https://docs.all-hands.dev/ logo
Join Slack
Powered by
<https://github.com/OpenHands/software-agent-sdk/i...
# questions
w
https://github.com/OpenHands/software-agent-sdk/issues/1825 https://github.com/OpenHands/software-agent-sdk/issues/1418 ^^ Do either of these issues still need to be solved? Was interested in trying if there's still interest
c
I'm not sure if the second is still happening, I don’t use MCPs with auth lately
Re: the first, basically I think it makes sense, the only quick thought on it is that it’s not clear to me if with Python 3.13 and 3.14 we’d still want to rely on GIL or look at other mechanisms. But this is just a tiny question in the back of my slow brain (I don’t yet know how Python without GIL works), not a problem.
w
Ok, so the first one is still worth trying out? I don't understand it, but have been trying to understand the SDK codebase over the weekend and wanted try out one of the "good first issues"
c
If you’d like, this one might be pretty good too, to play with how the tools are defined In that case, I'd suggest to try out TaskTracker or Terminal, whose description is pretty big (not all at once, unless the agent is doing it 😊 )
But yes, Event too, as you want
We haven’t kept up very well labeling issues, sorry about that Maybe this one is interesting if you’re interested in hooks
Found another candidate Another option, if you’d like, is to make a little script using the SDK (or use the CLI) and just figure out what you don’t like - and you’re always welcome to fix it! Either way, thank you!
w
thanks for sending these. I'll try this one since it makes the most sense to me at the moment: https://github.com/OpenHands/software-agent-sdk/issues/2129
gratitude thank you 1
should i condense the tasktracker/terminal prompt?
c
Sorry, what do you mean by condense, in this context?
w
like is it too long and I should trim it?
c
The prompt goes to the LLM, so any change to it could affect performance or behavior of the agent. So I don't think we should make changes, at least not in this issue: it should work exactly the same as before, it's a refactoring
w
Sorry, I think I may have misunderstood when you said "to try out TaskTracker or Terminal, whose description is pretty big" to mean that I should condense the prompt. I was missing the agent-spec file which is why the PR failed, but I think it should be ok now. I simply moved the prompts with their context variables to the .j2 files in the templates/ folders for each of the tools
c
Oh, sorry, I meant you could have chosen to do only one or two, in a PR, for example those most annoying to have description inside python code All good, no worries, thank you!
It's a bit easier to review, too, if we figure out the essentials for a single tool or two
and to test 🙂 How does it work now?
w
I ran the precommit tests and the pytests for the tools and those passed, but I can try it out a bit as a sanity check too.
At work rn, but I can run some of the demos in the evening and just make sure that everything is fine
thanks on a stick 1
I decided to do a few PRs just to understand the codebase better, but my major interest is in trying to integrate self-improvement, automatic tool-rewriting into the agent since I think OpenHands could be smart enough to re-write some of it's parts better during test time: https://openreview.net/forum?id=pUpzQZTvGY
c
Oh, that's awesome! FWIW, I was just making a Skill for self-adaptation: https://github.com/OpenHands/extensions/pull/66
👍 1
w
As an update, I'll try to get this one done over the weekend. Got a little bit occupied with studying for an interview, but will try to have it done and tested by 3/1.
c
Thank you! No worries, but I think maybe it's worth considering to try an alternative with only one or two tools, so we can make sure ... or maybe not, doesn't matter a lot, just to read easily the change pattern
4 Views
Open in Slack
PreviousNext
Powered by