Support multi-image edits for gpt-image-2

[adambalogh]

Summary

Enables image-to-image editing and multi-image compositing (e.g. "add this logo to this photo") for OpenAI gpt-image-2, which — unlike ByteDance — does editing on a separate /images/edits endpoint that takes reference images as multipart file uploads rather than a JSON image field.

What changed

• model_registry.py — Added ModelConfig.image_edit_endpoint. Reference delivery is now a single per-model choice:
  • unset → references ride inline in the JSON image field of /images/generations (ByteDance Seedream/Seedance, unchanged);
  • set (a path) → references are uploaded as repeated multipart image[] files to that endpoint.
  • gpt-image-2 now sets image_supports_reference=True + image_edit_endpoint="/images/edits".
• image_generation.py — generate_images picks the delivery path from config. Text-to-image (no references) still uses /images/generations exactly as before. Only inline data: references are uploaded; a plain-URL reference is skipped rather than dereferenced inside the enclave (SSRF), and if nothing is uploadable the request falls back to a plain text-to-image generation instead of failing. Up to 10 references per request.
• Tests — added coverage for the multipart upload shape, the no-reference generations path, and the non-inline fallback. All 30 image tests pass; make lint clean.

Notes

• Multi-image support is now: gpt-image-2 ✅, ByteDance Seedream/Seedance ✅. xAI grok-2-image and Z.ai glm-image do not support reference images and are unaffected.
• Pairs with the chat-app change that forwards all attached images and tells the Studio agent which models can composite.

🤖 Generated with Claude Code

https://claude.ai/code/session_01224b8BAcBisBUCX2EzrbG8

---

Generated by Claude Code

[Copilot]

Pull request overview

This PR adds provider-specific support for image-to-image editing and multi-image compositing for OpenAI gpt-image-2 by routing reference-image requests to a dedicated multipart /images/edits endpoint, while keeping existing /images/generations JSON behavior unchanged for text-to-image and for providers that carry references inline.

Changes:

• Extended ModelConfig with image_edit_endpoint and enabled reference-image support for gpt-image-2 via /images/edits.
• Updated generate_images to choose between JSON inline references vs multipart image[] uploads based on model config, with safe fallback when references aren’t uploadable.
• Added tests covering multipart edit uploads, no-reference generations routing, and non-inline reference fallback behavior.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
tee_gateway/test/test_image_generation.py	Adds test coverage for multipart /images/edits uploads and fallback/generations routing for gpt-image-2.
tee_gateway/model_registry.py	Introduces image_edit_endpoint in ModelConfig and configures gpt-image-2 to support reference edits via /images/edits.
tee_gateway/image_generation.py	Implements multipart upload path for edit-capable models and preserves JSON generations behavior otherwise.
CLAUDE.md	Updates documentation to reflect the new per-model reference delivery paths (inline JSON vs multipart edits endpoint).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.