0.0.2

Added

• Per-workspace conversations (existing global conversations migrate automatically)
• GGUF models auto-load on first message with a "loading model" card in chat
• llama.cpp server uses random free ports with retry on bind failure

Changed

• Composer dropdowns (model picker, mode menu) now position themselves within the viewport and work in edit mode
• All composers share one selected model and mode
• Auto model selection hidden for now; first enabled model is the default

Fixed

• Production error: Cannot find package '@huggingface/hub' (runtime deps now resolved via file URLs)

Removed

• MCP tool marketplace