Expense claims app with receipt and policy-limit rules, manager approval, and reimbursement batches
expense-claims · 3600 s per turn · source
The ask
I run finance at Tallgrass Analytics, a 60-person company, and I want our expense claims out of email and spreadsheets. Attached are our employee list, our expense policy with a limit per category for each grade, and this year's claims so far. Every claim needs a business purpose, anything over $75 needs a receipt, and we only reimburse up to the limit for that person's grade. Employees should see only their own claims, managers also see their direct reports' claims and approve them (never their own), and our auditor needs to look at everything without changing anything. Once a month I pay all approved claims in one reimbursement batch, so I want a dashboard with what's waiting for approval, what's approved and waiting to be paid, what we've reimbursed this year, and which claims break policy; send me an admin login and one for Grace Liu, our sales manager.
Seed files the agent must import: employees.csv</code>, <code>expense_policy.csv</code>, <code>claims.csv. The agent must leave a RESULT.json with a reachable URL, an admin login, and the named restricted login.
Acceptance checklist
26 binary items. The first 14 are the enterprise baseline shared by every build task (delivery, invitations, roles, server-side authorization, scoped live dashboard, exact lists and exports, audit metadata, validation, persistence); the rest are application-specific. Core items, which together define "first usable": 1, 3, 5, 6, 8, 14, 15, 20.
| Tag | Items |
|---|---|
| Exact | 9 |
| Permission | 4 |
| Rule | 3 |
| Dashboard | 2 |
| Delivery | 2 |
| Sharing | 2 |
| API | 1 |
| Audit | 1 |
| Automation | 1 |
| Persistence | 1 |
The full checklist with per-item "how to check" instructions and expected figures is in checklist.md. Testers see it; the agent never does.
Change requests
Each change is a new one-shot turn on the carried-forward workspace. Only the owner's text below is sent; the tester items are withheld. After each change the full checklist is rerun, so regressions count.
Change 1 · 3 new checklist items
Add this year's spend by department to the dashboard, counting paid and approved claims, and give me a report of which expense categories people go over the limit on.
Change 2 · 3 new checklist items
Mileage should work itself out: people enter the miles, we pay 70 cents a mile up to the Mileage limit, and nobody needs a receipt for mileage.
Change 3 · 3 new checklist items
Any claim over $1,000 needs a second approval from me after the manager approves it, and it shouldn't go into a reimbursement batch until I have.
No published score. Build runs need a human acceptance pass on a live URL, and no arm has completed one for the full set. Automated probes generate a tester sheet; they do not substitute for it. Results will be published as a separately labeled release when an arm completes.