Foodster

A self-hosted dinner log and meal suggester for a single household. Record what the family actually ate, and — later — let the app propose seven dinners drawn from that history.

The interface is in Finnish. The code, comments and documentation are in English.

See PRD.md for the full specification.

Status

Stage 1 — eating history: in development. The meal catalog and the daily log come first, because the suggester is worthless until there are a few weeks of real history to weight against.

Working:

• Kirjaa — log a dinner: pick a dish, tick sides, save. Dishes are ordered and sized by how often they are eaten, so the likely answer is the biggest target. The history sits on the same page underneath: every day back to the first entry, unlogged days shown as explicit gaps, and every row a link that loads that day into the logger above it.
• Ruuat — add, edit and delete mains and sides, or import a whole bundle by paste or file upload. Grouped by category and alphabetical inside, since this is a list you manage rather than one you pick from. Deletes are soft, so old log entries keep showing the dish they used.
• Light / dark, remembered per device, dark by default. The button shows the theme that is on — moon while dark, sun while light — not the one a click would bring.

Still to build:

• Live search as you type, and paging for the history and catalog lists once years of entries make them long. Both via Datastar.
• Category icons instead of plain colour dots — colour and shape together, so a red blob and a yellow blob are told apart by more than hue.
• Edit and delete as icons in the catalog rows, and a confirmation step before a delete actually happens.
• A background for the header. Something subtle; the palette gets overhauled later.
• Stage 2: the seven-meal suggester, which starts once there is history to weight against.

Stack

One static Go binary. No Node.js, no bundler, no separate database server.

Language	Go 1.27
HTTP	stdlib net/http + http.ServeMux
Views	templ, server-rendered
Interactivity	Datastar — signals and DOM patching in one ~11 kB script
Styling	hand-written CSS, light-dark() for themes
Database	SQLite via modernc.org/sqlite (pure Go)
Auth	HTTP Basic, one shared household password
Runtime image	FROM scratch

Branches

main holds released versions only. Every release tag points at a commit on main, so its history is the deployment history. Nothing is committed to main directly — it moves only by fast-forwarding dev into it.

All development happens on dev. main is protected on the remote: it takes no direct pushes, so a release arrives through a pull request.

git switch dev                              # where the work happens
# ... commits ...
make check                                  # lint, unit tests, smoke
git push origin dev

tea pr create --base main --head dev        # or open it in the forge
# merge the pull request, then:

git switch main && git pull --ff-only
make release                                # builds, tags vYYYYMMDD-N, pushes the image
git push origin --tags

make image additionally refuses to run from any branch but main, so a release tag can never point at a commit that was not released. That check lives locally because it has to: tags and images are built before anything reaches the remote, so protection there cannot catch it.

Quick start

cp .env.example .env     # then edit it
make run                 # http://localhost:8080

make on its own lists every target:

make fix        gofmt, templ fmt, go mod tidy
make lint       go vet, gofmt check, golangci-lint when installed
make test       go test ./...
make build      ./foodster
make seed       import a dish bundle (SEED=seeds/testi.json)
make vendor     re-download the Datastar client
make image      build and tag vYYYYMMDD-N (creates a git tag)
make push       push the newest tag and :latest
make release    image + push
make up/down/logs   compose

Importing dishes

The Ruuat tab takes a bundle of mains and sides: paste the JSON or upload a file, and the app reports row by row what it did.

{
  "mains": [{"name": "Kanacurry", "categories": ["chicken"], "has_sides": true}],
  "sides": [{"name": "Riisi"}]
}

Categories are meat, chicken, fish and vegetarian — stored in English even though the UI is Finnish. A dish may list several, which is how build-your-own meals like tortillas cover every category at once. has_sides defaults to true when omitted.

Import is best-effort, never atomic. Valid rows land; invalid ones are skipped and named (tuntematon kategoria, ei kategorioita, jo listalla).

Names are normalized to sentence case on the way in — whitespace collapses and the first letter is capitalized, the rest is left as typed, because Finnish capitalizes only the first word of a phrase. keitetyt perunat is stored as Keitetyt perunat, while BBQ-kylkeä and Kotipizza keep their capitals. Duplicates are caught case-insensitively, so re-importing a bundle is safe.

The same importer runs from the command line when you just want to repopulate a scratch database:

make seed                      # or: SEED=seeds/other.json make seed

Icons

cmd/foodster/static/favicon.svg is the browser-tab icon and follows the system theme. The home-screen icons are PNGs, because a home-screen icon cannot be transparent and must not change with the theme; they are rasterised from assets/icon.svg, which is opaque and keeps the artwork inside the central 80% so Android can mask it to any shape.

make icons      # rsvg-convert, then optipng -o7

The PNGs are committed so the build needs no rasterizer. Re-run make icons after editing assets/icon.svg.

Migrations

Numbered SQL files in cmd/foodster/migrations, embedded in the binary and applied in filename order at startup. Each runs in a transaction and is recorded in schema_migrations, so it applies exactly once.

Adding one means dropping a new NNNN_what_it_does.sql into that directory. Never edit a migration that has already shipped — a released file has run on a live database and will not run again.

Configuration

Everything is environment variables. .env is gitignored; start from .env.example.

Variable	Default	Purpose
FOODSTER_PASSWORD	required	Shared password. The app will not start without it.
FOODSTER_DB	./data/foodster.db	SQLite file path; the directory is created if missing.
FOODSTER_UID / FOODSTER_GID	1000	Host owner of ./data, for the bind mount.
TZ	Europe/Helsinki	Used for every calendar-day calculation.
FOODSTER_REPO	required to build	Image repository, no tag.
FOODSTER_TAG	latest	Tag to run under compose.
FOODSTER_PORT	8080	Host port to publish.

Set TZ in development too. Under UTC the date rolls over three hours late, which is exactly when dinner gets logged.

Deployment

Images are built with Podman and run under Docker Compose on a LAN server. They are OCI images, so either engine works.

make release             # build, tag, push
# on the server:
docker compose pull && docker compose up -d

Versions are CalVer — vYYYYMMDD-N, where N is the Nth build that day. The running version is served at GET /healthz, which is the one route outside authentication.

There is no database container. SQLite lives in ./data, bind-mounted into the container, so a backup is cp -r data and you can inspect the file with any sqlite client without going through the engine.

That directory must exist and be owned by the user compose runs as — make up creates it, and FOODSTER_UID/FOODSTER_GID in .env tell the container who that is. Get them from id -u and id -g.

If the app exits with cannot open /data/foodster.db ... unable to open database file (14), the ownership does not match. Docker creates a missing bind-mount directory as root, and the container is not root:

ls -ldn data                      # whose is it?
sudo chown -R 1000:1000 data      # match FOODSTER_UID / FOODSTER_GID
docker compose restart

Security

Access is a single shared password over HTTP Basic — no accounts, no sessions. Credentials are compared in constant time over SHA-256 digests, so neither the password nor its length leaks through timing.

The app is served on a public hostname behind Traefik, which terminates TLS, so the credentials are encrypted in transit. That leaves the password as the only thing between the internet and the app, so wrong guesses are rate limited per client address: five in a burst, then one per ten seconds, answered with 429. Requests carrying no Authorization header are not charged — that is the handshake every browser session begins with, and counting it would lock the household out for simply opening the app.

X-Forwarded-For is trusted only when the connection came from a private address, meaning it arrived through the proxy. A client connecting directly could otherwise forge a new address per attempt and skip the limiter.

None of this replaces a strong FOODSTER_PASSWORD. Rate limiting removes brute force as a practical route; it does not make a guessable password safe.

Mockups

mockups/ holds standalone HTML design studies. Open them directly in a browser; they are references, not part of the build.

License

MIT.