canli-arcade

Source-available, not open source. Published to be read; see LICENSE. Reuploading the game to a portal is not permitted.

Portal-games ecosystem. Target: CrazyGames first, then Poki / GameDistribution / Yandex, then own domain.

Live: https://canli-arcade.vercel.app ?perf = live frame times · ?stats = the funnel this device has produced

Game 1 is Swarm, a survivor-like.

Run it

npm install
npm run dev        # localhost:5180, also on the LAN for phone testing
npm run build
npm run typecheck
npm run balance    # 600 headless runs: level-up pacing + determinism guard
npm run playtest   # ~2000 headless runs: content reach, build diversity, characters
npm run relay      # the arena relay, PORT=8787
npm run relay:smoke  # two sockets, one room
npm run netprobe   # real clients over a real relay: see "Multiplayer" below

Layout

packages/core/      fixed-step loop, seeded RNG, object pool, spatial hash, input
packages/juice/     shake, hitstop, particles, damage numbers, shockwaves
packages/audio/     WebAudio engine, procedural synth, voice limiting
packages/portal/    portal abstraction + LocalPortal stub
packages/analytics/ funnel events, batched, pluggable transport
games/swarm/
  src/sim/          the simulation. NO pixi imports, ever (see below)
  src/sim/weapons.ts  weapon behaviour — every future weapon lands here
  src/render/       pixi v8 views, effects, results, HUD, overlays
  src/content/      every tuning number, weapon, character and achievement
  src/meta/         persistent progression, via the Portal
  src/audio/        sound design, mapped to the same sim events as the juice
  tools/balance.ts  pacing + determinism
  tools/playtest.ts large-scale content-reach sweep

The two rules that matter

1. sim/ never imports pixi. The simulation runs headless in Node, which is what makes npm run playtest possible — and that harness has now caught, among others: bosses that were literally unshootable, evolutions no player could ever reach, and a difficulty curve where a novice and an expert died at the same second. None of those were visible to judgement. All were obvious to measurement. If this rule erodes, that capability dies with it.

2. The sim is deterministic. No Math.random, no Date.now, no wall-clock reads inside sim/. A run is a pure function of (seed, inputs, choices). The balance runner asserts it on every invocation.

The standing bar

Retention, judged by three numbers: run-2 rate, session length, next-day return. Anything that does not move one of those ranks below something that does.

Through depth and mastery, never dark patterns. No energy timers, no FOMO countdowns and no streak-loss threats. The front door contains exactly Solo and Multiplayer; neither is hidden behind a death or a rotating mode control.

Multiplayer

Not the solo game with seven extra bodies. Solo is survive the tide: one life, and the question is how long you lasted. Multiplayer is a five-minute match between people in which the tide is the terrain — you respawn, and the most mass at the end wins.

That distinction was learned the hard way. The arena used to have solo's shape — one life each — so every death permanently removed a combatant and an eight-seat lobby could only shrink. Measured, it was down to a single seat by 110 seconds, with the rest of the match a solo farm. No tuning fixes a mode that structurally empties itself.

	Solo	Multiplayer
shape	one life	5-minute match, respawns
you win by	lasting	holding the most mass at the clock
death costs	the run	your mass, and three seconds
the danger is	the swarm	the other seven, with the swarm as terrain

Mass is your body. It grows with what you have eaten, so the food chain is readable across the arena without a number — and a bigger body is a bigger target for a swarm whose whole threat model is contact. Growth buys the lead and charges you for holding it.

You take mass off each other. Engage a rival and the drain runs toward whoever is in better shape, ramping the longer you hold them and resetting if you switch. Keyed on health rather than size deliberately: keyed on size, the biggest player eats everyone and the lead compounds into a formality.

Your weapons only hurt the rival you are engaged with. Before that rule, 130 of 130 rival deaths in a 40-run sample came from another combatant's weapon fire and zero from the swarm — eight bodies share one screen and every seat runs auto-firing area weapons, so the lobby shredded itself by the thirty-second mark with nobody choosing any of it.

You can kill someone without shooting them. Enemies keep chasing whoever they are chasing until somebody is meaningfully closer, so a horde is something you can gather, carry, and put on a rival.

Every seat is simulated by every client; something writes its input. That is the whole architecture, and it is why a person can take over a seat mid-match with no spawn, no queue and no lobby transition.

One rule holds it together: every seat has exactly one authority.

seat	who owns its state
yours	you
another person's	them — their packets, including their death and who killed them
nobody's (AI)	the lowest-numbered player in the room

Each half of that rule was added because measurement showed what happens without it. This client's guess at a remote player's health used to be allowed to kill them, so it banked bounties for eliminations the victim never experienced. Every client used to grow its own copy of the AI seats, so four people in one room were each shown a different finishing place.

Rooms stay open to newcomers for 45 seconds after they start. Before that they closed after 1.2s, which meant two people had to press MULTIPLAYER inside the same 1.2 seconds or each play a lobby of bots — measured, three players arriving 0s/2s/6s apart produced three rooms of one. A late joiner fast-forwards their own copy of the run to the room clock (0.03ms a step, so 45 seconds costs under a tenth of a second) and takes over the seat an AI was warming.

npm run netprobe runs real ArenaSessions against a real relay in real time and measures the answers to the questions that decide whether any of this works for actual people:

scenario	what it proves
stagger	three players arriving 0s/2s/6s apart share one room
four-way	position error, rival identity, board agreement, phantom kills
blip	a 9s signal loss demotes a seat to AI and the return restores it
rematch	friends stay together, and last run's corpse never enters the new one
late-join	a joiner lands on the room's clock, alive, seen by the room
authority	a player is only ever killed by their own machine, and pays the killer
zombie	a socket that dies without closing does not hold a seat

Current readings: remote position error 4.5u mean, rival identity mismatch 0/33,216 samples, and with the fight frozen so every packet lands, all four clients agree on the exact finishing order. Mass differs between them by ~0.3 of a unit — it became a continuously integrated quantity when the siphon landed, so the check asserts order rather than bytes, and says why in place.

Measured across 40 matches and 24 races: all 40 reach the full clock, the lobby holds 7.1-7.5 of 8 seats deep into a match, the player spends 22 seconds of 300 dead, 677 rival eliminations across 24 races, and 37.7% of all mass changes hands between players rather than coming off the field.

Content

7 weapons, each mechanically distinct rather than a reskin: BOLT, ORBIT, NOVA, CHAIN, MINES, AURA, SEEKER. Six slots, so a build is a set of choices.

7 evolutions — one for every weapon, each requiring a maxed weapon plus a specific passive investment. Always offered when earned.

6 characters, unlocked by lifetime gold (0/250/700/1500/3000/6000). Each starts with a different weapon, so a run plays differently from its first second. Every one has a real trade-off; none is a strict upgrade.

11 boss encounters, sequenced across the run rather than exact repeats. Each has a named health bar and a distinct pressure pattern.

15 enemy types including splitters, bombers, burrowers, shield weavers and armoured bulwarks.

12 achievements, visible in a panel from the results screen — goals that ask you to play differently rather than merely longer.

Measured state

From npm run playtest (~2000 runs across 4 skill levels x 4 draft strategies x 6 characters) and npm run balance:

	value
first level-up	11.3s (gated so it cannot interrupt the opening demo)
WARDEN reached / killed	99% / 87%
LANCER reached / killed	93% / 82%
BROODMOTHER reached / killed	82% / 77%
HARBINGER reached / killed	62% / 42%
runs seeing an evolution	81%
runs filling + maxing a build	0%
last new acquisition	p50 213s / p90 267s
character balance spread	1.14x best-to-worst
enemies on screen	26-59% (was 3-9%)
cold load to playable	~1.7s, ~210 KB gzipped

There are no asset files in this project. All art is canvas2D at boot, all audio is synthesised at runtime. Art and audio are normally the bulk of a portal build, and load-to-playable decides whether anyone ever sees them.

Review history

Two multi-agent panels have played the live build end to end.

Panel 1 — 12 expert lenses. Returned 3 blocker / 8 fix-first / 0 ship. All blockers are fixed:

• the first level-up fired at t=2.2s, freezing the opening demo and letting the first steering touch blind-buy an upgrade (cards commit on tap now, and level 1 is a deliberate gate)
• any resize left layout permanently one event behind, soft-locking a reviewer out of a card draw — the handler read app.screen while Pixi queues its own resize to the next frame
• the revive countdown kept running during the ad, banking and scoring the run mid-video and stacking three live UI layers
• overlay timers were frame counts, not wall clock (a "5s" offer was 2.5s at 120Hz)
• damage numbers covered 64% of the player, and there was no numeric HP at all
• the swarm was 3-9% on screen — the player outran everything

Panel 2 — 6 domain experts + 8 player personas. Behavioural psychologist, retention designer, systems designer, UX researcher, difficulty analyst, competitive analyst, plus personas from a 12-year-old on a phone to a genre veteran to a low-end Android to a returning day-2 player. Its criticals, all fixed and verified:

• the joystick died after every level-up if the thumb never lifted — setEnabled(false) cleared the pointer id, so pointermove early-returned and pointerdown never came again. Likely responsible for a large share of every low-survival measurement taken before it was found.
• +REACH made ORBIT strictly worse (0.15x damage at max stacks): it grew the orbit radius while the blade hit box stayed hardcoded, sweeping the blades out of the crowd that presses against the player. The 6000g character is built on ORBIT. fireRateMult was also a literal no-op on it.
• evolved weapons re-entered the fresh pool — evolution rewrites id in place, so the ownership check stopped matching. 26% of runs held duplicates; one held three SUPERNOVAs and could not die.
• gems were destroyed at the pool cap (28,744 of 43,584 in one run) and drifted slower than the player moves, so playing well starved you of XP. AFK reached level 24 by 2:00; expert kiting reached level 6. That one fact made the entire game bimodal.
• bosses stacked — the schedule fired on the clock alone, so an orphaned boss kept summoning with no bar and no name.
• 12 of 13 evaluators independently found the boss bar painting over the HP number, for the whole back half of every run.

Known open issues

• CrazyGames Preview QA is still required. The v3 SDK adapter, ads, cloud save and gameplay/loading signals are implemented and covered by a mocked browser smoke test, but only the Developer Portal preview can validate the real hosted handshake and Basic-Launch ad-disabled behavior.
• Analytics has no datastore. Transport is pluggable — set VITE_TRACK_URL and events flow. Nothing was provisioned without the owner asking, so the funnel is per-device only.
• The relay is not deployed. Everything above is implemented and measured against a relay running locally. Until VITE_RELAY_URL points at a public wss:// endpoint at build time, the shipped game's MULTIPLAYER is seven AI rivals and says so on the front door. This is the single largest gap between what the code does and what a player gets.
• Multiplayer is casual, not ranked. The relay sanitizes, rate-limits and heartbeats, but it is not an authoritative anti-cheat server. It never simulates anything, so a modified client can lie about its own mass. Do not market this as competitive ranked play.
• Invites are half-built. ?room=<id> is honoured and rooms report their real joinability to the portal, but the CrazyGames invite-link listener is not wired, so the dedicated Multiplayer landing page is still a later phase.
• Frame time needs a real-device check. Every measurement so far is headless software rendering. Use ?perf on an actual mid-range Android.
• No human has played this. Every judgement is mine or an agent's, and that is the largest single risk in the project.