Keep the identity of the original action
Record the proposal, request, app run, loop case or checkout reference shown by the supported interface. Keep the intended account and Mind with it. A new request with similar wording is not necessarily the same operation.
If the client retained a retry for the original attempt, read what it says may be retried. Do not change the recipient, source or payload while treating it as the same uncertain attempt.
Read the actual state
For a document change, inspect the original proposal and current saved record. Held means review is waiting; applied means it saved. If the source changed after the proposal, compare the current wording before preparing another edit. Review and readback explains the checks.
For an app or loop, open the existing run or case and look for the last completed step, waiting human decision and output. A running case may be waiting deliberately. Do not start another recurring job to replace a case you have not inspected.
For an invitation or payment, inspect the original account-side result before sending or paying again. A confirmation screen and the later joined-member or ready-Mind state remain separate checks.
Inspect before starting again
A lost response leaves the result uncertain until checked.
Keep the original operation reference and inspect its current state. If it already completed, verify the saved or delivered result. If it is waiting, continue the original review or run. If it failed, repair the named cause under current authority. An unavailable read leaves the outcome unknown; it does not authorise a duplicate effect.
Retry the cause that is safe to repeat
A read can usually be tried again after a transient connection error. A write or external action needs its original state reconciled first. Follow the specific procedure’s retry contract and keep the same operation identity when the interface requires it.
If access was revoked, a session expired or the source changed, obtain the current authority or content before retrying. Do not use an old approval as permission for new bytes or a different recipient. Network delivery makes that distinction explicit for sharing.
Deal with a named failure
Use the reported reason: missing grant, exhausted budget, unavailable provider, invalid input or conflicting source. Correct that cause through the authorised owner or maintainer. Repeatedly rewording a request does not repair a failed permission check.
Pause affected recurring work if repeated failures are creating noise, cost or conflicting results. Keep the existing run and output references for diagnosis. Resume only after a small supported check demonstrates the relevant repair.
Report what remains uncertain
State what is known: submitted, waiting, saved, refused or outcome unknown. Include the reference and the observation that supports it. Do not turn an unavailable result into either success or an empty queue.
If the normal interface cannot establish what happened, hand the reference and bounded error to the authorised operator. They can inspect the original operation before deciding whether a new attempt is appropriate.
Related
Practical guides for people and agents