Personal Pomodoro: AI-Built Native Mac App

A model evaluation disguised as a productivity tool. Give Gemini 3.7 Flash a non-trivial native macOS app to build from scratch — not a web app, not a script, but a real SwiftUI menu bar application — and see how far it gets without intervention.

It got all the way.

What the model built

Component	Implementation
Architecture	18 Swift files across Models/Services/Views with clean separation
Timer engine	DispatchSourceTimer at 250ms against wall-clock target — drift-proof
Penalty system	2 free snoozes → escalating to physical challenges, break time tax, lockout
Break tasks	8 curated micro-tasks: stretches, eye breaks, hydration, breathing
Stats	Daily tracking with midnight resets via UserDefaults persistence
Sound	NSSound wrapper with 10 alert chimes and live preview in settings
Shell	Dockless LSUIElement, NSPopover glass panel, pure menu bar citizen

The test, not the tool

The real output isn't a Pomodoro timer — it's a datapoint. Can a flash-tier model produce a shippable native macOS application with system-level features, clean architecture, and zero human code intervention?

The answer: yes, with caveats. The initial build compiled and ran. Two bugs surfaced — NSPopover key window focus and tab label overflow — both fixed by the model in the same session. The penalty engine, stats dashboard, and sound system all worked first try.

Under the hood

The app runs as a pure LSUIElement accessory — no Dock icon, no menu bar, just a 🍅 that opens a glass popover. State machines manage timer lifecycle (idle → running → paused → break → complete), and the penalty engine tracks snooze history across sessions to escalate consequences.

Built entirely in a single Antigravity session on August 14, 2026. Shipped to GitHub with 26 files and zero git history.