akillness/jeo-skills/github-repo-candidate-quality-gate

GitHub Repo Candidate Quality Gate

Use this skill when broad GitHub search output is too noisy for merge-ready recommendations.

When to use

• You already ran gh search repos (or equivalent) and need recommendation-grade keeps.
• You must preserve broad discovery evidence while filtering keeps deterministically.
• You need lane-level status + degraded-cause reporting for unattended cron review.

Minimum keep metadata

For every kept candidate, capture all fields:

• full_name
• url
• license
• pushed_at (or updated_at fallback)
• archived
• stars
• one-line fit rationale
• provenance label

Quality gate order

1. License gate: drop missing/unknown license unless explicit exception rationale.
2. Archive gate: drop archived repos by default.
3. Freshness gate: drop if pushed_at older than 24 months unless exception + risk.
4. Fit gate: drop keyword-only mismatches (low-fit).
5. Signal gate: keep broad discovery evidence, but recommendation-grade keeps require traction (default: stars >= 3) or explicit adoption rationale.

Reporting contract

• Keep and publish full discovery counts even when keeps are sparse.
• Emit lane_status: pass|degraded.
• Emit degraded_causes counts using: license, stale, low-fit, archived, low-signal, no-results.
• When a lane ends with raw_count == 0 after documented recovery, include no-results explicitly to keep degraded reports deterministic.
• Include lane health metrics: kept_count, raw_count, median_stars_raw, zero_star_raw.
• State provenance labels used (direct page retrieval, indexed snippet, etc.).

Rollback-safe operation

• Never delete raw evidence files.
• If stricter gates reduce keep counts too far, preserve output but mark lane degraded.
• Allow exception keeps only when rationale + risk note are explicit.

Example

• Input: 30 raw hits, many zero-star repos.
• Output: discovery evidence retained, recommendation keeps filtered by metadata/freshness/signal gates, degraded causes documented for reviewer.