Your work should leave a record, not a streak.

Streaks, badges, confetti for opening the app, a dashboard that measures tapping instead of work. The metric a productivity app optimises for is retention, not competence.

Monolith optimises for something else. A focused session should leave a visible mark, and the mark should be proportional to the effort. So the app has no streak to protect and no celebration to fire — just a grid that fills, and a block of stone that grows a layer every time you do the work.

The three existing categories each fail in a way that is hard to patch around. Streaks punish rest. Dashboards measure the wrong thing. Cloud trackers hold your record hostage.

None of them are wrong about wanting engagement. They are wrong about what engagement is for.

The centre of the app is a block of stone. Each identity — Code, Writing, Movement — is its own block. Every recorded session or manual log deposits a layer whose thickness is derived from the effort it took: a 47-minute session is a thick band, a 3-minute log is a hairline.

Layers stack in time. Recent work stays distinct at the top; older work compresses into a base that still accounts for every minute. Tapping a slice opens the layer inspector, so the geometry is never decoration — you can always read it back.

Opening the app is not the point, so the Grid leads: minutes today against your goal, the block of stone, 26 weeks of history, identity filters, and today’s tasks. Start focus from there.

The Focus screen is a count-up timer. It keeps the screen awake, survives backgrounding, and drafts an unsaved layer into the stone as it goes, so you can see what you are about to bank. Finishing records the session exactly once; the layer then settles into the stack. History lets you inspect any day, edit a duration, change an identity, remove a record, or undo the whole thing — and every edit keeps the original task, date, and completion time.

Everything is on the device. Activity records live in AsyncStorage, state is persisted with Zustand, and secrets — if any — go to SecureStore. The app is fully functional offline; there is no request to make before the first session.

Supabase is there but off by default. The repo carries a schema, row-level security policies, and an auth-profile trigger migration, and the email sign-in UI is built. What is not built is the sync itself: no mutation queue, no conflict resolution, no realtime. Turning it on is a decision, not a default.

Settings can export every record as JSON and delete all local data behind a confirmation. Both exist because local-first software has to be honest about who owns the data.

The app animates like stone settling, not like a slot machine. Presses spring. The tab indicator slides. A new layer drops into the block and finds its place. Nothing loops, nothing pulses for attention, and there is no sound.

Every animation respects the system’s Reduce Motion setting: the geometry swaps for fades rather than moving, because a layer that slides into place is the whole point of the model — and a reduced-motion build that still slides is worse than one that does not move at all. Haptics fire on the same layer drop, so the stone lands with weight.

The local-first loop is implemented end to end and verified on the iOS Simulator through both Expo Go 57 and a Debug development build. Android is verified through a production export, and the web build is used for browser QA. TypeScript and ESLint are clean, and 32 Jest tests across 4 suites cover the date, grid, and XP utilities.

The honest part: this is a pre-release MVP. Nothing is on the App Store, and verification so far covers simulators and exports rather than physical devices.

Designing a reward you cannot lose took longer than designing a streak. The first attempts kept wanting a celebration, and every celebration turned into a nag. The grid only works because it is quiet.

The second lesson was making effort legible. A layer whose thickness is derived from real minutes is a far better signal than a checkbox, but it means the geometry has to be data-driven rather than decorative — recomputed from the record, not drawn by hand. That decision is what makes edit and undo cheap: change the record and the stone changes shape.

The third was scope. It is much easier to ship a local-first MVP that admits it does not sync than a cloud product that pretends otherwise.