fake-star-audit— agent skill

This skill audits a GitHub repository's stargazers for signs of fake-star injection and returns an explainable risk verdict. It wraps audit.py, a single-file, zero-dependency, anonymous-API tool (no token, no file writes).

When to use

Trigger when the user asks anything like:

• "is owner/repo fake-starred?" / "are these stars real?"
• "check the star authenticity of this GitHub repo"
• "this repo has 5k stars but looks sketchy — verify it"

How to run

audit.py lives at the repository root (one directory above this skill). Run:

python3 /path/to/fake-star-audit/audit.py --repo <owner>/<name> --json

• Requires only Python 3.10+ and outbound HTTPS to api.github.com.
• Uses the anonymous GitHub API (60 req/h). Each audit costs 3–4 requests.
• If you only have the owner/name from a URL like https://github.com/a/b,

pass --repo a/b.

How to interpret the JSON

Key fields:

• risk_verdict: LOW / MEDIUM / HIGH.
• axes: the 5 detection axes, each with flag (bool) and evidence (string,

annotated by which window — earliest = bootstrap window, latest = recent drip).

• extended_signals: booleans like fork_star_inverted, single_repo_mass_injection,

trusted_org_parasitism (any hard one forces HIGH).

• warnings: caveats (e.g. repo too large to page to newest stars).

Verdict logic: HIGH = 3+ axes or a hard signal; MEDIUM = 2 axes or 1 axis + a signal; LOW = 0–1 axes and no hard signals. It is conservative by design.

How to respond to the user

Give a short, plain-language summary, then the reason:

HIGH risk. 100 stars landed in the first 33 minutes after the repo was created, with near-sequential account IDs — a bootstrap-injection pattern, not organic growth.

Always cite the specific flagged axes / evidence. If the verdict is LOW, say so plainly ("stars look organic on the sampled windows") and mention it's a page-1 sample, not a full-history proof. Never present a verdict as definitive proof of fraud — it's a heuristic signal; point the user to the evidence.

On errors

• rate_limited: the anonymous 60/h budget is spent; tell the user to retry in

up to an hour.

• not_found: repo is private or doesn't exist.
• pagination_capped / warnings mentioning size: very large repo; the recent

window falls back to oldest stars — note this caveat in your answer.

This skill only audits public repositories using public data.