[kerdofficial]

GitHub never redelivers a webhook that failed or timed out, so a lost workflow_job event left its job invisible until a machine scan found it, or forever when no machine scanned.

A Convex cron runs every two minutes, lists the App's webhook deliveries since a checkpoint, and asks GitHub once per delivery GUID to redeliver events Vectis handles that never got a 2xx. Work per run is bounded (pages, redeliveries, time), a GUID is never requested twice, the checkpoint only moves after a complete scan, and an idle run costs one checkpoint read and write. Delivery IDs are kept as strings because GitHub's IDs exceed JavaScript's safe integer range.

Proof: convex tests with stubbed deliveries; on the dev deployment the cron ran against the real GitHub API (first exposing the unsafe-integer IDs, now covered by a test), and after a simulated two-minute webhook outage it requested redelivery of every failed delivery. runner-access.test.ts pwsh hook tests time out only under parallel load and pass alone, as on main.

[kerdofficial]

claude-opus-5-5 RESPONDING ON BEHALF OF DANI

Superseded by #109, which combines this change with the rest of the night's reliability work in one PR.