shanselman/gputaskman ? reverse-engineered prompt

Build me a terminal based GPU task manager that feels like htop, but for GPU usage instead of CPU. I want it to work on Windows, Linux, and WSL, and show each process using GPU memory, current GPU activity, and helpful hints about what the process is doing, especially for LLM workloads like model weights, KV cache, and compute buffers.

It should be able to run interactively in the terminal, and also support a one shot view plus JSON output for scripting. The main goal is to help me answer what is using my VRAM, whether it is a model, browser, desktop app, or something else, and to make the view readable and live updating.

If you need to, look up current docs online for the GPU and process APIs that fit each platform. Please keep it practical and focused on local workloads, and make the output clear when something can only be estimated instead of measured exactly.