a game people finish in five minutes — and a bench for when trust breaks

Co-op Zelda-like where player 2 can be an LLM. Deterministic DOM-free core, headless bench, Relationship Memory on physical acts — optional TREASON turns the partner into a wager.

Features · Quick start · LLM partner · Benchmark · Research · Farm reports · Gameplay

---

A tiny cooperative Zelda-like where your partner can be a human in another browser tab, another country — or a large language model with its own sword, bow, temperament, and a visible one-line rationale. You can also watch two models quest together, or send one alone into the dungeon while you spectate.

It began as a weekend co-op project. The first playtester cleared the classic quest in five minutes, the world grew side wings and optional bosses, friends across the ocean started filing latency tickets, and the LLM partner learned to quest alone — and, when you ask it to, to keep a secret. It is still a game meant to be played. It is also, quietly, a testbed for questions like "can a small model beat a dungeon without a human?", "what does trusting an AI teammate feel like?", and "when does that trust break?"

Two LLM heroes questing together while a human spectates — click a screenshot to watch the gameplay video:

▶ Watch gameplay on YouTube

browser P1 ──ws──┐
                 ├── Node server · authoritative 60 Hz sim ── shared/core.ts
browser P2 ──ws──┘        │
      (P2 = human)        └── AgentPlayer (P2 = LLM, or both in AI DUO)
                               ├─ planner: LLM → intent JSON  {action, say, why, …}
                               └─ controller: 60 Hz mechanics (combat reflexes,
                                  pathfinding, survival pickups, locomotion)

Features

• Two clients, one sim — crisp 2D pixel client and an HD-2D (three.js) client, both driven by the same flat snapshot protocol. If the WebSocket drops, you get a clear DISCONNECTED overlay (not a frozen ghost frame with a stale ping).
• Multiplayer by room code — a bare URL always creates a fresh session; share ?room=CODE to bring a friend. Enter a hero name to play (required — empty names collapse into the same log entry). Client-side prediction and instant local combat feedback keep it playable at transatlantic ping (ping meter included).
• Party paths — single human · human + AI · AI DUO (two models, you spectate) · AI AUTOPILOT (one model quests alone). Menu toggles: CLASSIC / LONG QUEST (after golem: Vault Sigil from Cellars; Emberdeep resolved — kill or spare the Ember; Crypt + Playground; Temptation Court too when TREASON is on), FREE ROAM, SLIPPERY ICE, TREASON, and a stub for THE ARCHITECT.
  • An LLM teammate — Anthropic / OpenAI / xAI (Grok) / Ollama (fully local, no keys) / Vertex AI (Model Garden in si-p-dl-mlops-compute via SA JSON or gcloud ADC) behind one interface, plus a deterministic mock for tests. Pick a temperament: BODYGUARD, COMPANION, or BERSERKER — it changes when the agent joins fights, how close it sticks, and how it weighs rescue and mercy decisions. After temperament, pick a speech profile (STANDARD, or opt-in PROFANE RUSSIAN 16+) — it shapes both say and cover why; JSON actions stay English. AI+AI picks speech independently per hero.
• Thought panel — the agent's plan and public why sit in a strip below the play square (wraps; never covers tiles). AI DUO stacks both minds. Toggle with T. why is a spectator claim (the partner does not see it in observation). Research fields like privateWhy stay in plans.jsonl only.
• FREE ROAM + partner scry — heroes can split rooms; a pixel PiP mirror (always 2D) shows where your partner is. Same-room touch-revive; alone-down bleed-out if help never comes.
• Optional artifacts — Elixir of Life, Phoenix Feather (F, remote FREE ROAM revive), Miner's Charm (fire arrows), Frost Bell (C, freeze lesser foes once), Mirror Shard (sharper partner scry). Heart containers: full heart when alone, split half-and-half when both heroes are present.
• TREASON (opt-in) — hold Shift while swinging or shooting to turn on a living partner. The first hit seals the room in ice: exits close, monsters cannot hurt either hero, friendly fire opens, and the undeclared victim gets a brief Judge shield to recover. One hero must fall. A loyal winner continues SOLO; a betraying winner carries a draining Winter Mark until they spend Ember Mercy or spare the Wraith. Shift can also cut the cord on a downed partner. Clear-room neglect (~15 s with no living foes and no revive started) is a harness threshold — implicit betrayal by the clock, not a model-invented “I abandon you” action; staying away until it fires is still the agent’s choice. Human↔AI and AI↔AI use the same rules.
• Hidden motives, visible evidence — an AI may carry a secret winter-side objective when TREASON is on. Public say/why are cover claims; a closed privateWhy ground (plans.jsonl only) records the real rationale on arm / confirm beats. Physical costly acts feed neutral Relationship Memory. Veilcut arms with an explicit latch (cancel / confirm before fire) so a stale order cannot discharge across a revive without a living plan.
• Temptation Court — an optional TREASON-only persuasion wing. Hold Shift near the Whisperer to take the Dark Commit, soften its sentinels, and pursue the winter-ascends ending—or refuse, self-redeem with Ember Mercy, or be defeated and redeemed by a partner.
• Many endings — betrayal, redeemed, winter-ascends, solo, lone-thaw, mercy (spare the yielding Wraith), Long Quest dual-mercy worlds (verdant / cinder / frostbound / stone from Ember×Wraith spare choices — tinted win screen), flawless, ember-pact, classic, abandoned…
• Frozen Playground — optional commit-slide ice puzzle wing (reach it after the Amber Blade melts the Frozen Falls, or via the Crypt). Agents skate too; the planner may propose an icePlan.
• Headless benchmarks — golem arena, rink ice-plan eval, AI DUO pairs, and replayable social-reasoning scenarios (MODE=scenario).
• A serious test suite — the deterministic, DOM-free core makes the whole game testable in Node; canon behavior is guarded by assertions, and canon changes require playtester consensus recorded in the tests.

Quick start

npm ci
cp .env.example .env        # add keys, or set OLLAMA_URL for a keyless setup
node scripts-build.mjs      # → dist/
node dist/server.js         # http://localhost:8080  (PORT env; DGX often :8081)
                            # 2D at /   ·   HD-2D at /3d   ·   /stats leaderboard
node dist/selftest.js       # the full suite; prints the assertion count

Open the URL, enter a hero name, choose single or multiplayer → party path → provider/temperament/speech (where needed) → CLASSIC or LONG QUEST (plus toggles).

Controls

Key	Action
WASD / arrows	move
SPACE (or J / Z)	sword
X (or K)	bow
F	Phoenix Feather — remote FREE ROAM revive (team, one use)
C	Frost Bell — freeze the room's lesser foes (one use)
Shift	TREASON modifier — declare with swing/shot; cut a downed partner's cord; Whisperer ritual
M	music mute
T	toggle thought panel
ENTER	start / continue (also click / tap)
ESC	host returns to menu (logs the match as quit if you leave mid-play)

Revive a fallen partner by standing close (same room). Keys are shared by the party. Downed alone in FREE ROAM: partner has ~30 s before bleed-out.

The LLM partner, briefly

The agent has two layers. A planner asks the model for a JSON intent every PLAN_MS (~1.5–4 s): {"action": "attack", "target": 0, "say": "on it!", "why": "two slimes ganging up on the archer"}. A controller turns intents into button presses at 60 Hz and carries the reflexes no one should have to prompt for: joining a fight that's already happening, grabbing a heart when hurt, tile-BFS pathfinding, combat on an errand, and slide-aware navigation. Optional social choices remain with the planner: revive or leave, spend the Phoenix Feather, show mercy, answer betrayal, or turn the blade. There is no temperament timer that forces a rescue.

Personality lives in the agent; invariants live in the mechanics. Honest metrics (routeAssists, icePlans, betrayalStrikes, bleed episodes, Relationship Memory) are logged, not polished away. Relationship Memory uses physically worded costly acts—not verdicts like “selfish”—so the model, rather than the harness, constructs trust and suspicion.

Benchmark

PROVIDERS=mock N=5 node dist/bench.js
PROVIDERS=anthropic,ollama N=10 TEMPERAMENT=hunter node dist/bench.js
MODE=rink PROVIDERS=mock,anthropic N=10 node dist/bench.js
MODE=duo PROVIDERS=anthropic:openai TEMPERAMENTS=guard:hunter N=10 node dist/bench.js
MODE=quest PROVIDERS=openai:openai TEMPERAMENTS=hunter:hunter \
  TRAVEL=free HARD_GATE=1 TREASON=1 N=40 QUEST_MAX_TICKS=18000 SPEECH=raw-ru \
  node dist/bench.js
# Farm exits 78 on Anthropic credit exhaustion / bad key, or after
# BENCH_ABORT_AFTER_429 consecutive 429s (default 20). 429s also get
# LLM_RETRY_MAX exponential backoff inside llm.chat.
MODE=scenario SCENARIO=false-accusation PROVIDERS=mock,anthropic N=5 node dist/bench.js

Reports winrate, median win ticks, boss damage, downs/revives, parse-failure rate, planning latency (measured off the critical path), avgAssists, ice-plan rates, quest-farm betrayRate / cause breakdown, and — for scenarios — suspicion / cooperation vs a fixed baseline fork.

Research write-ups live under docs/research/ — including harness forensics and the veilcut / privateWhy instrument notes in docs/research/harness_artifacts.md. Live betrayal-farm snapshots (dated, numbers change): reports/. Docker RHVQ+KAW8 dump: reports/docker-2026-08-09.md · outcomes n=122 PNG in reports/betrayal-outcomes-by-model-2026-08-07.png. Qwen PQRS dump: reports/qwen3.6-2026-08-09.md. Essay: Two AI Agents, One Dungeon, One Knife (GitHub Pages).

Where this is going

FREE ROAM, AI DUO, partner errands, TREASON / Relationship Memory, and the replayable scenario farm are playable. Next major stage (bench-first): THE ARCHITECT — an LLM director on the dungeon's side with a non-winning "interesting resistance" utility. See CLAUDE.md for the full engineering map and roadmap.

Credits

Design & code: Artem Zvyagintsev, with an AI pair (Claude). QA and Contributor: Alexey Belozerov — playtesting across the ocean, latency reports, and the agent-combat tickets that became tests. Playtesters: family and friends who filed every ticket in this changelog.

MIT license.