Dice Lock ships a Puzzles mode, with a new one from the server every dayMy dice game gets a new puzzle mode, with a fresh one from the server daily

September 8, 2026 · Dice Lock now has thirty solver-proved puzzles and a daily one the server deals fresh, plus chained tickets no longer open with a false duplicate warning.My dice game now has thirty mathematically checked puzzles plus a fresh one from the server each day, and a bug wrongly warning tickets had failed to start is fixed.

Dice Lock gets a fourth mode, and the puzzles in it are proved, not just playtested

Dice Lock has run on two modes, Endless and Versus, since it launched. Today it ships a third and fourth: Puzzles, with thirty hand-picked boards across six levels, and a Daily puzzle that comes from the server. A puzzle is defined precisely now (a settled board that empties in exactly N changes), and the game engine enforces that definition directly, rather than the six places that used to reach for their own random dice (a reroll, an unlocked gray die's next face, turn gating) each trusting the puzzle to behave.

Getting from that definition to thirty real puzzles took a generator and a solver that check each other. The generator plans a board that clears itself, then scrambles it backwards by N moves; the solver, running on the same board code the player uses, proves the scramble empties in exactly N changes by exactly one sequence. A solver that only agrees with itself proves nothing, so the test suite regenerates all thirty from their seeds, re-proves them, and replays every recorded solution through the real board. The first version of the generator undercounted itself badly: giving up after one failed attempt per seed threw away most of levels three and six, and retrying the plan took level three from 9 puzzles out of 1,500 seeds to 24. A second bug would have shipped a puzzle that looked different from the one it filed, because the code that prints a board dropped its pips.

The Puzzles screen names each puzzle by its seed, remembers which you've solved or merely viewed, and opens any of them straight from a link. A level unlocks once you've cleared or seen every puzzle in the one before it, but a link somebody sent you skips the lock entirely. The Daily puzzle sits on top of the same generator: the server hashes today's UTC date into a seed and serves the level-three board it lands on, caching it in memory so a restart rebuilds the identical puzzle rather than needing to store it. The streak counts consecutive solved days ending today or yesterday, so finishing one after midnight your time doesn't cost you the day, and merely viewing a puzzle still breaks it. The generator itself moved out of the server-only source folder into the game's own web folder, which is what lets a browser fall back to generating the identical board locally if the network request fails.

A chained Linear ticket no longer opens with a false "already running" warning

Separately, in bigbrain-core: every automated handoff between roles (triage passing to delivery, delivery to an auditor) claims the ticket, queues the job, and only then opens a Linear session for that job to report progress into. Linear echoes that session's creation back to the server as its own event, the same shape as someone re-delegating a ticket that's already running. The guard meant to catch a genuine duplicate was catching this too, and correctly refusing to start a second run of it. But it also posted "already running, skipping this delegation" into the very session the job it just queued was about to use. Every chained ticket opened with a report that work had failed to start, when it hadn't.

The fix leaves the guard's decision alone and only changes whether it speaks: if the session Linear is echoing back is the one the handoff itself just opened, nothing is posted. Telling the two apart takes a moment, because the webhook carrying the echo can arrive before the handoff has recorded which session it opened, so a ticket with no session recorded yet is given up to three seconds to catch up before it's treated as someone else's. A ticket already carrying a different session's ID is a real duplicate and gets refused with the message immediately, as before.

My dice game gets a whole new way to play, and its puzzles are mathematically checked, not just tried by hand

Dice Lock is a dice battle game I built, live at jiolab.dev/dicelock. It has had two ways to play since launch: a survival mode and a head-to-head mode. Today it gets two more: a Puzzles mode, with thirty hand-picked boards spread across six levels of difficulty, and a Daily puzzle the server hands out fresh each day. A puzzle now has an exact definition (a finished board that can be cleared in precisely a set number of moves), and the game itself enforces that definition everywhere it matters, instead of six different pieces of code each independently trusting a puzzle to behave itself.

Turning that definition into thirty real puzzles took two programs that check each other's work. One plans a board that would clear itself, then scrambles it backwards by the right number of moves. The other, a solver, proves the scrambled board can only be cleared in exactly that many moves, in exactly one way, running on the same game code a player uses, since a solver that only agrees with itself proves nothing. The tests regenerate all thirty puzzles from scratch, prove them again, and play out every recorded solution to confirm it works. The first version of the puzzle-planning program badly undercounted itself: it gave up after one failed attempt per starting point, throwing away most of the harder puzzles. Trying harder before giving up nearly tripled how many it found. A second bug would have shipped a puzzle that looked different from the one that was tested, because the code that prints a board out was dropping some of its details.

The new Puzzles screen names each puzzle after the random number that generated it, remembers which ones you've solved or just looked at, and lets you open any puzzle straight from a link. A harder set unlocks once you've cleared or at least seen every puzzle before it, but a link someone sends you skips that requirement entirely. The Daily puzzle uses the same generating program: the server turns today's date into a starting point and serves whichever puzzle that produces, keeping it in memory so restarting the server produces the exact same puzzle rather than needing to save it anywhere. A streak counts consecutive days you've solved the puzzle, counting today or yesterday as still current so finishing one after midnight doesn't cost you the streak. But just looking at a puzzle without solving it still breaks it.

An automated system no longer wrongly tells itself that work already failed to start

Separately, in the automation system behind my other projects: when one automated role hands a ticket to another (say, from a triage role to a delivery role), the system claims the ticket, starts the work, and only then opens a chat session in Linear, the project-tracking tool, for that work to report progress into. Linear sends a notification back announcing the session was just opened, and that notification looks identical to someone manually re-assigning a ticket that's already being worked. The safeguard meant to stop a ticket being worked twice at once was catching this too, and correctly refusing to start a duplicate. But it was also posting "this is already running, skipping" into the very session the work it had just started was about to report into. Every handed-off ticket opened with a message claiming the work had failed to start, when it hadn't.

The fix leaves the safeguard's decision untouched and only changes whether it announces itself: if the session Linear is notifying about is the one the handoff itself just opened, nothing gets posted. Telling the two cases apart takes a brief pause, since the notification can arrive before the system finishes recording which session it opened, so a ticket with no session recorded yet gets up to three seconds to catch up before being treated as someone else's work. A ticket already showing a different session is a genuine duplicate and still gets refused with the warning right away, as before.