Review failed HVAC workflow events before replaying them

Collect a bounded event trail for the integration owner

When an office integration appears to fail, gather approved diagnostic metadata for a small time window: source system, record reference, event topic, occurrence time, delivery attempt, processing result and downstream reference if available. Exclude credentials, full customer payloads and unnecessary personal information. AI may organize the trail and point to gaps; it should not operate the integration.

Keep this separate from duplicate customer or service-request cleanup. The question is whether an event was delivered, processed and acted on, not whether two customer records happen to look similar. Preserve source timestamps and the integration owner’s actual event-identity rules.

Test a fictional timeout without repeating the action

Suppose an event handler created draft task T-82 but its acknowledgement timed out. A repeated delivery arrives for the same underlying event. The review sheet should identify the recorded task reference and ask the integration owner whether the retry was handled according to the system’s design. It must not recommend creating another task just because delivery was reported unsuccessful.

Jobber documents at-least-once webhook delivery and warns that separate events can also arise from a single user action. Therefore, matching record ID and topic alone is not enough to declare every nearby event a duplicate. Use the provider’s documented payload and the developer’s verified processing rules; do not invent an event ID that the source does not supply.

Leave replay to an authorized technical owner

Use review states such as processing confirmed, downstream effect unresolved, candidate retry and separate source event. Have the integration owner inspect authoritative logs and downstream state before choosing a replay or recovery action. If an earlier action might have sent a message or changed a record, hold the case until that consequence is established.

The assistant’s output should include the exact evidence, missing observation and owner question for each case. It must not call endpoints, reset markers, edit identifiers or erase logs. Rehearse the analysis with fictional repeated, out-of-order and partially completed events. Measure incorrect duplicate classifications and unresolved effects. A queue that reports every event delivered is not proof that the resulting business records are correct.

Review checklist

• Use minimal approved metadata and real source references.
• Separate delivery from downstream effects.
• Distinguish repeated delivery from separate source events.
• Require technical-owner verification before any replay.

Source context and limits

Jobber documents at-least-once delivery, idempotent handling and separate events from one user action. Our reconciliation sheet is a proposed review process, not production code or evidence that an existing integration correctly implements those controls.