Skip to content

About

Install agent skills for one or more coding agents with reviewed version pins

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Use this GitHub action with your project
Add this Action to an existing workflow or create a new one
View on Marketplace

Latest commit

 

History

8 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Setup Agent Skills

Install agent skills with the gh skill CLI from a newline-separated list of <owner/repo> <skill>[@pin] entries, for one or more agents in a single step (e.g. GitHub Copilot and Claude Code).

gh skill install writes upstream provenance (metadata.github-*) into each installed SKILL.md, so no lockfile is required — checked-in skills are self-describing and update-agent-skills or gh skill update --all picks up drift natively.

Inputs

Name Description Required Default
skills Newline list of <owner/repo> <skill>[@pin]. @pin is an optional tag/branch/SHA. # comments and blank lines are allowed. ✅ —
agents One or more agents passed to gh skill install --agent, separated by whitespace, commas, or newlines (e.g. github-copilot claude-code). Each skill is installed once per agent. See gh skill install --help for the full list of supported agents. ❌ github-copilot
scope Value passed to gh skill install --scope (project or user) ❌ project
gh-version Minimum required gh version (must support gh skill) ❌ 2.90.0
github-token GitHub token exposed to gh as GH_TOKEN ❌ ${{ github.token }}
experimental-rate-limit-retry Opt in to five install attempts with waits of 60, 120, 240, and 240 seconds (eleven minutes total, plus command runtime). Off keeps the shared fast retry defaults. See Rate-limit retries. ❌ false

Rate-limit retries

GitHub recommends waiting at least one minute after a secondary rate limit when response headers are unavailable, followed by exponential backoff. The opt-in follows that fallback policy, with a four-minute cap per wait and five total attempts. A successful install returns immediately; an exhausted retry sequence fails with the CLI's exit status. The policy applies to all failed skill-install commands, since the CLI does not expose the response headers needed to distinguish rate-limit classes reliably.

This option does not guarantee recovery from a primary hourly quota or honor a server-provided reset deadline that the CLI does not expose. Keep it disabled until real burst recovery is evaluated for the consumer, as tracked in #288.

Outputs

Name Description
installed-skills Newline-separated list of source/skill-name[@pin] entries that were installed (listed once per skill, regardless of how many agents it was installed for)

Usage

steps:
  - uses: actions/checkout@<full-commit-sha> # vX.Y.Z
  - uses: devantler-tech/setup-agent-skills-action@<full-commit-sha> # vX.Y.Z
    with:
      # Install every skill for both Copilot and Claude Code in one step.
      agents: |
        github-copilot
        claude-code
      skills: |
        github/awesome-copilot git-commit
        fluxcd/agent-skills gitops-knowledge
        devantler-tech/agent-plugins plugins/engineering-practices/skills/ways-of-working@dc62d7fc09d62602d45da43e31cb1c3998bf05c5
        # Pin with @<tag|branch|sha> or omit to track the upstream default branch.

Omit agents to install for GitHub Copilot only (the default), or set it to any single agent (e.g. claude-code).

Agent selection

The agents input accepts one or more supported agents. The same installed SKILL.md content works across the agents supported by gh skill.

Requirements

  • A trusted runner-provided gh with gh attestation support. If it is older than 2.90.0 (and therefore lacks gh skill), this action downloads the required GitHub CLI release archive on Linux or macOS (amd64/arm64) and verifies its GitHub artifact attestation before installation. Windows runners must pre-install gh >= 2.90.0.

Release recovery

Releases run from protected main. To preview the next release without writing anything, dispatch the Release workflow on main with dry-run enabled:

gh workflow run release.yaml --ref main -f dry-run=true

A dry run previews semantic-release's next version or reports that no release is needed. It does not recreate a missing GitHub release for an existing tag. If a release fails before creating its tag, repair the failure and rerun the workflow from main.

If the tag was pushed but its GitHub release is missing, first verify that the existing tag resolves to the intended reviewed commit on protected main and that its CI passed. In the repository's Releases page, choose Draft a new release, select that existing tag, generate and review its release notes, and publish. Select Publish this Action to the GitHub Marketplace if the release also needs to be listed there. Confirm the published release's source commit and Marketplace version afterward. Do not move or delete the tag, or try to edit an immutable release; use a new patch release at the same reviewed commit if a published immutable release needs a separate Marketplace listing.

About

Install agent skills for one or more coding agents with reviewed version pins

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages