trove-lint

A memory that isn't inspected quietly rots.

When to use

• Periodic health check (after a stretch of captures/ingests, or weekly).
• After an import, to see what arrived unlinked.
• When recall quality feels off.

Process

Step 1 — run

lint

trove://lint

orphan_node

missing_evidence

duplicate_title

dangling_edge

Step 2 — triage

• Orphans: real pages nothing links to. For each (or the top N),

  recall

  its title to find candidate hubs; propose

  connect { predicate: "mentions" | "part_of" }

  connections.
• Missing evidence: agent-captured claims without citations. Propose evidence backfill (

  annotate

  against a source span) or an explicit inference note via

  remember

  .
• Duplicate titles: read both; propose merging (repoint edges to the survivor, then the loser is soft-deleted server-side) — surface, don't auto-merge.
• Dangling edges: endpoints deleted; propose

  forget

  .

Step 3 — propose, then apply

Step 4 — confirm

Notes

• Smoke-test debris (actors ending

  -smoke

  ) is cleaned with

  npm run db:clean:smoke -- --apply

  in the repo, not by hand.
• Jobs pending/failed counts come with the lint report on the dashboard's health card.

Anti-patterns

• Don't auto-fix without proposing — especially merges.
• Don't link orphans to hubs on title similarity alone; read the node first.
• Don't treat every warning as urgent; a leaf note with no links can be legitimate.