feat(vercel-provider): add OrqAiImageModel with /images/generations +…

[hamza-adami]

… /images/edits support

Replace direct use of OpenAICompatibleImageModel for image generation with a dedicated OrqAiImageModel that targets the Orq AI proxy correctly across all upstream providers (OpenAI, fal, Google Imagen, Leonardo, etc).

Problems fixed

1. OpenAICompatibleImageModel hardcodes response_format: "b64_json" in the request body. Newer OpenAI image models (gpt-image-1 and later) reject this parameter with Unknown parameter: 'response_format', and the Orq proxy forwards that error verbatim. Image generation through Orq was broken for any gpt-image-1+ model.

2. The upstream model's response schema only accepts data[].b64_json. Several Orq-routed upstreams (fal, Leonardo, dall-e-2 default) return the image under data[].url instead — either as a data URI or as an http URL. These responses failed schema validation with AI_TypeValidationError: data[0].b64_json: Invalid input: expected string, received undefined.

3. The upstream model did not support image editing via Orq. Calling generateImage({ prompt: { images: [...] } }) had no effect and Orq's /images/edits endpoint was unreachable through this provider.

Changes

• New OrqAiImageModel implementing ImageModelV3 (specificationVersion "v3", maxImagesPerCall = 10), wired up via provider.imageModel(id) in createOrqAiProvider.

• Request: drops response_format so each underlying upstream uses its own default — gpt-image-1 returns base64 inline, dall-e-* defaults to a URL, fal/Leonardo return data URIs. User-supplied extras (quality, style, background, etc.) pass through via providerOptions.orq.

• Response schema accepts every shape documented in orq-node's CreateImageData ({ b64_json?, url?, revised_prompt? }) and an optional usage envelope (input_tokens, output_tokens, total_tokens).

• normalizeImageEntry handles three response shapes uniformly:

  • b64_json field -> returned as-is
  • url data URI -> strips the prefix
  • url http(s) URL -> fetches and base64-encodes Always returns base64 strings, matching the ImageModelV3 contract.

• New /images/edits branch (multipart form-data) activates when files != null && files.length > 0. Each file is converted to a File with a .png / .jpg / .webp filename so Orq's middleware doesn't reject the part as application/octet-stream. The mask parameter is dropped with a warning because Orq's edit schema does not carry a mask field.

• Inline copy of defaultOpenAICompatibleErrorStructure because @ai-sdk/openai-compatible exports the type but not the value. Orq's proxy returns OpenAI-shaped error envelopes, so the schema matches 1:1.

• createOrqAiProvider now throws if apiKey is empty, and the URL builder no longer wraps the path in a redundant new URL(...).

Spec compliance verified against:

• @ai-sdk/provider's ImageModelV3 type (return shape: images, warnings, response { timestamp, modelId, headers }, optional usage).
• Upstream OpenAICompatibleImageModel in vercel/ai for body / schema / warning conventions.
• orq-ai/orq-node's CreateImageRequestBody, CreateImageEditRequestBody, CreateImageData, and CreateImageUsage for body + response shapes.