are all these skills being loaded into the context window every time? should these be more opt-in?

They're AgentSkill format, so I think only a short description is loaded, for each of them

Actually, not all skills are loaded in the context window. We have two type of skills: • always loaded context • trigger loaded context The former are always loaded into the system prompt, while the latter are only triggered when a keyword appears in a user message. That said, I’m not sure if all the skills in your screenshot are loaded into the system prompt. Note: In the Always-loaded skills there is a subset called semantically triggered skills, as the agent can invoke/fully disclose them if the skill description is semantically similar to the user’s message.

These are Agentskill format, I think, they're: https://github.com/OpenHands/skills/tree/main/skills We could say they're three types... it's getting a bit confusing, but this is the ones for which only the description should be loaded in the prompt https://docs.openhands.dev/sdk/guides/skill#progressive-disclosure-agentskills-standard

Thanks for the clarification 🙂

I think maybe we could fix the docs... because you are correct, Vasco, they should (tm) be two, but maybe we could present them like: • always in context • agentskill extended standard ◦ has a description in context ◦ (optional) trigger

actually there are two but with subtypes I will say • always loaded skills: ◦ the whole skills loaded in context (e.g. memory) ◦ just description loaded in context and then progressive disclosure (e.g. pr-review) if the agent asks for it • trigger loaded skills

We could do it like that, it has some disadvantages though: • agent tools out there have the notion of AGENTS.md / claude / gemini, and AFAIK they don't name them skills • AgentSkill is an emerging standard, so combining it with AGENTS.md seems more difficult to make clear to people • we did intend to extend AgentSkill, with triggers: so it seems an easier mental model to name them "extended AgentSkill standard" More importantly, it seems to me super-cool that an "extended AgentSkill" as we define it, means: • triggered, not always in context • the user can trigger it, if you define a trigger • the agent can trigger it, since it knows a short description that makes the LLM decide to use it.

agent tools out there have the notion of AGENTS.md / claude / gemini, and AFAIK they don't name them skills

I see your point. But actually having Memory is a skill and it would make sense to have it incorporate in the skill machinery. But this is just my taste 🙂

triggered, not always in context

What do you mean with that? Can you give an example of such skill?

Those in

skills/

repo, I think maybe they all have

triggers

defined. Most of them come from old Openhands format ("keyword-based") where they were only loaded on user message, and the agent couldn't run/load them on its own

I think it would be really useful to have a solid overview of how skills are implemented in the repo so that we can best attack this epic.