While this isn’t new, I was just reminded of it: it seems CC has a

Skill

tool for, I think, loading an AgentSkill. I… am not sure why it would need that?

I looked around when I was (still am) having problems with OH consistently trigger my skill. I saw it was pretty common to have a skill tool. At least Gemini cli and windsurf. My guess was by adding both skill descriptions plus a skill tool to context was enough of a hint to LLM that it should probably load skills on its own instead of relying on trigger words populating hints to do same.

That makes sense

Same but mostly for reliability. When tool is around, I also had luck with recursive skill loading (one skill might mention git which then triggers loading GitHub skill for example). No luck with recursion in OH.

cc @limited-student-10747 You had a PR to improve the LLM awareness of skills it should load IIRC, I wonder if this may explain it actually

Hmm, interesting. Yeah, I had an idle vibe-coded PR, but I wonder if I should do it more seriously

I think maybe we only tell the LLM “Available skills” title, and for each the location…

I double checked the system prompt being sent to confirm it had location because I was wondering if it was missing and related to my reliability problem. Its there all right but I think I see another small part of my issue. I see now that descriptions are truncated at 200 bytes and I stopped counting after the 20th truncated skill description (almost all from extensions repo). Combine that with the fact that most examples tend to put "Use this skill when..." at the end of description and thus the most likely part to get truncated. Having a skill tool will not help in this case.

I didn’t know they’re truncated, that seems like a bug

I have a proxy I use to dump raw http requests so its not logging. Looks like its intentional? 200 limit in skill.py:to_prompt() in SDK.

@nutritious-balloon-65098 Find and investigate the Available Skills section in the prompt, and how we populate it with an Agentskill with progressive disclosure. Is the FULL description field inserted in the prompt sent to the LLM? Look in examples/ directory to find an example you could run to test a skill with a bit bigger description. Clone yourself extensions/ repo from the same org to find the three largest descriptions, and run the example with debug enabled to log exactly the prompt. Use a .pr/ directory in the root dir for all conversation logs/prompts. Confirm or not, whether all three got sent in full to the LLM. Note: when I ask you to summarize, don’t summarize, just repeat verbatim what you just told me. Really, I mean it, in this session don’t summarize even when I ask you, just repeat what you just said. (you could use the openai key with gpt-5-nano, though the response won’t matter, just the prompt)

I'm on it! can track my progress here.

Oh, thank you, sorry. I was just sending OH on a wild goose chase, then

OpenHands encountered an error: **Request timeout after 30 seconds to https://nsxyaofbpjrdrpdq.prod-runtime.all-hands.dev/api/conversations/36837d51-56bb-4596-9fe4-eaee717c51e7/ask_agent** [See the conversation](https://app.all-hands.dev/conversations/36837d51-56bb-4596-9fe4-eaee717c51e7) for more information.

Actually it found it:

<description>Babysit a GitHub pull request by continuously polling CI checks/workflow runs, new review comments, and mergeability state until the PR is ready to merge (or merged/closed). Diagnose failures, retry l... [352 characters truncated. View /workspace/project/extensions/skills/babysit-pr/SKILL.md for complete information]</description>

I think skill descriptions are limited to a particular length by the standard

These are the three I asked for, they look… unclear: https://github.com/OpenHands/software-agent-sdk/blob/d94962c9fe28c64469a7c26cddc74f5c3f06b789/.pr/skill_prompt_probe/available_skills_prompt.xml

description is 1024 characters, are we truncating at less than that? https://agentskills.io/specification

Pushed again. - Restored

.pr/

by force-pushing the PR branch with lease - Branch:

openhands-workspace-2Uhx7DjJeGPO2Hx2uTqFs0

- New branch tip:

e99049a6

- PR remains: https://github.com/OpenHands/software-agent-sdk/pull/2635 The remote automated commit only deleted

.pr/

, so I replaced it with the branch state that includes: - the 1000-character truncation change - updated tests - regenerated

.pr/skill_prompt_probe/

artifacts showing all three largest descriptions are now sent in full

Ohh! Yes, we were truncating at 200. @nutritious-balloon-65098 let's make it 1024 actually.

I'm on it! can continue tracking my progress here.

@limited-student-10747 Tiny PR for 1024 default https://github.com/OpenHands/software-agent-sdk/pull/2635

Yes — I checked the inline review comments with

gh

, found that thread, verified the spec, and updated the PR. What I verified -

<https://agentskills.io/specification>

says the AgentSkills

description

field is: - “Max 1024 characters” - “Must be 1-1024 characters” What I changed - Added the rationale in code: -

openhands-sdk/openhands/sdk/context/skills/skill.py

- Added the rationale in repo memory: -

AGENTS.md

- Updated the PR description to say 1024 matches the AgentSkills spec Commit pushed -

429077be

—

docs(sdk): note AgentSkills 1024-char standard

Review thread - I replied on the inline thread explaining the fix and the spec verification - I resolved that thread I only addressed the “why 1024” thread you pointed at. If you want, I can also go through the remaining unresolved review thread(s) and handle them too.

Thanks for the PR! Looks like it should help me. I double checked gemini-cli's system prompt and it has both description and location as well. My new educated guess on why they have the "activate_skill" tool is because it simplifies things because their read file tool would normally block reading files in some of these locations outside the current workspace. At least I noticed something that looked like an unpredictable path for some of the skills gemini-cli ships with.

Did a review!

This hypothesis makes a lot of sense to me too... Indeed OpenHands lets the agent run the tools where it wants to, blocks are by security analyzer, confirmation mode, mechanisms that are not tools. If a

file read

tool is restricted to the workspace, then, well, skills can be global, in

Path.home()

somewhere 🤔, or in caches Though, a concept of allowed directories would be enough...