"Challenge Kestrel, the keeper" holds a seat under the keeper's name,
writes the call to the ledger, says so at the table, and raises a
Sentry issue fingerprinted to the room — one ring per call, with the
room's link in the message — which is the alarm the keeper's phone
listens for. The keeper answers by the link: taking the seat, and
leaving a word if the game must wait. For that word, and for any table
to settle when to start, the lobby now shows its talk and carries a
composer; a joiner sees what was said before they sat. Three calls per
address per hour, one per room.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jm2auWk6RP71CjaAb4FMoG
Connection: while the socket is down a notice says so, the hand, board,
and action buttons go quiet, a card left selected stays selected, and
table talk stays in its box. A command that could not leave is named
as not sent; one that left before the drop is named as unconfirmed
until the board answers. A command on its way shows as sending beside
End turn and confirms when the table replies. Talk shows as sending
until the table echoes it, and comes back to the composer if the
socket drops first.
Rematch: anyone at a finished table may call one. The server makes a
new room hosted by the caller, with the same clockwork at the same
tiers and moods (a mystery stays a mystery) and the same expansion
setting, and keeps seats for the last table's other wizards, who find
the call on the finished table and in their lobby ledger and sit with
one click. The old table's replay stays where it was. A solo table
against the clockwork keeps "play again".
Reading the table: an ongoing spell's chip opens its card with a plain
account — what it does, who cast it on whom, and when it ends, in the
rulebook's own terms. Each score row counts the treasures a wizard has
carried home, and marks a wizard one carry from winning.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jm2auWk6RP71CjaAb4FMoG
Engine: pit-rim exits live in one helper (pitRimExits) shared by the
resolver and the automaton; chargeReach names the six-stride cap; dust
sight is inDust + dustAtEnd; PushPending is exported once; the force-field
counter builds its events in one list; drawUnderSlowDeath says what it
draws under. Test groups are named for the rule they pin, not the
revision that introduced it.
Client: one smoothstep; one edgeScar per battle-scarred edge; one
board-zone rule (which also seats the caption strip on phones); the
compass names live in SIDE_NAMES; net.flash() is the one toast; MediaQuery
replaces two hand-rolled matchMedia listeners; sprites carry their sizes
as plain numbers and their CSS in one rule each; canvas helpers are
formatted like the rest of the tree. Comments that told how a cel or a
fallback used to look now say what it draws. The actions-over notice no
longer blames a pickup when slime ended the turn.
Server and ops: dataRoot() is the one data directory; headlineOf builds
its deeds without a cast; the rollup and skills read the same way the
code behaves.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jm2auWk6RP71CjaAb4FMoG
Idle rooms already went back to their ledgers, but a boot replayed
every ledger before the first sweep could put them to bed again — most
of a restart's fifteen seconds spent rebuilding games nobody was at.
A room put to sleep now leaves a stub beside the ledgers, with the
ledger's size at that moment; at boot a stub whose size still matches
answers for the room and the ledger stays unreplayed. Replayed rooms
take their eviction clock from the ledger's last line, so the first
sweep after a boot writes the stubs for everything long idle.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jm2auWk6RP71CjaAb4FMoG
Two blind reviews of everything since the last pass (afa0e17), every
finding checked against the code, no behavior changed: all thirty
scene goldens match without a re-bless and the engine suite is
untouched.
Reel and renderer: the camera's look-down rule is stated once, beside
LOOK_DOWN, instead of twice in the effect; the empty aim branch that
stood where a cutaway used to be is gone (the guard it implied is now
explicit); the pit events and the punch are handled by their own
types, not through "in" casts; smoothstep is one export used by every
tween instead of eleven inline copies; the two floor rings share one
painter; project() takes a Billboard instead of a third hand-typed
copy of its fields; the strides-left figure and the web rim no longer
shadow the reel's steps and the pane's fx; the die card's verdict is
built from events, not by matching an emoji; the workshop asks for the
hover cue by name instead of passing an empty click handler.
Server and engine: one requestBase() for the origin, one slug pattern
in store.ts gating both the clip page and its files, one 404 for both;
LOOPBACK sits above its only caller; doCounteract names what a counter
is played against once; fearCells sits beside its own docblock rather
than between sightedCellsFor and its.
Deploy: chromiumExe, the private server, ffmpeg, and the reel rewind
live in deploy/lib/harness.mjs, shared by the gate, the recorder, and
the card cutter instead of pasted three times; the recorder drops its
duplicate frame counters and names its poster settle; the card uses the
gallery's exact gold; the one-time Sentry URL bootstrap leaves
deploy.sh; the backup comment states the rule rather than the incident.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jm2auWk6RP71CjaAb4FMoG
Each clip page already carried its own title, blurb, poster and
og:video — but the assets behind the tags let the unfurl down. The mp4s
were tail-indexed, so previewers in iMessage, Discord, Slack and
Facebook gave up before finding a player; the tags carried no video
dimensions, without which those same sites refuse to embed one; and the
poster was a 736x578 screenshot of the workshop chrome, cropped to
nothing by any share sheet.
deploy/clips-prep.mjs fixes the set before clips-publish.sh ships it:
a lossless faststart remux of every take, a 1200x630 card per clip cut
from the poster's first-person viewport with the title set over it (an
HTML card screenshot through the harness's chromium — ffmpeg here has
no drawtext), and the fpv frame size stamped into clips.json. The page
emits the card with its dimensions, og:video:width/height, and a Share
button that opens the phone's share sheet or copies the link.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jm2auWk6RP71CjaAb4FMoG
?fpv&script=<name> plays a screenplay — a seeded deal, cards slipped
into hands as props, then a command list run through applyCommand, so
every event on screen is a legal move under the real rules. Ten scenes
ship in the catalog: wormhole heists, a fireball threaded through a
warp, an ambush that watches its corridor through a mouth, the leap
over a conjured pit, and the wall between two homes coming down.
The server grows a public gallery to hang them in: /clips lists the
catalog, /clips/<name> pages each scene in two takes — through the
wizard's eyes and from the board — with og:video unfurls and
Range-served mp4s (Safari refuses video without it). Files live in
the clip vault beside the rooms, shipped by deploy/clips-publish.sh;
the name gate is the security.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
Server: turnFacts joins actingSeat's full priority order (a hanging
ward or slow-death window now reads as the lobby's attention, stubs
included), summarize and stubOf share one baseSummary, the stranded
doc comment returns to summarize, the protocol contract names
feedbackList, Sentry init moves below the imports it cannot precede,
randomBytes joins its group, the feedback path hoists once with the
operator-fields note. Engine: the safe's sight check trusts castSight
alone, the locked-safe predicate becomes lockedSafeAt, the doomed-grab
test drives both seats and must reach endTurn. Web: the twin safe
handlers merge, paPoints joins the reset ritual with a derived burnable
range, five decorative empty handlers and a dead class go, the hand
earns real list semantics, fitHandCards cites Card.svelte's tokens,
two wrap-sites reindent, feedbackReports moves home as a named type,
and the chronicle's stay-mounted reason is stated. feedback-reply.sh
adopts its siblings' set/usage conventions and speaks Node. Two test
screenshots leave the repo root.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
Reports gain ids; replies live in the same JSONL, folding onto their
report when read (legacy reports answer to their timestamp). The lobby
asks with the same seat-token proof as the games ledger — wake-free —
and lists each report with its reply and status, or 'the wizards are
studying the moment'. deploy/feedback-reply.sh answers one from here.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
The masthead's room code is now a button: one click copies the /join
link, which seats newcomers in a waiting room and the gallery once the
boards flip. And a 🐞 report opens a two-question slip — what happened,
what was expected — that lands in feedback.jsonl beside the rooms with
the room code, move number, round, and rules rev attached, so any
surprise can be replayed to its exact moment.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
The lazy room wake handed caller-supplied ids to fileFor; an 8-char
traversal like '../../aa' fit through the length trim. Only [A-Z0-9]
codes reach disk now — reads refuse quietly, server-minted writes
fail loud.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
Memory now carries only live tables — a sweep puts finished rooms to
bed after 30 minutes and anything untouched after a day, and getRoom
wakes a sleeping room from its ledger the moment anyone asks (the
seed plus the log IS the game). The room cap counts ledgers on disk,
not just rooms awake. Share-card PNGs render once per share id and
cache immutable. Rate-limit kicks use terminate() so an abuser cannot
hold the socket open by ignoring the closing handshake.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
Third pass, scoped from 7a32370. Three blind reviewers (engine, web,
server/tools) each concluded the work is coherent engineering with
seam-level tells; every finding was verified before touching a line.
Session biography left the comments: the bot brain's heuristics no
longer cite the opponent who taught them, the RNRX coma parenthetical
and the thief-chase citation are gone, the seat-wallet comments state
their invariants without the war stories, and process-named test
groups now name the behaviors they pin. The incident record lives
where history belongs — commit messages and the ledgers.
Structural dedup: one VISIONSTONE one-edge-sight loop serves both
LOS paths; one creature-arrival touch handler serves walking and
warp-stepping (error text aligned); one facingWedge helper draws both
keymap ribbons; one spriteVisibleInCol rule serves the draw pass and
the hover test (which also stops re-sorting per pointermove); and
deepestFacing joins the director, replacing four copied scans.
Test hardening exposed real rot the tells were hiding: the tight
CreatureState cast caught two literals with a bogus field masking
three missing ones; the number-hoarding rig had NEVER run (its
column didn't exist on seed 42 — it now carves its own geometry);
the bank-guard rig now drives the whole table to an arrival
assertion; the bent-trace test walls off straight sight so the bend
must answer. Silent `return`-on-rig-failure became loud throws, and
can-never-fail assertions were removed.
Sweep-up: the eyeTurn ghost comment, the stacked leave() doc
comments (leave now delegates to leaveLocal), the dead ternary in
the seat client, the kick handler's name-coercion drift, kick ledger
lines gain timestamps, archiveRoomFile reuses fileFor, the RULES_REV
alias retires in favor of the engine constant, hitTest un-exports,
the NUL-sentinel hover shape becomes an honest "none" variant, and
the steering holds get named constants. The bezel's stride cluster
also centers per the table's note.
288 tests, 24 ledgers verified, all workspaces typecheck.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
Two lobby-lifecycle gaps from tonight's duels (a token lost to a
client bug left an unkickable ghost seat holding a chair in VQJW).
kickSeat: the host removes any other seat from an unstarted room —
the kick is a ledger line so restarts replay it, and a kicked live
socket is set adrift with notice. abandonRoom: the host dissolves a
lobby or a finished game; every member is notified and detached, and
the ledger is archived to rooms-abandoned/ — moved, never deleted.
The lobby shows hosts a ✕ per seat and an abandon-room control.
Proven live: kick with notice, kick surviving a server restart, and
abandonment archiving its ledger.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015RCWSTnb1KYTPyL4GmhGnF
Tiers degrade resources and repertoire, never judgment — the design
constraint was that no tier may ever look dumb. The APPRENTICE draws
one card a turn instead of two (a poorer wizard, not a worse one),
spends counters only on heavy hits (thrift, not blindness), and
carries a modest spellbook: damage, summons, stones, keys, and
treasure play, with no afflictions, amplifies, ambushes, guarding
tricks, or deja-vu. The ADEPT draws fully and knows everything except
ambushes and amplify. The ARCHMAGE is the full curriculum. Every card
any tier plays, it plays correctly.
The lobby workshop gains a tier picker beside the temperaments
(default adept); the tier persists with the seat, shows in roster and
scoresheet ("⚙ apprentice mystery"), and rides the drive loop.
Measured where it should matter: in berserker combat mirrors the
archmage beats the apprentice two to one, while pure treasure races
stay honest — the handicap lives in the card exchanges a human
actually feels, pinned deterministically in the tournament suite.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
"Random" read as random ACTIONS, and a revealed random pick is barely
a mystery anyway. The fourth workshop button is now "mystery": the
server rolls the temperament and keeps it — the join line records it
for replay, the drive loop plays it, but the room tells the table
only "mystery". Roster and scoresheet show ⚙ mystery; the machine's
behavior is the only tell.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The automaton now plays number cards where they matter: attached to
value-scaled attacks (waterbolt, powerthrust) choosing the biggest
for damage, and played for movement when the goal is just out of
stride. It raises its monsters when no wizard is in spell range and
commands them every turn — creatures march at enemies by the same
hazard-shy BFS and maul whoever shares their square. Targets are
picked by lowest life.
And it has moods. The HUNTER plays the classic game: gold first,
violence when convenient. The BERSERKER hunts wizards over treasure
and finishes games by last-wizard-standing. The WORRIER paths around
enemies, counters at a lower threshold, shields against chaos, and
never brawls. Hosts pick the temperament when seating one (or the
workshop assigns at random); it persists with the seat, shows in the
roster, and the tournament harness proves every pairing finishes.
Across ten seeds of berserker versus hunter: six treasure wins to
four kills — the moods play genuinely different games.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Phase 4 begins. The automaton is a pure function in the engine —
automatonCommand(view) — playing from its own redacted GameView, the
same information a human seat receives: hidden hands stay hidden from
the clockwork. It ranks simple damage spells, counters what hurts
(full shield at 3+, reflection at 4+, blunt at 2+), discards its
worst cards by a value order, BFS-pathfinds to enemy treasures and
home again, refuses to path through hazards, brawls when there is
nothing to steal, and always has a safe fallback; the server's drive
loop steps any bot-held seat through the same runCommand path as
humans, so bot commands log, persist, replay, and broadcast like
anyone's.
Hosts seat them pre-start with "⚙ seat an automaton" (Automaton,
Automaton II, ... V); bot seats persist as tokenless join lines and
restore on boot. Proven three ways: bot-vs-bot engine games conclude
across seeds in 8-14 rounds (~100 commands — human-scale), a full
four-automaton table finishes, and a live websocket game of human
vs. automaton ended with the clockwork carrying two treasures home
through a do-nothing opponent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A say-box under the chronicle sends table talk to the room; messages
land in the event log itself, styled as parchment asides — banter
between the battle lines, the way it happens at a real table. Talk is
seat-authenticated, stripped and capped at 300 chars, rate-limited by
the existing bucket, persisted as chat lines in the room file (the
game replay ignores them; the room keeps its last 500), and replayed
to anyone joining. Async games get unread badges in the lobby ledger
(💬3), cleared by watching the table, and catch-up replays speak the
lines at the step where they were said. Public to the whole room, no
whispers — and no chat in hotseat, where the table talks for itself.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
makeRoomCode checked only the in-memory map, which covers every live
and restored room — but a file that failed to restore stays on disk
with no map entry, and a new room taking its code would append a
second game into the orphaned file, corrupting both. The generator
now rerolls on disk presence too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
"Not applicable in a 2-player game." — the card face. New games deal
from deck revision 2, which removes it when exactly two wizards sit
down; three or more keep it. The revision travels in GameConfig and
the persisted start line, and stored games without one replay against
the original build — changing an existing game's deck composition
would scramble its deterministic replay. Test pins all three cases:
removed at two, present at three, present in legacy two-player games.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Instead of re-scanning every room's log on demand, stats.json now
accumulates beside the room files: each room is remembered at its
highest counted stage (created → started → finished), so boots and
replays reconcile without double-counting, and the tally will survive
any future pruning of old rooms. Finished games contribute their
moves, table-time, and manner of victory exactly once, at the moment
of victory.
And with an accumulator to receive them, hotseat games finally count:
each local game mints an anonymous id, pings "started" with its
player count, tracks its own between-moves clock, and on the final
move reports counts only — commands, minutes, players, win reason.
No names, no moves leave the device. The server dedupes by id, clamps
everything to sane ranges, and the booklet's tally now shows how many
of the games were hotseat tables.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Waiting rooms now show the six wizard standees: tap to claim yours,
claimed ones gray out with the claimant named for screen readers, the
roster shows each player's chosen standee (or "choosing..."), and
conflicts are refused first-come. Choices sync live to everyone in
the room, resolve to first-free defaults for the undecided when the
host flips the boards, persist in the start line of the room log (so
restored games keep their colors), and flow through the same engine
config field the hotseat picker uses. Also fixed: the static-serving
realpath guard crashed the dev server when no client build exists —
it now 404s static requests instead. Verified end to end: claim,
conflict refusal, second pick, start, and in-game colorIndex 5/2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Raw seat tokens no longer touch disk or long-lived memory: rooms store
sha-256 hashes, joins compare timing-safely, sessions keep the raw
token they authenticated with only for minting transfer phrases, and
legacy plaintext room files still load (hashed on read). Verified:
rejoin and transfer both work, and the room file contains only
hostTokenHash — no raw token anywhere.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every room is now an append-only JSONL file (data/rooms/CODE.jsonl):
a birth-certificate meta line, then every join, start, and command.
Because the engine is deterministic, seed + log IS the game — on boot
the server replays each file and reconstructs the exact state, hands,
deck order, and chronicle. Verified live: create, join, play, KILL
the server, restart ("restored 1 room(s) from disk"), rejoin with the
seat token — identical positions, deck count, and hand, with the
chronicle history redelivered redacted per player. The client
remembers its seat (name, room, token) in localStorage and walks back
to the table automatically on connect, clearing the memory if the
seat is stale; a quiet "leave table" control forgets it on purpose.
This is the foundation phase 2 (play-by-turn) sits on: games now wait
indefinitely for their players — and debugging restarts cost nothing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
