The Tower's puzzle rungs are playable now, not just scoredThe Tower's puzzle rungs are playable now, not just scored

October 9, 2026 · Pressing PLAY on a Dice Lock Tower puzzle floor now deals the real solvable board instead of a scored one, checked by a probe that drives an actual browser.Tapping play on a puzzle floor in Dice Lock's Tower now gives the real puzzle, not a scored board, confirmed by a test that clicks through a browser.

Dice Lock's Tower has twenty puzzle rungs scattered through its climb, each one a board proved to clear in an exact number of moves rather than an ordinary points target. Until today they could not be played. towerStep() resolved what a rung was asking for, but main.ts only ever built fights, so pressing PLAY on a puzzle rung dealt the floor's normal board and ended the step on a score bar nobody had asked for.

PLAY on a puzzle rung now regenerates the proved board from the goal's own frozen seed and hands it to the same code that runs all eighty of the game's standalone puzzles: Undo, Reset, Solve, the move dots, the first-encounter card. Routing the rung through that shared path rather than writing it a second time is also what let an old bug surface: three guards in the game were checking which mode the player was in instead of which board they were actually holding, so resuming into a puzzle left Undo dead under live dots.

What a cleared rung pays needed settling too. The spec had left it open, three stars for a board solved unaided, one star if the player presses Solve and is shown the answer, and no state between the two, with a later unaided clear upgrading a one-star mark without wiping anything. That meant the top bar over a puzzle rung could no longer show a score, since a puzzle board doesn't score at all. It now shows rings instead, one per move in the budget, filling as moves are spent, with a single line, SOLVE IN 2, reused unchanged across the map's tag, the die's sheet, the floor's rules and this bar, replacing a points figure and fill bar that nothing in a puzzle ever reached.

None of that would have been trustworthy on the strength of the test suite alone, which is why it was checked by a new script that drives an actual headless browser: clicking the rung on the map, pressing PLAY on its sheet, clicking through the reveal card, and clearing the board by the exact moves its seed is proved to clear in. That matters because this project has twice shipped a green suite covering code that read right and drew nothing: a board that generated empty, a module that silently failed to load, a cash-out screen the game never actually displayed. A script that presses the buttons a player presses catches that class of bug where the first kind of test cannot. It passed all eighteen checks it runs, confirming both the unaided clear (three stars, no leftover mark) and the Solve path (one star, the sheet correctly reading "shown" off the stars alone, with no second field anywhere tracking the same fact twice).

Separately, a dig step's tag turned out to say the opposite of what it meant. A dig asks a player to clear the board down to a line, and its tag read as a row count, DIG 4 ROWS, DIG 5 ROWS, and so on. But the actual rule is a board height ceiling, so the rows a player has to clear is the board's depth minus that number, which runs backwards from how the tag reads. Across all twenty dig rungs the tag's printed number went 4, 5, 6, 5, 5, 5, 4, 4, 5, 5, 4, 4, 4, 3, 3, 3, 3, 3, 3, 3 while the actual difficulty climbed, so the easiest dig in the whole tower carried the biggest number on its tag and the hardest carried the smallest. It was caught because a player reading it in the field, misread it the same way the tag itself was wrong: on floor 1 the number happened to match the true row count by coincidence, which is exactly the kind of agreement that hides a backwards rule until a harder floor breaks it. The tag now reads DIG TO ROW n, naming a destination instead of a count, which reads as getting harder when the number drops, matching how the difficulty actually moves.

Dice Lock is a game where you climb a tower, floor by floor. Scattered through that climb are twenty special floors that are not ordinary score-based challenges but puzzles: boards built so that they can always be cleared in an exact number of moves, no more, no fewer. Until today, those floors did not actually work. Tapping PLAY on one dealt the tower's normal kind of board instead and asked you to beat a score target that had nothing to do with the puzzle at all.

That is fixed now. Tapping PLAY on a puzzle floor deals the real puzzle board, built fresh from the exact setup that was designed to be solvable, and hands control to the same screen the game's eighty other, non-tower puzzles already use: the same Undo button, the same Reset, the same hint feature, the same dots tracking your moves. Reusing that existing screen rather than building a second one also turned up an old bug: a few places in the game were checking what kind of screen you were on instead of what board you were actually holding, so resuming a game you had left partway through could leave the Undo button doing nothing while the move counter kept moving.

The game also had to decide what a cleared puzzle floor should reward, which had been left undecided. The answer: three stars for solving it yourself, one star if you use the built-in Solve button and get shown the answer, with a later unaided clear able to upgrade a one-star floor to three. Because a puzzle floor like this does not have a running score the way an ordinary floor does, the bar at the top of the screen that used to show points now shows a row of small rings instead, one per move you are allowed, filling in as you spend them, next to a line like SOLVE IN 2 that now says the same thing everywhere it appears: on the map, on the floor's info sheet, and on this bar.

None of this could be trusted just because the game's automated tests passed, which is why it was also checked with a new kind of test that controls an actual browser the way a person would: clicking the floor on the map, tapping PLAY, clicking through the pop-up that reveals the puzzle, and clearing the board using the exact moves it was built to be solved with. That matters because this project has twice before shipped tests that passed while the actual feature was broken on screen, for instance a puzzle that generated nothing, or a results screen the game never actually showed anyone. A test that clicks the real buttons catches that kind of failure in a way that reading the code cannot. All eighteen checks passed, covering both solving it yourself and using the hint button, and confirming the game never tracks the same fact about whether you got help in two different places at once.

Separately, one of the tower's other challenge types, a "dig" floor where you clear the board down to a line, had a label that said the opposite of what it meant. The label showed a number of rows, like DIG 4 ROWS, but the actual rule was a height limit, so the number of rows you really had to clear was the board's total height minus that number, which runs backwards from how the label reads. Across all twenty dig floors in the tower, the label's number went 4, 5, 6, 5, 5, 5, 4, 4, 5, 5, 4, 4, 4, 3, 3, 3, 3, 3, 3, 3 while the floors actually got harder, so the single easiest dig floor in the entire tower displayed the biggest number, and the hardest displayed the smallest. It got caught because someone playing the floor read the label backwards too, in exactly the way the label itself was wrong, and only the very first floor happened to show a number that matched reality by coincidence. The label now reads DIG TO ROW n instead, naming the line you are digging toward rather than a count, so a smaller number correctly reads as a harder floor.