skill-leaderboard

Fetch active forks via the GitHub API, retrieve each fork's configuration file, and diff it against upstream defaults. Counting only forks whose config diverges from defaults avoids the tautology where every untouched fork reports identical values.

A fork is CONFIGURED if its aeon.yml differs from upstream defaults on any skill's enabled, model, var, or schedule fields, or if it contains fork-only skill directories. Forks with zero differences are classified as TEMPLATE and excluded from leaderboard math.

It uses gh api to list forks filtered by pushed_at within 30 days, then one recursive git-trees call per fork to enumerate files, and a contents call to fetch aeon.yml only when present. Rate-limit 403 responses trigger a 60-second backoff and one retry.

The notification is gated on having at least two configured forks. When fewer than two forks diverge from defaults, the leaderboard has no meaningful denominator, so the run logs a status code and stays silent to avoid training the operator to ignore alerts.

Deltas come from a persisted JSON snapshot at memory/topics/skill-leaderboard-state.json, not from parsing the previous article. Skills moving three or more ranks are flagged as rising or falling, and snapshots older than 14 days are treated as stale.