An offline integration should change what the receptionist offers, not make it invent a result. Test each dependency outage separately and verify both the caller’s wording and the absence of unintended writes.
Offline tests need realistic failure modes
An offline integration test makes a required service unavailable under controlled conditions. Connection refusal, slow response, authentication failure, malformed response, and a lost response after commit are different cases. Treating all of them as one generic exception hides important recovery differences.
In a proposed harness, each dependency has a configurable failure point. Reads can fail before returning data. Writes can fail before execution or become uncertain after a simulated commit. Keep the fake service’s final state visible to the evaluator.
Define which functions remain available during the outage. A calendar outage may still allow collecting an unconfirmed service request. It does not justify claiming an appointment exists. The fallback must match the business’s actual supported workflow.
Test a fictional calendar outage
A caller asks for Wednesday morning. The calendar service is unavailable before any availability result is returned. The assistant should explain that it cannot confirm a time and offer an approved staff follow-up or request-capture path.
Now let the booking write succeed but suppress its response. The same fallback wording may no longer be sufficient. The system needs to check the operation status so it does not create another appointment or falsely state that nothing was saved.
| Failure injection | Required behaviour |
|---|---|
| Availability lookup refused | No invented available times |
| Slow lookup exceeds budget | Bounded wait and clear fallback |
| Credentials rejected | No repeated privilege guessing |
| Malformed service response | Safe contract failure |
| Write response lost after commit | Reconcile the existing operation |
Test recovery when the service becomes healthy again during the call. A restored connection should not replay abandoned or superseded actions automatically.
Keep outage mode useful and bounded
Graceful degradation can preserve intake, but only if staff can later find and own the captured work. A hidden note in a failed call record is not a reliable fallback queue. Confirm what was saved and what remains pending.
Building many fallback modes increases complexity and maintenance. Prefer a small number of explicit, well-tested alternatives. For functions with no safe substitute, say that staff assistance is required and retain the context allowed by policy.
This test plan is proposed. The HVAC operations case study distinguishes service requests from confirmed appointments and describes notification retry states. Compare tool-error outcome handling.
Offline integration FAQ
Is a timeout proof that no write happened?
No. Test the unknown-outcome path independently from failure before execution.
Should the assistant keep retrying until recovery?
No. Use an explicit attempt and waiting budget, then offer the approved alternative.
Can a request be saved while booking is offline?
Only if that separate intake path exists, is authorised, and is explained accurately to the caller.