Summary

Adds an OpenAI-compatible /v1/images/generations route backed by a swappable ImageGenerationGateway interface.

The default gateway uses the existing Codex /responses transport with the image_generation tool, then converts streamed image events into OpenAI-style image generation JSON with b64_json.

This also adds optional reference image support through a project-specific images field.

Changes

• Add ImageGenerationGateway interface for deployment-specific image generation backends
• Add CodexResponsesImageGenerationGateway default implementation
• Add /v1/images/generations server route
• Convert Codex image generation SSE image events into OpenAI-style JSON responses
• Add optional images?: string[] request field for reference images
• Map reference images into upstream input_image blocks
• Preserve image data URLs and wrap raw base64 as data:image/png;base64,...
• Cap generated image count and reference image count at 10
• Export image generation gateway types
• Add deployment examples for Docker, systemd, and reverse proxy usage
• Document text-to-image and reference-image request/response usage
• Add tests for custom gateway routing, SSE conversion, reference image mapping, and invalid image lists

Verification

• bun run --filter openai-oauth test -- server.test.ts
• bun run typecheck
• Targeted Biome check for changed files

Notes

The images field is a project-specific extension to /v1/images/generations, intended for reference images such as character consistency, style reference, and image-guided editing.