Seeking feedback on User-level skills in GUI: Background: When loading skills in GUI, OpenHands will look for User-level skills in

{your git provider}/{username}/.agents/skills/

Questions: • Do you like this approach? Do you use it? • For users in corporate/enterprise environments (or anyone with experience in enterprise environments), does this approach work? Why or why not?

I was just thinking about this... Maybe worth noting, it doesn't work in V1. Q1. When I work with the CLI, and/or other local interfaces, they load my skills from

~/.openhands/

or

~/.agents/

. If I then work on the Cloud: I don't have those skills. So loading from the personal repo means I could continue to use my skills on the cloud. Otherwise, oops, they were there and now they're not. This assumes that I'll keep the personal git repo updated from the local version. thinky spin

I guess that's why I've been having trouble using my personal skills repo lolsob I like the idea of customization. Maybe: • In GUI + Cloud, a user can define 0 or many Skill sources that they want to make available to new convos? Like your example of

zarazhangrui/frontend-slides

• In GUI+Cloud, a user can customize the location of their

save source

. Maybe it defaults to the way things were (

{your git provider}/{username}/.agents/skills/

), but can be customiable? Curious your thoughts!

Little detail: • in the personal repo, I put a skill

adapt-yourself

• (cloud) I tried to ask the agent if it knows the

adapt-yourself

skill • when I have a conversation with OpenHands/* repo, it doesn't know it • when I have a conversation with an enyst/* repo, it does know the

description

in context • but apparently didn't get the skill: https://app.all-hands.dev/shared/conversations/30f97557a78b44e8ba40b65a00ed1322 It looks like a bug to me / missing something

I heavily use <org>/.OpenHands for skills - it allows a few different things 1. I can provide skills at a massive scale without having to maintain a bunch of versions 2. I can implement skills that assist with security concerns I may have Org level skills is probably one of my favorite things about OH - because I'm constantly testing and adapting new skills and once verified pushing them out at a large scale - and I don't have to worry about “are people making sure they are up to date with our skills, etc…” - I just rely on OH to implement them for new conversations

If this worked I would use it 🙂

Coming back to this @calm-artist-7206 I think it's still worth supporting user/org skills over a plugin because I have three conditions to solve for • GUI • CLI • GitHub Actions All of which should have access to the same skills And having to write scripts to init plugins each time doesn't sound fun 😆

Just to note, related, so the current direction is this https://github.com/OpenHands/software-agent-sdk/issues/2272#issuecomment-3992608091

Nice, makes sense - wonder if it’s possible to lift https://docs.litellm.ai/docs/skills It's probably hard coded to Anthropic but a good api baseline

Oh, thanks for the reminder. I noticed that too just the other day, and haven't had the time to dig into it. As you well know, we've done the implementation/support so far ourselves

Yep! Something new everyday

Re: user-level skills. I see people are trying to use user-level skills, although this is not from a repo, it's from local home: https://github.com/OpenHands/OpenHands/issues/12858 With the Web UI, there are problems and workarounds making them available inside docker.

BTW this is ready for re-review, which is also targeted at fixing that issue (I wanted to use my own personal skill marketplace): https://github.com/OpenHands/software-agent-sdk/pull/2253/changes

Yes, thank you! ❤️ That’s what my agent and I are looking at, sorry I’m dense and was talking to it to make sure it works in a smooth way

Ah, I see, thanks. Looking at alternatives.

If we explicitly treat them as skills… maybe we can?