Solutions
What the agents cost, and what that bought.
Agent spend is usually a line on a card statement with no decomposition behind it. This rebuilds it from the transcripts on the machine: per model, per week, and per avoidable cause.
Rebuilt, not reported
No proxy, no telemetry endpoint, no account lookup. The figures come from the same local files the agent already wrote, so nothing new sits in the traffic path.
Measured and modelled, kept apart
Each finding says which it is. Where a number would be a guess, none is given — and each is priced on the meter the traffic was actually billed at.
Budgets that bite
Session, daily, and a tighter unattended one, enforced in the hook before a call runs, sized from this machine's own median session and median day.
The number is API-equivalent, and that matters
It is what the traffic would have cost at list price. If the work runs under a subscription, that is not how it was metered — so this is the right number for comparing two ways of working, and the wrong one for reconciling an invoice. Stating that is more useful than a dashboard that implies otherwise.
Where the money actually goes
Cold caches, context carried past the point where it is doing anything, the same result read twice in one session, an instruction file that is re-sent on every turn, thinking as a large share of output. Each is counted separately and priced on its own meter — carried context billed as fresh input would overstate it tenfold, which is the mistake this field exists to prevent. The nine detectors.
Right-sizing is arithmetic
The comparison prices traffic that already ran against the next model down the ladder. It is not advice: the note beside each swap says what the transcripts suggest about whether the work would survive it, and the view says plainly that a smaller model changes the answers as well as the bill.
A cap says it is about cost
When a budget trips, the prompt says in as many words that this one is about money rather than safety — confusing the two teaches people to dismiss both. And the cap layer fails open: an unreadable policy or a half-written tally lets the call through, while the safety checks above it stay fail-closed. How the budgets work.