Make Background Work Retryable by Design

A background task should carry enough structured input to be resumed after a worker failure. Hidden process state and prompt-only instructions make retries ambiguous and invite duplicate side effects.

Define a stable task envelope, record success and failure, and use bounded backoff rather than immediate retry storms. For operations that mutate state, pair retries with an idempotency key or a deduplication check.

The useful question is not whether the first attempt worked. It is whether the second attempt is safe, whether the operator can see the failure, and whether exhausted work has a recoverable dead-letter path.