A release spans a fix, a feature and some chores at once, so no single
conventional type describes it honestly. release: makes main's log one line
per deployment, which is what that branch is for, with the pull request body
as the notes.
A dozen conventions had accumulated that existed only in my head and in the
commit log: Finnish interface against English code, no infrastructure detail
in the repository, migrations immutable once shipped, interactions that patch
rather than navigate, ponytail comments marking deliberate shortcuts, and
asserting what a response does rather than only that it responded — the one
that let the /ruoat 404 ship.
Commit messages take a type from here on: feat, fix, chore and the rest. The
body still carries the weight, since these commits are the only design record
this project has. Pull request titles get the same treatment, because main is
squash-merged and a title becomes a commit message on it.
With dev and prod open side by side in the same browser, the tabs were
indistinguishable. ENV is written into the title of every page unless it says
prod, so "dev · Foodster" picks itself out. The value is used verbatim, so
ENV=staging labels itself too, and prod and production both count as unmarked
so a stray capital cannot tag the real instance.
Environment variables lose their prefix: PASSWORD, DB, ENV, ADDR, HOST, REPO,
TAG. The container namespaces them already, and this matches how the other
services here are configured.
PUID/PGID are the exception rather than UID/GID. UID is read-only in bash, so
a value set in .env would be silently replaced by the invoking shell's own and
compose's user: would ignore what was asked for.
Breaking for a running instance: the deployed .env has to be rewritten in the
same deploy, or the app will refuse to start on an unset PASSWORD.
Kirjaa now works like the catalog: opening a day, picking a dish, saving,
deleting, cancelling and "Näytä lisää" all patch the list where it stands.
Nothing loads a page, so the scroll position never moves.
One builder serves all three paths. buildLog takes what the screen should
show — the day, an open dish, whether the entry is being changed or a delete
confirmed — and the page render, the patch and the post-write response all go
through it. After a save it is called with only the date, so the day comes
back closed rather than reopening the sides step it was just submitted from.
Links stay links and forms stay forms, with data-on:click__prevent and
data-on:submit__prevent layered over them, so it all still works with
JavaScript off. Each response patches a single element, so plain text/html is
enough here; the SSE writer is only needed by the catalog, where the list and
both forms have to move together.
The anchors added in the previous attempt are gone. They could never have
worked: the browser positions an anchor without knowing where the page was
scrolled, so it jumped regardless.
Deleting a dish partway down the list sent the browser back to the top. The
first attempt at fixing it used anchors, which cannot work: the browser
positions the element with no knowledge of where the page was scrolled, so it
still jumps. Datastar was already loaded and doing nothing but search.
Every catalog action now patches. The bin, the pencil and Peruuta stay real
links; the forms stay real forms. Datastar intercepts them with
data-on:click__prevent and data-on:submit__prevent, and the same handlers
redirect when the Datastar-Request header is absent, so none of it requires
JavaScript. Posting with {contentType: 'form'} sends the enclosing form as
FormData, which means the handlers keep reading r.FormValue and no input had
to be rewritten as a signal.
The response is one SSE event carrying three elements: the list and both
forms. They have to move together — opening an edit form also has to clear a
delete that was mid-confirmation — and a text/html response can only replace
one element. The writer is twenty lines rather than re-adding the SDK and the
four modules it brings for a generator we would otherwise never call.
Smoke checks assert the wire format: that these answer with an event stream
carrying all three elements, that the deleted dish is absent from the patched
list, and that a header-less post still redirects.
Adding, editing or deleting a dish redirected to /ruoat, which stopped
existing when the tab was renamed to Ruuat. Every one of those actions ended
on a 404. Shipped in v20260905-4.
The rename was done with a scripted replace across views.templ, main.go and
the smoke script; handlers.go was not in the list.
The tests did not catch it because they asserted only that the response was a
303. A redirect to a dead URL is still a 303. They now assert the target.
The deployment is a public hostname behind Traefik rather than a LAN-only
box, which changes two things.
TLS is now terminated by the proxy, so Basic credentials are no longer in
cleartext. The container publishes no ports: doing so would leave an
unencrypted copy of the app on the host, bypassing the proxy. The hostname
lives in .env rather than compose.yaml, so no infrastructure detail is
committed and the MIT publication option stays open.
The password is now the only thing between the internet and the app, and a
500 ms sleep is not a defence at that exposure. Wrong guesses are rate
limited per client address: five in a burst, then one per ten seconds,
answered with 429.
Two details that decide whether this works at all:
- a request with no Authorization header is not charged. That is the
handshake every browser session opens with, and counting it would lock the
household out for simply opening the app a few times.
- X-Forwarded-For is believed only when the connection arrived from a private
address, i.e. through the proxy, and then only its last entry, which is the
one the proxy observed. A direct client could otherwise forge a fresh
address per attempt and walk past the limiter entirely.
None of this substitutes for a strong password. It removes brute force as a
practical route, nothing more. PRD §3, §9 and §10 are updated: "no external
internet exposure" is no longer true.
main carries released versions only, so its history is the deployment
history and every release tag points into it. Development happens on dev
and merges into main when a release is cut.
Three pieces of chrome that were missing or misleading.
A header with the bowl mark and the app name, on every page. The page title
bar below it is no longer sticky: on a phone a tall sticky header eats the
screen, and the bottom tab bar already handles navigation.
A theme toggle, remembered per device in localStorage, because a shared
instance with no accounts should let the kitchen phone and the laptop
disagree. Dark by default, and the default lives in the server-rendered
markup so it survives with JavaScript off and never flashes. The button shows
the theme that is on — moon while dark, sun while light — rather than the one
a click would bring, and its label names the state before the action. Both
icons ship on every page and CSS picks one, so the server never needs to know
what this device chose. Following the system was considered and dropped: a
third state costs a control harder to read than the choice is worth.
The checkbox chips no longer hide the native control behind opacity: 0. With
only a background colour to go on there was no way to tell "Tarjoillaan
lisukkeiden kanssa" on from off. Colour is not a state indicator; a checkbox
is. This covers the side, category and has_sides chips alike.
Smoke checks cover the parts that would regress silently: that dark is the
no-JS default, and that both theme icons are present for CSS to choose from.
The favicon is a bowl of soup: this household's catalog is half keitto, so it
is at least honest. favicon.svg follows the system theme in the tab strip.
Home-screen icons cannot do the same. They must be opaque and must not change
with the theme, so assets/icon.svg is a separate fixed-colour source, with the
artwork inside the central 80% for Android to mask to any launcher shape. The
same 512 is declared maskable in the manifest.
`make icons` rasterises with rsvg-convert and compresses with
`oxipng -o max --zopfli`, which beat `optipng -o7` at every size — 5860 bytes
against 6109 for the three files. Plain `oxipng -o max` was not an upgrade: it
won the two small icons and lost the 512, ending up larger overall. All output
verified pixel-identical to the rasterizer. The PNGs are committed so the
build needs no rasterizer.
Two things that only fail on a real device, so both are covered by smoke
checks: Go has no mime entry for .webmanifest and served it as octet-stream,
which browsers ignore; and a manifest fetch carries no credentials by default,
so behind Basic auth it needs crossorigin="use-credentials" or it 401s.
The catalog could only be filled by importing JSON, which is a poor way to
add the one dish you are about to eat.
Ruoat now covers PRD §7.3 in full: add and edit mains with their categories
and has_sides, add and edit sides, and delete either. Deletes are soft, so a
log entry keeps resolving the dish it used and the freed name can be reused.
Validation messages are Finnish and the rejected form comes back filled in
rather than blank. Bulk import moves into a details element, since it is now
the occasional path rather than the only one.
Kirjaa gets the same ability without the detour: a search that finds nothing
offers to add what was typed, and saving creates the dish and continues
straight to the sides step. An empty catalog shows the same card instead of
dead-ending on a link to another tab, and the search box is no longer hidden
behind the empty state.
The importer's own insert is gone; it and the UI both go through createMain
and createSide, so duplicate detection lives in one place and reason() can
match on errNameTaken instead of poking at driver strings.
Stage 1's point is collecting eating history, so the logging path is the one
that has to be frictionless: pick a dish, tick sides, done.
Kirjaa:
- dishes ordered and sized by how often they have been eaten, so the likely
answer is the biggest target on the screen
- picking a dish opens the sides step; a dish with has_sides false says so
instead of offering an empty list
- saving redirects, so a refresh cannot double-post
- a day already logged shows the entry with edit and delete. Editing reopens
the sides step with the existing sides ticked, which makes editing and
creating the same screen
- day switcher and a server-side search over the catalog
Historia walks back day by day to the oldest entry, so a day nobody wrote
down appears as an explicit gap rather than quietly missing.
No JavaScript is involved: every interaction is a link or a form, and the
checked-chip styling is :has(input:checked). Datastar stays loaded but unused
until an interaction genuinely needs to avoid a page load.
Also fix a Makefile ordering bug: lint did not depend on generate, so `go vet`
could run against templ output that was being rewritten. check now runs its
phases as sub-makes so `make -j` cannot interleave them.
Records two features that are specified but not built: dish CRUD in the UI
(§7.3, only import exists today) and a per-device light/dark switch (§7.4).
The switch must show the state that is active, not the one clicking produces.
SQLite writes three files — the database plus its -wal and -shm companions.
Putting them in one directory means a deployment mounts a single path and a
backup copies a single directory.
- FOODSTER_DB defaults to ./data/foodster.db, and openDB creates the parent
directory on startup rather than failing on a fresh checkout
- compose bind-mounts ./data instead of using a named volume, so the file can
be listed, copied and opened with any sqlite client without going through
the container engine
- the image runs as UID 65534, which cannot write to a host directory owned
by someone else, so compose now sets user: from FOODSTER_UID/FOODSTER_GID
- `make up` creates ./data first: left to the engine it appears root-owned
and the app silently cannot write to it
Bring up the stage 1 skeleton described in the PRD, enough that the app
builds, serves, and can be populated with dishes.
- net/http server with shared-password Basic auth, /healthz outside it,
graceful shutdown, and TZ-aware calendar days
- SQLite via modernc (pure Go, static binary), opened with WAL and a single
connection
- migration runner: numbered SQL files embedded and applied once each inside
a transaction, recorded in schema_migrations
- bundle import (PRD §7.3) as a live feature on the Ruoat tab: paste JSON or
upload a file, get a per-row Finnish report. The same importer is reachable
as `foodster -import` for repopulating a scratch database
- templ views and hand-written CSS with light-dark() theming; the Datastar
v1.0.3 client is vendored, since the Go SDK ships no browser asset and a
CDN would break an offline LAN
Names are normalized to sentence case rather than title case: Finnish
capitalizes only the first word of a phrase, so "Keitetyt perunat" is right
and "Keitetyt Perunat" is not. PRD §6 and §7.3 are amended to match.
Testing is behind make targets rather than ad-hoc commands: `make check` runs
lint, unit tests and scripts/smoke.sh, which exercises auth, static assets
and every import path against a scratch database on a spare port.
Foodster is a self-hosted dinner log and meal suggester for one household.
Stage 1 (eating history) is in development; stage 2 (the suggester) follows
once there is enough history to weight against.
Replace the PRD's original stack (Nuxt, Postgres, Drizzle, Pico CSS) with
Go 1.27, net/http, templ, Datastar and SQLite — one static binary, no
Node.js in the build. Sections 3, 4, 5, 9, 10 and 11 are rewritten to match.
Record the decisions made while reviewing mockups:
- the UI is written in Finnish; PRD, code and comments stay English
- access is one shared password over HTTP Basic rather than open on the LAN
- images are CalVer vYYYYMMDD-N, built with Podman, run under Docker Compose
- registry coordinates live in .env, so no infrastructure detail is
committed and the MIT publication option stays open
mockups/ holds standalone HTML design studies. log-fi.html is the current
one; the others are superseded exploration kept for reference.
Esa Kataja