Add native support for Open WebUI's image generation API as a new
image provider option. Open WebUI exposes an OpenAI-like
/v1/images/generations endpoint but with key differences that
require special handling:
- Response is a bare JSON array instead of {"data": [...]}
- Image URLs are relative paths (e.g. /api/v1/files/.../content)
- File downloads require the same Bearer auth token
The implementation uses raw HTTP calls via aiohttp rather than the
OpenAI SDK to handle these differences. No model parameter is sent
since Open WebUI manages the image model in its own admin settings.
Backend changes:
- New OPEN_WEBUI enum value in ImageProvider
- generate_image_open_webui() method in ImageGenerationService
- Environment getters/setters for OPEN_WEBUI_IMAGE_URL and
OPEN_WEBUI_IMAGE_API_KEY
- UserConfig model and config loading/saving pipeline updated
Frontend changes:
- New "Open WebUI" option in image provider dropdown
- Settings UI with URL and optional API key fields
- Validation, field mappings, and config persistence
Docker:
- OPEN_WEBUI_IMAGE_URL and OPEN_WEBUI_IMAGE_API_KEY added to all
docker-compose service definitions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Added a new method `_generate_codex` to LLMClient for generating responses using the Codex Responses API.
- Refactored tool handling to support structured tool calls and improved error handling during Codex interactions.
- Updated the client initialization for Codex to streamline the generation process.
- Enhanced message processing to accommodate new input formats and recursion for tool calls.
- Introduced CodexLLMAdapter to handle structured and unstructured LLM calls via the Codex Responses API.
- Implemented methods for converting tool formats and building request bodies for Codex.
- Refactored LLMClient to utilize CodexLLMAdapter, enhancing tool call handling and response processing.
- Updated LLMToolCallsHandler to support Codex as a valid LLM provider.
- Enhanced schema utilities for better JSON schema handling.
- Implemented a new method to create an AsyncOpenAI client specifically for Codex.
- Added functionality to convert Codex event responses into a dictionary format for easier consumption.
- Updated the streaming methods to utilize the new Codex client and handle API errors gracefully.
- Improved logging for debugging purposes during Codex interactions.
- Refactored message handling in the slide content generation function for clarity and efficiency.
- Added Codex authentication endpoints and logic for OAuth flow.
- Updated Docker configuration to expose port 1455 for Codex callback.
- Enhanced user configuration model to include Codex-related fields.
- Integrated Codex selection into the LLM provider UI.
- Implemented token management and refresh logic for Codex.
- Added utility functions for handling Codex OAuth tokens and state management.
