Register a PreToolUse hook on Bash that rewrites each command through rtk,
so `git status` runs as `rtk git status` and the output is compressed on the
way back. The rewrite is transparent; commands are still written normally.
Measured on this repo, a full `git diff HEAD~3` drops from 521 KB to 28 KB,
a recursive grep over app/src/main from 117 KB to 19 KB, and `git log -20`
from 20 KB to 5.3 KB. Small outputs grow marginally because rtk adds a
header, which at those sizes is not worth avoiding.
The container starts empty, so a SessionStart hook installs the binary first.
It pins RTK_VERSION because the installer's default is to resolve the latest
tag through api.github.com, which the proxy answers 403 for every repository
but this one. Outside the sandbox the hook installs nothing and only prints
the command to run, and every failure path exits clean: with rtk absent the
rewrite no-ops and the only cost is verbose output.
Verified that rtk leaves what must not change alone -- `rtk read` returns
GameSetupScreen.kt byte-for-byte, commit messages keep their quoting, and
already-rewritten commands are not wrapped twice.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Urn6eEBA6XzrnfoibL7faC
The update channel is now a property of the installed APK rather than a
setting that could point a phone at the other channel's builds, and the
"check on launch" switch is wired to a check that speaks only when there
is something to install.
The switch has been in Settings since the updater landed and has never
been read by anything: there was no launch-time check to turn on or off.
It is now wired to one.
It speaks only when there is a build to install. An app that is already
current says nothing, and neither does one that could not reach GitHub —
opening the app is not asking a question, and a phone with no signal
should not be told off for it once a launch. The button in Settings *is*
a question and still answers every time, up to date and failures alike.
That rule is `launchOffer`, next to the channel rule and tested with it.
Two things it is careful about:
The setting is read straight from the store rather than from the state
the screens are drawn with. That state begins as the defaults — where
checking is on — and is replaced a moment later by what is stored, so a
check driven off it would have fired on the default before ever learning
it had been switched off.
The prompt waits for the splash to clear before appearing, while the
check itself runs behind it. The look costs no time and cannot land on
top of the tartan.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DCuRESmSZeJoJond43itw2
The updater took the channel from a setting, so a production install
could be pointed at dev builds and vice versa. That setting was the only
thing standing between a phone and the wrong line of builds, and it was
a chip you could tap by accident.
The channel is now a property of the running APK and nothing else. The
app reads the -dev suffix in its own versionName, asks that channel and
no other, and then checks again at the point of decision — a release
belonging to the other channel is refused whatever came back from the
fetch. There is no preference, no key in the DataStore, and no way to
cross over.
Crossing over was never an upgrade in the first place. The two channels
are counted by different workflows and their numbers are unrelated: dev
is on 69 while production is on 40, so "newer" means nothing across the
divide. They are separate apps under separate applicationIds, so having
both is a matter of installing both, and each then keeps itself current.
The rule is now a pure function with no Android in it —
update/UpdatePolicy.kt — so it is unit-tested on the JVM like the rules
engines are, rather than being buried in a class that needs an Activity.
UpdatePolicyTest holds the wall: a dev install is refused a production
build and a production install a dev one, at every version number
including the ones where the two sequences overlap.
The settings screen states which build is running and where its updates
come from. It no longer offers a switch, because there is no longer one
to offer.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DCuRESmSZeJoJond43itw2
Adds cribbage — the crib, the cut, the play and the show, for two head to
head, three playing singly, or four in partnerships — and fixes the hold
that left every end-of-hand scoreboard impossible to dismiss.
Four-handed is the only table where the crib in front of somebody else
is still yours, and the lay-away is where that has to be read. Two AI
seats facing each other take more than half the games against two
random ones, which a side giving away half its own cribs would not.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DCuRESmSZeJoJond43itw2
The pause between hands never ended. driveIdleSeats asks
awaitsConfirmation about the state, and the state does not change while
the players are reading it — so confirmAdvance restarted the loop, the
loop stopped on the same scoreboard, and the next hand was never dealt.
The screens hid their own dialog on the way out, so the table looked
released and then simply sat there.
The release is now something spendable rather than a question re-asked
about an unchanged state: confirmAdvance grants one, and the loop
consumes it to walk past the hold exactly once.
This is not new with cribbage — Golf's card, Wizard's bids against
tricks and the Crazy 8s round score all stop on the same hold and all
end there. Cribbage's show is what made it impossible to miss, because
a game of cribbage is nothing but hands one after another.
Covered by MatchControllerScoreboardTest, which plays a whole game of
one-move hands and insists every one of them stops to be read and then
goes on.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DCuRESmSZeJoJond43itw2
The whole game: the crib, the cut, the play and the show, pegging on a
board drawn with both pegs so the last score is visible as well as the
total.
Three tables rather than one. Two play head to head with six cards and
lay two away; three play singly with five and lay one, with a fourth
card dealt straight into the crib because three hands cannot fill it;
four play in partnerships pegging on one score, where your partner's
crib is your crib. Every table keeps four cards and every crib holds
four, which is what lets one engine serve all three.
The rule that shapes the code is that a game ends on the point that
reaches the target, mid-count if that is where it falls. Every award
goes through one function that stops the game there and refuses
everything after it, so a non-dealer who counts out wins before the
dealer's hand and crib are counted at all — the show reports those as
never counted rather than adding them on.
A go is called by the engine rather than tapped. The count only rises
within a series, so a hand that cannot play now cannot play later in
it, and there is nothing to decide; the log says who went past.
The opponent works out its lay-away rather than guessing: every way of
keeping four is scored against all forty-six cards that could still be
cut, and what goes to the crib is valued separately and added or
subtracted depending on whose crib it is. It takes 114 games out of 120
against random legal play, from either seat.
teamForSeat now takes the seat count, because cribbage is only a
partnership game at four — reading seat % 2 at a table of three would
have put the first and third players on the same side.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DCuRESmSZeJoJond43itw2
The series a build is named for was written out three times — in the
build file and in both workflows' tag lines — so bumping it meant
changing all three and a release could be tagged one thing and built as
another. It now lives in gradle.properties as plgVersionSeries, which
the build reads when it stamps the APK and both workflows read when they
name the tag.
Set to 1.1.
versionCode is deliberately left as the CI run number alone, which is
what makes the bump safe: the run number keeps climbing across a change
of series, so 1.1 never looks like a downgrade to Android and an OTA
update is not interrupted. The series is a name for people; the run
number is what the machines compare. The updater already read only the
trailing number, and its comments now say so rather than naming 1.0.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_019Pr6B8xDDtJJcb6WJMw7Ud
A generated code takes away half the game. Whether to repeat a colour,
whether to leave one out, whether to do the thing this opponent will not
expect — those are real decisions, and no random number generator makes
them.
So a game now opens in a SETTING phase: each player chooses the code
their opponent will have to break, one after the other, and the guessing
starts once both are down. They choose in turn rather than at once
because the engine only ever has one seat on the clock, and it costs
nothing — neither can see the other's, so going second is not an
advantage.
The same move carries both: while the codes are being set it is your
code, and after that it is a guess at theirs. Which one it means comes
from the phase rather than from a flag on the move, because the phase is
the thing that actually knows — a flag could disagree with it.
An empty code used to mean one thing and now means two — not chosen yet,
or chosen and hidden from the device this copy was sent to — so the
state carries a flag per player. Whether the other player is ready is
not a secret, and a screen has to be able to say "waiting for them"
rather than sit blank.
The computer sets a code at random and does not try to be clever about
it: a code chosen to beat a particular solver has a pattern in it, and a
pattern is the one thing a code should not have. That single choice is
the only random in the app drawn from a source nobody else can see —
every other one is seeded from the position so a host and a client
agree, but a code seeded from anything on the wire could be worked out
by the player who is meant to be guessing it.
A code set on a guest's device reaches the host as a move and stops
there: moves are intents sent to the host and are never relayed to the
other player, so the only copy that reaches an opponent is redacted.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_019Pr6B8xDDtJJcb6WJMw7Ud
Verified green on the feature branch before merging: build, unit tests
and Android lint (ci.yml run 51). Nothing in the catalogue is marked
coming soon any more.
The four games that were listed but not written. Nothing in the
catalogue is marked "coming soon" any more.
Checkers is English draughts on the thirty-two dark squares, numbered
the way the game is written down. Captures are compulsory and a multiple
jump is one move object, so a player cannot stop halfway through one —
true by construction rather than by a rule that has to be enforced
somewhere. The move generator matches the published perft counts at one
to four ply (7, 49, 302, 1469), which fold in compulsory capture,
crowning and the men's forward-only rule all at once.
Backgammon's hard rule is not how a checker moves but that a player must
use as much of the roll as they can, and that cannot be checked one move
at a time: a move can be illegal precisely because playing it strands
the other die. So the generator works backwards from whole turns —
memoised depth-first over the remaining dice — and offers only the first
steps of the longest plays. Recomputing it after each checker is what
makes the rule true by construction. Every die comes from the table's
seed and a count of rolls made, which is in the state rather than beside
it so a game is reproducible from what crosses the wire. There is no
doubling cube; the rules document says so plainly.
Mastermind is played as a duel — both players are set a code and both
are breaking one — because the pencil-and-paper game gives the codemaker
nothing to do. It ends at the foot of a round, never mid-round, or going
first would be worth half a guess. It is the one game here with anything
to hide, so viewFor does real work: the host sends each device only its
own code, and a device that was never given a code refuses to score a
guess against it rather than quietly answering "nothing".
Pirates and Bulgars is Sackson's theming of Asalto, and the rules
document is explicit about which parts are attested and which had to be
chosen — the book could not be obtained. Two hunters who take by
jumping against twenty-four who cannot take at all, on a thirty-three
point cross whose drawn lines come from the same adjacency table the
moves come from, so the board and the rules cannot disagree.
130 unit tests across the four engines, all run locally as well as in
CI: game/ has no Android in it, which is what makes that possible.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_019Pr6B8xDDtJJcb6WJMw7Ud
Nine pieces each on twenty-four points: place them, then move them, and
every line of three takes an enemy piece off the board. Two pieces left,
or no legal move, and the game is over. See docs/RULES-morris.md.
The board is numbered ring * 8 + index, so its geometry falls out
arithmetically rather than being typed in as a table of twenty-four
rows: neighbours around a square are the index either side, the ladders
between squares run through the odd indices alone, and a mill is three
consecutive points centred on an odd one. A table would say the same
thing and could be wrong in one row.
A turn is one move object — the piece placed or stepped, and the enemy
piece taken when that closes a mill. Generating the removal as part of
the move rather than as a turn of its own means there is no moment where
a player is halfway through a turn, so a device that drops off the
network between closing a mill and choosing what to take cannot leave
the table stuck, and the search counts one ply for what is one decision.
Three rules about mills are easy to get backwards and all three are
tested: the piece that arrives is what closes a mill, so stepping out
takes nothing and stepping back in pays again; a slide along the line
being completed is not a mill, because the third piece it needs is the
one that just moved; and a piece in a mill is spared unless every enemy
piece is in one, without which a player could never be reduced and the
game would not end.
The AI is negamax with alpha-beta and iterative deepening over the same
rules engine, with a node budget so the placing phase does not stall a
phone. It has no quiescence tail and does not need one: there are no
exchanges in Morris. Its evaluation knows that being down to three is
dangerous even with flying on — without that term the search reads three
pieces going anywhere as freedom and starts steering towards losing
pieces to get there, which is a real game it would throw away.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_019Pr6B8xDDtJJcb6WJMw7Ud
The same sheet, drawn as the paper it replaces: a column per player with
their name at the head, a row per round running down the page, and the
running total ruled off at the foot.
The bottom row is always the one being filled in. Finishing it writes it
down and opens a fresh empty row above the totals, so the place to type
is always in the same spot. The head and the totals sit outside the
scrolling rounds and share one horizontal scroll state with them, which
is what keeps a column and the name over it from ever coming apart —
every row is built by the same composable and is therefore exactly the
same width, which is what lets them share it.
Reordering moves to where a spreadsheet puts it: drag a name sideways
and the whole column travels with it. Tapping a name renames it or
removes that player, tapping a round number down the left edge opens
that row for correction, and a + column at the right adds a player. The
mode toggle and the round dialog are gone with them.
Points still go in as typed text until the row is finished, so a
half-typed minus is not a number yet. The ± button flips the last number
typed: not every keyboard offers a minus on its number pad, and taking
points away has to work on all of them.
adjustResize is now stated in the manifest rather than left to the
system to resolve. This screen types into the bottom row with the totals
and the finish button below it, and they have to stay above the keyboard
rather than being panned off the screen.
The model is untouched: ScoreSheet and its tests did not need to change
for any of this, which is the point of it having no Android in it.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_019Pr6B8xDDtJJcb6WJMw7Ud
A pencil, reachable from the main menu: say how many are playing, name
them, and write down what each scored every round. The tally sits at the
foot of the screen because that is the one number anybody looks up.
Points are recorded against a player's id rather than their position,
which is what lets the players be dragged into a different order without
their scores moving with them. Ids are never reused, so somebody sitting
down where somebody else got up starts at nought instead of inheriting a
stranger's score. Removing a player does take their points with them: a
score nobody can see is a score nobody can correct.
The drag leaves the sheet alone until the finger lifts. Rows move on
screen as the drag passes them, but the positions the gesture is working
from cannot shift underneath it, and the handlers read the sheet through
rememberUpdatedState so a drag cannot be applied to a stale copy and
throw away a name typed since.
Rounds are not written in ink — tapping one reopens it for correction or
strikes it out — and the sheet is saved as it is edited, so a phone call
does not lose the game. It lives in its own DataStore rather than with
the settings: a game in progress is not a preference.
score/ScoreSheet.kt has no Android in it, for the same reason game/ does
not, and every rule above is unit-tested on the JVM.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_019Pr6B8xDDtJJcb6WJMw7Ud
Adds CLAUDE.md documenting that nothing the sandbox does — pushing,
opening a PR, merging one — triggers a workflow run, and that
workflow_dispatch is the only mechanism that does.
Verified on this branch: dispatching ci.yml on a feature branch runs
build, unit tests and lint while skipping the staging and publishing
steps, which are gated on refs/heads/dev. Run 31128906621 passed all
three and skipped both publish steps.
Opening this pull request through the API produced no workflow run, so
the previous version of this file was wrong about the API route being
the fix. The suppression covers everything the sandbox does — pushes,
pull request opens, and merges — because it follows the integration, not
the account the UI attributes the action to.
workflow_dispatch is the only mechanism that works, because it creates a
run directly rather than relying on an event.
Record the loop that follows from that, including the part worth having:
ci.yml gates only its staging and publishing steps on refs/heads/dev, so
dispatching it on a feature branch runs build, tests and lint while
publishing nothing. That is a real pre-merge gate with no YAML changes.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01YEsqYAMJKUJoM7UAavm169
A git push from a Claude session is made with a bot token, and GitHub
creates no workflow run for it — the branch moves and CI stays silent.
Today that put a change on dev and then main having only ever been built
after the fact, and both branches needed a manual workflow_dispatch to
produce anything.
Record the working rule that avoids it: push the feature branch, then
open and merge the pull request through the GitHub API, whose calls are
attributed to the repo owner and do trigger workflows. ci.yml already
listens on pull_request for main and dev, so the PR is what builds the
change before it lands.
Also note what a merge actually does here — dev publishes a dev-channel
prerelease, main publishes a production release — and that Gradle cannot
run in the sandbox, so CI is the only verification available.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01YEsqYAMJKUJoM7UAavm169
A dev build now installs under its own applicationId, org.prolibertate.games.dev,
so it is a separate app to Android and sits beside the production copy instead
of replacing it. Each channel still updates itself normally, because an update
within a channel keeps the same package.
Three things move together off one isDevBuild flag, because a dev package
wearing the production name and colours is exactly the mix-up that having both
installed is meant to prevent:
- applicationId gets a .dev suffix;
- the launcher label becomes "Pro Libertate Games DEV";
- the cloth becomes Wallace Hunting, the same threadcount with green where
the clan sett has red.
The name and the colours reach the resources through resValue, so app_name,
tartan_field, splash_background and ic_launcher_background are generated by the
build rather than declared in res/values. The icon vectors reference
@color/tartan_field instead of a literal, which is how one set of vectors serves
both colourways without a second copy of the paths. BuildConfig.DEV_BUILD
carries the same flag into Kotlin, where AppSett picks the Compose sett for the
splash, and WallaceHuntingSett is derived from the same threadcount builder as
WallaceSett so the two cannot drift.
versionName now tracks isDevBuild rather than the CI stamp alone, so an
unstamped local build cannot claim a channel its applicationId contradicts.
Separate packages also mean a channel switch can no longer be refused as a
downgrade; the settings copy and the README said it could, and now say what
actually happens.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01JDuAayWXLRoUNTsfE79Mg1
Multiplayer is now Wi-Fi (LAN) only. The BluetoothTransport, its RFCOMM
plumbing, the BLUETOOTH/BLUETOOTH_ADMIN/BLUETOOTH_CONNECT permissions and
the runtime consent prompt are all gone; the transport abstraction stays,
so LanTransport is simply the only implementation of it now.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01YEsqYAMJKUJoM7UAavm169
Multiplayer now runs over Wi-Fi (LAN) only.
- Delete BluetoothTransport and drop it from LobbyController's transport
list, leaving LanTransport as the sole transport.
- Drop TransportKind.BLUETOOTH and the now-unused RFCOMM_UUID constant.
- Remove the BLUETOOTH/BLUETOOTH_ADMIN/BLUETOOTH_CONNECT permissions and
the android.hardware.bluetooth feature from the manifest, along with
the runtime permission request in MainActivity and the
requestBluetoothPermissions hook on AppEnv.
- Update lobby copy so an empty transport list and an empty host list
talk about Wi-Fi rather than pairing, and refresh the comments and
README sections that described two transports.
The transport abstraction itself is kept: everything after the handshake
was already transport independent, so adding another one later needs no
changes above LanTransport.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01YEsqYAMJKUJoM7UAavm169
Two things reach production together.
Ta Yü is a new board game: 1x3 river tiles laid out from the centre of an
18x18 board, two or four players, scored by multiplying the exits you reach on
one edge by the exits on the opposite one. The game is out of print and its
rulebook could not be obtained, so docs/RULES-tayu.md separates what is
attested from what was reconstructed, and names the three things to check
against a real copy.
The other is the networking fix already on dev: a guest saw nothing until the
next state push, and with more than one guest every device decided it was
sitting in the same seat. Both are worth having in a production build more
than Ta Yü is, since they affect every networked game rather than one new one.
Verified as the merged tree rather than as two separate branches: CI run 26
built, tested and linted e17272b clean, and the published dev APK was opened
to confirm it carries both.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015pyhC1KyFkQk6hUnnX2Tof
The dev prerelease steps were gated on the event as well as the branch, so a
manual run from the Actions tab built, tested and linted dev and then skipped
the publish. That matters because a push does not always fire the workflow:
anything pushed by an automation token is suppressed by GitHub's own recursion
guard, which leaves the dev channel stuck on an older build with no way to move
it from the Actions tab.
Gated on the branch alone now, which is how release.yml has always gated main —
the inconsistency was between the two pipelines, not a deliberate rule. Nothing
about the safety property changes: publishing still only happens after the same
job has built, tested and linted clean, and a pull request is still excluded
without being named, because a PR build runs against refs/pull/N/merge and
never refs/heads/dev.
PLG_CHANNEL moves to the same condition, which is the part that would bite if
it were missed. It decides whether the APK is stamped with its run number, and
the publish steps decide whether that APK ships. Leave them out of step and the
dev channel gets an APK claiming versionCode 1 — an "update" the installed app
can never see as newer than what it already has.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015pyhC1KyFkQk6hUnnX2Tof
The Ta Yü merge landed on dev without a CI run behind it, so the dev channel
still offers the build before it. Nothing in the tree needs to change for a
new one: versionCode and versionName are derived from GITHUB_RUN_NUMBER and
PLG_CHANNEL at build time, so the version iterates on its own and there is no
constant to bump — this commit exists to give ci.yml a push to fire on, which
is the only event the publish steps are gated on.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015pyhC1KyFkQk6hUnnX2Tof
Rivers laid out from the centre of an 18x18 board on 1x3 tiles, two or four
players, against a computer opponent held to measured numbers.
Clean merge: the Ta Yü work touches only its own package, its screen, and the
three wiring points every game goes through, so it does not overlap the guest
resync work already on dev. PlayerSlot.deviceId arrived there in the meantime
but the GameRules contract itself is unchanged, so the new engine needed no
adjustment for it.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015pyhC1KyFkQk6hUnnX2Tof
Rivers laid out from the centre of an 18x18 board on 1x3 tiles, each with
three mouths that must line up with whatever they touch. One side runs water
north and south, the other east and west, and a side's score is one edge
multiplied by the other — so reaching only one of your two edges is worth
nothing, which is the whole tension of the game.
Two things fell out of the geometry rather than needing to be tabulated. The
tile set is generated: three mouths chosen from the eight places one can sit
on a 1x3 tile gives 56 arrangements, turning a tile end for end pairs them
off with none fixed, so there are exactly 28 distinct tiles and the box's 84
is three of each. And because every tile joins the network by a mouth while
all three of its own mouths interconnect, the board is always one connected
river system — so scoring never traces connectivity, and a mouth at the edge
is an exit by construction. Both properties are asserted over whole games.
The placement check is one comparison doing two jobs: where a new tile abuts
an old one, both sides must agree. That refuses a river dead-ending against
stone and, read the other way, refuses a blank flank that would wall off a
river already down.
18x18 comes out under 20dp a cell on a phone, too fine to tap, so laying a
tile is a tap to line it up and a button to commit. Tapping the same cell
again cycles the other legal ways the tile could cover it, which is also how
it gets turned — a rotate control would have landed on illegal positions. A
tap on a cell the tile cannot cover is ignored rather than treated as a
deselection, so a fingertip one cell wide of the mark does not throw away a
placement.
Two AI settings, not three. A blocking-heavy variant was written and dropped
because it played the full setting dead even over 30 games; shipping it as a
step up would have been a claim that could not be backed. What ships is held
to measured numbers: full strength won 20 of 20 against random placement from
either axis, gentle beats random about three games in four and loses to full
strength about four in five, and random against random comes out even, which
is what makes the rest of those numbers mean anything.
The reconstruction's uncertainties stay called out in the rules doc rather
than being smoothed over now that the code exists: the 18x18 board reading,
the marked exits placed at 3, 8 and 13 by choice, and the three-player flood
variant left out because nothing found says how the flood scores.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015pyhC1KyFkQk6hUnnX2Tof
Ta Yü has been out of print since the 2009 Goliath reissue, so the rulebook
had to be reconstructed from secondary sources — and this machine's network
policy blocks everything but search summaries, so BGG's rules files were out
of reach too.
The tile set fell out of the geometry rather than needing a component list.
A 1x3 tile has eight places a river mouth can sit; three mouths chosen from
eight gives 56 arrangements, and turning a tile end for end pairs them off
with no arrangement fixed, leaving exactly 28 distinct tiles. The reissue
has 84 tiles and the 1999 original had 112 — 3x28 and 4x28. Twelve of the 28
touch three flanks of the rectangle, which is what the ringed centre stud on
the reissued tiles marks. Three independent numbers agreeing is why the
geometry is worth trusting.
What could not be pinned down is called out in the doc rather than smoothed
over: the board is taken as 18x18 because that is the reading under which
"the four centre indentations" is true, the marked exits are placed at 3, 8
and 13 by choice, and the three-player flood variant is left out entirely
because nothing found says how the flood scores, and inventing that would be
shipping a different game. Same posture as the Sequence board layout.
The catalogue entry is added with available = false, so it shows as coming
soon on dev builds and stays out of production until the engine lands.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015pyhC1KyFkQk6hUnnX2Tof
A guest saw nothing until the next state push, which on most tables means
until its own turn came round: no board, no cards, no scoreboard. The
cause is a gap in who is listening. A guest's link is opened in the lobby
and read there; the table only subscribes to it once the host's start
message has been through a state change, a recomposition and a
LaunchedEffect. The host publishes the opening position immediately, into
a link whose table has not attached yet, and incoming is a SharedFlow with
replay = 0 — so that first push went to the lobby's collector and nowhere
else. Not a rare race but the normal order of events, which is why it
looked like a rule rather than a bug.
The guest now asks. A new Resync message is sent from inside
onSubscription, so the request cannot leave before the collector is
attached and the reply cannot arrive before there is anything to receive
it. The host answers one peer with that seat's redacted view. A request
that arrives before the deal needs no answer, since a peer that has asked
is listening and the opening publish will reach it.
Second reason a guest could not see its hand, on any table with more than
one guest: guests worked out their own seat by matching PlayerSlot.peerId,
which is the address the host's transport invented for the incoming link —
an IP on Wi-Fi, a MAC over Bluetooth. A guest has never seen that value,
so no guest ever matched it and every one fell through to "the first seat
held by somebody remote". With one guest that is right by luck; with two,
both claimed the same seat and the other drew a hand belonging to a seat
it was not sitting in, which redacted is no hand at all.
The host now records the id a guest introduced itself by, as
PlayerSlot.deviceId, kept separate from peerId because the two answer
different questions: peerId is how the host recognises a link, deviceId is
how a guest recognises itself. The lookup moved out of LobbyScreen into
seatForDevice, where it can be tested, and a dropped player stops claiming
its old seat.
Neither change needs a protocol bump. Resync is ignored by a host that
does not know it, deviceId is an unknown key to a guest that does not read
it, and the old fallback is kept for exactly that case.
Tests reproduce the real start-up order — host publishes, then the guest
subscribes a frame later — rather than the order that happens to work, and
cover a board, a dealt hand, and a guest that is on the clock at the deal.
Both regressions were confirmed to fail before the fix: the guest had no
state at all, and two guests both came back as seat 1.
216 JVM tests pass.
Leaving a multiplayer game deadlocked the main thread in
BufferedReader.close(), which wants the lock readLine() holds for the
whole of a blocking read — so a game of chess could not be ended at all.
Chess also gains an End game action, the other device is told when a
player leaves, and leaving no longer lands back in a live lobby.
Also carries the fixed-name dev APK change from dev.
Leaving a networked game deadlocked the main thread in
BufferedReader.close(), so a multiplayer chess game could not be ended.
Also adds an End game action, tells the other device when a player
leaves, and stops a leave landing back in a live lobby.
GitHub's /releases/latest/download/<name> redirect needs a file name
that stays put, and the release asset is named after its version. Each
production release now carries the same build twice: the versioned copy,
and a plain ProLibertateGames.apk, which makes this link permanent:
.../releases/latest/download/ProLibertateGames.apk
Only the production channel gets one. /releases/latest skips
prereleases, which is exactly what stops a production install being
handed a dev build, so there is no equivalent dev link to offer.
The updater now picks its asset by name rather than taking the first
.apk it sees. Both copies are byte-identical, so nothing would have
broken either way, but relying on upload order for which file gets
installed is not something to leave lying around.
Leaving a multiplayer game froze the app for good. Closing a link closed
the BufferedReader first, and BufferedReader.close() wants the very lock
readLine() holds for the whole of a blocking read — a read that only
closing the socket underneath it can ever end, which is the step the
close never reached. Both callers run on the thread drawing the screen
(AppRoot.pop and PlayScreen's onDispose), so the main thread waited on
itself and the app stopped responding. runCatching was no help: nothing
was thrown, it simply never came back.
StreamConnection.close() now closes only the raw streams, which is what
actually ends the read; the reader coroutine then closes its own wrapper
on its own thread, where it owns the lock. Leaving takes single-digit
milliseconds instead of never returning.
It bit chess because chess is the one game here worth leaving in the
middle of. A card hand is over in minutes; a game of chess between two
people who have had enough of it has no natural end, and there was
nothing to end it with, so:
- Chess has an End game action, with a confirmation, so a table can be
walked away from without hunting for a way out.
- A link that drops on its own now reports itself on the same channel
as everything else, so the device still at the table is told the
other player has left rather than waiting on a move that is not
coming. Bye was in the protocol and never sent; nothing needed to
change on the wire for this.
- Leaving a table no longer lands back in the lobby that started it. A
lobby starts advertising the moment it is drawn, so "leave the game"
used to put the player straight into another one.
Two crashes of the same shape on the same path, both in coroutines
launched in the screen's own scope where nothing catches anything:
- The host's AI applied a move chosen before it took the lock, checking
only that the seat had not changed. applyMove throws on a move that
is no longer legal. Harmless while an AI answered instantly, which
stopped being true when the chess search moved off-thread and started
taking seconds. It is now checked against the position as it stands,
and a stale choice is made again rather than forced.
- A table the rules reject, or a state a client cannot decode, took the
app down. Both are now reported.
Covered by MatchControllerNetworkTest, which plays a game over a real
socket pair and leaves it from either end. The leaving test runs on a
daemon thread and asserts it returns, so a regression fails the build
rather than hanging it — verified by reinstating the old close(), which
hangs the suite until it is killed.
Four more games, taking the app from four playable to eight. All four
are listed in a production build, so a release now offers Euchre,
Kaiser, President, Golf, Wizard, Crazy 8s, Sequence and Chess.
Also carries two changes the chess search forced: the AI now thinks off
the caller's thread rather than on the screen's scope, and offline setup
can seat the local player somewhere other than seat 0 — which is what
lets you play Black.
Mastermind, Backgammon and Checkers remain dev-only placeholders.
The full game: castling with every condition, en passant including the
case where taking would expose your own king, promotion to any of the
four, pins, checkmate, stalemate, insufficient material, the fifty-move
rule and threefold repetition.
Move generation is checked against the published perft counts rather
than against hand-written cases — the opening array to depth four
(197,281 positions) plus four standard test positions chosen precisely
because they break generators. Those counts are exact, so matching them
leaves very little room for a rule to be subtly wrong.
The opponent is a negamax search with alpha-beta, iterative deepening
and a capture-only quiescence tail, at three strengths. The quiescence
tail is the part that stops it hanging pieces: a fixed-depth search that
stops mid-exchange scores the position as though the recapture never
comes.
Two supporting changes:
The AI now thinks off the caller's thread. A card AI answers instantly,
but a chess search runs for a second or more, and the match controller's
scope belongs to the screen — searching there would freeze the board it
is drawn on.
Offline setup can seat the local player somewhere other than seat 0,
which is what lets you play Black. Seat 0 is White by definition, so
choosing a colour has to choose a seat.
FEN is read and written, so a table can start from a set position.
203 engine tests pass locally.
Three more card games, engine to screen.
Kaiser: 32-card prairie trick-taking with bidding, the five of hearts
worth five and the three of spades costing three. Bidding is capped at
ten because ten is all a hand can be worth, and a losing score gives the
game a floor — without one a table that keeps bidding and going down
never reaches a result.
Crazy 8s: eights are wild and name the suit in force, which is why the
table states that suit in words rather than leaving it to be read off
the pile. Blocked rounds are scored rather than hanging.
Wizard: the 60-card deck, one more card dealt each round until the pack
runs out, bid your tricks exactly. The shared card model gains WIZARD
and JESTER ranks, so every ordinary deck now builds from Rank.standard
rather than walking the whole enum.
Also fixes a lobby bug this would have hit: hosting seated the
catalogue minimum rather than the table the host had just set up, so a
six-handed game would have been rejected by its own rules engine.
164 engine tests pass locally.
Two separate faults. The pattern was wrong, and the scale was broken.
Pattern: the sett is published as K/4 R32 K32 Y/4, the slashes marking
pivots, which reflects out to K4 R32 K32 Y4 K32 R32 — 136 threads, about
half black and half red with a yellow line worth 3%. I had the structure
inverted: the yellow overstripe belongs down the centre of a wide black
band, and the narrow black guard belongs between two red blocks. Mine put
yellow between narrow blacks and the wide black elsewhere, which is why it
read as a generic check rather than as Wallace.
Scale: thread width was a fixed fraction of the screen while the sett grew
from 84 threads to 216, so a repeat became 933dp on a 360dp phone and only
0.39 of one was ever visible. The cloth had zoomed into a few enormous
blocks and stopped being a weave at all — that is the part that got worse,
not the colours. Thread width is now derived from the sett, so a couple of
full repeats are on screen whatever the threadcount says, and changing the
count rescales instead of silently magnifying.
The icon is generated from the same numbers at one repeat across the
adaptive safe zone.
Threadcount cross-checked against two independent sources rather than
guessed; recorded in Tartan.kt and the README so the next change starts
from the published figures.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015qjfMiPUDshdKEUa6vxqzN
The old sett was three flat stripes in roughly equal measure, which read
as a generic red-and-black check rather than as Wallace. Rebuilt against
a reference swatch.
- Setts are now written as a half-sett and reflected about their pivots,
the way a threadcount is actually published, so the data stays in the
form it is quoted in instead of a flat list kept symmetric by hand.
- Wallace is gold 6, black 6, red 36, black 24, red 36, black 6: a red
ground at about two thirds, black bands alternating wide and narrow,
and one thin gold overstripe per repeat. The reflection puts black on
both sides of the gold line, which is what the reference shows.
- Colours taken from the swatch: a brighter scarlet and a stronger gold
than before.
- Both the cloth and the icon are turned 45 degrees. The renderer draws
well past the canvas edges so the corners stay covered once rotated.
The icon vector is generated from the same numbers rather than written
out by hand, so the icon and the splash cannot drift apart — changing
the sett rebrands both.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_015qjfMiPUDshdKEUa6vxqzN
Claude