rawOpensaasContext is a one-shot promise: a boot before the database is reachable poisons createAuth for the process

[borisno2]

The problem

The generated rawOpensaasContext is a one-shot promise. If the process boots before its database accepts connections, the promise rejects once and stays rejected for the life of the process, and everything built on it — notably createAuth, which wraps it in a lazy Proxy — rejects on every subsequent request.

Found on PR #1375 (issue #1170): the Playwright harness started Next before the sidecar was reachable and every /api/auth/* route failed until restart. The harness now serves under opensaas dev, which sequences the database first, so the e2e suite no longer hits it. A production cold start racing a database blip would.

What to do

Decide the contract and make it true:

• Either the promise retries construction on failure (a rejected first attempt does not poison later ones — memoise only success), or
• The generated context refuses to be created before the database is reachable and says so, and the guidance tells deployers to sequence accordingly.

Whichever is chosen: a test that rejects the first connection attempt and confirms the second succeeds (or confirms the documented refusal), and packages/core/CLAUDE.md / the deploy guide stating the behaviour.

Context

• PR Convert starter and starter-auth to Prisma 8 under opensaas dev, with the e2e suites green in CI (#1170) #1375, where it surfaced.
• Spec: Runtime boundary — client construction, the Engine stamp, the Unsafe surface and the Test context #1122 — build spec 2, which owns the runtime boundary and client construction.
• ADR-0014 (the sync-escape-hatch record) describes the lazy Proxy pattern this sits under.

---

🤖 Generated with Claude Code