- Register Chrome built-in AI as a local text provider with availability checks.
- Add explicit setup/download gating and auto-enable safeguards.
- Preserve structured prompt roles through initialPrompts and stream tests.
- Keep no-key and local model sources disabled until explicit user configuration exists.
- Tighten default text and image model enablement checks.
- Cover environment-driven defaults for local providers.
- Add custom request header parsing and provider metadata resolution
- Add quick switching controls for text model workspaces
- Keep model UI labels and custom header output tidy
- Align Cloudflare model defaults, metadata, and connection labels across core and UI
- Add transient retry handling for Cloudflare image requests
- Harden reference prompt normalization and extend related regression coverage
- Backfill builtin model API keys when restoring stored model configurations
- Keep model manager defaults aligned with the new image model loading flow
- Add focused manager test coverage for missing-key recovery
- add a Workers AI image adapter for FLUX.2 [klein] 4B generation and editing
- add a Cloudflare text adapter with dynamic model discovery and a Qwen3 default
- load Cloudflare token and account defaults from environment variables for both providers
- cover registry integration and default config wiring with unit tests
(cherry picked from commit 3ef9bdbd187ef150d331ef6b9b46f9d7ad3ef2af)
Allow Docker and other runtime environments to inject JSON request parameters for suffixed custom models so provider-specific defaults like NVIDIA NIM thinking flags persist outside the UI. Add parsing, sanitization, tests, and docs for the new VITE_CUSTOM_API_PARAMS_<suffix> flow.
- Replace LangChain dependencies with native OpenAI and Google Generative AI SDKs
- Rewrite LLMService to use direct API calls for OpenAI and Gemini
- Update core package dependencies to remove LangChain-related packages
- Modify message handling and streaming logic to work with new SDK implementations
- Simplify model configuration and instance management
- Update integration and unit tests to support new SDK approach
- Remove LangChain-specific message conversion and streaming methods
- Remove redundant integration and unit test files from web package
- Update import paths in remaining test files to use local source
- Remove console logs and unnecessary comments in test files
- Simplify test structure and remove duplicate test scenarios
- Add dotenv dependency to core package for environment variable support
