beevibe-capabilities

The community-maintained data layer that powers Beevibe's find_repo MCP tool.

Beevibe agents call find_repo({ goal }) to discover which open-source repo to borrow for a task. The ranker reads two signal sources from here:

• data/registry.json — community-promoted learned skills. Populated when a beevibe team approves a skill_outcome and the publishing flow lands a PR here. Adds +30. This is the curation that builds itself from real proven outcomes, not opinion.
• data/trending-daily.json / -weekly.json / -monthly.json — GitHub repos with recent star velocity, refreshed daily via the cron in .github/workflows/refresh.yml. Adds +25 if a candidate appears in the daily or weekly window. The hybrid query (pushed:>X stars:>Y sort:stars-desc) catches both brand-new repos AND established repos hitting a star spike.

Beevibe fetches these via raw.githubusercontent.com — no API hosting, no auth, no rate limit on the read side.

No curated boost list (intentional)

The repo used to ship a hand-curated boost-list.json (pdfplumber, yt-dlp, FFmpeg, transformers, etc.) that bumped well-known mature tools above raw GitHub search. We removed it. Reasons:

• Those tools are already in every LLM's training data — agents know about them. The boost added no information.
• "Mature, stable" is the opposite of "trending" — by definition the boost list misses what just shipped this week.
• Hand-curation creates opinion drift. v1 of the boost list contained a hallucinated repo (nicowillis/spreadsheet-intelligence) that didn't exist, and ranked it equally with real ones.

If a tool repeatedly proves itself across beevibe instances, the publish flow promotes it to registry.json automatically — that's the path where curation comes from real proven outcomes, not opinion.

File contracts

Each JSON file in data/ is consumed by beevibe's find_repo tool. The shapes are pinned — don't break them without a coordinated change to packages/api/src/tools/find-repo.ts.

registry.json

{
  "version": "1.0.0",
  "skills": [
    {
      "repo_url": "https://github.com/foo/bar",
      "goal_pattern": "extract tables from a PDF",
      "invocation": "python -m foo.cli --input <pdf>"
    }
  ]
}

trending-{daily,weekly,monthly}.json

{
  "fetched_at": "2026-05-19T14:00:00Z",
  "period": "daily",
  "source": "github-search:pushed:>2026-05-17 stars:>100 sort:stars-desc",
  "repos": [
    {
      "repo_url": "https://github.com/owner/name",
      "owner": "owner",
      "name": "name",
      "description": "What it does, one sentence.",
      "language": "python",
      "stars": 2843,
      "rank": 1
    }
  ]
}

Refresh cycle

A GitHub Actions cron at noon UTC daily runs scripts/fetch-trending.ts against the GitHub Search API (auth'd with the workflow's GITHUB_TOKEN for 5000-req/hr capacity), regenerates the three trending JSON files, and commits the diff to main.

The fetch can also be triggered manually:

GITHUB_TOKEN=... pnpm tsx scripts/fetch-trending.ts

A future "dogfooding" path will route the daily refresh through a beevibe agent task instead of GitHub Actions — exercising the Capability Network on a real, ongoing maintenance job. The Actions cron stays in place as the production safety net.

Promotions from skill_outcome to registry.json

When a beevibe team approves enough skill_outcome reviews for a saved learned_skill, the publish flow opens a PR here adding the skill to registry.json. Reviewer signs off, merge. The skill becomes available to every beevibe instance globally.

The threshold + automation lives in packages/api/src/routes/learned-skills.ts.