• fix: lay out the AI provider picker as a 2x2 grid
• fix: detect vision models on local OpenAI-compatible servers
