github-repo-vitals

GitHub Repo Vitals Skill

The repository is where a project's real pulse shows — registries say what shipped; the repo says whether anyone's home. GitHub's REST API serves the vitals keylessly (60 requests/hour anonymous — enough for a handful of checks, and the skill budgets accordingly): last commit, release cadence, issue responsiveness, contributor spread. The interpretation discipline matters here too: stars measure attention once, commits measure life now, and an unanswered issue tracker measures the thing adopters actually care about.

What This Skill Produces

• The verdict — alive / coasting / life-support / abandoned, with reasoning
• The vitals table — last commit, release recency/cadence, open-issue dynamics, contributor concentration
• The responsiveness sample — how recent issues/PRs actually got treated (the tell no summary stat carries)
• The commands — each vital's curl, within the anonymous rate budget

Required Inputs

Ask for these if not provided:

• The repo — owner/name; from a package check, the registry metadata's repository URL (chain from [package-health](../package-health/SKILL.md))
• The stakes — building on it, contributing to it, or evaluating for a fork: the verdict calibrates ("coasting" blocks adoption of a framework, not of a finished parser)
• How many repos — the anonymous budget is ~60 calls/hour; a comparison of five repos is fine, a screening of fifty needs a token (say so rather than degrade)

Framework: The Vitals and the Budget

1. The core calls (4–6 per repo): repo overview: curl -s "https://api.github.com/repos/OWNER/REPO" (pushed_at, open_issues_count, archived flag — archived is the verdict, stop there) · latest commits: .../commits?per_page=5 · releases: .../releases?per_page=5 (cadence from the dates) · recent issues: .../issues?state=all&sort=created&per_page=10 (responsiveness sampling) · contributors: .../contributors?per_page=5 (concentration).
2. Read pushed_at against the project's nature: a spec implementation quiet for a year ≠ a web framework quiet for a year — same rule as package-health, applied at the repo layer. The archived flag and a README deprecation note outrank every other signal.
3. Sample responsiveness, don't just count: open_issues_count alone is noise (popular ≈ many open issues); the tell is the sample: of the last 10 issues, how many got any maintainer response, and how fast? Ten unanswered issues in a row is the loudest abandonment signal there is — louder than commit age.
4. Bus factor from contributor spread: one contributor with 95% of commits is a fact worth naming (with the nuance that solo-maintained excellence is common — it's a risk note, not an indictment); an org-backed repo with recent commits from several people reads differently.
5. Rate-limit honesty: anonymous = 60/hour; the skill batches to ~5 calls per repo, reports when the budget shapes the depth, and never silently fails into fabrication — a 403 rate-limit response is reported as exactly that, with the try-later or use-token options.

Output Format

Repo Vitals: [owner/name]

Verdict: [alive / coasting / life-support / abandoned / ARCHIVED] — [two sentences].

| Vital | Value | Read |

|---|---|---|

[Last push · release cadence · responsiveness sample (n of last 10 answered, typical lag) · contributor spread · archived flag]

[Comparison mode: repos × vitals]

Source: GitHub REST (anonymous, rate-budgeted) · as of [date] · rerun: [the curls]

[If rate-limited: the honest report + options]

Quality Checks

• [ ] The archived flag was checked first and short-circuits everything
• [ ] Responsiveness came from a sample, not the open-issues count
• [ ] Activity age is read against the project's nature
• [ ] Contributor concentration is noted as risk-with-nuance, not indictment
• [ ] The call budget stayed anonymous-friendly and rate limits are reported, never papered over

Anti-Patterns

• [ ] Do not read stars as health — attention once ≠ maintenance now
• [ ] Do not count open issues as a negative — sample how they're treated instead
• [ ] Do not burn the rate budget on one repo's full history — five calls tell the story
• [ ] Do not fabricate around a 403 — rate-limited is a reportable state
• [ ] Do not condemn quiet-but-finished projects — the nature test applies before the verdict