I built a task app that stores everything in one plain SQLite file — and I don't regret skipping Core Data

Phinn — Tue, 22 Sep 2026 12:30:34 +0000

This is the engineering story behind KinetTask, a Mac/iOS task app I shipped last week. It's a Show HN-adjacent story but dev.to gets the parts HN would skewer me for. Numbers and regrets included.

The decision everyone told me was wrong

When I started building a local-first task app, every senior Swift dev I asked said the same thing: "Use GRDB. Or SwiftData. Writing SQLite by hand is masochism."

I did it anyway: the SQLite3 C API, directly, in Swift 6.1. Not because I'm tough — because I wanted exactly three things and none of them needed an ORM:

The database file is the export format. Copy tasks.sqlite3 to another machine, it just opens. No "export" feature needed.
Every agent action lands as an append-only journal row — so when the AI planner retries a task, I can diff what actually changed instead of replaying blindly.
Checkpoints store a hash of the normalized artifact. Hash the normalized artifact, not the raw output — timestamps will gaslight you into thinking two identical results differ.

The actual surface area (it's small)

// The whole "ORM":
// - open with WAL + busy_timeout
// - one prepared statement per query
// - one user_version pragma per schema migration step

That's it. A few tables, WAL mode, prepared statements, PRAGMA user_version for migrations. The entire persistence layer is ~600 lines including the journal and checkpoint logic. GRDB would have been 150 lines of my code plus a dependency I'd have to reason about during every weird concurrency bug.

The honest cost: I wrote two use-after-free bugs against sqlite3_stmt in the first week. C API gives you exactly enough rope.

The bug that justified the whole approach (v1.2.1)

My iOS widget was silently losing every checkbox tap. Root cause, the fun kind:

Main app's DB lived in its sandbox Application Support
The widget wrote to the App Group container — a different physical file
Both files happily existed forever. No crash. No error. Just… widget taps vanishing.

Fix: unify the data root (App Group on iOS), migrate legacy DBs, turn on WAL + busy_timeout(2s) because now two processes share one file, and add a snapshot-reconciliation pass when the main app becomes active — because the widget writing to SQLite doesn't magically invalidate the main app's in-memory cache, and its next UPDATE would happily overwrite the tap with stale state.

None of this was exotic. All of it was invisible until a real user compared widget state with list state. With Core Data this specific bug would've been harder to create (shared store coordinator handles cross-process hints) — that's the fair counterargument, and I'd still make the same trade, because I can sqlite3 tasks.sqlite3 on a user's machine during support and see everything.

What on-device LLM actually feels like

The AI features (natural-language capture, "what should I focus on") run on-device: Apple's foundation model where available, a GGUF model via llama.cpp in-process otherwise.

The real numbers nobody puts in marketing:

Cold inference on a 3B model: noticeably dumber than any cloud model. For "parse this sentence into a task with due date" it's fine. For "reorganize my projects" it's… enthusiastic.
The failure mode isn't hallucination, it's over-triggering: the model wants to help when it shouldn't. Fix wasn't a better model — it was an anti-filler gate: if the parse adds no fields beyond what you typed, don't create filler structure.

What I'd do differently

Ship the widget after verifying both processes write the same file. Obvious in hindsight; it always is.
The journal + hash checkpoint design came from my agent-engine side project, and it's the only reason I could debug the retry-drift cases at all. If your app has any automated writes, an append-only log is cheap insurance.
I still don't have sync. The file is yours — iCloud Drive it if you want. v1.2 added a .kinettask backup format for machine moves, which covers 90% of what people actually ask for.

Free for 50 tasks, one-time $4.99, no subscription, no account, no telemetry: App Store