Portable Agent Skills for HYGON-AI software, infrastructure, training, inference, operator and general engineering workflows.
The default path is simple: a skill lives in this repository and ships in one pull request. Mirroring from a product repository stays available as an explicit opt-in, for teams that want a skill to evolve in the same repository as the code it documents.
After the repository is published, browse or install skills with the standard skills CLI:
npx skills add HYGON-AI/skillhub --list
npx skills add HYGON-AI/skillhubInstall one skill into a specific agent without prompts:
npx skills add HYGON-AI/skillhub --skill skillhub-contributor --agent claude-code --yesPass --agent more than once to install into several agents, or --agent '*'
for every agent the CLI detects:
npx skills add HYGON-AI/skillhub --skill skillhub-contributor \
--agent claude-code --agent codex --agent cursor --yesThe pinned CLI installs into any agent it recognizes -- claude-code, codex,
cursor, windsurf, gemini-cli, github-copilot, zed, trae and around
seventy others. Run npx skills add HYGON-AI/skillhub without --agent to pick
from the agents detected on your machine. Skills in this catalog are portable
and are not written for one agent.
Scaffold a local skill, fill in the TODO markers, and open one pull request:
python3 scripts/new_skill.py my-skill-name \
--owner "Owning team" \
--description "What it does, when it triggers, and the nearest case that must not trigger it." \
--license Apache-2.0 \
--category "Developer Tools"Pass --repo HYGON-AI/<product> instead to opt into a remote product source.
See the quick start for the full walkthrough and
CONTRIBUTING.md for the normative rules.
The repository separates candidate content, source registration, published skills, evaluation evidence, and generated metadata:
| Path | Purpose |
|---|---|
skills/ |
Flat catalog of published, independently installable skills |
staging/ |
Catalog-owned SKILL.md.candidate files that cannot be discovered |
components.d/ |
One reviewed registration per component, local or remote |
templates/ |
Non-discoverable contribution scaffolds |
assets/ |
Repository-level README media; never skill content |
docs/ |
Architecture, admission, evaluation and release policy |
Every direct child of skills/ is one catalog identity. Published skills must
not contain nested SKILL.md files or depend on sibling skills. See the
normative repository layout and
admission policy.
| Product | Description | Skills |
|---|---|---|
| SkillHub | Author, validate, onboard, and publish portable Agent Skills across HYGON-AI projects. | skillhub-contributor, torch-trace-operator-profiler |
2 skills across 2 categories.
| Skill | Product | Description |
|---|---|---|
skillhub-contributor |
SkillHub | Create, review, and onboard portable Agent Skills into Hygon SkillHub. Use when adding a new SKILL.md to the catalog, registering a local or remote component in components.d, preparing a SkillHub contribution, or diagnosing catalog validation and synchronization failures. |
| Skill | Product | Description |
|---|---|---|
torch-trace-operator-profiler |
SkillHub | Analyze a torch.profiler Chrome/Perfetto JSON trace to attribute time across Python scopes, ATen operators, GPU kernels, runtime API overhead and memory copies. Use when diagnosing a slow PyTorch operator, custom extension, Triton kernel or submodule from a captured trace. |
A local skill, which is the default:
- The skill is written under
skills/<skill-name>/in this repository. - A
components.d/<component>.ymlfile registers it withlocal: true. - Admission review checks ownership, licensing, self-containment, routing data, and behavior evidence.
- Validation checks naming, frontmatter, resources, evaluation data, secrets, and generated catalog drift.
- One pull request lands the content, its registration and the regenerated catalog.
A remote component, when a product team opts in:
- The product team merges the self-contained skill in its own HYGON-AI repository.
- A
components.d/<component>.ymlfile records the repository, ref and source path. - Synchronization mirrors the registered content and records the resolved commit and digest.
- The same admission and validation gates apply before the mirror lands.
Catalog maintainers can run:
python3 scripts/validate_skills.py
python3 scripts/validate_agent_skills_spec.py
python3 scripts/generate_catalog.py --check
python3 scripts/sync_sources.py --check --component <component>See CONTRIBUTING.md for both paths.
Catalog-owned candidates start under staging/. Remote product candidates stay
in their product repositories until admission; staging/ is not a second
product mirror. A candidate entrypoint is named SKILL.md.candidate until its
reviewed promotion into skills/, preventing deep-discovery clients from
installing staging content.
The catalog publishes reviewed content; it does not make arbitrary third-party skills trusted. Consumers should still review executable scripts and permissions before installation.
A local skill is reviewed here: its integrity rests on Git history,
protected branches, required checks, CODEOWNERS review and DCO sign-off. It has
no .skillhub-lock.json entry and no remote content digest.
A remote component additionally records its repository, ref, and source path
in catalog.json, with synchronized commits and tree digests in
.skillhub-lock.json. See
supply-chain integrity for what each mode does
and does not prove.
CLI discovery proves format compatibility only. Published status additionally
requires the owner, license, source and lifecycle recorded in skill-card.md,
plus positive, negative and behavioral cases under evals/evals.json.
The catalog additionally enforces exact remote commit/digest provenance and a
pinned Agent Skills reference-validation pass; neither check alone proves that
a Skill's operational behavior is correct.
Product repositories remain the source of truth for mirrored skills. The catalog preserves upstream authorship and license terms, records each source repository, ref, and path in catalog.json, and does not treat an unchanged third-party skill as a HYGON-AI adaptation.
Repository code and catalog-owned skill content are licensed under the Apache License 2.0 unless stated otherwise. Mirrored skill content remains under its source license, and imported skills must carry a license compatible with public redistribution.