StarSling Skills

Public agent skills from StarSling. Each skill lives in its own self-contained
directory under skills/<name>/ and installs individually.

Today this repo ships one skill: ci-speedup.

ci-speedup: measured CI audits for GitHub Actions

ci-speedup audits a repository's GitHub Actions workflows and produces a
root-cause-analysis report of what actually makes your CI slow, on two
measured axes:

• Developer wall-clock wait: the merge-gating critical path, the slowest
  checks a pull request waits on before it can go green. This is the ranking
  axis, the thing engineers feel.
• Runner-minutes: the cloud bill. Sized per finding and reported as a
  secondary axis (a fix can cut the bill while adding developer wait, so the two
  are never blended into one number).

It works from real run history, not estimates: per-job P50/P95 timings
sampled over the gh API, the critical path re-derived from the actual job
graph. Findings come from a
catalog of 70+
optimization patterns across 14 categories (missing caches, redundant setup,
sleep-based readiness, unsharded test jobs, full-history checkout, dead env
vars, build-cache misconfig, queue time, and more) plus a structural track
routed from the measured long pole.

ci-speedup diagnoses; it does not prescribe the fix. A generic tool can't
see a file's intent or whether a "waste" is deliberate, so instead of a
baked-in diff, every finding ships a ready-to-paste agent prompt that hands
the measured root cause to your own coding agent, which reads the real logs and
the file's git history before shaping a safe change. Detection, ranking, and
every measured number are deterministic; the only place an LLM steps in is a
log-grounded gap-fill when a drilled pole matches no catalog detector.

Install

npx skills add starslingdev/skills

Select the skill, your agent (Claude Code, Codex, Cursor, …), and an install
scope. The skills CLI (built by
Vercel) is fetched fresh via npx, so you always get its latest version.

Then invoke it with your agent:

# Claude Code
/ci-speedup

# Codex
$ci-speedup

First run

Requirements:

• An authenticated GitHub CLI (gh auth login): the run-history data pass
  reads the audited repo's Actions run/job/log data through read-only gh API
  calls. A token with read access to the repo's Actions is enough.
• python3 (3.9 or newer) and PyYAML: the bundled scripts are
  stdlib-only apart from one third-party dependency, PyYAML, which the
  scanner uses to parse your workflow YAML. Install it with pip install pyyaml
  (or python3 -m pip install pyyaml).

What it does: it defaults to the repo you're in (and confirms the target with
you first), samples your recent CI runs over the gh API, and closes by leading
with the biggest measured lever, the slowest check gating your merge and how much
developer wait it costs, then lets you pick what to fix. The full markdown report
is opt-in: it is rendered and integrity-checked internally on every run, and
"Save the full report" is one of the offered options. Pick it and the
verified report is written to ./ci-speedup-findings-report.md in your working
directory (a generated artifact you can gitignore or delete).

Cost / timing (measured on a mid-size OSS repo): the data pass made 314
gh API calls in ~51s (the committed pallets/flask example). Larger repos
sample more (the microsoft/playwright example: 824 calls in ~3m 07s; a
very-high-volume monorepo like next.js measured ~1,800 calls in ~8m); the pass
is frugal by design
(one workflow list, one total-count per workflow, one log per pole, and an
adaptive two-pass job-list sample). It sends nothing to StarSling or any
third party. See SECURITY.md for the data-handling model.

Learn more

• docs/methodology.md: how the numbers are measured
  (P50 over sampled runs, the merge-gating critical path, measured vs modeled,
  adaptive sampling, why it diagnoses instead of prescribing).
• skills/ci-speedup/SKILL.md: the full skill
  contract.
• skills/ci-speedup/references/optimization-patterns.md:
  the pattern catalog.
• skills/ci-speedup/references/wall-clock-methodology.md
  and savings-methodology.md:
  the in-skill methodology references.
• PROVENANCE.md: how the detection catalog was developed and
  validated.
• examples/: a sanitized sample report showing the output shape.

For maintainers

The skill's self-improvement loop infrastructure (the gap→catalog, transcript,
and dogfood loops) lives outside the installable skill, under
maintainers/ci-speedup/, so the skills CLI never
copies it into an install. It runs locally via Claude Code only, never as a
GitHub Action. See maintainers/ci-speedup/MAINTAINERS.md.

This repo's CI runs on StarSling Runners. Fork PRs
run the identical suite on GitHub-hosted runners
(ci-fork.yml); untrusted code never executes
on the self-hosted runners. See CONTRIBUTING.md.

License

MIT