BuilderIO/agent-native @agent-native/core@0.145.1

on GitHub

Patch Changes

• b242acf: Cut idle database round trips on serverless deployments.
  • The Netlify keep-warm scheduled function is now opt-out and its cadence configurable (AGENT_NATIVE_DISABLE_KEEP_WARM, AGENT_NATIVE_KEEP_WARM_SCHEDULE), and the expensive background-function warm can be dropped on its own (AGENT_NATIVE_DISABLE_KEEP_WARM_BACKGROUND). Previously a once-a-minute wake was hardcoded, so a scale-to-zero database (Neon, Aurora Serverless v2, paused-compute Supabase) never autosuspended and burned its compute quota with zero users online. Defaults are unchanged; an unparseable cron fails the build rather than silently reverting to the old cadence.
  • Thread ACL checks no longer load the conversation blob. resolveThreadAccess was loading the full chat_threads row — including thread_data — for the access decision, discarding it, and reading the same row again, so every agent-chat request downloaded the conversation twice. The share dialog and collab routes got the same projected-load fix.
  • The SQL-backed SSE relay now backs off from 500ms toward 2s after a run has been quiet for several polls, and probes reapIfStale on its own 5s cadence instead of the 500ms status cadence — a reap cannot match a row younger than 15s, so it was issuing ~30 rounds of round trips before it could do anything. Streaming cadence is unchanged.
  • The in-process backstop sweep timers can be turned off with AGENT_NATIVE_DISABLE_INPROCESS_SWEEPS where a durable scheduler already drives the same recovery. On serverless these timers run per warm container, so their query rate scaled with instance count rather than with load.
  • Domain auto-join no longer probes organizations on every authenticated request. It short-circuits free email providers (which can never match, since an org may not claim one as its allowed_domain) and caches a no-match per domain, invalidated when allowed_domain is written. Previously this ran once per request for every account not already in a domain-matched org, with no fixed point.
  • Session-token and org-membership resolution are cached across requests behind short TTLs plus write invalidation, instead of one round trip each per request. getAllSettings is memoized per request and now seeds the single-key settings cache.
  • Added shared/ttl-cache.ts as the single bounded-TTL cache primitive and moved the hand-rolled one in triggers/condition-evaluator.ts onto it. Failed reads are never cached anywhere, so an unreadable table can't be served as "absent".
  • On-demand ensureTable() schema probes are answered from one batched introspection pass per database (two queries) instead of one query per table, column, and index — up to ~390 serial round trips per cold start before this. The snapshot is keyed per database client, so the hosted multi-app gateway can never be answered with another app's schema, and an unreadable information_schema falls back to per-object probes rather than being read as "the schema is empty". AGENT_NATIVE_SKIP_ENSURE_TABLES=1 skips the probe-and-DDL machinery entirely for deployments that run real migrations.
  • Default resources are seeded once per database instead of once per process, behind a durable marker. Previously every cold start re-issued ~10 INSERT … ON CONFLICT DO NOTHING writes plus two migration scans for rows that had existed since day one. Note the behavior change: a default resource the user deletes is no longer silently recreated on the next cold start.
  • Expired agent-scratch cleanup no longer blocks resource reads. resourceGet/resourceGetByPath/resourceList each awaited a DELETE before their own SELECT, which doubled the round trips, took row locks inside a user-facing request, and could fail the read outright.
  • useDbSync now applies actionInvalidatePredicate to the framework-prefix invalidate too. An app's opt-out list was silently conditional: honoured for batches carrying an action event, ignored for batches carrying only db/collab/settings/screen-refresh.