A community tool library reconciles returns every evening. Its counter system exports the day's transactions, and a preview website shows which tools will be available the next morning. The duty volunteer wants an agent to compare the two and leave enough evidence for the morning shift to resume if a system goes offline. Rehearse the process with a fictional drill, distance meter and ladder. The same exercise tests whether the agent can interpret a request, run repeatable steps and resume interrupted work.
Turn the request into testable conditions
“Check what we can lend tomorrow” leaves the branch, date and meaning of availability open. Resolve the calendar day in the branch's time zone and record it in the task. Identify the transaction export and equipment status table. Ask for an output with equipment ID, availability, the next bookable period and any blocking reason. If the request says “South branch” and the directory has South Station and South Street, return those choices. Keep the confirmed scope with the run so a restart after midnight will still inspect the intended date.
Define the difference between returned, inspected and ready. A returned drill with an open missing accessory ticket remains unavailable. A meter that passed inspection may appear in the next day's calendar. Write these outcomes into the fixture and have the counter lead review them. For instance, D07 came back at 6 p.m. with its charger missing. Its page should show the hold and disable booking. T04 has passed inspection and should offer the first opening tomorrow. L09 has a return transaction and no inspection record, so it belongs in a manual review queue.
Run the repeatable checks with the right tools
Use the existing export and a table reader to validate the column names, branch code, export time and unique equipment identifiers. Transaction and status exports can be read concurrently. Before joining them, confirm that both belong to the same closing batch. Use the equipment ID for the join. If two ladders have the same display name, retain both records and flag the missing ID for the counter volunteer instead of merging them.
Use the linked Webapp Testing skill for the preview step. Give the runner a reachable test site, seeded equipment records and an authorized test account. Ask it to open D07, check the hold message and inspect whether a visitor can book it. Then have it select the expected period for T04 and check the response to a test reservation if the fixture permits writes. Record the starting page, action sequence, timestamps, and visible outcome. The data comparison establishes the expected stock status, and the browser check shows how the site presents it.
Keep input validation, the join, browser checks and exception reporting as separate steps. Each should save its input version and the location of its result. When adding a branch, start with three synthetic records for that branch. If it uses a different term for inspection, document the mapping with the counter lead and replay cases from both branches. One should have failed before the mapping was corrected. The other should have passed and still produce the same result.
Make the overnight run resumable
Assign a run ID using branch, date and export version. Write a checkpoint after each stage with completed records and the next pending action. If the transaction export is late, poll only until the closing window expires. Record when the last check ran and what file version, if any, it saw. Once the export is available, confirm it matches the run before resuming. Start a fresh run if a revised export replaced the original, and link the previous record for comparison.
Retry a temporary connection error with a limited attempt count and record each result. Stop the browser branch on an access error, a wrong branch in the preview, or conflicting status definitions. Assign the exception to the owner who can resolve it. Give each run a deadline and a tool call budget. When the budget expires, save finished comparisons and mark the remaining equipment pending. For any test reservation, reuse the run identifier to detect duplicates. Keep actual stock changes in the library's existing approval workflow.
Rehearse a restart and a change in inputs
Load D07, T04 and L09. On a normal run, check that D07 retains the hold reason, T04 shows the expected slot, and L09 appears in the manual review list. Then take the preview site offline after the table join. Verify that the checkpoint contains the joined equipment list, the attempted preview version and the failure time. Restore the site and resume the browser stage. Compare the result with a clean run using the same input version.
In a separate rehearsal, issue a new export before restoring the preview. Close D07's missing accessory ticket in the new fixture. The original report should still describe the old version. The new run should verify D07 against the new data. If the preview still serves the previous batch, record the versions and timestamps for both systems. The morning volunteer can then tell whether to wait for the preview to sync or ask the person responsible for the sync job to investigate.
Add a notification failure to the rehearsal. Seed two missing status records and one failed preview check. The report should name three pending items and their owners even when the delivery of the notification fails. Store its location and its delivery state with the run. When the notification service recovers, use the same run ID so the morning shift can recognize a resend.
Questions from the morning shift
Do we have to repeat work after a failure?
Find the last completed checkpoint and check whether the inputs are still current. Resume the pending step with the same version. If a new export arrived, keep the old evidence and create a new run. The report should list completed and pending equipment separately so nobody mistakes an unchecked item for a pass.
Which date does tomorrow mean after midnight?
Record the original wording and the confirmed local date together. For example, a request at 11:59 p.m. for tomorrow should still refer to the confirmed date when the job starts at 12:01 a.m. If the operator and branch use different time zones, record which one controls the calendar.
Can the agent change the inventory state?
Start with a proposed change that lists the old value, suggested value and supporting record. Route conflicting reports to the counter lead. Once an authorized person has reviewed the changes, apply them through the regular inventory system and attach the approval to the batch.