A refused account retries elsewhere now, and Endless swapped math for cardsA backup account now kicks in the moment the main one gets turned away, and a game screen got simpler
September 14, 2026 · The fix that lets a refused Claude account retry on another one got its first live use today, alongside a redrawn Endless mode picker.The fix that lets automation switch to a backup account the moment one is refused got its first real use today, alongside a simpler game screen.
The budget gate retries a refused account instead of waiting on it
A job's third hop today spawned on the default Claude account at 05:27 and the API refused its first turn: the weekly window was full. Ten minutes earlier the budget gate's probe had cleared that same account's weekly halt on a reading of 36%, a number it read off a login and Keychain item that, on this machine, actually belong to the backup account's config directory. The probe has no way to see the token a default session actually spends. So when the refusal came back, nothing halted the account for it, and the retry planner refused to try again until the window's own reset, even though backup was sitting at 4% of its week and could have taken the job immediately.
The fix treats the API's own refusal as a reading in its own right: it's the one answer guaranteed to be about the exact account and window that was actually spent. haltOnRefusal() now halts that account the moment a session is refused, until the window's own reset, and no probe reading can clear it early. The retry planner then checks whether another account has measured headroom and, if so, sends the retry there instead of waiting. That's exactly the case backup at 4% should have caught the first time.
The same day surfaced a second gap in the same system, this time in a live conversation rather than a commit: the auditor kept failing against the backup account with a credits_required error, because that account's org plan didn't include the fallback model, Fable, at all, so every Fable call on it was rejected outright. Upgrading the plan mid-session fixed it; a test call on the backup token completed with no rejection where it had failed before. Two follow-up bugs came out of that same conversation and are still open: the gate will still send a Fable job to an account whose plan doesn't support it, and the backup account's login token had already expired twice that day without anything noticing. Both fixes were proposed and are waiting on a decision before they're made.
Endless mode's picker lost its rate math and gained real cards
The Endless difficulty screen used to list Easy, Medium and Hard as three sentences of computed rate math (something like "70% speed, 8.4 rows/min rising to 25.2 over 4 min") for each one. It was honest, pulled straight from the game's config, and also a paragraph of arithmetic sitting in front of three words a player already understands. Today's redraw drops the sentences and replaces them with the mode's own scene as a hero image and three plain cards, each with one muted best score, laid out in a 3-column grid. The numbers didn't leave the game: the in-run status panel still reports rows/min every second, and the spec and its probe script still check the board against the same curve.
The screen stopped being a picker in the process. A card used to highlight the speed you'd selected, waiting for a separate action to start; now pressing a card starts that run directly, so there's nothing left to highlight between runs. The keyboard path still opens on whichever speed was played last, so Start, Enter, Enter replays it, but that's now just a starting position for the focus ring, not a saved choice. A speed with no runs yet shows a dash rather than a zero, since zero is a score someone could actually earn. The change was checked against the project's full probe suite (pixel layout, both color themes, all eight target viewports, and the rate-curve assertions), all passing.
A backup AI account now takes over the moment one gets turned away
One of the automated jobs I run today spawned onto my main Claude account and got turned away outright: it had used its whole week's usage. Ten minutes earlier, the system that watches usage had checked that same account and decided it was fine, reading it at 36%, but that reading came from the wrong place, a login file that on this machine actually belongs to my backup account instead. The checking system has no way to see the usage the main account actually racks up. So when the account got turned away, nothing marked it as unavailable, and the retry system refused to try again until the following week, even though the backup account was sitting at only 4% used and could have picked up the job right away.
The fix treats getting turned away by the AI service itself as trustworthy information, since that is the one signal guaranteed to be about the exact account and the exact limit that was actually hit. The system now marks an account as unavailable the moment it gets turned away, until that account's weekly limit resets, and a routine usage check can no longer clear that mark early by mistake. The retry system then checks whether a different account has room and sends the job there instead of waiting, which is exactly what should have happened with the backup account sitting at 4%.
The same day turned up a second, separate problem in the same area, found through a conversation rather than a code change: an automated review task kept failing against the backup account because that account's payment plan didn't include the specific AI model, Fable, it needed to fall back to, so every attempt was rejected. Upgrading the plan mid-conversation fixed it right away; a test run on the backup account went through with no rejection where it had failed moments before. Two more problems came out of that same conversation and are still unresolved: the system will still send that kind of job to an account whose plan doesn't support it, and the backup account's own login had already expired twice that same day without anything catching it. Both fixes were suggested and are waiting on a decision before anyone makes them.
A game screen dropped its speed math and became three simple cards
A screen in a game I'm building, in its Endless mode, used to describe its three difficulty levels, Easy, Medium and Hard, as full sentences of calculated numbers, something like "70% speed, 8.4 rows per minute rising to 25.2 over 4 minutes," for each one. Those numbers were accurate, pulled straight from the game's own settings, but they were also a paragraph of math in front of three words any player already understands. Today's redesign removes the sentences and replaces them with the mode's own animated scene as the main image, plus three plain cards, each showing just one faded best score, arranged in a row of three. The numbers didn't disappear from the game: the screen shown during an actual run still reports speed every second, and the underlying design document and its automated check still test the game against those same numbers.
The screen also stopped working like a menu you choose from and confirm. A card used to light up to show which speed you had picked, waiting for you to confirm it separately; now pressing a card starts that run immediately, so there's nothing left to highlight between runs. Using a keyboard still opens on whichever speed was played last, so pressing Start then Enter twice replays it, but that's now just where your selection starts rather than something remembered as chosen. A speed with no runs played yet shows a dash instead of a zero, since a zero is a real score someone could earn by playing badly. The change was checked against the project's full set of automated checks, covering pixel-perfect layout, both color themes, all eight screen sizes the game supports, and the speed-math checks, and all of them passed.