Unify AI connections and fix GPT image generation

[jamesgraves52]

I get the "No Image models Found" Error when trying to enable model generation. I know that GPT-5.6-sol has image generation capablility can we Simple Chat configured to accept GPT-5.6-Sol as an image generation model???

[paullizer]

Updated implementation scope

The agreed direction is one shared AI Connections framework, reusing the existing model-endpoint registry rather than creating a second image-only endpoint manager.

• Configure an endpoint and its credentials once; chat and image-generation defaults reference the connection/model by stable IDs rather than copying configuration.
• Wire chat and image generation now. Keep the existing Image action and one image-model selection; do not require a second image-backend picker.
• Add capability-aware discovery, selection, and runtime validation. Preserve existing chat/agent/workflow/vision behavior when image-only models enter the shared registry.
• Automatically import existing direct/APIM image settings, preserving the active model, credentials, routing, and existing connections. Do not force the irreversible chat multi-endpoint switch or publish imported image connections as new chat models.
• Correct the GPT image-generation API path, unify generation callers, and address the coupled image-persistence and misleading-error problems. Actual Azure image-tool availability/default routing still requires confirmation against the configured resource.
• Establish extension points for future embeddings, transcription, speech, computer use, and other capabilities; their runtime integrations and new controls are explicitly deferred.

Priority remains P2, assigned to Paul Lizer (@paullizer). Size is L for the expanded framework, compatibility, migration, and UI scope. The implementation plan is being revised; application code has not been changed yet.

[paullizer]

Implementation ready for review

The shared AI Connections framework is implemented in the working branch paullizer-gpt-image-generation for 0.261.102. This is not a deployment or a claim of live Azure availability.

• One connection/credential registry now supports independent chat and image model defaults, capability-aware discovery/selection, and reusable operation adapters.
• Existing image settings import automatically with active-model/auth/API/APIM preservation, scoped staged credentials, conditional settings writes, and recovery without erasing retained settings.
• GPT image requests use Azure v1 Responses; dedicated image requests retain the appropriate Images API. Image mode does not require a usable text model. Shared image edits preserve configuration/refusal/rate-limit statuses.
• Classic and V2 use one grouped image-model picker. Image-only models are excluded from text consumers, unverified model variants cannot inherit image-tool support, and duplicated connections retain image operation profiles.
• Framework documentation, administrator guidance, generated documentation inventory, release notes, and the application version are updated. Embeddings, transcription, speech, computer use, and other runtime integrations remain deferred.

Validation

• All changed functional tests plus route policies: 324 passed, with 457 subtests passed.
• Final Classic/V2 browser workflows: 28 passed.
• Classic/V2 model-choice scripts, TypeScript typecheck, and production V2 build passed.
• Documentation inventory/coverage: 7/7; site quality: 6/6.
• Real OpenAI 1.109.1 request construction was exercised with mocked HTTP, including complete registered /api/chat image requests. Reported code-review findings were corrected and regression-covered.

Deployment boundary

No live Azure inference or resource provisioning was performed. Actual image-tool entitlement, regional availability, and any service-required backing deployment remain properties of the configured Azure resource and must be confirmed there. The application does not invent deployment bindings or silently switch models.