Add image quality tier support for image generation models

[adambalogh]

Summary

Adds a quality request parameter (low | medium | high) to image generation endpoints, allowing callers to control output resolution across multiple providers. Each model maps the simple quality tier to its provider-specific size format via a new image_quality_sizes field in ModelConfig.

Key Changes

• Model Registry (model_registry.py):

  • Added image_quality_sizes: Optional[dict[str, str]] field to ModelConfig to map quality tiers to provider-specific resolution strings
  • Configured quality mappings for ByteDance Seedream/Seedance (1K/2K/4K), Z.ai GLM-Image (1024x1024/1280x1280/2048x2048), and Gemini Nano Banana (1K/2K/4K)
  • Added VALID_IMAGE_QUALITIES constant and resolve_image_size() helper function to validate and map quality tiers to model-specific sizes

• LLM Backend (llm_backend.py):

  • Updated get_chat_model_cached() to accept image_quality parameter and pass resolved image size to Gemini's image_config
  • Updated generate_images() to accept quality parameter and apply size overrides to provider payloads (Z.ai, Seedance, Seedream)
  • Models without resolution control (xAI Grok, Gemini 2.5 Flash Image) silently ignore the quality parameter

• Chat Controller (chat_controller.py):

  • Added validation of quality field in CreateChatCompletionRequest with clean 400 error response for invalid values
  • Threaded quality parameter through both streaming and non-streaming image generation paths
  • Updated request serialization/deserialization to preserve quality across session tracking

• API Models (create_chat_completion_request.py):

  • Added quality field to CreateChatCompletionRequest Pydantic model

• OpenAPI Spec (openapi.yaml):

  • Documented quality parameter with enum values and provider-specific resolution mappings

• Tests (test_image_generation.py):

  • Added comprehensive test coverage for quality tier mapping across providers
  • Tests verify correct size presets are sent to each provider's API
  • Tests confirm models without resolution control ignore the quality parameter
  • Tests validate error handling for invalid quality values

• Documentation (CLAUDE.md):

  • Added section explaining image output resolution selection and per-model quality mappings

Implementation Details

• Quality resolution happens at request time via resolve_image_size(), which validates the tier and looks up the model's provider-specific size string
• The medium tier is configured to match each model's previous hardcoded default, ensuring backward compatibility when quality is omitted
• Models without image_quality_sizes configured (e.g., xAI Grok) return None from resolve_image_size(), causing the provider default to be used
• Validation occurs early in the request handler for clean error responses before any provider calls

https://claude.ai/code/session_01ENGmJDbJJHAM89HSGQuoeF