Anionex/codex-vision-proxy ? reverse-engineered prompt

Build me a small Python tool that sits in front of Codex and makes a text only model handle images without breaking. I want it to catch image requests, send the image to a vision model I configure with env vars, then return a clear text description so Codex can keep working like it has multimodal support.

It should work for pasted images and for images passed through the built in view_image flow, and it should cache repeated image descriptions so the same image does not get reprocessed every time. If there are multiple images in one request, handle them efficiently. Please also add an optional lightweight vision toolkit with simple command line helpers for asking questions about an image, doing OCR, and locating an object or UI element in a screenshot.

Make it easy to install into Codex as a skill, and include a clear setup and verification guide. If you need current docs for Codex skills or the API shape, look them up online first.