HEIST//ONE

An observable stealth game where Jev supplies guards' split-second judgments
while deterministic code remains in control of the world.

Watch the 37-second film · Inspect the live-run evidence · Read the technical specification

Why a heist game?

Stealth is a legible test bed for decisions under ambiguity. A guard rarely has perfect information: a door opens, a light goes out, footsteps come from an empty room, or someone with the right jacket carries the wrong badge.

HEIST//ONE makes those judgments visible. Select any guard to inspect the evidence it received, its probability distributions, confidence, proposed intent, applied intent, latency, and fallback state. You are not fighting hidden scripts; you are manipulating what the guards have reason to believe.

The design follows TypeSafe's recommended shape for Jev: ask narrow, typed questions, then let ordinary software validate and act on the answers. Jev is used as a tactical judgment engine—not as the game engine.

The boundary that matters

Jev proposes	Deterministic code owns
Whether local evidence suggests a threat (Noul)	Authoritative 30 Hz simulation
How suspicious a guard should be (Score)	Visibility, collision, navigation, and physics
Which legal tactical intent fits (Choice)	Legal actions, access control, and side effects
Which observed entity deserves attention (Choice)	Inventory, alarms, win/loss, and lockdown
Probabilities and confidence for each judgment	Deadline handling, stale-response rejection, and fallback

Every provider response is a proposal. The server rejects illegal targets, stale context revisions, and policy-incompatible actions before they can affect the simulation.

flowchart LR
    P["Player inputs"] --> S["Authoritative simulation"]
    S --> O["Per-guard local observations"]
    O --> B["One batched typed request"]
    B --> J["Jev or scripted adapter"]
    J --> V["Policy + freshness validation"]
    V --> S
    S --> D["Decision Lens + JSONL trace"]

Loading

What is in the MVP

• One deterministic museum with six guards, three access-controlled doors, disguises, credentials, light control, noise decoys, one artifact, and two exits.
• An authoritative server simulation; the browser renders snapshots and sends inputs over a WebSocket.
• Event-driven decision scheduling with a bounded refresh, one request covering all four typed judgments for every guard.
• A live adapter for the official @typesafe-ai/sdk and a deterministic scripted adapter for credentials-free play and tests.
• A Decision Lens that exposes the complete applied judgment—not merely the winning label.
• Append-only JSONL evidence traces for requests, responses, material world events, outcomes, latency, usage, fallbacks, and deterministic checkpoints.
• A full browser-controlled clean-extraction test and a reproducible Remotion launch-film project.

Verified live run

On 17 September 2026, the complete jacket → badge → vault → artifact → exit route ran against the TypeSafe sandbox through JevDecisionEngine.

Observed measure	Result
Outcome	Clean extraction (won)
Batched requests	12
Typed judgments	288
Successful responses	12 / 12
Decision errors	0
Fallbacks	0
Stale decisions	0
Median adapter latency	259.9 ms
p95 adapter latency	344.1 ms
Provider usage	63,349 input / 12,181 output tokens
Exact provider cost	Unknown—not exposed by the response

This is evidence of one end-to-end run, not a benchmark or a claim that Jev is better than scripted game AI. The exact session identifier, checkpoint hash, measurement scope, and deliverable hashes are recorded in docs/LIVE-JEV-VERIFICATION.md.

Quick start

You need Node.js 22.13 or newer and pnpm 11.

git clone https://github.com/AbdelStark/heist-one.git
cd heist-one
corepack enable
pnpm install --frozen-lockfile
pnpm dev

Open http://127.0.0.1:4173, read the briefing, then enter the museum. Scripted mode is the default, so no account or API key is needed.

Controls

Input	Action
WASD or arrow keys	Move
Shift	Sprint
C	Crouch
E	Interact with doors, pickups, the breaker, and exits
Q	Throw a noise decoy
Click a guard	Pin that guard in the Decision Lens

There is a clean route, but the game does not prescribe it. Find useful credentials, manipulate light and noise, steal the artifact, and reach an exit before the guards can confirm the theft.

Run with Jev

Early-access credentials are available through the TypeSafe Console. Keep the key on the server:

cp .env.example .env

Then edit .env:

TYPESAFE_API_KEY=your_key_here
HEIST_DECISION_MODE=jev

Start the app with pnpm dev. The simulation—and therefore provider usage—does not begin until you leave the briefing. Live calls are event-driven, bounded by HEIST_LIVE_CALL_LIMIT, and protected by a deadline. A missing key or failed decision falls back to the scripted policy without exposing the credential to the browser, logs, traces, fixtures, or media.

To capture another full live run after the app is running:

pnpm capture:live

The capture script drives a complete route and fails unless the result is CLEAN EXTRACTION.

Architecture

apps/
├── server/       Fastify authority, Jev adapter, scheduler, trace writer
├── web/          React + Phaser client and live Decision Lens
└── video/        Remotion launch film, thumbnail, and source recording
packages/
└── game/         Pure simulation, world projection, policy, scripted adapter
scripts/
└── capture-live-gameplay.ts
tests/
└── e2e/          Browser interaction and full-heist acceptance tests

The simulation imports the small DecisionEngine interface, never the TypeSafe SDK. That seam keeps provider access server-side and makes offline tests and recorded replays independent of a live credential.

For the full decision schema, stale-result rules, trace envelope, and acceptance criteria, read SPEC.md. Product scope and non-goals live in PRD.md.

Development

Command	Purpose
pnpm dev	Run web and server in watch mode
pnpm check	Lint, type-check, unit-test, and production-build everything
pnpm test:e2e	Run the Playwright browser suite in scripted mode
pnpm capture:live	Record an opt-in Jev gameplay run
pnpm video:studio	Open the Remotion composition
pnpm video:render	Render the H.264 launch film

Generated traces and local credentials are ignored by Git. Do not commit either. See CONTRIBUTING.md before changing the decision boundary or adding a live-provider test.

Project status

v0.1.0 is a complete, playable vertical slice. The next useful work is depth, not breadth:

• a standalone trace/replay inspector;
• scenario packs that probe distinct uncertainty patterns;
• predeclared, repeatable comparisons between decision policies;
• accessibility and input remapping;
• a deployable hosted demo with isolated, rate-limited provider access.

Issues and focused pull requests are welcome. Please use SECURITY.md for private vulnerability reports and read the CODE_OF_CONDUCT.md before participating.

Credits and license

HEIST//ONE was designed and built by Abdel Stark. The game code and original media are available under the MIT License; third-party dependencies and fonts retain their own licenses. Asset provenance is documented in docs/ASSETS.md.

HEIST//ONE is an independent community experiment. It is not affiliated with or endorsed by TypeSafe AI. Jev, TypeSafe, and System One are referenced only to describe the integration built with TypeSafe's public SDK and service.