github_install_assistant.py is a platform-aware installer assistant for GitHub repositories.

What it does:

• Probes the local machine first: OS, Python, Git, GPU, VRAM, NVIDIA driver, CUDA runtime, CUDA toolkit, and processor count.
• Sends the installation goal, repository URL, and detected system profile to the OpenAI Responses API.
• Uses a planner model to research the install path, then runs installation commands in a second visible terminal window.
• Mirrors every command and its output back into the original terminal with ANSI colors, including the exact prompt text sent to the LLM.
• Re-queries the LLM after each command result to decide the next command.
• Requires a dedicated virtual environment for dependency installation and pushes the planner toward explicit venv-aware commands.
• If the repository appears to need PyTorch, CUDA inference, RealtimeTTS, faster-whisper, faster-qwen, or similar GPU stacks, it probes nvidia-smi, nvcc --version, and CUDA_PATH, then asks the LLM to research a compatible torch wheel instead of blindly using the newest CUDA build.
• Prints a final structured installation report with successful commands, virtual environment details, test/example discovery, and useful follow-up notes.

Requirements:

• Python
• OPENAI_API_KEY in the environment
• Network access for the OpenAI API and the repository being installed
• A terminal launcher supported by your platform Windows: cmd.exe macOS: Terminal.app via open Linux: one of x-terminal-emulator, gnome-terminal, konsole, or xterm

Usage:

$env:OPENAI_API_KEY="sk-..."
python .\github_install_assistant.py https://github.com/owner/repo
python .\github_install_assistant.py https://github.com/owner/repo --goal "Install RealtimeTTS with faster-qwen3-tts support and CUDA-enabled PyTorch. Create a new venv. Install torch first via the correct CUDA index URL before other dependencies."
python .\github_install_assistant.py https://github.com/owner/repo --target-dir .\installed_repositories\repo --max-steps 25

Notes:

• The default planner model is gpt-5.4-mini and for loop model it's gpt-5.4-nano.
• The worker window is intentionally limited to install commands and command output.
• The assistant only treats the installation as finished after a verification command succeeds and the LLM returns finish.
• There is also a hard stop via --max-steps, plus a repeated-command guard to avoid infinite loops.
• When torch is required, the planner must return a researched torch_install_command, and the orchestrator will force that command to happen before later dependency-install commands.
• After the last step, the original window shows a detailed documentation block describing what succeeded and how to continue using the installed repository.