# StarSling Skills

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

This repo ships three skills: **`ci-speedup`** ([overview](https://starsling.dev/ci-speedup)), **`ci-score`**, and **`ci-secure`**.

## 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](skills/ci-speedup/references/optimization-patterns.md) 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.

## ci-score: a best-practice grade for your CI