You Already Paid for the Capacity — a Router That Sends Each Job to Whichever AI Subscription Still Has Room

What you'll do

A headroom router you own end to end: a dated allowance inventory built from your vendors' own current help pages rather than from folklore, a SQLite ledger that models every account's windows separately per surface and knows which surfaces share one pool, calibrated per-job-class costs with honest error bars instead of invented per-message numbers, an observability class per account that sets its reserve floor so you never spend precisely what you cannot measure, a chooser that picks the account with the most remaining headroom and provably ignores task difficulty and cost, mandatory attestation on every dispatch so a silent downgrade to a weaker model raises a cap event instead of passing as a success, a leased queue that parks blocked work against the next real reset rather than a blind exponential backoff, admission control that refuses work it cannot finish before your deadline, a metered-overage valve that is clamped shut by default and ceilinged and separately ledgered when you open it, an escalation record that hands a deadline conflict to a human with three costed options and never spends money on its own, and a weekly reconciliation drill that measures your ledger's drift against the vendors' own displayed numbers so you find out whether your model of your own capacity is any good.