EJRyan89/coding-agent-skills

Portable agent skills for Claude Code, Codex, and GitHub Copilot CLI, with a guarded deployer that renders machine-specific configuration into immutable templates. Skill instructions are authoritative in Claude format; Codex and Copilot read them through generated adapters that map tools and models with the shared runtime-compatibility.md.
Included skills
| Skill | What it does |
|---|---|
| analyze-skill-cost | Audits an agent skill for token, tool-call, delegation, adapter, and model-selection efficiency. |
| audit-ai-config | Audits a repository's AI-agent configuration, read-only and deterministically. |
| curate-agent-memory | Audits a project's Claude Code auto-memory and, with approval, moves durable rules to where they belong. |
| github-activity-report | Reports one user's pull requests, commits, and reviews in one GitHub organization, month by month. |
| repo-cleanup | Performs guarded Git repository housekeeping. |
| update-coding-agent-skills | Fast-forwards the clone the skills were deployed from to origin/main and redeploys them. |
| dotnet | An opt-in bundle installed together: dotnet-format, which runs formatting and analyzer checks against changed C# files, and dotnet-dependencies, which answers why a package or assembly is in a solution, where it is used, and which versions conflict. |
| code-review-operations | A bundle installed together: review-prs, update-pr-dashboard, review-insights, flag-review-finding, and review-document. |
Skills explains how to start each skill, its arguments with examples, and what it needs installed. Most skills run fully on Claude Code, Codex CLI, and GitHub Copilot CLI; Runtime support lists the exceptions and what each one loses: the code reviews on Codex and Copilot, and user-only skills in a headless Copilot session.
Supported platforms and runtimes
| Platform | Where it stands |
|---|---|
| Windows | Supported. |
| Linux, WSL included | The deployer and the whole validation suite run there, and the deployable workflow below deploys onto a fresh Ubuntu runner and checks discovery. Not yet supported: no model has run a skill on Linux through the runtime canary. |
| macOS | Not supported. The deployer stops before changing anything. |
Install the runtimes you use. None of them is needed to deploy, so you can use the skills from Codex or Copilot without installing Claude Code; Runtime support says which skills run there in part.
| Runtime | Maintainer's machines | Fresh Windows runner | Start a skill with |
|---|---|---|---|
| Claude Code | 2.1.291 | 2.1.289 | / |
| Codex CLI | 0.160.0 | 0.160.0 | $ |
| GitHub Copilot CLI | 1.0.98 | 1.0.91 | / |
The maintainer's machines ran the runtime canary and real skill and review runs. The fresh runner is the deployable workflow, which runs weekly and on dispatch; its last dispatched release run passed on 2026-10-15: on a clean GitHub-hosted Windows runner it installs the prerequisites and each CLI, deploys, checks that Codex and Copilot find every skill, uninstalls, and deploys again. Claude Code reads the deployed files directly and cannot list its skills without a session, so for it the runner checks the installed version and that every deployed skill and agent is in place. It starts no model, so it does not show a skill running; that stays a manual check through the canary. The workflow runs the same versions the same way on a fresh Ubuntu runner, where the same dispatched release run passed. See Releasing.