Cool! The DX is tricky to nail, when combined with LLM's tendency to hallucinate. I asked it to extract some text from an image, which it dutifully tried to do

Cool! The DX is tricky to nail, when combined with LLM's tendency to hallucinate. I asked it to extract some text from an image, which it dutifully tried to do. However the generated python kept throwing errors. There's no image -> text tool yet, so it was trying to use the image segmenter to generate a mask and somehow extract text from that. It would be super helpful to: 1) Have a complete list of available tools (and / or a copy of the entire prompt given to the LLM responsible for generating python). I used prompt injection to get a partial list of tools and checked the Github agent PR for the rest, but couldn't find `<<all_tools>>` since it gets generated at runtime (I think?). 2) Tell the LLM it's okay to fail. E.g.: "Extract the text from image `image`. If you are unable to do this using the tools provided, say so." This prompt let me know there's no tool for text extraction. Update: per https://huggingface.co/docs/transformers/custom_tools https://huggingface.co/docs/transformers/custom_tools you can output a full list of tools with `print(agent.toolbox)`