lunch-notifier/CLAUDE.md

# Lunch Notifier

Single-file Flask app that scrapes the Mariposa (Kyrene school district)
lunch menu and pushes each day to a Home Assistant calendar as an all-day
event. Runs as one Docker container on the user's Unraid home server.

**2026-09-03: pivoted from SMS to calendar-only.** Originally texted phone
numbers via email-to-SMS carrier gateways; that whole path (Gmail SMTP,
phone number management, carrier gateway settings) was deleted at the
user's request in favor of Home Assistant calendar events exclusively. If
you find references to texting/SMS/phone numbers anywhere (docs, old
commits), they describe the pre-2026-09-03 app, not the current one.

**2026-09-08: pivoted from Nutrislice to a Claude-read PDF, and from
weekly to monthly pulls.** The district discontinued Nutrislice (the old
scraper hit `kyrene.api.nutrislice.com`, a clean per-day JSON API) in
favor of publishing one monthly menu PDF per grade band on
`kyrene.org/family-student-resources/food-services/school-menus`. See
"Menu source" and "Monthly pull scheduler" below for how that's handled
now. Anything describing a weekly Nutrislice pull describes the
pre-2026-09-08 app.

## Architecture

- **`app.py`** — everything: Flask web UI/API, `schedule`-library job
  scheduling, and the Home Assistant webhook push. The HTML/CSS/JS frontend
  is embedded as a Python string constant (`HTML`) rather than a separate
  template file — no build step, no template engine.
- **`scraper.py`** — fetches the currently-published K-5 lunch menu PDF and
  asks Claude (Anthropic API, `claude-opus-5`, structured JSON output via
  `output_config.format`) to read it into `[{"date": ..., "items": [...]}, ...]`
  for every day of whatever month is published. Days with no posted menu
  (weekends, breaks, holidays) come back with an empty `items` list rather
  than being omitted — same contract the old Nutrislice-based scraper used,
  so `app.py`'s push logic didn't need to change.
- **Menu source (Nutrislice → PDF, 2026-09-08):** the district's page links
  a *stable* Finalsite "resource manager" permalink
  (`kyrene.org/fs/resource-manager/view/32bb3f8b-5a50-4b38-ad68-a5d3108de571`
  for the K-5/"Standard" menu — there's a different UUID per grade band/
  campus on that page) that 302-redirects to whichever monthly PDF is
  currently live (`scraper.MENU_PAGE_URL`). **Don't hardcode the PDF URL
  itself** — it's a Finalsite CDN path that changes every month (verified:
  the September 2026 PDF and August's page both linked through the same
  permalink); only the permalink UUID is stable, and only changes if the
  district restructures that page entirely. The PDF itself is a single-page
  graphic calendar grid (mascots, colored cells, day numbers laid out
  visually) — not text-extractable in any reliable way (raw text extraction
  jumbles day numbers away from their menu items), which is why this reads
  it via Claude vision/document understanding rather than `pdfplumber`/
  regex. Cost is negligible: one ~370KB single-page PDF per month, well
  under a cent of Opus 5 tokens.
- **`ANTHROPIC_API_KEY`** is now required (server `.env`, alongside
  `HA_WEBHOOK_URL`) — `scraper.get_month_menu()` fails without it, same
  fail-soft-into-`_status`-error behavior as a Home Assistant push failure.
  `GET /api/integrations` reports both keys' configured/not-configured
  state.
- Settings (calendar pull schedule) persist to `/data/settings.json`, not
  environment variables — env vars (`CALENDAR_RUN_TIME`,
  `CALENDAR_RUN_DAY_OF_MONTH`) are defaults used only the very first time,
  before a settings file exists. `load_settings()` merges the file's
  contents *over* `_DEFAULTS` rather than returning the file verbatim —
  this matters because the server's settings.json predates several schema
  changes (SMS era, then the weekly-schedule era) and is missing whatever
  keys the current schema expects; without the merge, the scheduler would
  silently come up with zero jobs after every deploy until someone happened
  to re-save the Schedule card in the UI.
- **No vacation/no-push date ranges (removed 2026-09-03, briefly existed).**
  A day with no posted menu already produces an empty `items` list from
  `get_month_menu()`, and `_push_calendar()` already skips pushing when
  `items` is empty — that's sufficient for "don't create an event for a day
  with nothing happening." A separate vacation-range setting to *additionally*
  suppress days the school did post a menu for was judged unnecessary
  complexity and deleted at the user's request; don't re-add it without
  being asked.
- **Home Assistant integration (built 2026-09-02/03, calendar-only pivot
  2026-09-03):** `HA_WEBHOOK_URL` is required — it's the only way this app
  does anything externally-visible. On the HA side
  (`automation.lunch_menu_family_calendar`, `http://192.168.86.207:8123`),
  a `local_only` webhook trigger runs `calendar.create_event` against
  `calendar.family_2` (the "Family" Google Calendar entity — there are two
  entities both named "Family" in this HA instance, `calendar.family` and
  `calendar.family_2`; this integration deliberately targets `_2` per user
  choice). `calendar.family_2` only supports `create_event`, not
  `delete_event` — a bad push can't be cleaned up via the HA API, only
  manually in Google Calendar. No duplicate-prevention: re-pushing the
  same date's menu creates a second event — this is why the default
  schedule is a single day/month, not daily (see below).
  **`mode: parallel`, not the default `single` (fixed 2026-09-03 after a
  real bug):** a push fires one webhook POST per school day back-to-back
  with no delay. Under `mode: single`, HA rejects any webhook trigger that
  arrives while a previous run is still executing (`calendar.create_event`
  itself takes ~600-700ms) — most days in a batch were silently dropped as
  `failed_single`, so only one day's event ever actually got created per
  push. Found via `ha_get_automation_traces` after the user reported a
  manual test push not showing up on the calendar. Each invocation is
  independent (one `calendar.create_event` call, no shared state), so
  `parallel` is the correct mode per HA's own guidance ("independent
  per-entity actions → parallel"), not just a workaround. The webhook id is
  a random token that lives only in the server's `.env` (`HA_WEBHOOK_URL`)
  and in the automation's trigger config — never committed here. If the
  calendar push stops working, check both ends: `GET /api/integrations`
  on this app reports whether `HA_WEBHOOK_URL` is set at all, and
  `ha_get_automation_traces` on `automation.lunch_menu_family_calendar`
  shows whether HA is even receiving the webhook.
- **Monthly pull scheduler (weekly → monthly, 2026-09-08):** calendar pulls
  run on their own schedule (`calendar_run_time`/`calendar_run_day_of_month`
  in settings, `schedule` jobs tagged `"calendar"`), completely decoupled
  from anything else — the tag/isolation pattern
  (`schedule.clear("calendar")` instead of a bare `schedule.clear()`) is
  deliberate so a future second schedule wouldn't stomp this one. The
  `schedule` library has no native day-of-month trigger, so
  `apply_calendar_schedule()` registers a *daily* job
  (`_maybe_push_month`) that no-ops unless `datetime.now().day` matches
  the configured day — `push_month_to_calendar()` only actually runs on
  that one day per month. `next_calendar_run()` computes the real next
  firing directly from settings (handling month/year rollover and clamping
  the configured day to short months, e.g. day 30 in February → the 28th)
  for the UI's "Next Scheduled" field, since the underlying `schedule` job's
  own `next_run` is meaningless here (it's always just "tomorrow"). Default
  is day 25 at 18:00 — picked because the district tends to post next
  month's PDF a few days before month-end, so pulling then usually catches
  the *upcoming* month rather than the one ending. Whatever month is live
  at the permalink when the job fires is what gets pushed — there's no
  logic forcing "next month specifically."
- **Manual actions:** "Push This Month to Calendar" (`POST /api/push-calendar`
  → `push_month_to_calendar()`, same function the schedule calls) and
  "Preview This Month's Menu" (`GET /api/preview` → `get_month_menu()`
  directly, pure dry run, no Home Assistant push, no side effects at all —
  use this to check the scraper before trusting a scheduled push).

## Local development

```bash
pip install -r requirements.txt
python3 app.py
```

Needs `ANTHROPIC_API_KEY` set for anything menu-related to work at all
(including Preview, since that calls the same Claude-backed scraper) and
`HA_WEBHOOK_URL` set to actually push anywhere — without the latter the
app still runs and previews fine, `GET /api/integrations` just reports
`ha_webhook_configured: false` and pushes resolve to `"not configured"`.
Visit `http://localhost:2323`.

## Deployment

**Target:** Unraid server, container path `/mnt/user/appdata/lunch-notifier`,
built from `Dockerfile`/`docker-compose.yml` in this repo. Container name
`Lunch-Notifier`, port `2323`. No reverse proxy — accessed directly via
`http://<server-ip>:2323` on the LAN.

**Deploy mechanism: Forgejo push-to-deploy**, same pattern as the other
three projects on this box (`card-grader`, `n64-tracker`, `arr-summary`) —
see `card-grader`'s `CLAUDE.md` for the full pipeline mechanics (runner
config, HTTPS-not-SSH git auth, known tradeoffs). In short: push to `main`
on the `forgejo` remote (`https://git.hippofam.com/ninja_hippo/lunch-notifier.git`)
triggers `.forgejo/workflows/deploy.yml`, which syncs into
`/mnt/user/appdata/lunch-notifier` (excluding `.git`, `.env`, `data/`) and
runs `docker compose up -d --build`. **`.env` changes need an explicit
`docker compose up -d` on the server afterward** — env vars only load at
container start, a deploy that doesn't change code won't pick up a
manually-edited `.env` on its own.

**Persistent state, never touched by deploys:** `data/settings.json` (the
volume-mounted `/data` inside the container) holds the calendar pull
schedule — this is live user data, not deploy-managed. `.env` holds
`HA_WEBHOOK_URL` and `ANTHROPIC_API_KEY` and is server-only, gitignored,
never committed.

**Container timezone (fixed 2026-09-08):** the container previously had no
timezone configured — `python:3.12-slim` defaults to UTC and the
`/etc/localtime` bind mount in `docker-compose.yml` was commented out — so
`calendar_run_time` (and the `datetime.now()` call in `scraper.py` that
picks which week to fetch) were evaluated against UTC, not the user's
local clock. This happened to be masked because Arizona (no DST) and
Pacific Daylight Time are both UTC-7 right now, but would have silently
drifted the actual fire time by an hour once PDT becomes PST in November.
Fixed by uncommenting the `/etc/localtime:/etc/localtime:ro` bind mount
and installing `tzdata` in the `Dockerfile` — without `tzdata`, the
host's `/etc/localtime` (typically a symlink into `/usr/share/zoneinfo/…`)
would dangle inside the container since that directory wouldn't exist,
silently falling back to UTC again. Found while investigating why the
first automatic Sunday calendar push after the 2026-09-03 pivot appeared
not to run — that specific miss turned out to be unrelated (the school's
Nutrislice menu was published empty for the whole week, so `_push_calendar`
correctly skipped every day), but this latent TZ bug surfaced during the
investigation and was worth fixing regardless.

**Migration note (2026-08-25):** this container originally used a Docker
**named volume** (`lunch-notifier_lunch-settings`) instead of a bind mount.
That meant Unraid's Appdata Backup plugin — which only knows about
`/mnt/user/appdata` and `/mnt/cache/appdata` as backup sources — silently
skipped it entirely; the settings data was never included in any backup.
Migrated to a bind mount (`./data:/data`) with the existing settings copied
over first and verified intact before the old named volume was deleted.