Native Gemini image generation through the TEE

[adambalogh]

Adds native, provider-side image generation routed through the TEE, so image output rides the existing OHTTP/HPKE + signing pipeline instead of the privacy-breaking DALL-E tool.

Model registry

• Register gemini-3.1-flash-image ("nano banana 2") — the latest Gemini image model and the one the chat app now routes to — with image_output=True and per-token pricing ($0.50/MTok in, $60/MTok out; an image is ~1120 output tokens ≈ $0.067).
• Also register gemini-2.5-flash-image (image_output=True, ~1290 tokens/image ≈ $0.039) as the prior image model.
• The Google backend requests the IMAGE modality for image_output models.

Chat controller

• Split generated image content blocks out of the text and surface them on the response message (non-streaming) and the SSE final frame (streaming).
• Images are carried out-of-band and are NOT folded into the signed output hash — the OHTTP/HPKE channel already guarantees end-to-end confidentiality and integrity, while the TEE signature continues to cover the request and any text output.

Billing

Image generation is charged exactly like any other completion: Gemini reports the image as output tokens in candidates_token_count, which langchain-google-genai folds into usage_metadata.output_tokens → completion_tokens → output_price_usd. The image bytes themselves are never metered by size. This path (and the no-usage fail-open caveat, where a missing usage_metadata means the client isn't charged) is documented inline in the chat controller and covered by regression tests in test_image_billing.py.

Also folded in

Merges claude/nice-brown-d8XRm, which registers gemini-3.5-flash as a text model ($1.50/$9.00 per MTok). Note: 3.5 Flash is text-output only — it is not an image model, despite the name; image generation uses the *-image models above.

PCR note: changes here require an enclave rebuild + measurements.txt refresh; uv.lock is untouched.

[Copilot]

Pull request overview

Adds native Gemini image-generation support through the existing TEE chat pipeline, including model registration, Google backend modality selection, response image extraction, streaming/non-streaming surfacing, and billing regression coverage.

Changes:

• Registers Gemini image-output models and Gemini 3.5 Flash in the model registry/pricing tests.
• Requests image response modalities for Google image models.
• Splits generated images from text responses and carries them out-of-band in chat responses/SSE final frames, with billing tests for image token accounting.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tee_gateway/model_registry.py	Adds image-output model config support and new Gemini model entries.
tee_gateway/llm_backend.py	Configures Google image models with TEXT and IMAGE response modalities.
tee_gateway/controllers/chat_controller.py	Extracts generated image blocks and includes them in non-streaming/streaming responses.
tee_gateway/test/test_image_billing.py	Adds regression tests for Gemini image token folding and cost calculation.
tests/test_pricing.py	Adds pricing resolution tests for new Gemini models.
CLAUDE.md	Updates documented Google model routing list.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.