main
Add pytest setup and remove standalone test script
Fail agent runs that finish without producing output
Make extract_assets degrade gracefully on provider + partial failures
Batch image editing and background removal tools in parallel
Enforce Anthropic’s 2000px limit for requests with many images
Add Claude Opus 5 model variants with effort-based configuration
Improve screenshot asset extraction accuracy and robustness
Pin @babel/standalone so generated React actually renders
Move capabilities endpoint into dedicated route module
update model for edits and format for prompt
Add persistent design systems for prompts
Serve local assets correctly in evals
Support text-brief eval sets alongside image sets
add caching-related tools and remove prompt_cache_key
Support script image URLs in export
Add Gemini 3.6 Flash model variants with pricing and tests
Attach tool-output images to all providers (Gemini, Anthropic, OpenAI)
Clarify image upscaling guidance in prompts and tool definitions
Convert localhost URLs for all Replicate image tools
Refresh text-create model mix from judged text evals
Add GPT-5.6 Terra as an available low-thinking model
Set GPT 5.5 image detail to original
Add an asset extraction preference for code generation
Add eval sets, sessions, and per-session matrix view
Summarize prompt history refactor
Make backend dev fall back to an open port
Support Replicate key in settings (#613)
Fix select-and-edit losing the element reference on imported projects
Finalize extracted assets directly; stop double-saving them
Fix WebSocket disconnect cleanup