User repository on GitHub
River shooter game in Python, inspired by Atari's River Raid, played by a TypeSafe AI pilot
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
River Run TypeSafe
A pygame-ce river shooter whose AI pilot splits the river into 20 lanes and asks Jev one parallel System One request per decision for lane, fire and speed.
- Stars
- 0
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- ashaazami
- Last verified
- Oct 2026
About this project
A vertical-scrolling river shooter in Python with pygame-ce, plus an AI pilot that plays it using TypeSafe's System One model, Jev. Inspired by the 1982 Atari 2600 game River Raid by Carol Shaw (Activision). This is an independent project, not affiliated with or endorsed by Activision; all code, graphics and sounds are original. The TypeSafe pilot playing in real time (with sound): Fly up the river. Touching land, a ship, a helicopter, a jet or a standing bridge costs a life.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Code measures each reachable lane for time to line up, open water ahead and what drifts into it; the model only picks.
- The bottom strip shows lanes live: outlined is the plane's lane, filled the chosen one, the tick the exact target.
- Real time runs about 8 requests per second at roughly 115 ms each, and --lockstep makes the game wait instead.
- river_run/game.py holds the rules with no pygame import, exposing step and observation for automated players.
Watch out
Every decision is a hosted System One request, so the pilot needs an API key, network access to api.typesafe.ai and Python 3.10 or newer for typesafe-sdk 0.6; real-time play runs about eight requests per second, which is worth knowing before you leave it going. The README is careful that its scores and latencies come from the author's own runs and are not an official benchmark. Steering is deliberately tunable through top or blend, sum or max block scoring, a blend probability floor and a dead-end pixel count, so a good run is partly a function of those defaults rather than the model alone. Rendering and sound need pygame-ce, and --record additionally needs ffmpeg installed. It is an independent project inspired by River Raid, not affiliated with or endorsed by Activision, with original code, graphics and sounds. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Both, deliberately. river_run/game.py carries the rules with no pygame dependency and exposes a step action plus a plain-data observation, which is what lets automated players and tests drive it; rendering and keyboard play live elsewhere. It ships real rules with fuel, bridges, scoring and extra lives, and --human lets you fly it yourself with no API key, which is also how you record your own runs.
Real time issues requests back to back at roughly 8 per second and 115 ms each, so a slow call costs you frames rather than failing outright. --lockstep makes the game wait for each decision instead, which is the honest mode for watching what the model actually chose. --headless with a frame count runs no window and prints stats, so you can get numbers without a display.
Code turns them into controls. By default blend groups likely neighbouring lanes into blocks and flies to the probability-weighted centre of the chosen block, with sum or max picking the block, a floor probability that splits blocks, and a dead-end pixel count so lanes in front of an island do not merge with the channels beside it. --steering top gives you the single most likely lane instead.
For the pilot, yes: every decision is one System One request to Jev, the SDK defaults to jev-latest at api.typesafe.ai, and a missing or wrong key fails loudly with a specific SDK error rather than silently degrading. For the game itself, no --human plays without a key, and the tests run against the pygame-free rules module under pytest.
You fly up the river, and touching land, a ship, a helicopter, a jet or a standing bridge costs a life. Fuel drains constantly and you refill by flying over a depot, or shoot it for 80 points. Every section ends with a bridge, and destroying it makes that point your respawn for the next section. Ships score 30, helicopters 60, depots 80, jets 100, bridges 500, with an extra life every 10,000 points.
White corner brackets mark the plane, and the strip along the bottom of the river shows the lanes: outlined is the plane's lane, filled is the chosen one, and the tick below is the exact steering target. Text at the top reports the chosen lane with its confidence, the fire probability, the chosen speed and the time the last request took.
Yes, through the SDK's environment variables rather than anything in the game. TYPESAFE_DEFAULT_MODEL defaults to jev-latest, TYPESAFE_BASE_URL defaults to https://api.typesafe.ai, TYPESAFE_LOG_LEVEL turns on SDK logging, and TYPESAFE_API_KEY is the only required one. The README is also explicit that its scores and latencies come from a few of the author's own runs, not an official benchmark.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPython42GitHub stars
Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
JoshuaSP/open-jevPython
JoshuaSP/open-jevPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.