[leviFrosty]

GET /notes-import/status is about two thirds of prod traffic. Over the last 7 days in Sentry its response time was p50 302 ms, p90 1.16 s, and p99 1.74 s. There were two causes:

• OpenRouter probe on the request path. The provider-health cache expired after 60 s at each edge, so about 29% of requests (5,620 probe spans out of 19,264) waited on OpenRouter's endpoints API. The gpt-6-luna probe has p95 973 ms, and its p99 hits the 2 s deadline.
• KV reads in sequence. The kill-switch, then the provider cache, then sometimes the probe plus a KV write, and only then allowances and minimum version. At a quiet edge each read is a round trip to KV's storage.

Change

• KV notes-import:provider-health now holds {healthy, checkedAt} for 5 minutes. A result under a minute old is served as is. An older one is still served while waitUntil re-probes and rewrites it. Only a missing result is probed inline, and that cache write also moves to waitUntil. The bare up/down values from before this change expire within 60 s, so they're treated as fresh.
• The kill-switch, provider cache, allowances, and minimum version are now read in parallel. The kill-switch still decides first, and the existing fail-open rules are unchanged.
• The status, kickoff/legacy gate, and transcribe routes pass the request's waitUntil. Without it (unit tests), a stale result is re-probed before answering, as before.
• Trade-off: a provider outage now takes a minute plus one request to show, instead of a minute. AGENTS.md documents this.

Verification

• pnpm test (741 passing) and tsc --noEmit. The new tests cover fresh, stale + background refresh, a failed refresh keeping the stale value, the no-waitUntil path, the deferred write, and all four reads starting before any of them answers. Against the old status.ts, 6 of them fail.
• Verify loop with a real OpenRouter key: doctor ok and e2e 44/45 (1 skipped). Status route over HTTP:
  • nothing cached: probed inline, 130 ms, stored {healthy:true, checkedAt}
  • fresh: 11 ms, no probe
  • 2-minute-old down: no_provider served in 14 ms, then rewritten to healthy:true in the background; the next request was available
  • old-format up: served without a probe

The Cloudflare wall-time p99 (30 s) and p999 (300 s) are a separate effect: long-lived Buddies live sockets and Notes Import progress streams count their whole lifetime as wall time. This PR doesn't target those.

🤖 Generated with Claude Code