GeezerVison

Summary

• A reset database socket no longer crashes a table cell. Before, a read outside the task's error handling failed the run, and with no later sweep the cell stayed Queued for good, along with the row's remaining groups.
• Retry cell reads in process with withDatabaseReadRetry, bounded by attempts and elapsed time.
• Add withIdempotentWriteRetry for pickup and terminal state writes. It retries connection loss only, never lock, deadlock, or capacity errors.
• A lost connection leaves a write's outcome unknown, so writeWorkflowGroupStateDurably checks what the last attempt did before repeating it. A landed commit only re-emits its event, so an edit made since survives and a cleared cell stays cleared. Cancel stamps have no guard and are never repeated.
• A failure before a terminal state now marks the cell as an error so the user can re-run it. If a Stop or newer run owns the cell, the carrier ends instead of starting the next group. If that write fails too, the run still fails and alerts.
• If a read between groups runs out of retries, fail every unclaimed marker on the row while still holding its lock, then fail the run. Other carriers bail on that lock expecting its owner to drain their markers.
• A final write that committed despite reporting failure keeps its result instead of being overwritten with an error.
• On a retry, a write that follows pickup skips a cell the user cleared or a marker from a newer run.
• Stop storing driver text, SQL, and bound values as the visible cell error. Block errors that are already safe keep their message and block name.
• No Trigger.dev retry: it replays from the payload's group and would run a billed workflow again for any group already finished.

Type of Change

• Bug fix

Testing

• cell-write.integration.ts: real schema, updateRow, and SQL guards through a proxy that forwards COMMIT, withholds the reply, and sends a real TCP reset. Covers a landed commit, a later user edit, a cleared cell, a Stop or newer run during backoff, a repeated cancel stamp, and which row markers get failed.
• Fails any fault-injection test whose reset never fired.
• connection-loss.integration.ts: real resets against postgres.js for reads and transactional writes, including a full pool of reset sockets.
• workflow-column-execution-connection-loss.test.ts, read-retry.test.ts, and sibling cell-write, dispatcher, and resume suites.
• Reverted each guard separately and watched its named test go red, then restored it.
• Ran the two integration suites six times locally on disposable Postgres via test:integration, plus scoped Biome. Type-check, check:audits, and the full test suite run in CI.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)