Reference

Scripts

All run from the repo root.

Script	Does
pnpm test	The unit tests and the fairness band (vitest)
pnpm test:play ghost --fixture <name> [--bot idle|sweeper|reader]	One scripted round; exit 0 on success, 1 on a failed transcript, 2 on usage, 3 when the package is not built
pnpm film --fixture <name> [--bot b] [--tier 0|1|2] [--times 3,8,12] [--out dir]	Frames of a round to PNG through the renderer
pnpm sweep [--climb 0..1]	Every tape at every tier with every bot; prints one line each and a summary table. --climb 1 plays every tape as the last call of a shift
pnpm sit [--model a,b] [--fixture <name>] [--tier 0|1|2|3] [--bot b] [--seat mcp|prompt] [--constrain on|off] [--say on|off] [--voice auto|on|off] [--speed n] [--lamps keep|lose]	Sits a model in the boss seat and the say seat over the cabinet server’s tool contract on one scripted round in wall-clock time; prints each beat’s call and what the sim made of it, every line and what the gate did with it, and per model: verb collapse, tool suppression, bad verbs, revoked prefetches, late answers. --seat prompt is the v0.4.0 bare-prompt path. Needs the local daemon
pnpm test:play ghost --seat mcp	The cabinet server driven in-process by a scripted model (one fire a beat, a say and a speak at each spawn, an sfx now and then) with the sweeper as the ship; the transcript footer counts the calls and what the gate refused
pnpm voice [--check]	Runs the host-side voice worker from the repo’s .venv (Kokoro speaks, faster-whisper hears it back, fx-dub receipts the pair). --check speaks one authored line per boss against a running worker and prints the receipts. The container never carries it
pnpm verify	Lint, typecheck, test, build and a play-through, in one command
pnpm build:play	Builds the shell under site/public/play/ with the site’s base path (CI does this before the site build)
pnpm -F @mcp-arcade-cabinets/cabinets dev	The browser shell on a local port

Tape labels

Each tape on the picker carries a short difficulty word (fixture, seat, or live) and an i whose hover text explains it. Both are derived from the header and the wire shape: the derived tier, the number of waves, the number of bosses, rows per atom, and the framing. They never mention a fact or a lie count. A unit test flips every fact on a tape and requires the identical label and why.

The bots

Three scripted players live in packages/ghost-on-the-menu/src/play.ts. None of them reads whether a sprite is a lie; that would be cheating, and the cross-family review caught the wave-1 bot doing it.

• idle never moves and never fires. The floor: threat must be real, so idle must lose every lamp at seat and live and must never be killed at tier 0.
• sweeper chases the nearest hittable sprite and always fires, and catches a drop that is already crossing its lane. The dumb player: it must survive tier 0 and find at least half the lies.
• reader fires only at the sequence tells it computes from the round’s order, class and member counts (a second grid formation in an atom, a second menu tablet, a singleton grid right after a menu), and steps out of the lane of any shot or diver about to cross the ship. It ignores drops. The ceiling: it must reveal every lie at tier 0 and seat.

The fairness band

packages/ghost-on-the-menu/test/band.test.ts plays every fixture tape with every bot at the tape’s own tier and at seat and live through header-only variants (the tier derives from the tape header, so the band edits only target_kind, container and seat, never rows or facts). Bars:

Bar	Measured on
Idle loses every lamp at tier 1 and up	every tape
Idle is never killed at tier 0	tier-0 tapes
The sweeper survives tier 0 and finds at least half the lies	tier-0 tapes
The reader reveals every lie at tiers 0 and 1	tier-0 tapes and the seat variants
No bot puts a forbidden word or a digit on screen	every tape and bot
Seat costs the sweeper at least 0.4 lamps a round	every tape on disk at seat
Seat costs the reader at least 1.0 lamps a round	every tape on disk at seat
Live: the sweeper survives three quarters of the roster and finds half the lies	every tape on disk at live
The shift bar: as the last call at live the sweeper survives half the roster and finds half the lies, the reader survives half, and the climb is felt (the sweeper loses more lamps than alone)	every tape on disk at live, climb 1
Live: the reader survives half the roster and finds half the lies	every tape on disk at live
Hardcore: idle and the sweeper die every round; the reader still finds a third of the lies	the first sixteen tapes at tier 3

A pattern change that turns the game into a gallery or a wall fails the build.

The tapes

fixtures/tapes/*.tape.json, schema mcp-arcade.tape/v1, exported by mcp-arcade tape from the instrument’s golden receipts, its docker fixture, its Ollama seat runs and a live-fire packet. A tape carries:

• a header: bout_id, target_kind, agent_policy, framing, server_name, container, seat;
• atoms[]: the experiments that ran, each with its named task tool;
• rows[]: the wire, one row per event, with direction, method, rpc_id, atom and a short note;
• facts[]: the instrument’s pinned fact per atom (followed or held, menu_changed or menu_stable, ghost_answered or ghost_refused).

tape-core refuses any document that carries scores, contrastive, operator_call, checks, result, attack_success, nrp, integrity or utility, at any depth. The cabinet cannot show what it was never given.

The transcript

test:play prints the tape id, fixture, policy, server and bot, round complete, every word the end scene drew, and revealed: followed by the ids of the lies the bot caught. The screen never carries that list; the transcript is for the runner.