets me decide ahead of time if we're going to add tests • it says "ready for review" • you can push to a branch • but weirdly not create a PR?
There's a repo view which I love, I can see all the work going on there
This is kind of bizzarre though 😅
my latest repo is chosen by default on the new project page. we should do that
c
Hah.
but weirdly not create a PR?
Gemini's default "safety" settings didn't let it create a PR in the past, because it was dangerous stuff or something. I think even with them at None, I had to fight some older models when I was testing them 😅 . I was telling them all the time that btw it's just a draft PR, we can do those lolsob
2.5 no longer has that behavior (at least on None). Apparently Jules didn't get the update 😅
lolsob 1
p
no moat
Oh, this would be nice in OpenHands. Or maybe I don't understand how OpenHands stops a request at the moment? I read this to mean "pause" as in like a breakpoint in the current run, so I, as a human can keep up. With an accompanying
Resume
button to continue
Ideally, we can tweak the trajectory before unpausing (i.e.: edit any prior instruction and or prior assistant response)
c
openhands has a Stop button in the web UI, which I used as a pause, since, well, it's a pause really, it doesn't permanently stop after all. It stops the agent loop, and we can "resume" for example by telling the agent "ok continue"
👍 1
p
any reason why it takes a long time to "resume", it feels like it kills the runtime container(?)
c
I don't know, I don't think so, but I rarely use docker lately. How are you running to see a long time?
p
takes about 5 seconds to stop it (maybe just slow UI feedback?), then maybe another 10s to continue feels unresponsive when I want to stop it, and so I don't really know if it's gone wild eating up credits 😄 in the background
🤔 1
c
Ah, you mean on the saas
p
no, locally, inside WSL docker (not Docker Desktop), I have it plugged in into an Azure gpt-4o deployment
c
Ah, okay. I don't know what might be up, I was trying actually to see locally the current behavior, with docker on Mac, but I ran into another issue lolsob
😅 1
I stopped it twice: once it was instant, the second time it took an unusual time, just as you say. It looks like it was waiting for the LLM response to arrive. I think there is no case when it would make a new step, it just might not cancel everything at once