The daemon’s own view of whether it is capturing and indexing. The
watcher, the outbox drain and the repair loop report what they see; each
report re-assesses the state, so GET /health only clones the result.
The drain thread: applies the index outbox every few seconds and, when
the outbox is quiet and the interval has passed, repairs the store with
the writable indexes it owns. On-demand repairs arrive over a channel so
the same thread services them; two writers never exist.
The one-time compaction after the stored-content redaction pass. The pass
rewrites rows, but SQLite keeps their old bytes in free and split pages
and tantivy keeps deleted terms until a segment merges away; a VACUUM and
a forced merge are what take that plaintext off the disk.
Keeps agent writes off the SQLite writer while store maintenance holds
it. A write that arrives then is saved to the write queue and answered
202; once maintenance ends, writes keep queueing behind the saved ones
until the replayer has sent them all, so none lands out of order.
Replays the writes the MCP server queued under queue/ while this
daemon was unreachable. Each entry is the request the MCP server would
have sent; it goes through this daemon’s own router, so auth, redaction
and every handler apply exactly as for a live call.
Extract the entity names a query refers to, using the same gazetteer
extraction the ingest path applies to transcripts. Matching extractors on
both sides is what lets query entities resolve to stored entity ids, so the
graph term in recall is conditioned on the query rather than raw
connectedness. Shared by the live recall route and the eval harness so both
measure the same ranking.