memory-health
embedding-worker telemetry. memory_units commit with embedding=NULL and the worker fills in vectors on a throttled cycle (default ~1 RPS, 5 rows per 30s). errored rows back off exponentially, capped at 10 minutes — a recovered dhara picks them up on the next cycle without intervention.
coverage
workspace-scoped. recall vector search only ranks embedded rows — non-embedded units fall back to tag/matter filtering.
per-matter embed progress
1 matter(s)| matter | embedded | total | coverage |
|---|---|---|---|
| active-engagements | 0 | 33 | 0.0%still embedding |
worker
live · refreshed just nowerrored sample
0 rows · oldest firstno errored rows. either the worker is hitting every embed cleanly or there are no live rows with embed_attempts > 0 right now.
gardener
daily pass: disabledlast fired neveryield: 0 suggestion(s) · 0 digest(s)no gardener checkpoint deposited yet — the daily pass deposits digests only on matters with findings, and only on Fridays.
the daily pass never judges — this is the human-triggered, supervised run. dry-run reads the band-candidate pool with zero LLM calls; judge run spends up to judgeCallCapPerRun BYOK calls and writes only judge-confirmed contradictions.