hermes - 💡(How to fix) Fix Expose image editing as first-class Hermes tool

Problem

Hermes currently exposes image generation as a text-to-image workflow, but models/providers such as OpenAI Codex + GPT Image 2 can also edit an existing input image. Agents have no first-class tool for passing an input image or optional mask through the image generation provider layer.

Proposal

Add an image_edit tool alongside image_generate that accepts:

• prompt: edit/enhancement instructions
• image_url: HTTP(S), data URL, or local absolute path input image
• aspect_ratio: same normalized aspect ratio enum as generation
• mask_url: optional backend-specific mask

OpenAI Codex should map this to a Responses image_generation call with action=edit and an input_image content part. Local images should be converted to data URLs before dispatch.

Use cases

• Background-only enhancement for App Store screenshots while keeping UI/text deterministic
• Inpainting / masked replacement
• Reference-image edits without requiring an OpenAI API key when Codex auth is configured

Acceptance criteria

• image_edit is registered in the image_gen toolset
• OpenAI Codex provider accepts image_url/mask_url and dispatches edit requests
• Existing image_generate behavior remains unchanged
• Tests cover request shape, local file conversion, mask forwarding, and schema integrity