Rolling back an action does not undo it
Separate restoring action definitions from compensating for an external refund, email or file change that already happened.
Rolling back a definition changes what subsequent invocations use. It does not reverse the effects of earlier invocations. An email remains sent, a refund remains a provider operation, and a file change remains in repository history.
Anlyon agent-version rollback repoints the agent to an earlier immutable version and its pinned action definitions. It does not restore every policy, prompt, budget or external resource involved in the workflow.
Split the incident into two jobs
The first job is to prevent further incorrect operations. Halt new dispatch if appropriate, inspect the changed definition, and select the intended version. The second job is to identify already dispatched effects and decide whether a separate corrective operation is possible.
| Existing effect | Recovery consideration |
|---|---|
| Sent email | Cannot be recalled through the send adapter |
| Stripe refund | Do not assume a reverse-refund operation exists |
| GitHub file update | A new reviewed update may restore intended content |
| Declared HTTP action | Compensation depends on that API and your business rules |
A corrective operation deserves its own identity, approval and receipt. Avoid hiding it as an internal retry of the original request.
Check the restored definition
Review the selected action version, its target, declarations and approval requirement. Confirm any credential it references is still valid and appropriately bound. A historical definition does not resurrect a deleted secret or make a previously valid target current.
Test the restored definition in a safe environment or against a controlled target. Do not use production side effects merely as a health check. Then resume deliberately, with attention to pending work created before the rollback.
Keep an effect inventory
List affected invocations by time and version, separating confirmed, acknowledged, pending and unknown outcomes. Reconcile uncertain effects before deciding what compensation is necessary. A compensating change based on an assumed failure can create a second incident.
Use version promotion and rollback, incident evidence and multi-step workflow design.