The agent fleet now works through a spent usage limit instead of stoppingMy AI helpers keep working when their budget runs out

August 24, 2026 · Sessions now reroute to the one model with confirmed headroom instead of draining the queue until reset, proven live at midday on a halt that had held since dawn.Background work used to stop whenever the AI budget ran out; now it switches to a model with room left, ending a halt that had held since dawn.

The fleet on this machine now works through a spent usage budget instead of parking the queue until the window resets. A budget halt used to have exactly one move: drain mode. Nothing new claims, everything queued waits, and the wait is however long the window says it is. As of this morning there is a second kind of fallback provider, FALLBACK_PROVIDER=fable, which stays on the same Anthropic subscription but spawns every rerouted session pinned to claude-fable-5, model aliases included, so a subagent asking for "sonnet" cannot resolve its way back into the exhausted window. There is no key to manage; the CLI's own login is the credential.

Because it spends the same account, it engages only where that actually helps. A per-model window that is not Fable's own: reroute. The account-wide weekly window was the harder call, and I put a condition on it: only if there is a way to confirm Fable's specific weekly budget. There is: the usage payload carries a Fable-scoped weekly window. So a weekly-all halt reroutes only when a live reading actually contains that window and it sits under its own line, re-checked before every spawn; a reading with no such window is not confirmation, and the queue drains the way it always did.

The proof came the same day. The fleet had been halted since 03:42 with the weekly window at 98% and six jobs queued behind it. At midday the probe read weekly-all at 100% and Fable's own window at 16%, the drain lifted, and within seconds two queued jobs were claiming on Fable; the reckreation commit two paragraphs down landed five minutes later. Wiring this in also fixed two latent bugs in the existing MiniMax reroute: a halt whose message promised "the fleet keeps running" was in fact running only the Architect, and rows on the exempt tier (the one tier a halt never applied to) were being sent to the fallback instead of spawning normally.

The evening added the other half of staying up: spawned sessions now carry a durable credential of their own. From 17:30 every dispatch died in seconds with "OAuth session expired". Under launchd the spawn environment has no Keychain, so the CLI fell back to the credentials file, which an interactive login had quietly zeroed when it rotated the refresh token out from under it. My own terminal kept working the whole time, which is precisely why the outage was invisible from one. The fix overlays a long-lived setup token from the control plane onto every spawned session, and like the API key it is deleted from the third-party overlay so it can never travel to another provider's endpoint. Twenty minutes of dead dispatches, and the morning had already tripped over that same credentials file once, when the budget probe went blind because the file had fallen behind the Keychain. That defect is now its own filed ticket.

Over in reckreation, one of the day's rerouted sessions closed out the sign-in audit, and what remained was recording decisions rather than fixing code: the audit's two real defects had been fixed nine minutes after it by an adjacent verification commit. So the README now says that Google sign-in stays on (my call) and what that costs later: a dev instance runs Google on Clerk's shared development credentials, which do not exist in production, so the eventual reckreation.com cutover needs a Google OAuth client of my own on top of the live keys. The same session surfaced the day's best stray fact: this repo has no git remote, and the obvious GitHub name is already taken by an unrelated private project of mine last pushed in 2018. Six commits exist only on this machine until I pick a name, and that is now a ticket too.

My computer runs a team of AI workers in the background, and they spend from a subscription with usage limits. Until today, hitting any of those limits meant everything stopped and waited for the meter to reset, sometimes for a day or more.

Now the system checks whether one particular model still has room in its own separate allowance. If a live reading confirms it does, the work switches to that model and keeps going; no reading means no switch, and everything waits as before. It got its proof the same day it was built: work had been stopped since before four in the morning, and at midday the switch flipped and jobs were running again within seconds.

The evening supplied the matching half. Every background task began dying seconds after launch, because the saved login they all relied on had been quietly cancelled when I logged in from my own terminal, which kept working, so from where I sat nothing looked wrong at all. Background tasks now carry a long-lived pass of their own that my logins cannot invalidate. The outage lasted twenty minutes.

And in one of my side projects, a worker running on the switched-over model wrote down two decisions before they could evaporate: sign-in with Google stays on, plus a note about the extra setup that choice will cost once the project moves to its real web address. It also noticed the project has never been backed up anywhere off this machine, because the natural spot for it online is occupied by an unrelated project of mine from 2018. Picking a new name is now on my list.