Contents

Gardener

Monthly catalog hygiene. Orchestrator runs inside a forked subagent. Walks every skill and agent across the marketplace, scores them in parallel, detects collisions via Bedrock Cohere v4 embeddings, runs an HDBSCAN taxonomy check, and composes a delta report against the prior audit. Proceeds autonomously — planning is reversible. Gates only at the point of writing fixes back to SKILL.md files, which it does not do: it proposes.

How it works

Read references/orchestrator.md and follow its phases. Short version:

1. Inventory — walk the marketplace, extract frontmatter and file stats, write audit/YYYY-MM/inventory.md.
2. Collide — run scripts/embed-catalog.py once (Bedrock Cohere v4), then scripts/pairwise-collisions.py and scripts/cluster-taxonomy.py. One interpreter Agent opens the flagged SKILL.md files, verifies the collisions by reading, and writes audit/YYYY-MM/collisions.md + audit/YYYY-MM/clusters.md with recommended actions.
3. Score — launch one parallel Opus Agent per skill. Each Agent reads the skill's SKILL.md, references/, and templates/ in full, applies the shared rubric, and cites file:line evidence in every rationale. Collision and cluster data from Phase 2 are passed in as cross-reference context — where to look, not substitute for reading.
4. Report — one foreground critic Agent reads inventory + all scorecards + collisions + clusters + the prior month's report, writes audit/YYYY-MM/report.md with health score, delta, quarantine list, and top-5 action items.

Output layout

plugin-gardener/audit/
  YYYY-MM/
    inventory.md
    scores/
      <skill-name>.md         # one per skill
    vectors.npz               # cached Cohere v4 embeddings
    collisions.csv            # raw pairwise output
    collisions.md             # interpreted, ranked, with actions
    clusters.md               # HDBSCAN clusters, outliers, straddlers
    report.md                 # the deliverable

Each month's audit is a standalone artifact; deltas are computed by reading the prior-month report inline.

What it does not do

The gardener proposes. It does not:

• Edit SKILL.md or agent files.
• Open pull requests.
• Delete files.
• Change marketplace.json.
• Publish to skills.sh or any external registry.

Those actions are the user's to approve. Every suggestion in report.md includes the exact file path and the proposed change, ready for you to apply manually or via a follow-up /rewrite-descriptions invocation.

Cost note

One monthly run against 40–50 skills: ~50 parallel Opus Agents for scoring, 1 embedding call (single Bedrock invocation, all descriptions fit under the 96-item cap), 1 critic Agent for the report. Significant — budget for it as a monthly line item, not a casual run.