Dice colors now survive color blindness, and usage tracking stopped going darkDice colors now survive color blindness, and usage tracking stopped going dark
October 24, 2026 · Dice Lock's palette now survives color blindness and a full-screen board mode, while a quiet eleven-hour gap in usage tracking got a fallback that closes it for good.Dice Lock's dice now work for colorblind players and the board can fill a phone screen, while a quiet eleven-hour gap in usage numbers got a backup that closes it.
last updated
Dice Lock's board can fill the screen, and its colors now hold up for color blindness
Settings in Dice Lock grew two new switches today, both off by default and both aimed at players the fixed-size board and six-color palette were quietly failing. LARGE BOARD, on a phone only, pulls the top bar out of the normal layout so it carries just the Menu button in the safe-area band, hides the deck entirely, and drops the ability meter's 52px track to zero, so the die cell can grow into the space: measured in headless Chrome, it goes from 48px to 60px on a 390x844 screen and from 43px to 53px on a 375x667 SE. With the deck gone, swipe-up-to-raise and double-tap-to-reroll switch from optional settings to the only way to raise or reroll, and a line under the PLAY card says so. Turning the switch off hands both controls back.
The six-color dice palette had never been checked against color blindness, and measured today it does not hold up: run through a deuteranopia simulation in OKLab color space, face 1 (blue) and face 5 (purple) come out only 1.4 units apart, indistinguishable, and faces 2, 4 and 6 crowd within 8 to 9 units of each other. Protanopia collapses faces 1 and 5 to 6.6 apart, tritanopia faces 4 and 7 to 4.0. A sighted player's own closest pair sits at 12.5, for comparison. Roughly one man in twelve cannot tell these faces apart by color alone. A new COLOR BLINDNESS switch under READING THE DICE swaps in a seven-color palette where every pair is at least 15 units apart under all three forms of color blindness tested, every face still meets a 4.5:1 contrast ratio against its pips, and no face has drifted more than 34 degrees of hue from the color it replaced, so a 1 still reads as blue and a 7 still reads as magenta. A test file computes all 84 pairwise readings from the actual color values rather than trusting numbers typed in by hand.
Usage tracking went blind for eleven hours overnight, and now has a fallback
Separately, in the control-plane repo: the backend that serves /board/usage and the terminal's usage readout starts before anyone logs into the Mac, and every attempt it then made to read the login token from the keychain failed with the same locked-out error, through a restart, for eleven straight hours. A separate background agent reading the identical keychain item ten minutes later worked every time, because it runs after login. The result during that window was not a wrong number but no number: both accounts showed as blind.
That background agent now does double duty. Each run it fetches both accounts' usage directly and writes the result, percentages and reset times only, no token, to a file readable only by its own user. The usage probe falls back to that file whenever it cannot read the login itself, but only if the file is under ten minutes old; past that it still goes blind rather than show a stale number, and says why. The agent's own run interval dropped from ten minutes to three to keep that file fresh enough to matter. A second, smaller fix landed in the same commit: three of the test suite's own probes had been spawning the real command-line tool and waiting out a 25-second timeout on every single run; they're now skipped under the test harness, and the suite's count moved from 5090 to 5098 tests, all passing.
Dice Lock's board can fill the screen, and its dice colors now work for colorblind players
Dice Lock, the dice game, got two new options in its settings today, both turned off unless a player turns them on, and both aimed at players the game had been quietly failing. The first, LARGE BOARD, only shows up on a phone. Turn it on and the top bar shrinks down to just a menu button, the row of pieces at the bottom disappears, and the little gauge that shows your special move goes away too, so the dice themselves can grow bigger to fill that freed-up space. Tested on a couple of common phone screen sizes, the dice grew from 48 pixels across to 60, and from 43 to 53. Since the row of pieces is gone, the only way left to raise the stack or reroll a die is a swipe or a double tap, and the settings screen says so. Turning the option back off brings the old controls back.
The six colors the game uses for its dice faces had never actually been checked against color blindness, and when checked today, they failed. Run through a simulation of the most common form of color blindness, the blue face and the purple face come out looking nearly identical, and three other faces crowd together into a similar problem. Two rarer forms of color blindness each cause a different pair of faces to blur together too. Roughly one man in twelve has some form of this, and for them, telling the dice apart by color alone was close to impossible. A new COLOR BLINDNESS switch swaps in a second set of seven colors, chosen so that every single pair of them stays clearly different under all three kinds of color blindness that were tested, while each one still contrasts well enough against its white dots to read easily, and each new color stays close enough to the old one that a blue die still looks basically blue and a magenta die still looks basically magenta. A test file in the code checks all 84 possible pairings of these colors against real color-blindness math, rather than trusting that someone picked reasonable-looking colors by eye.
Usage tracking went blind for eleven hours overnight, and now has a backup
Separately, in the system that runs this whole setup: there's a background service that reports how much of each Claude account's usage allowance is left, shown on an internal status page and in the terminal. That service starts up automatically before anyone has logged into the Mac, and once it starts that early, every attempt it makes afterward to read the saved login from the Mac's secure credential store fails with the same error, for the rest of the day, even across a restart. A separate, smaller helper program that reads the exact same saved login ten minutes later works fine every time, because it only runs after someone has actually logged in. During an eleven-hour stretch overnight, this meant the usage service didn't show a wrong number, it showed nothing at all, for both of the accounts it tracks.
That smaller helper now does extra work. Every time it runs, it also fetches both accounts' current usage numbers, just the percentages and reset times, no login information, and saves them to a file only its own user account can read. The usage service now checks that file whenever it can't read the login itself, but only trusts it if it was written in the last ten minutes; if it's older than that, the service still reports that it can't get a number, rather than showing something stale without saying so. The helper now runs every three minutes instead of every ten, so that saved file stays fresh enough to lean on. A smaller fix rode along in the same change: three of the automated tests had been starting the real Claude command-line program and waiting a full 25 seconds for it on every single test run. Those are now skipped during testing, and the total number of passing tests went from 5090 to 5098.