GameForge is an open game-development harness. The solver gets a complete disposable project and freedom to use code, shell, image, and official engine tools. A separate evaluator builds, launches, replays, records, and verifies the result.
“Do not prescribe how the model builds the game. Do not let the model decide whether the game works.”
The solver owns exploration and implementation. The harness owns the reproducible project boundary, engine lifecycle, execution profile, receipts, and independent evidence.
The model can inspect, improvise, call the engine, capture feedback, and change strategy. Final acceptance remains outside that loop.
A game brief or change request enters a disposable full workspace.
The model chooses its own files, tools, engine commands, and validation strategy.
The output is a complete Unity or Godot project, not a chat answer.
A solver-invisible evaluator builds, launches, injects input, and captures evidence.
Receipts separate solver, engine, host, runtime, evaluator, and gameplay outcomes.
Frozen suites and paired baselines show how the same open harness performs on repeatable game-development tasks.
Empty or minimal projects. Natural-language game briefs. No required node names, source layout, API, tool order, intermediate gate, or solver workflow.
Twenty genres from empty projects, with blinded quality parity against local Official.
View evidenceTwenty genres from minimal projects, with 40 of 40 valid evaluator screenshots.
View evidenceAnimated previews link to full MP4 replays, with input traces and machine-readable receipts in the repository.
Godot · engine-input replay
Tower Defense
Placement · waves · combat · speed control
Godot · engine-input replay
Rhythm Game
Four lanes · score · accuracy
Godot · engine-input replay
Fishing Challenge
Cast · hook · reel · tension
Unity · OS-native input replay
Deckbuilding Duel
Cards · shield · damage · turn progression
Godot · engine-input replay
Brick Breaker
Paddle · collisions · score · lives
Godot · engine-input replay
Arena Survivor
Movement · aim · fire · dash
Solvers, engines, and authority profiles are independent choices behind one game workspace and evidence contract.
Use a saved Codex subscription login or a standard OpenAI Responses API tool loop.
Unity and Godot share workspace, budget, receipt, backend, and result contracts.
Choose maximum native freedom, supervised engine transport, or external isolation.
Bring Python 3.11 or 3.12, uv, Git, and the engine you want to use. Authentication can come from a local Codex login or an API key kept in the host environment.
$ git clone https://github.com/AlbusChen/GameForge-Harness.git
$ cd GameForge-Harness
$ uv sync --locked --all-groups
$ cp config/model.subscription.example.yaml \
config/model.local.yaml
$ uv run gameforge run-game-workspace \
--project /path/to/game \
--engine auto \
--request "Add a playable dash mechanic." \
--model-profile config/model.local.yaml