xRyul/pi-nvidia-nim — reverse-engineered prompt

Build me a small extension for the pi coding agent that adds NVIDIA NIM as a provider, so I can use the models from build.nvidia.com inside pi without any extra manual setup beyond adding my API key. It should work with either NVIDIA_NIM_API_KEY or NVIDIA_API_KEY, show the models in the normal model picker, and also let me choose them from the CLI.

I want it to support a solid curated set of chat models, then also discover more available NIM models automatically at startup. Please make reasoning or thinking mode just work when I change the thinking level in pi, including models that need custom request fields to turn thinking on or off. Use the safer message role that avoids NIM server errors, support streaming responses and OpenAI style tool calling, and ignore non chat models like embeddings or safety models. If a discovered model is unknown, give it conservative defaults. Add clear setup notes and basic tests, and look up current NVIDIA docs online if you need to.