# puppy-tracker

A tiny offline-first PWA for tracking your puppy's sleep, walks, meals, pees,
poos, weight, and training.
The browser is the primary client; a small Go server provides a shared
source-of-truth and sync between devices.

## How sync works

- Each event has a UUID and an `updatedAt` timestamp.
- Mutations (add / edit / delete) happen against `localStorage` first, so the
  app keeps working when offline. Deletes are recorded as tombstones so they
  can propagate.
- On app load, on `online`, on every mutation (debounced), and every 60 s, the
  client POSTs its full event list to `/api/events/sync`. The server merges
  it with its own copy using last-write-wins on `updatedAt` and returns the
  merged set.
- The server keeps its copy in a SQLite database (`puppy.db`); events and the
  shared profile are separate tables, and last-write-wins is enforced by the
  upsert itself. On first start it auto-imports any legacy `events.json` /
  `config.json` sitting alongside it, renaming them to `*.imported`.
- Service worker bypasses cache for `/api/*` so writes always hit the server
  when online; static assets are still cached for offline use.
- Training exercises (name + how-to instructions) are their own synced
  collection with the same contract as events (UUIDs, last-write-wins,
  tombstones) via `POST /api/exercises/sync`. Training sessions are ordinary
  events (`type: "training"`) referencing an exercise by id, so they ride the
  event sync unchanged.
- The puppy's name and birthday are a per-account profile stored on the host
  (`GET`/`PUT /api/config`), so a new device picks them up automatically instead
  of being configured per-client. The client caches the last-seen values in
  `localStorage` for offline/instant paint and reconciles with the server by
  last-write-wins on `updatedAt`. The age shown in the header (in weeks and
  months) is derived from the birthday.
- All data is scoped to the signed-in account (see [Accounts](#accounts)): every
  event, profile and photo carries a `user_id`, and `localStorage` is namespaced
  per user so two accounts on one browser never mix.
- A day can be marked **not counted** (see [Days that don't
  count](#days-that-dont-count)). The mark is itself an event
  (`type: "day-excluded"`, timestamped at noon), so it syncs and un-marks by
  tombstone like everything else.

A status pill in the header shows `syncing…` / `synced 2m ago` / `pending` /
`sync error` / `offline`. Tap it to force-sync.

## On screen

The panels are grouped into five tabs — **Today**, **Sleep**, **Walks**,
**Habits** (pee/poo/meal timing and counts) and **Growth** (weight, training,
notes) — so each screen holds one subject instead of all fourteen panels in one
column. The day bar and the log buttons sit above the tabs and stay put on all
of them, because logging has to be one tap from wherever you are.

- The day bar carries up to two timers, each of which logs the boundary that
  ends what it is counting when tapped: the asleep/awake one, and — only while
  a walk is running — the walk. They are frozen at the top on every tab, and
  they are the only place a timer appears. There used to be a big card at the
  top of Today as well, with the pills standing by until it scrolled out of
  sight; a timer you have to scroll to, on one tab in five, is not doing the
  job a timer is for. Having one place to render them is also what lets them
  carry seconds — they are the display now, not a summary of one.
- **The day picker is a month grid of the app's own**, not the browser's. The
  native one is a sheet covering the screen, and the reason to change day is to
  see what the figures did on it — so this is a small panel under the bar, with
  the overview still visible and updating as you move. `←` and `→` stay in the
  bar for the common ±1 day; the grid handles jumps and carries *Today*, which
  is what freed the width to fit two timers on one row. The hidden
  `<input type="date">` remains the value everything reads; only its own picker
  is no longer opened.
- The bar can still wrap, and does below about 300px. Its height changes when
  it does and the tab bar sticks to that height, which is why `--day-bar-h` is
  kept current by a `ResizeObserver` rather than measured once.
- Each tab is a `.tab-panel` wrapper around the existing sections. The
  **wrapper** is what gets hidden, never the sections: `walk-timeline` and
  `walk-trend` carry their own `hidden`, set by `renderWalkPatterns` once a walk
  exists, and hiding them directly would clobber it.
- `render()` still draws every panel on every pass, including the tabs you
  can't see. Nothing measures layout — the charts scale through their `viewBox`
  — so drawing into a hidden wrapper is safe, and it means a tab is never
  briefly stale when you arrive on it.
- Tapping a panel's heading still folds it away, remembered across reloads, and
  composes with tabs: tabs group, folding tunes what shows within a group. The
  chosen tab is remembered the same way (device-global, like the theme).
- **Back returns to Today**, from any tab, in one press; a second press leaves
  the app. Exactly one history entry is ever live — armed on leaving Today and
  spent on returning, whether that return came from the back button or from
  tapping the tab. An entry per switch is what a browser does unaided, and is
  why tabbed apps get a reputation for trapping you: flick between tabs fifteen
  times and it takes fifteen presses to escape. Two presses, always, from
  anywhere.

## Layout

```
puppy-tracker/
├── flake.nix         # packages (server, static, default), devShell, nixosModule
├── module.nix        # systemd unit, StateDirectory, hardening
├── server/
│   ├── go.mod
│   ├── go.sum
│   ├── main.go       # SQLite store, LWW sync, static file serving
│   ├── auth.go       # accounts, sessions, invite-gated registration, guest links
│   ├── reminders.go  # reminder rules, the evaluation loop, push subscriptions
│   ├── webpush.go    # VAPID + RFC 8291/8188 message encryption
│   ├── pedigree.go   # SKK lookup, background crawl, per-dog cache
│   └── htmlutil.go   # scraping helpers for the pedigree crawl
└── src/              # the web app
    ├── index.html
    ├── app.js
    ├── style.css
    ├── sw.js
    ├── manifest.json
    ├── changelog.json
    ├── icon.svg
    └── icon-180.png, icon-192.png, icon-512.png
```

## Run locally

```sh
nix run                       # http://localhost:8080, data in $XDG_DATA_HOME/puppy-tracker
PUPPY_ADDR=:9000 nix run      # custom port

# Registration needs an invite code (see Accounts). Set it in the environment:
PUPPY_INVITE_CODE=letmein nix run

# Hot-iterate (data in /tmp):
nix develop -c sh -c 'cd server && go run . -static ../src -data /tmp/puppy.db -invite-code letmein'
```

## Accounts

The app is multi-tenant: each person signs in and sees only their own puppy's
events, profile and photos.

- **Sessions.** Passwords are hashed with bcrypt; login mints a random session
  token stored (hashed) in the `sessions` table and set as an `HttpOnly` cookie.
  `/api/*` (except `login`/`register`/`logout`) requires a valid session.
- **Registration is invite-gated.** Sign-up requires the shared secret passed via
  `-invite-code` / `PUPPY_INVITE_CODE`. With no code set, registration is
  disabled (existing accounts can still log in). Share the code with whoever you
  want to give an account.
- **First account adopts existing data.** When accounts are introduced on a DB
  that already had single-tenant data (or that imported a legacy `events.json`),
  the first account to register inherits all of it — events, profile and photos.
- **Self-service deletion.** Settings → *Delete account* removes the signed-in
  account and everything it owns (`DELETE /api/me`, re-confirming the password):
  events, profile, sessions and the photo directory are all wiped.
- **Serve over HTTPS in production.** Session cookies are only marked `Secure`
  when you pass `-secure-cookies` (enable it behind a TLS proxy), so passwords
  aren't sent in the clear.

## Days that don't count

Not every logged day is equally trustworthy. A day someone else had the puppy —
a sitter who forgets half the pees, a stay at kennels — leaves a thin record
that reads exactly like a real one, and then drags the averages down and puts a
misleading trough in every chart. **Not counted**, in the overview panel's
heading, takes the day you're looking at out of the aggregates.

- **Nothing is deleted or hidden.** The day's overview, history and sleep/wake
  list are unchanged — just dimmed and labelled. Navigate to it and it is all
  still there.
- **What stops counting** is the behaviour: sleep hours, timeline and trend,
  walk minutes and patterns, pee/poo/meal counts, food, by-hour, the training
  grid, and the Timing panel's typical gaps.
- **Marking a day is not the same as deleting its events**, and only two panels
  are handed a filtered list (by-hour and training, which bucket individual
  events and care about neither day boundaries nor spans). The rest take the
  whole log and exclude days themselves, because three things break if the
  events simply go: "how long since the last pee" is a question about *now* and
  answered from the wrong event; the trend curve for the day you are *looking
  at* collapses to zero; and a nap from 23:00 on a marked day to 07:00 on the
  next loses its `sleep-start`, leaving a dangling `sleep-end` and costing the
  next day — which isn't marked — seven hours it really slept.
- **What keeps counting** is weight and notes. A weigh-in and a vet note are
  records of fact, not behaviour a sparse logger distorts, so they stay on the
  weight curve and in the Notes log.
- **Charts keep the day's slot**, drawn as a hatch rather than a bar. Dropping
  it would make consecutive bars stop being consecutive days, and an empty bar
  would read as "the puppy barely slept" — the exact misreading being fixed.
- **Gaps that reach across a marked day are discarded, not measured.** With the
  day's events gone, Tuesday's last pee sits next to Thursday's first, and
  subtracting invents a thirty-hour gap that would blow out the Timing panel's
  "longest" far worse than the sparse day did. Sleep and walk durations need no
  such care — `sleepMsInRange` / `walkMsInRange` already clip to the day being
  measured, so a nap running in from a marked day contributes only its counted
  part.
- **Owner-only.** A guest can't decide their own thin day shouldn't count, nor
  take a good one out of the averages; the server drops `day-excluded` events
  arriving on a guest session and the client hides the control.

## Guest links

A dog sitter needs to log a pee; they do not need your password. **Settings →
Guest access** mints a link that does exactly the first thing.

- **It is a session, not an account.** Opening `/guest/<token>` mints an ordinary
  session row against *your* `user_id`, tagged with the link it came from. Every
  data path downstream — sync, photos, the profile — is scoped by `user_id` as
  before, so a guest simply is you as far as the data is concerned. Only the
  capability checks differ.
- **What a guest gets.** The whole app to read: every panel, every chart, all
  history. They can log new events freely, and edit or delete the ones they
  logged themselves. What they don't get is anything under Settings that belongs
  to the account — the puppy profile, the pedigree id, reminders, other guest
  links, and deleting the account. Those routes are behind `requireOwner` and
  403 for a guest; the client hides the matching UI. The two device-local
  preferences (dark mode, confetti) stay, since they are the guest's own browser
  and not your account.
- **A guest cannot change your logs.** The upsert in `Store.sync` only lets a
  guest update rows carrying their own link's id, so a sitter can fix up their
  own entries and cannot rewrite or delete a single one of yours — including
  everything logged before guest links existed. The check is on the link *id*,
  not its label, because two links can easily both be called "Sitter". You keep
  full control either way and can edit anything on your own account, theirs
  included. The exercise library is the owner's for the same reason: a guest
  logs training sessions against it but the server drops any exercise a guest
  sends. In the app a guest opening someone else's entry gets a read-only view
  rather than a form that would throw away what they typed.
- **It expires, and you can revoke it.** You pick the last day the link should
  work; it stops at the end of that day in your own timezone. Sessions minted
  from a link are capped at the link's own expiry, so one can never outlive it,
  and every request re-checks that the link is still live — so revoking kicks
  whoever is already using it out on their very next request, not whenever their
  session happens to lapse. Revoking also deletes those session rows outright.
- **The URL stays available.** Settings lists each live link by label, expiry
  and when it was last used, with the URL and a *Copy* button, so a link can be
  re-sent without minting a new one and stranding whoever holds the old. That
  means the token is stored, not just its hash — a deliberate trade, and not the
  one you would make for a password or a session token: a guest link grants a
  subset of what the same database already holds in plaintext, so whoever can
  read `puppy.db` gains little from it, and it expires and can be revoked
  besides. The lookup column stays a hash; the secret sits beside it.
- **Events say who logged them.** An event created through a link carries that
  link's label (badged in the History log) and its id (which is what authorises
  changes). The server stamps both from the session on insert and never reads
  them off the wire, so neither can be forged; both are left out of the update
  path, so a later edit by anyone keeps the original attribution.
- **A link is a bearer token — serve over HTTPS.** Anyone holding the URL can
  redeem it until it expires. Send it over something private, and run behind TLS
  (`-secure-cookies`) as above. When a link ends, the guest's browser drops its
  cached copy of your history rather than keeping it around.

## Reminders

Opt-in push notifications for the two things that are easy to lose track of:
"time to sleep" and "nothing logged for a while". Turn them on per rule in
**Settings**.

- **Evaluated on the server.** A closed PWA has no timers, so the browser cannot
  remind you of anything on its own. The server already holds the event log
  (clients sync on every mutation), so a goroutine re-checks every enabled rule
  once a minute and pushes the ones that have come due.
- **Two rule shapes.** `sleep` measures from the last `sleep-end` and fires only
  while the puppy is awake. `pee` / `poo` / `eat` measure from the newest event
  of that type. Each rule has its own interval and repeats at that interval while
  it stays overdue.
- **Quiet while the puppy sleeps.** The event rules are suppressed whenever the
  latest sleep boundary says "asleep", which is what keeps them from nagging all
  night — and means an overdue rule fires promptly on waking instead. The server
  derives sleep state exactly the way `currentSleepState()` does in `app.js`,
  tie-break included, so both sides always agree.
- **One notification per rule.** Every push carries a `tag`, so a repeat replaces
  the previous notification instead of stacking another one on the lock screen.
- **Late syncs cancel a reminder retroactively.** Rules measure from the event's
  own timestamp, not from when the server heard about it, so a pee logged offline
  at 03:10 and synced at 03:40 resets the clock as if it had arrived on time.
- **Web Push is implemented directly** (`server/webpush.go`): RFC 8291 message
  encryption in the RFC 8188 `aes128gcm` content encoding, authorized with an
  RFC 8292 VAPID token. It is stdlib-only, and checked against the RFC 8291
  test vector in `webpush_test.go`. Subscriptions the push service reports as
  `404`/`410` are deleted.

### Requirements

- **HTTPS.** Push needs a secure context — the same reverse proxy you need for
  `secureCookies`.
- **On iOS the app must be added to the Home Screen** (16.4+). Safari tabs have
  no `PushManager` at all; the app detects this and says so instead of showing a
  toggle that cannot work. iOS also drops subscriptions periodically, so the
  client re-subscribes and re-registers its endpoint on every launch.
- **A VAPID key.** Generated into `vapid.json` next to `puppy.db` on first start,
  or supplied via `-vapid-key` / `PUPPY_VAPID_KEY`. Browsers pin this key at
  subscribe time: replacing it invalidates every existing subscription. If no key
  can be established the server logs a warning and comes up without reminders —
  the `/api/push/*` and `/api/reminders` routes are simply not registered, which
  is also how the client knows to hide the UI.

Settings has a *Send a test notification* button, which is the only practical way
to tell "never subscribed" apart from "subscribed but not delivering" — push
failures are invisible from the browser side, especially on iOS.

## Pedigree lookup

Set your dog's SKK chip or registration number in **Settings** (it rides the
synced profile, next to name and birthday). Once set, a 🌳 button appears that
opens a page rendering that dog's ancestry as a tree.

- SKK has no public API, so the server drives the interactive site the way a
  browser would: it resolves the id to SKK's internal dog id, fetches the
  pedigree page (7 generations per request), and follows each generation's leaves
  deeper. A lookup returns the first generations immediately and keeps crawling in
  the background; the client polls and fills the tree in as ancestors arrive.
- Because a deep crawl is dozens of sequential upstream requests, finished trees
  are cached per dog in the `pedigree_cache` table (pedigrees don't change), and
  the id→dog resolution is memoised, so a dog is only ever crawled once and repeat
  opens hit SKK zero times. The client also mirrors the finished tree in
  `localStorage`, so the page paints instantly and shows the last-known tree even
  offline.
- The lookup is behind auth like the rest of `/api/*`; the first trace of a new
  dog needs to reach SKK, but after that it works from cache (including offline).

## Use it on NixOS

In your system flake:

```nix
{
  inputs.puppy-tracker.url = "path:/path/to/puppy-tracker";

  outputs = { self, nixpkgs, puppy-tracker, ... }: {
    nixosConfigurations.my-host = nixpkgs.lib.nixosSystem {
      system = "x86_64-linux";
      modules = [
        puppy-tracker.nixosModules.default
        {
          services.puppy-tracker = {
            enable = true;
            port = 8080;
            openFirewall = true;
            # Registration secret, kept out of the Nix store. The file holds:
            #   PUPPY_INVITE_CODE=some-shared-secret
            inviteCodeFile = "/run/secrets/puppy-invite-code";
            # Optional. Without it the server generates and keeps its own Web Push
            # key in /var/lib/puppy-tracker. The file holds:
            #   PUPPY_VAPID_KEY=base64url-p256-private-key
            vapidKeyFile = "/run/secrets/puppy-vapid-key";
            # Enable once you terminate TLS in front of the service.
            secureCookies = false;
          };
        }
      ];
    };
  };
}
```

The server runs as a `DynamicUser` systemd unit. Data is stored in a SQLite
database at `/var/lib/puppy-tracker/puppy.db` via `StateDirectory` (with photos
alongside it under `photos/`, and a generated `vapid.json` if no `vapidKeyFile`
is set).

## Notes

- Accounts gate access, but there is no built-in TLS. If exposing publicly,
  terminate TLS with a reverse proxy in front (Caddy / nginx / Tailscale Funnel)
  and set `secureCookies = true`. Without HTTPS, passwords and session cookies
  travel in the clear.