The Tower is on Dice Lock's front page as a fourth way to playThe Tower is on Dice Lock's front page as a fourth way to play

September 19, 2026 · Dice Lock's Tower finally shows up on the homepage, three weak abilities got cut for good, and my dashboard now shows how much Claude usage is left.Dice Lock's climbing mode finally shows up on the homepage, three weak abilities got cut for good, and my work dashboard now shows how much AI usage is left.
last updated

The Tower is now on Dice Lock's front page, and three weak abilities are gone

The Tower has been playable since it shipped a day earlier, but only from the in-game Start menu. Nobody landing on the homepage could tell it existed. Today the front page's modes band goes from three scene cards to four, with a new one drawn in the game's own tower silhouette, and the heading above it now reads "Four different ways to play" instead of three. That number has been wrong twice before: once when the tutorial got counted as a mode, once when Puzzles shipped and was left out of the count. Getting it right forced a layout change too. Between 769px and 1099px, three scene cards used to stay three across a two-column band on purpose, because three cards split into two columns leaves an orphan. A fourth card removes that problem, so the three-card exception is deleted and both bands now share one 2×2 rule instead of two that had to agree.

Separately, three of Dice Lock's nine abilities (Row, Degrade and Empower) are gone outright, my call from playtesting the Tower the day before. They weren't disabled behind a flag; they're deleted, along with the code paths only they used: a damage-multiplier field, a reroll-risk ladder, a cross-arena effect callback. None of that machinery had another caller, but each piece still type-checked and had tests, which is exactly how dead code survives: it looks alive until someone proves otherwise. Cutting it means the ability meter now deals evenly across six abilities instead of nine, so every one that's left gets picked more often. A quick simulation afterward (280 matches per ability) shows six of the seven survivors within a few points of a 50% win rate, except Shield at 43.9%, noticeably behind, though at that sample size the gap is suggestive rather than proven. That's a question for the next balance pass, not resolved here.

My dashboard now shows how much Claude usage each account has left

BigBrain, the tooling that runs my automated work, has five Claude accounts it can dispatch jobs to, each metered against its own usage window. Until today, checking how close any of them was to running out meant opening one of three separate admin pages, none of which is the page I actually open day to day. That page, /board, just gained a fifth view for exactly this: each account's percentage against its own halt line, how stale that reading is, and which account would get picked next.

The obvious way to build it (probe live on every page load, cache for five minutes) would have been wrong twice over. The usage API needs a permission scope none of the stored credentials carry, so the real usage check starts an actual Claude CLI session and waits up to 25 seconds; running that on every page load of the page you open when something already looks broken is its own outage. And a five-minute cache would have been a second staleness rule competing with the one the budget gate already keeps, so the page instead just reads whatever the background worker last recorded and shows its age honestly.

The page also had to decide what to do with Gmail, which has no usage number to show at all. Its account, used for the email-organizer project, has no quota endpoint the app is allowed to read, and separately both of its stored login tokens had already expired. Rather than leave Gmail off the page, which would read as "Gmail is fine," it shows up explicitly marked as unmeasured, dated, so the gap is visible instead of silent.

Dice Lock's new climbing mode is finally on the homepage, and three weak abilities are gone for good

Dice Lock's new mode, the Tower, has been playable since it shipped the day before, but only if you already knew to click Start and find it. Anyone just visiting the homepage would never have known it existed. Today the homepage's row of game modes grows from three cards to four, with a new one for the Tower drawn in the game's own tower shape, and the heading above it now correctly says "Four different ways to play" instead of three. That count has been wrong twice before in this project's history: once when a tutorial got miscounted as a mode, once when a real mode called Puzzles was left out. Fixing the count also meant fixing how the cards resize on medium-width screens: with only three cards, splitting them into two columns left one card alone on its own row, so the site used to force three cards to stay in a single row on those screens as a workaround. A fourth card fixes that automatically, so the workaround gets deleted, leaving one simple rule instead of two that had to be kept in agreement.

Separately, three of Dice Lock's nine special abilities, called Row, Degrade and Empower, are removed from the game entirely. This was a decision I made myself after playtesting the Tower the day before, and they weren't just turned off, they were deleted, along with the supporting code that only those three abilities used. That's a meaningful cleanup: unused code that still passes its checks looks exactly like code that's actually being used, until someone takes the time to prove otherwise. Removing three abilities means the game's ability system now spreads its chances evenly across six abilities instead of nine, so each remaining one shows up more often. A quick simulated test afterward, running 280 practice matches per ability, shows six of the seven remaining abilities winning close to half their matches, except one called Shield, which won only 43.9% of the time. That's a real gap, though the test size is small enough to call it a lead rather than a settled fact.

My work dashboard now shows how much AI usage is left before it runs out

BigBrain is the automation system that runs my day-to-day AI-assisted work, and it can send jobs to any of five different Claude accounts, each limited by its own usage cap that resets on its own schedule. Until today, checking how close any account was to hitting that cap meant visiting one of three separate technical status pages, none of which is the page I actually check day to day. That page, called /board, just got a new section built for exactly this: each account's percentage used against its own limit, how old that reading is, and which account the system would use next.

The straightforward way to build this (check live usage on every page load, and remember the answer for five minutes) would have caused two problems. Checking live usage needs a permission the saved logins don't have, so the real check starts an actual AI assistant session behind the scenes and waits up to 25 seconds; doing that on every load of the one page you open when something already looks wrong would just add a second problem to the first. Remembering the answer for five minutes would also have created a second, competing rule about how fresh a number needs to be, on top of the rule the usage system already follows elsewhere. So the page just shows whatever number was last recorded in the background, and honestly says how old it is, instead of pretending to be more current than that.

The page also had to decide what to show for Gmail, which has nothing to display. That project has no way to check its usage limit, and its saved logins had already expired besides. Instead of leaving Gmail off, which would look fine, the page marks it "not measured," with a date, so the gap stays visible.