HYGON-AI SkillHub

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.

Quick start

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/SkillHub

Install one skill into a specific agent without prompts:

npx skills add HYGON-AI/SkillHub --skill skillhub-contributor --agent claude-code --yes

Pass --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 --yes

The 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.

Add a skill

Create a branch, then prefer importing an existing self-contained skill directory. Its SKILL.md, references, scripts, assets and bundled license material are copied without modifying the source. The command asks who will maintain the SkillHub copy; existing source authorship remains preserved separately:

python3 scripts/contribute.py import ../my-existing-skill

Create a skill from scratch only when there is no existing skill directory to import. The command prompts for the owner, description, category, runtime requirements and permissions, registers the skill locally, and defaults original contributions to the repository's Apache-2.0 license:

python3 scripts/contribute.py new my-skill-name

For either path, review the generated Skill Card (lifecycle already defaults to published), then run all local catalog checks with one command:

python3 scripts/contribute.py check

Review the resulting diff, commit with --signoff, push your branch and open a pull request in the browser. The helper never creates branches, commits, pushes or pull requests. Remote synchronization remains an opt-in advanced path; see the quick start and CONTRIBUTING.md for the rules.

Repository structure

The repository separates candidate content, source registration, published skills, validation tooling, 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 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.

Skill catalog

Product	Description	Skills
SkillHub	Author, validate, onboard, and publish portable Agent Skills across HYGON-AI projects.	skillhub-contributor, torch-trace-operator-profiler

Skills by category

2 skills across 2 categories.

Developer Tools

Skill	Product	Description
skillhub-contributor	SkillHub	Create, review, and onboard portable Agent Skills into HYGON-AI 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.

Performance and Profiling

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.

How publication works

A local skill, which is the default:

1. The skill is written under skills/<skill-name>/ in this repository.
2. The helper registers it in components.d/skillhub.yml with local: true.
3. Admission review checks ownership, licensing, self-containment, trigger boundaries, and runtime permissions; any behavior claims must match available evidence.
4. Validation checks naming, frontmatter, resources, Skill Cards, licenses, secrets, and generated catalog drift.
5. One pull request lands the content, its registration and the regenerated catalog.

A remote component, when a product team opts in:

1. Select a self-contained skill from any maintained GitHub repository; see the external import guide.
2. A components.d/<component>.yml file records the repository, ref and source path.
3. Synchronization mirrors the registered content and records the resolved commit and digest.
4. 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 unfinished prototypes may use staging/; it is not required for the normal local contribution path. 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.

Trust model

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 and has no .skillhub-lock.json entry or remote content digest. Its safeguards are Git history, review and DCO sign-off; protected branches and required checks/CODEOWNERS reviews must be configured separately on GitHub to enforce the merge policy.

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 discoverability, not operational correctness. The catalog requires owner, license, source, lifecycle, and non-empty runtime/permission information in skill-card.md; a link to documented SKILL.md requirements is allowed. No separate eval dataset or Validation section is required. The generated published value is metadata, not review approval or a release. 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.

Source attribution

Source repositories remain the source of truth for mirrored skills, including personal and third-party GitHub repositories. The catalog preserves upstream authorship and license terms and records each repository, ref and path in catalog.json. Imports require quality checks and maintainer review; inclusion does not imply HYGON authorship or HCU adaptation.

License

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.