A booking tool times out after sending a request. The agent retries, and two appointments appear. The error was treating “no response received” as proof that nothing happened.
A timeout describes your observation. It does not establish the remote system's final state.
Introduce an unknown outcome instead of a false failure
Separate validation rejection, confirmed success, confirmed failure, and uncertain completion. For a write that may have reached another system, uncertainty needs a reconciliation process.
| Observation | Safe interpretation |
|---|---|
| Input rejected before submission | No remote write was attempted |
| Remote result includes a confirmed record | Save and return that reference |
| Network timeout after submission | Outcome unknown until reconciled |
| Explicit remote rejection | Handle the documented failure |
Do not collapse every exception into a generic “try again” instruction. The caller-facing response should reflect whether the operation is still unresolved.
Create the operation identity before sending the write
Give the intended business action a stable reference. Persist its request details and processing state in a durable place, scoped to the appropriate customer or organisation.
Use a provider-supported idempotency mechanism when the operation offers one. For example, Stripe documents idempotent API requests. Its semantics apply to supported Stripe operations; they should not be assumed for unrelated calendar or CRM APIs.
If the remote system accepts an external operation reference, it may help reconciliation. Otherwise define a documented lookup strategy and a manual review path. Searching by approximate time and customer name alone can produce ambiguous matches.
Do not claim an operation log creates a cross-system transaction
Your process can crash after the remote write commits but before the local receipt is updated. A local status of “pending” is therefore not proof that the remote operation never happened.
Reconciliation should check for an existing result, confirm its relationship to the original request, and only retry when the provider's guarantees or your workflow make that safe. Keep uncertain cases visible rather than resolving them through optimistic wording.
For a fictional appointment, the receptionist can explain that confirmation is pending and offer the approved follow-up. It should not say the booking failed definitively while a confirmed appointment might already exist.
Test the commit-before-timeout case
Use a controlled integration fixture that saves the remote record but drops the response. Then trigger a retry, restart the integration worker, and inspect record count and operation state.
Also test an actual rejection and a request that never leaves your application. Those cases should not all follow the same recovery path.
The duplicate-booking guide covers repeated writes more broadly. This article isolates the uncertainty created by a missing response.
Tool-timeout questions
Does cancelling a request undo its side effects?
Not reliably. Verify the remote operation's state.
Is automatic retry always wrong?
No. It needs guarantees appropriate to that specific operation.
What should support staff see?
The operation reference, submitted values, last evidence, and next owner. Bring those records to a reliability review.