Files
routstrd/src
redshift dbf4197adb perf(daemon): reuse bootstrap evidence and share cold model population
Startup ran provider bootstrap, then immediately forced the same Nostr filters
and every provider's model refresh again. Bootstrap already refreshes expired
provider, review, mapping and curated-model evidence (concurrently), so the
second pass paid another relay round trip plus a full provider fan-out for
nothing, and left a redundant model pass running in the background.

- Drop the immediate forced relay refresh. Scheduled and explicit refreshes
  still force network discovery.
- Share one cold population pass between readers and refreshes so concurrent
  /models requests and startup coalesce instead of each fetching every
  provider. A failed pass clears, so the next request retries.
- Keep a pass that started genuinely cold authoritative until it settles: the
  SDK now publishes providers as they land, and a partial cache must not look
  like a complete model list to onboarding (it would install bare stubs).
- Let startup reach the SDK's per-provider TTL checks without forcing healthy
  providers to refetch, while ordinary readers keep serving persisted models
  immediately instead of waiting on the refresh.
- refreshModelsAndIntegrations takes explicit options and injectable deps, so
  startup warms models without rewriting client configs when automatic refresh
  is disabled. Scheduled/manual defaults stay forced.
2026-10-13 20:55:34 +08:00
..
2026-03-19 16:25:57 +00:00