A budget halt can now reroute work instead of stopping everythingA budget halt can now reroute work instead of stopping everything
August 23, 2026 · The spending guard gained a second answer (rerouting sessions to a fallback provider rather than draining the queue) and stopped treating a busy week as a crisis.When spending nears the limit, work can now switch to a backup service instead of stopping, and a weekly limit no longer shuts things down days early.
last updated
The fleet's budget guard learned a second answer today. Until now a halt meant exactly one thing: drain the job queue, stop dispatching, and wait for the spending window to reset. Now there is a second option. When a fallback provider is configured, a halt reroutes background claude -p sessions to it and the queue keeps moving. The trick is that the Claude Code CLI talks to whatever endpoint its environment points at, so a rerouted session is the same spawn as an ordinary one (same binary, same arguments, same logs), with a different environment laid over it just before launch. There is no second runner and no second queue.
Most of the day went into what a reroute must not do. The real API key is deleted from a rerouted session's environment, because pointing the base URL at a third party while the credential is still in there would ship it to that host on every request. Pinned model names are stripped, since the fallback endpoint owns its own model list and asking it for a Claude model produces a session that starts, runs supervised, and delivers nothing. A restart never re-drains a rerouted halt, which would freeze a fleet that was deliberately left running. And the whole mechanism is inert until a fallback key exists in the control plane's secrets file. Until then a halt drains exactly as it always has, and the health endpoint says why.
That same gate also stopped treating a normal week as a crisis. One 90% line used to apply to every spending window, and this week the weekly counter crossed it: the queue sat drained for three days with a tenth of the week's capacity unspent and nothing wrong with the account. The halt line is now a property of the window (a session window halts at 90%, a seven-day window at 98%), and a resume requires every window back under its own line, so a refilled session can't restart the fleet while the week it sits inside is still over.
Elsewhere, the CRM can now hold a tag on a contact. The ticket asked to infer groups from the message database, and there turned out to be nothing to infer from: no label or group column anywhere in the schema, an address book of 249 people and zero groups. So a tag is something I type, and the morning check-in takes --tag family and returns only those threads. Measured against the live data: three people tagged, three threads back, and the unfiltered run identical in every count. The filter deliberately neither hides nor re-ranks. It narrows to a subset I asked for by name.
Sign-in arrived on the games-night site. Clerk landed on the Next side (provider, hosted sign-in and sign-up pages, a header that knows who you are) with no route protected yet, which is honest while nothing writes. A second pass then proved last week's migration against a real scratch database rather than a green suite, and every constraint refused exactly what it exists to refuse. That pass caught what the tests could not: four documentation sentences claiming no auth dependency existed, falsified by the auth commit landing after them, and a documented setup that builds, passes tests, and answers its first request with a 500 because the auth keys were missing from the env template. Both are fixed, and a new test insists every env var the source reads appears in that template.
Over on the projects page, the Vintage Store Map is now listed as live (it has refreshed itself monthly since its Linear project closed), and the feasibility notes across four repos stopped docking points for code that is merely unwritten. Feasibility measures what stands between an idea and a working version; work not yet done is not that.
My projects run on a spending allowance for the AI that does the work, and a guard watches it. Until today, getting close to the limit meant a full stop: every queued job waited until the allowance reset. This week that fired on the wrong number: the counter for the whole week crossed its line, and the queue sat parked for three days with a tenth of the week's budget still unspent.
Two changes came out of that. The weekly counter now has its own, much later trip line, so a busy week no longer reads as an emergency. And when a halt does fire, the work can be sent to a backup service instead of stopping: same jobs, same queue, different supplier. Getting that safe took most of the day: the backup must never receive the key to my main account, and if no backup is set up, everything stops cleanly the way it always did.
The tool that helps me keep up with people learned tags. I had hoped it could work out groupings on its own, but there was nothing in the data to work them out from. My address book holds 249 people and not one group. So I type the tags myself, and the morning catch-up can show just the people marked family. Marking three got exactly those three conversations back. Everything else stayed put.
The games-night site got sign-in pages, and a recheck of last week's database work, against a real database this time, turned up a setup guide that produced a broken site on its first visit, plus documentation the new sign-in had quietly made untrue. Both are fixed. And the projects page now lists the vintage store map as live, which it has quietly been for a while: it refreshes itself monthly without me.