Dice Lock's two-player mode reaches its own homepage, and replies stop vanishingDice Lock's two-player mode reaches its own homepage, and replies stop vanishing
September 17, 2026 · Dice Lock's homepage finally tells visitors about its live two-player mode, and Linear stops swallowing replies typed into an open coding-assistant conversation.Dice Lock's homepage finally tells visitors you can play a friend, and the automated coding assistant stops ignoring follow-up messages typed into an ongoing conversation.
Dice Lock's two-player mode reaches its own homepage
Dice Lock has let two people play a live match since JIO-1161, over three weeks ago, but you would not have known it from the site. Today's audit of the homepage found the Versus card still reading "You and the computer, trading locked dice until one board fills", the same sentence it had before a second human could ever join, while the actual match endpoint had been answering live traffic the whole time. A feature can ship and stay invisible if nobody goes back and rewrites the page selling it.
The fix landed in two pieces. Four homepage mockups were drawn for how far to take it (a plain invite, a rivalry tracker, a leaderboard, a public ranked board), and two were ruled out on the spot: Dice Lock has no account system, no persistent player identity, and the one id a match does track is deliberately kept off screen, so a leaderboard mockup would have been a picture of features nobody had agreed to build. What actually shipped is smaller and honest: the landing page's hero lost its second, rotating headline (it was arguing with itself about whether to sell the daily puzzle or the game), every play button now opens in a new tab so a match in progress can't be lost to the back button, and the Versus card was rewritten to say "or a friend on a link."
Writing the honest version of that page turned up something the shipped feature itself had missed: the match room hands each player the other's raw visitor id, which is also the key their match history and public nickname are stored under. Right now that means one player in a match can look up the other's full history or overwrite their nickname, just by reading a value the game already sends them. It's flagged, not fixed.
Replies typed into an open conversation with the bot stop disappearing
Delegating a Linear ticket to the bot opens what Linear calls an agent session: a running conversation thread on the issue. Linear sends one kind of notification when that session opens, and a different kind every time someone types a follow-up into it afterwards. The code only recognized the first kind; every follow-up came in as the second, got logged as an unsupported event, and was thrown away. Counted over the three weeks between August 26th and September 16th, that happened fourteen times: fourteen people typing an answer or a follow-up question into what looked like an ongoing conversation, and getting silence back, with nothing to show for it but a log line on a machine nobody was reading.
About half of those replies (8 of the 14) also arrived as an ordinary issue comment, which looked like it could have been a fallback. It wasn't: that path exists to check whether a ticket is still open for comments, and on the one case the ticket was filed over, it correctly refused a comment on a ticket already being worked. Both paths were doing their job and the reply still went nowhere.
The fix answers a reply inside the same session it arrived in, rather than opening a new one that would leave the original thread looking abandoned, and it picks up the conversation's existing depth and the person already having it rather than restarting either. Running the full test suite to prove it also caught something unrelated: two branches, merged separately, had each added an identical line to the same file four lines apart. Git merged both copies without complaint, and the file stopped parsing. Nine test files died at start-up as a result, and the total silently fell from 3,934 tests to 3,522. That was the only reason anyone noticed a file had stopped running at all.
Dice Lock's homepage finally admits you can play against a friend
Dice Lock, the puzzle game, has let two people play each other live since early in the project, over three weeks ago now. You would never have guessed that from the website, though: the homepage card describing the two-player mode still said "You and the computer, trading locked dice until one board fills", the same sentence it had before playing against a real person was even possible. The feature worked the whole time; the page selling the game just never got updated to mention it.
Fixing that took two steps. First, four mockups were sketched for how far the homepage should go, from a simple invite link up to a public ranked leaderboard. Two were dropped right away, because Dice Lock has no accounts and no way to remember who a player is from one visit to the next, and the one internal id it does track is deliberately kept hidden from view. A leaderboard would have promised features nobody had agreed to build. What shipped instead is simpler: the homepage's big rotating banner was cut down to a single message instead of flipping between two, every play button now opens the game in a new browser tab so a match in progress can't be lost by hitting back, and the two-player card now says "or a friend on a link."
Writing the honest version of that page turned up a real problem in the two-player mode itself: when two people play each other, the game hands each of them an internal ID number identifying the other player, the same number their match history and public nickname are filed under. Either player can currently look up the other's whole match history, or overwrite their public nickname, just by reading a number the game already sends them. That has been written down as a problem to fix, but it has not been fixed yet.
Typing a follow-up reply to the automated coding assistant stops getting ignored
This project uses an automated assistant that gets assigned tickets in Linear, the tool used to track to-do items, and works each one inside a running conversation attached to the ticket. Linear signals this project's systems one way when that conversation first opens, and a different way every time someone later types a follow-up message into it. The code only recognized the first signal. Every follow-up came in as the second, got logged as something the system didn't know how to handle, and was thrown away unanswered. Counted over the three weeks between August 26th and September 16th, that happened fourteen times: someone typed an answer or a question into what looked like an ongoing conversation and got nothing back, with no record anywhere except one buried log line.
About half of those fourteen messages (8 of them) also showed up as an ordinary comment on the ticket, which looked like it might have quietly covered for the missing replies. It didn't: that comment-reading path exists for a different reason, to check whether a ticket is still open for new comments, and in the one case checked, it correctly said no because the ticket was already being worked. Both parts of the system were behaving correctly, and the message still went nowhere.
The fix now answers a follow-up inside the same conversation it arrived in, instead of starting a confusing new one, and it keeps track of who has been talking and how long the conversation has run. Running the full set of automated tests to check this fix also turned up something unrelated: two separate pieces of work, built and merged separately, had each added an identical line of code to the same file, four lines apart. The tool that combines code changes didn't notice they were duplicates and merged both copies in without complaint, which broke the file so badly that nothing in it could be read anymore. Nine separate test files failed to even start as a result, and the total number of tests running silently dropped from 3,934 to 3,522. That was the only reason anyone noticed a whole file of tests had stopped running at all.