feat(api): route Google AI Studio Imagen through /v1/images/generations

[danscMax]

Problem

gemini/imagen-4.0-generate-001 (and the -ultra / -fast variants) appear in GET /v1/models — they're surfaced live via Google's ListModels — but they are unroutable:

• On POST /v1/images/generations → 400 "Invalid image model: gemini/... Use format: provider/model", because gemini was not registered in the image registry at all (the message is misleading — the format was already correct).
• On POST /v1/chat/completions → 404 "is not found ... or is not supported for generateContent", because Imagen uses the dedicated :predict endpoint, not generateContent.

So the models are advertised but produce an error on every route.

Fix

Wire a gemini image provider (new google-imagen format) that targets the Imagen :predict API:

• Endpoint: POST {baseUrl}/{model}:predict where baseUrl = https://generativelanguage.googleapis.com/v1beta/models
• Auth: x-goog-api-key header (key stays out of the URL/logs), matching the gemini chat provider
• Request: { instances: [{ prompt }], parameters: { sampleCount, aspectRatio } } (sampleCount clamped to 1–4; aspectRatio via the existing mapImageSize)
• Response: normalizes predictions[].bytesBase64Encoded → OpenAI { data: [{ b64_json, revised_prompt }] }

Only imagen-* models dispatch here (isImagenModel guard) — gemini flash-image / nano-banana continue to route through /v1/chat/completions (generateContent), unchanged. Errors go through sanitizeErrorMessage.

Test

tests/unit/gemini-imagen-predict.test.ts — 7 cases: registry wiring, parseImageModel("gemini/imagen-4.0-generate-001") resolution, the isImagenModel guard (imagen ✓ / flash-image ✗), the :predict body shape + sampleCount clamp + aspectRatio mapping, and response normalization + empty/absent tolerance.

npm run typecheck:core clean; npm run lint clean (no new no-explicit-any); existing image tests still pass.

Validation limits (please note)

The request-builder and response-parser are pure and unit-tested, but the live Google call is not exercised: Imagen on the Gemini API requires a billing-enabled project — free-tier keys return 403 / quota 0. I don't have a paid Google key to run an end-to-end generation, so a reviewer with billing enabled should confirm one real imagen-4.0-generate-001 call before relying on it. The code path, auth, and shapes follow Google's documented Imagen :predict contract.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[diegosouzapw]

Nice fix — confirmed on release/v3.8.49 that gemini/imagen-4.0-* was advertised via /v1/models but genuinely unroutable (400 on /v1/images/generations, 404 on /v1/chat/completions), and the new google-imagen :predict wiring follows the existing image-provider handler pattern exactly (auth header, sanitizeErrorMessage, saveCallLog). Ran the new test file directly — 7/7 pass — and a full-project tsc --noEmit and eslint on the touched files come back clean (the 3 no-explicit-any hits in imageGeneration.ts are pre-existing/allowlisted, ~800 lines away from your change). Only note, which you already flagged yourself: the live :predict call needs a billing-enabled Google key to exercise end-to-end — worth one real-key smoke test before this gets wide traffic, but that's a merge-after-verify item, not a blocker. Thanks for the clear writeup and the honest validation-limits section.