Halfway through the rescue, the agent got sidetracked into thinking DeepWeather was an image generator — I had to pull it back. The AI Studio export came with three Gemini-powered tabs: the 3D weather globe (the actual app), a Veo 3.1 video studio, and an Imagen 4.0 prompt box. The agent latched onto the Imagen tab as if that was the product. It's not. It's a generic text-to-image input that has nothing to do with procedural weather visualization — just a corollary that got exported alongside the real feature. Ripped out ImagenStudio.tsx, its enum value, its service function, and its orphaned screenshot. The nav went from three tabs to two: Globe and Veo. The demo GIF capture also hit a wall — the Chrome DevTools MCP kept dropping the connection when interacting with WebGL pages, so the final recording used macOS screencapture -v as a fallback and got cropped/converted with ffmpeg. The GIF shows the default globe rotating but not a full city-search-and-transform cycle, because that required sustained browser interaction the MCP couldn't maintain. Shipped to GitHub as-is — the live app is the demo.