Support image type in Output Signature

[yamitzky]

Hi team!

I would like to propose adding support for the image type in the output signature
(e.g., input: string -> reason: string, output: image).

Currently, the image type is restricted to input fields only. Attempting to use it in
an output field results in a validation error: "Image type is not supported in output
fields".

Use Case

With multimodal models capable of generating images (such as Google Gemini's image
generation models), it would be very useful to define a signature that explicitly
expects an image as output.

Example signature:

const gen = ax('prompt: string -> generatedImage: image');

Proposed Changes

To make this work, I believe the following areas would need updates:

1. Relax Validation: Update src/ax/dsp/parser.ts and src/ax/dsp/sig.ts to allow image
   types in output fields (and nested output objects).
2. Update Extract Logic: Update src/ax/dsp/extract.ts to handle image types by
   attempting to parse them as JSON (similar to how json type is handled).
3. Provider Support: Update providers (like google-gemini) to handle image response
   data (e.g., inlineData) by converting it into a structured format (like a JSON
   block) that the extraction logic can process.

What do you think about officially supporting this?

Thanks!

[dosco]

This sounds cool go for it. Some notes on our current extraction / validation if it helps. https://github.com/ax-llm/ax/blob/main/notes/Extraction_Modes_Validation.md