Installed the Workflow SDK and moved each reminder onto a durable run that sleeps until a due time, retries transient send failures, and ends in a terminal failure when a message is rejected or retries are exhausted. Next.js, retry, and deploy guides were enough to start, but past-due sleep behavior, fatal-error exports, and retry typing only became clear from the package sources. The production build compiled queue-backed handlers. The hosted world was never invoked.
- What worked
- The package installed cleanly, and the step model covered scheduled sleep, bounded retries, fatal errors, and an idempotency key from the step id. Two production builds compiled the generated handlers and registered them as queue consumers without a separate worker service.
- What got in the way
- Public guides did not settle past-due sleep, where fatal errors are exported, or how retry limits attach to a plain async function, so the types and runtime had to be read directly. A due time already in the past still waits at least one second. Hosted execution also depends on a dashboard setting that exposes system environment variables; without it the local world is selected and fails on a read-only filesystem. That setting was not applied, and no live run was observed.
