The example is deliberately small. A customer asks for Tuesday at 3 p.m. The agent finds the slot, presents the time, and receives approval. At 10:14:08, it sends the booking request. At 10:14:18, its connection times out.

The timeout tells us what the client observed. It says nothing conclusive about the provider. The provider may have rejected the request, accepted it, or accepted it and delayed the response. Those cases require different next actions.

The first wrong move is calling the attempt a failure

If the workflow writes failed, a generic retry policy may send the same booking again. The second request looks like recovery inside the software and like a duplicate appointment to the customer.

I record the attempt as unresolved. That state carries two facts: a request left the system, and its effect has not been established. While those facts remain true, another booking request is forbidden.

Reconstruct the run without asking the model

The durable record for this run is short:

TimeRecorded fact
10:13:41The customer approved Tuesday at 3 p.m. for the named provider.
10:14:08The workflow sent one booking request with a stable attempt identifier.
10:14:18The client timed out before receiving a conclusive response.
10:14:19The action entered unresolved; another attempt was blocked.
10:14:24A read against the provider found the appointment and its confirmation reference.

At 10:14:24 the workflow can mark the booking confirmed. The evidence came from the provider’s state, not from the agent’s recollection or the wording of its final message.

What the customer should receive

During the ten-second gap, the accurate message is: “I sent the request and have not received confirmation yet. I am checking before I try again.” After reconciliation, the agent can report the appointment time and provider reference.

This wording is part of the operating behavior. A cheerful confirmation during the gap would give the customer false certainty. An apology followed by an immediate retry could produce the duplicate.

The same trace exposes a weak handoff

Suppose the provider offers no read path. The workflow cannot reconcile the attempt automatically. Sending a Slack message or opening a ticket does not finish the customer’s job. Someone must accept the case with the attempted time, provider, request identifier, last confirmed state, and the action that remains unsafe.

Until that acceptance is recorded, the original workflow still owns the unresolved booking. Notification alone cannot make the work ownerless.

Why this page is separate from the field guide

The behavioral-contract field guide explains how I define authority, evidence, and stopping conditions across a capability. This example shows one run and one repair. The distinction matters: the contract sets the rule; the trace shows whether the system obeyed it.