Life OS

I wanted one honest picture of my own week, and the data was scattered across eight apps that do not talk to each other. So I built the thing that joins them.

Scheduled jobs pull from Hevy and Strava for training, Toggl for time, Apple Health for sleep, HRV and nutrition, Apple Calendar and Reminders through a Swift bridge, and four banks by CSV drop. Everything lands in a single SQLite database that is the source of truth. Transactions are categorised by rules first and by a language model only where the rules are ambiguous, which keeps it cheap and predictable.

On top of that sits a generation layer that writes plain markdown into my notes, so the output is readable in the same place I already work rather than in another dashboard I would stop opening.

It is a real codebase rather than a script: 46 Python modules, 154 tests across 21 files, nine ingestion sources and nine renderers, uv and pyproject, a Makefile, a `life` CLI, and macOS LaunchAgents doing the scheduling on a fixed daily timetable. It has run since April 2026.

The part I would defend in a review is that every source is idempotent (re-running a sync must not duplicate a row) because a scheduled job that runs eight times a day will eventually run twice at once, and a personal dataset with silent duplicates is worse than no dataset.