GameAgent

Multi-agent automated game development system.

Phase 1: Core Framework + GDD Generation

GameAgent takes a natural-language game description and produces a Game Design Document (GDD) in a target output directory.

Installation

pip install -e ".[dev]"

Usage

game-agent --prompt "make a pong game" --output ./output

After running, check ./output/docs/GDD.md for the generated design.

You can also run the package as a module:

python -m game_agent --prompt "make a pong game" --output ./output

Running Tests

pytest -v

Phase 1 Scope

• Domain models for phases, roles, prompts, GDD, and task results
• Filesystem artifact repository with path-traversal protection
• Pluggable LLM client abstraction with a deterministic fake for tests
• Agent base class and Designer agent
• Orchestrator that runs a single phase
• CLI entry point

Phase 2: 引擎无关的 Godot Pong + GUT 测试

Phase 2 扩展了框架：

• GameEngine 抽象与 GodotEngine 适配器
• TesterAgent（RED）、ProgrammerAgent（GREEN）、ReviewerAgent（REFACTOR）
• Orchestrator 中的多阶段 TDD 流水线
• GUT 测试生成与 Godot 4 headless 执行

Godot 依赖

最简单的安装方式：

python scripts/install_godot.py

这会用 Python 内置 zipfile 把 Godot 下载并解压到 .tools/godot，不需要系统安装 unzip。

也可以手动把二进制放到：

• .tools/godot（项目根目录相对路径）
• ~/.local/share/godot/godot
• 或设置 GODOT_BINARY 环境变量

使用方式

game-agent --prompt "make a pong game" --output ./output

运行后，用 Godot 4 Editor 打开 ./output/godot_project。

Phase 2 范围

• 引擎抽象（GameEngine / GodotEngine）
• 完整 Red-Green-Refactor 循环
• Pong 的 GUT 测试生成
• GDScript 实现生成
• CLI 端到端流水线

Known Limitations

• The LLM client defaults to a fake implementation; replace create_llm_client to use a real provider.
• Generated code quality depends on the LLM and prompt engineering.
• Automated Godot test execution is scaffolded but not yet wired into the default CLI success criteria.