fix(dotAI): Support new OpenAI image models (gpt-image-1)

Description

Following the dotAI refactor that introduced support for multiple AI vendors/models (#34058), the image generation feature was broken when using gpt-image-1 (and related models: gpt-image-1-mini, gpt-image-1.5) as the configured image model. Two distinct bugs were identified and fixed.

---

Bug 1 — Backend: b64_json response not handled

Root cause:
OpenAIImageAPIImpl.createTempFile() expected a url field in the OpenAI image response, but the gpt-image-1 family always returns b64_json (base64-encoded PNG). These newer models do not support the url response format or the response_format parameter at all, unlike DALL-E 2/3.

Model	Response field
DALL-E 2 / DALL-E 3	data[0].url (default)
gpt-image-1 / gpt-image-1.5 / gpt-image-1-mini	data[0].b64_json (always)

Error thrown: "Image Response does not include URL" even though the request succeeded.

Fix:
Updated OpenAIImageAPIImpl.createTempFile() to detect the response format and branch accordingly:

• b64_json present → decode with Base64.getDecoder() → create temp file from ByteArrayInputStream via TempFileAPI.createTempFile()
• url present → existing behavior via TempFileAPI.createTempFileFromUrl()

Added AiKeys.B64_JSON = "b64_json" constant.

---

Bug 2 — Frontend: Wrong pixel dimensions sent for non-dall-e-3 models

Root cause:
The AI image prompt form used semantic orientation values (landscape, square, portrait) and a client-side mapper (DotAIImageSizeMapperService) to translate these to model-specific pixel dimensions. However:

1. The form component defaulted to dall-e-3 before loading the actual model from the config endpoint.
2. If the user selected an orientation before the config API response arrived, the mapper computed dall-e-3 pixel dimensions (e.g., 1792x1024 for landscape).
3. When the config then resolved the correct model (e.g., gpt-image-1 → landscape = 1536x1024), the form's size control was updated silently ({ emitEvent: false }) — so the NgRx store was never notified.
4. The store still held the stale dall-e-3 size and sent it to /api/v1/ai/image/generate.

Additionally, the DotAICompletionsConfig TypeScript interface used an incorrect field name imageModel (which does not exist in the backend response). The backend returns:

• imageModelNames — comma-separated list of configured image model names
• availableModels — array of { name, type, current } objects (mirrors com.dotcms.ai.model.SimpleModel)

Fixes:

• DotAICompletionsConfig: replaced imageModel: string with imageModelNames: string; typed availableModels as DotAISimpleModel[] (new interface matching SimpleModel.java)
• Model resolution now mirrors the text model approach in CompletionsResource: find availableModels.find(m => m.type === 'IMAGE' && m.current), falling back to the first entry in imageModelNames
• Removed { emitEvent: false } from sizeControl.setValue() inside loadImageModelConfig() so the NgRx store receives the correct model-specific size after async config load

---

Files Changed

Backend:

• dotCMS/src/main/java/com/dotcms/ai/AiKeys.java — added B64_JSON constant
• dotCMS/src/main/java/com/dotcms/ai/api/OpenAIImageAPIImpl.java — dual-path createTempFile()

Frontend:

• core-web/libs/dotcms-models/src/lib/dot-ai.model.ts — DotAISimpleModel interface; fixed DotAICompletionsConfig
• core-web/libs/ui/src/lib/components/dot-ai-image-prompt/components/ai-image-prompt-form/ai-image-prompt-form.component.ts — correct model resolution; emits size update to store
• core-web/libs/ui/src/lib/components/dot-ai-image-prompt/components/ai-image-prompt-form/ai-image-prompt-form.component.spec.ts — regression test for the race condition

---

Acceptance Criteria

• Image generation works end-to-end with gpt-image-1 (landscape, square, portrait orientations)
• Image generation continues to work with DALL-E 3 (no regression)
• The correct model-specific pixel dimensions are sent to /api/v1/ai/image/generate regardless of when the user selects orientation relative to config load
• No "Image Response does not include URL" error when using gpt-image-1

Related

• Epic: #34058
• Regression Test Task: #34075

Start with OpenAIImageAPIImpl.java and AiKeys.java to trace image response handling, then read dot-ai.model.ts and the image prompt form component for model resolution and size updates. Run ai-image-prompt-form.component.spec.ts and verify the listed acceptance cases, including base64 responses, DALL-E 3 compatibility, and model-specific dimensions.