Tootsies watches her own vitals, catches a service going down, and files her own fix, usually before anyone in the room notices.
The whole idea: every part of her reports how it's doing. She reads her own health on a loop, and when something craters she doesn't just go quiet, she trips a safety switch, leans on a backup, and orders a fix herself.
Safety switches (circuit breakers)
When an outside service starts failing, she doesn't keep slamming a dead door. A breaker trips, she stops calling it for a bit, and she leans on a backup feed, so she degrades gracefully instead of going dark.
Each outside service has its own breaker, so one going down never drags the rest with it.
Two brains, always one up
She thinks with one AI and keeps a second on standby. If her main brain goes down or runs out of credit, she hands off to the backup mid‑answer and keeps talking, and it works both ways. Every surface can run on either.
Neither outage takes her dark, a stalled brain just hands the mic to the other.
A backup for every feed
The live data comes from outside services, and any one can throttle or drop. So each kind has more than one source. When the main one fails she falls to the backup on her own, you just get the answer.
Odds, scores, music, prediction markets, none of it rides on a single service.
Budget watch
The paid feeds have monthly or daily limits. She tracks how much of each is left and flags one nearing its cap before it hits the wall, so a feed never silently runs out mid-game.
How she scales
The server is hundreds of thousands of messages, and she redeploys on every merge. She keeps every stat current without ever recounting from scratch. Three tricks carry it, each with a name.
A whole-server scan can't finish before the next deploy, so it finishes across as many as it takes.Weekly and monthly are one subtraction away, so the window costs nothing to compute.Proven in the tests, so a cheap daily catch-up is trusted to match a full rebuild exactly.
A twice-daily self-review
On top of the live watch, a routine runs twice a day that grades her own recent posts, checks her speed, and groups any errors, then files a tracked issue for anything off, so a slow drift gets caught even when nothing's outright broken.
She grades herself, with a rigged test
A set of quality checks run her real posts past a judge, is this grounded, in her voice, not invented? Each check ships with a known‑bad example that must get flagged. If the judge ever lets the bad one through, the test itself is broken, so the checks can't quietly rot. She reads the trend, not one off night.
A judge is just another model, so every check keeps a trap that proves the judge still has teeth.
Add it up and she's her own night-shift: two brains that cover each other, a backup for every feed, breakers for the outages, a budget watch, and checks that grade her own work, most of which fixes itself before the room ever sees a hiccup.
under the hoodRailway (runtime logs + deploys) · GitHub (files the fix-orders + issues) · Claude + GPT (two model backends, either can cover the other) · Apple Music/Deezer, SGO/The Odds API, Polymarket/Kalshi (backup data feeds) · her own circuit breakers, budget polls, evals & structured telemetry
mostly invisible by design · see also Order a feature
Tootsie's