- Plumb input modalities from model catalog through the openai model protocol. Default to text and image. - Conditionally add the view_image tool only if input modalities support image.