OK, for agent server this PR should allow per-conv...
# proj-agent
l
OK, for agent server this PR should allow per-conversation docker isolation: github.com/OpenHands/software-agent-sdk/pull/3403 This is somewhat reviving the “app-server”/“sandbox-server”, but imo the implementation is better and more sustainable than before, only about 800 lines of app code. I’d appreciate a look at the PR!
🙌 1
m
left a coupoe of comments. Should I take over the line this PR or are you taking care of that?
l
I’ll finish it up, I’ve already tested it pretty extensively
👍 1
w
Came in to see what the architecture was here and was horrified to see my own face 😂 I don't remember doing this
l
a lot has happened since May 😂
it’s based on your original PR, but I did a ton of polishing
w
😌
yeah it looks like all the commits are recent, maybe not at all what I originally proposed
l
Maybe my agent force pushed. It's spiritually the same architecture though. I also tried a separate one where actions were executed in the sandbox and the agent server lived outside but that was too complex.
w
Yeah this is great, thanks for pushing it over the line!
We're back to "nested runtimes" 😂
This is probably relevant to DigitalOcean
l
I know, ironically i was one of the people pushing against this originally, but I think that the design is clean enough now that we should be able to switch between the two with minimal extra work
🙌 1
And I was mentioning that it might be relevant to digital ocean on the thread, but if they’re just another backend then maybe it’s not relevant?
I thought that maybe they wanted per-conversation VMs (in which case it’d be relevant), but it sounds like that’s not the case
w
tbh I think per-convo isolation is a better use case
I'll raise w/ them
👍 1