After chat() returns, fire-and-forget background thread POSTs the (user
message + assistant response) as one episode to /episodes. Default extraction
(Sonnet). Errors logged, never raised — chat is not gated on the write.
Wall-clock cost in the background is ~20 min per episode against the
current ~4,300-entity graph. The chat experience is unaffected; the graph
catches up with a delay. Search_facts queries reflect new turns once the
sidecar has finished processing them.
Kill-switch: SKIP_GRAPHITI_CHAT_PUSH=1 in the api service environment
disables the push without code changes. Useful if dedup contention surfaces
under sustained load.
Companions to this commit: search_facts tool (e96bf40), orientation indexer
worker (e96bf40), FalkorDB vector index patches (d2ec20e, 313c0f0).
aaron
151c756b89