Type GameState for real instead of returning any from initialState(), which
made every game.state.* read in every component unchecked. apply() now
switches on payload.case so the oneof narrows, and is wrapped in try/catch so
a throw partway through a case cannot leave a half-mutated snapshot on
screen. Also key chain meanings by position instead of gloss, since the
dictionary gives no gloss-uniqueness guarantee, and gate the word field's
turn-seed effect on the same connection check `enabled` already uses so a
reconnect cannot seed a field the player cannot submit from.
Chat lines are keyed by a client ordinal rather than author plus
timestamp, which one seat sending a burst could duplicate within a
millisecond and turn into a Svelte runtime error in every tab. The nginx
snippet now sets X-Forwarded-For, without which naming that proxy as
trusted would let each client pick its own limiter key. Quick-match skips
a waiter whose connection has already ended instead of seating a ghost,
and a match the server could not open clears the waiting panel. A
near-miss suggestion is withheld when the word would not link or is
already used, so taking it cannot earn a second refusal. The builder's
delete-then-rename fallback is confined to Windows, where the rename over
an existing file fails; elsewhere a failed rename is reported, not made
worse by deleting the good database.
expvar counters for connections, rooms, games, submissions by rejection
reason, eliminations, chat, joins and bot moves, served on a separate
debug address so they never sit on the public mux, plus one structured
word_rejected log line per refused word carrying the normalized word and
its link. GET /readyz flips to 503 while draining; SIGTERM stops new
rooms, waits up to NOITU_DRAIN_TIMEOUT for live games, then shuts down.
GET /version and the startup log carry the build's git describe.
Fuzz targets for the frame decoder, the text sanitizer and Vietnamese
normalization; the last one found that composing before lowercasing
could leave a non-NFC result, now recomposed after lowering.
CI runs on dev as well as main, gates gofmt and golangci-lint, tracks the
buf major instead of an exact pin, and dependabot watches every
ecosystem. The lint findings that had been hidden by the default
per-issue cap are fixed.
A /rules page says, in one place, what nothing in the app said before:
the chain rule, the clock, what a dead end costs, elimination and the last
player standing, how a word is scored and the reconnect window. Linked
from the landing page and from the board and lobby headers.
The chat control moves above the chain as a pill with the unread count,
where it can be reached on a phone mid-game, and the lobby's folded chat
now carries an unread badge too.
ESLint with the Svelte and JSDoc plugins, run in CI; the real findings
it turned up (missing each keys, untyped timer handles) are fixed.
Three read-only reviews, one per surface: the solo turn loop, the online
room, and the global layout with its tokens and accessibility.
They are stateful records of what was found on 2026-09-10, not evergreen
authority — the contrast tables and the drift inventory describe the
palette as it was measured, before the fixes that follow.
Measured against the shipped corpus rather than estimated: viwiktionary is
already a third of what we ship, and undertheseanlp is bigger and denser but
carries an unlicensed Vietlex derivative in one of its three branches.
A room holds up to four people and needs two to start. Both numbers are
server constants sent to the client in RoomState, so the lobby draws
whatever the server allows and widening a room is a server change alone.
Failing a turn eliminates that player rather than ending the game. The
syllable and the used words survive them, the turn passes to whoever is
next, and the last player standing wins. Two seats is that same rule seen
from close up, which is why there is one implementation of it and not two.
A dead end still costs the first player to face it their own clock, as
before: they get their turn, and lose it. Everyone behind them has already
seen that board, so they go out together rather than each sitting out a turn
limit they cannot use — which leaves the player who closed the position
standing, the same outcome two players get.
A player who is knocked out keeps their seat. They watch the rest of the
game, chat included, with only the word input gone, and everybody lands back
in the same lobby when it ends. The result screen is the whole table, ranked
by who outlasted whom, with each score reported beside the place rather than
deciding it.
The turn clock is deliberately not paused for a seat that has dropped, so a
player who loses their connection on their own turn loses it the way anybody
else would. Their reconnect window decides only whether they are still in the
game afterwards. Any number of windows can be open at once, settled by one
timer armed for the nearest of them.
Starting waits for every guest, not merely the first: a room of four that
began on one yes would have dealt three people a turn they never agreed to.
Kicking names a seat and is still refused on a player who is ready, and on
the owner's own — leaving is what an owner who wants out does, and it hands
the room on. Joining stays a lobby thing: a room with a game running turns a
latecomer away even with seats going spare, because there is no way to hand
somebody a game already in progress.
BREAKING CHANGE: RoomState, TurnUpdate and GameOver lose the fields that
could only ever describe a second player, OpponentLeft is retired in favour
of presence on RoomState, and suggestions move to the new PlayerEliminated —
they describe the position that beat a player, which by the end of a longer
game is nobody else's position. ProtocolVersion goes to 2, so a client built
against 1 is refused with a readable error rather than decoding a frame that
now means something else.
Chat belongs to the room rather than to a game, so it works in the lobby
while they agree on one, during it, and in the lobby it ends in.
A player is replayed what was said while they held their seat. That is what
a refresh restores, and it is also the boundary: a room code is pasted into
group chats by design, so somebody who redeems one starts at silence rather
than reading what the last two people said. A seat records where the
conversation stood when it was filled; the room keeps twenty lines and no
more, so a room that lives all day cannot grow.
Text is untrusted input rendered in a stranger's browser, so it goes through
the filter nicknames already used — now with a cap on stacked combining
marks, which that filter admitted. Twenty runes made mark stacking a
curiosity; two hundred make it a glyph cluster tall enough to cover the
board, and it would sit in the history being replayed to everyone who
followed.
Talking is not playing. A chat message does not reset the room's idle clock,
or one open tab could hold a room and its code for the life of the process
by typing into it once every nine minutes. It does not spend the move budget
either, and a line to a player who cannot keep up is dropped rather than
allowed to close their session — losing a line is recoverable, losing a
session mid-game costs them the game. A history is not droppable that way:
it is the frame that corrects a whole panel, and there is nothing behind it.
When a seat is vacated its words stay and its author goes, name included,
and the player who stayed is re-synced rather than left holding a name that
the next person through the door could ask for.
A room used to be a wrapper around one game: joining started it, and the
room died with it unless both players accepted a rematch inside thirty
seconds. It is now a lobby that outlives its games.
Whoever created the room owns it and the other seat is the guest. The guest
readies and the owner starts; the owner has no readiness of their own,
because starting is the same statement. A finished game returns both to the
lobby, where the next one is agreed exactly as the last was — the readiness
that started a game is spent with it.
A guest takes their readiness back before leaving, which is deliberate
friction: a player the owner is waiting on should have to say so before
walking away. The owner can free the seat of a guest who is not ready, and
not of one who is — readiness is a commitment, not an inconvenience. An
owner who leaves hands the room to whoever is left, unreadied, because they
are the one who starts now.
Something has to bound a room that outlives its games: the last player out
closes it, as does ten minutes in a lobby nobody started a game in. A
dropped connection is still not a player leaving — the seat is held for the
reconnect window in the lobby as well as mid-game, so a refresh no longer
costs somebody their room, and a resume lands in the lobby it left.
The rematch handshake is retired, and RoomCreated and RoomJoined go with it.
All three described part of what RoomState now describes in full, and three
messages for one lobby is three ways for a client to hold a view of it the
server never had. One snapshot per recipient, broadcast from the one place
that knows an input is finished, so no handler can forget to send it.
One distroless image of about 25 MB carries the binary, the built frontend and
the derived dictionary. The 179 MB upstream release is downloaded in a builder
stage and never reaches the final image; the derived wordlist is copied in as
its own layer alongside its licence, attribution and notice, because CC BY-SA
4.0 applies wherever that data is distributed and an image is distribution.
FIXTURE_DICT=1 builds the same Dockerfile against the checked-in word sample,
so the image is built and smoke-tested on every push rather than only at
release. An image built only at release time is an image that breaks at release
time.
CI runs the Go suite under race detection, the frontend type check and tests,
the browser suite, and the image with its licence assertions. The wire contract
keeps its own workflow; the test steps it duplicated were removed from it.
docs/deployment.md covers configuration, the reverse-proxy settings that each
break the game in a way that looks like something else, and what a restart
costs.
The board, the home screen and the game-over panel, in Vietnamese, served by
the Go binary as a static single-page app.
The store is a reducer over ServerMessage and computes nothing. Validity, turn
order, scores and the result are read from the wire, which is what lets one
screen serve the bot now and online play later. Every Vietnamese string lives in
one module, including the map from RejectReason to a message, so the server can
send UI keys instead of prose.
The word field is uncontrolled. A Telex or VNI input method composes a diacritic
across several keystrokes, and writing the value back on each one cancels the
composition. It is read on submit and cleared only there.
The countdown is drawn against the server's clock, estimated from the ping round
trip, and settles 300ms early so the ring never claims more time than the server
allows.
The screen owns the socket and the game while it is mounted, and gives the
current game up on the way out. Asking for a game is stored intent rather than a
condition inferred from the board being empty: clearing the board for a rematch
is that same condition, so the inference sent a second StartBotGame and the
server built two rooms that then destroyed each other.
A test greps the built bundle for dictionary words and holds a size budget, so
the wordlist cannot reach the browser unnoticed. Another reads the error codes
out of the Go transport and fails when one has no Vietnamese message.
One goroutine owns each room and its engine. The room goroutine starts before
anyone is seated and seating is itself a message, so reading run() is a complete
proof of the concurrency contract rather than a convention to uphold. The bot
searches a frozen copy of the board instead of the live engine.
Reads carry no deadline; liveness is ping-based, because a read timeout cannot
distinguish a healthy player idling in the lobby from a dead socket.
The hub no longer binds a joiner to a seat before the room decides whether to
seat them. Anyone holding a room code could previously resign or play on a
seated player's behalf, and the room code is the only credential online 1v1 has.
cmd/noitu-server serves the API and, when NOITU_WEB_DIR is set, the built
frontend, with unknown paths falling back to index.html for client routes. All
configuration is environment-only and every variable has a working default.
proto/noitu/v1/game.proto is the single source of truth for every WebSocket
message. buf generates Go types into server/gen and JavaScript types into
web/src/lib/proto; both trees are committed so building needs no codegen
toolchain.
The Go suite emits binary fixtures into proto/testdata and the JavaScript suite
decodes the same bytes, so the two generated clients are checked against one
artifact rather than against each other's assumptions. CI lints the schema,
rejects breaking changes against main, and fails when the committed generated
trees drift from the schema.
game.NumRejectReasons and game.NumEndReasons let the mapping tests prove every
engine reason has a wire value without guessing where the enum ends.
The engine is transport-free: no sockets, no protobuf, and no clock of its
own. Callers pass the current time in and read the deadline back, so every
rule is testable without a timer. One goroutine owns a game.
Validation resolves the word before checking the chain link. Roughly a
third of dictionary aliases move the first syllable, so "sy hai" resolves
to "si hai"; checking the link against what the player typed would reject
legal moves. The used-word set is keyed on the canonical form, which also
stops the same word being played twice under two spellings.
An opening whose last syllable starts nothing is refused. It would hand
the first player a game already lost, with no move and no reason, that
resolves only when the turn timer expires and then reports a timeout.
Three bot difficulties, separated by how far they look ahead rather than
by how willing they are to win: random, one-ply greedy, and depth-limited
negamax with alpha-beta. Strategies see a read-only view of the board, so
a bot cannot bypass the same validation a human's move goes through.
Simulation on the real corpus, alternating sides across 60 games per
pairing: hard beats easy 98%, medium beats easy 87%, hard beats medium
65%. Decisions take at most 19ms against a 150ms budget.
Three defects surfaced only under simulation. Withholding the winning
move from the middle bot, as first designed, made it lose to the random
bot 97% of the time. The search evaluated leaves with an inverted sign,
so it hunted for positions where it was about to be trapped. And the
rate limit on taking an instant win did nothing, because declining the
shortcut let the search rediscover the same move.
Games against the hard bot end after about three moves versus fifteen
for two random bots: with 1,814 dead-end syllables an instant win is
usually available. Tunable, and flagged for playtesting.
Load the whole dictionary into maps at Open and close the database before
Open returns. The plan called for per-lookup SQLite, but a round-trip
benchmarked at 55us against 8.9ns for a map hit, and the hard bot in a
later phase explores hundreds of candidates inside a 150ms budget. The
in-memory form is also simpler: no connection pool, no prepared
statements, no tail latency. Costs ~70ms and ~7.8MB at startup.
Resolve returns the canonical word, never the spelling the player typed.
Canonicalization moves either end: about half the aliases differ in the
last syllable and more than a third in the first, so "sy hai" resolves to
"si hai". FirstSyllable and LastSyllable report the canonical's ends, and
the engine must chain on those or it will reject legal moves.
WordsStartingWith yields an iterator rather than the backing slice. A
caller could otherwise sort, shuffle or append into dictionary state:
verified that a write landed in the store, that most buckets have spare
capacity for append to scribble into, and that concurrent callers race.
Open validates what it loaded against the builder's recorded word count,
cross-checks every out-degree against the words actually indexed, and
rejects orphan aliases. A truncated database otherwise opens cleanly and
the server starts, rejects every word, and fails every room creation.
RandomOpeningWord picks from a pre-sorted slice by binary search instead
of rebuilding a filtered copy per call, cutting room creation from 374us
and 720KB to 18ns and no allocation.
Escape the database path when building the URI: SQLite reads # as a
fragment delimiter, so an unescaped path opens a different file and
reports a misleading schema error.
The store does not log. A library writing to the global logger fights
structured logging later, and the caller has WordCount, AliasCount and
License to state the CC BY-SA attribution itself.
Research the Vietnamese noi tu word-chain game and plan a 7-phase web
implementation: SvelteKit frontend, Go backend, WebSocket transport with
Protobuf framing, and a server-authoritative dictionary over SQLite.
The server validates every move so the browser never holds the wordlist,
which keeps player-vs-player cheat-resistant and lets the bot and PvP
modes share one rule implementation.
Records the validated decisions: words of two or more syllables linking
on first and last syllable, a 20s turn limit, user-typed nicknames, and
Docker deployment behind a reverse proxy.
tiennm99