halkyon/deploy/sidecars/README.md

# Sidecars

Services an instance may run beside itself (decided 2026-10-14), each switched on by
naming its compose file after the instance's own in `deploy/.env`:

```sh
COMPOSE_FILE=compose.yaml:compose.jellyfin.yaml
```

A sidecar's compose file adds its service and, if it is reached from outside, mounts
its site from this folder into Caddy's `/etc/caddy/sidecars/`, which the Caddyfile
imports. It joins the `edge` network beside Caddy and the server's public API, never the
internal one where Postgres is, and publishes no port of its own. Its process runs as a
user ID of its own, which no person, no other container and not the CI runner uses.
`redeploy.sh` deploys whatever `.env` switches on, and recreates Caddy when a sidecar's
site changes.

| Sidecar | Compose file | Site |
|---|---|---|
| Jellyfin, which people watch and listen to learn from | `compose.jellyfin.yaml` | `jellyfin.caddy` |
| Jellyfin at home, published through a WireGuard tunnel, in its place | `compose.jellyfin-home.yaml` | `jellyfin.caddy` |
| LanguageTool, which checks spelling and grammar | `compose.languagetool.yaml` | none |
| Halkyon Train's server, with BRouter for its routes | `compose.train.yaml` | `train.caddy` |
| Train's own map tiles, beside Train | `compose.maps.yaml` | `train-maps.caddy` (in Train's site) |
| Photon, which finds places for Train's route planner, beside Train | `compose.search.yaml` | none |

Both together: `COMPOSE_FILE=compose.yaml:compose.jellyfin.yaml:compose.languagetool.yaml`.

## Jellyfin

A Jellyfin server beside the instance, which the instance then suggests to its people
(`JELLYFIN_URL`, set by the compose file): the app's Jellyfin sign-in fills it in and
lists the people it shows on its login screen. Videos stream as stored, nothing is
transcoded. Caddy refuses every change to a user from outside, so an account without a
password (a demo's) stays everyone's; the admin signs in through an SSH tunnel only.

Steps, in order; ✱ marks a step as root.

1. ✱ **Its user and folders** (the operator, on the instance's server, as a user with
   sudo): the user and group ID 61096 are Jellyfin's alone, as the compose file runs it
   (`useradd` warns that the ID lies above the system range, which is meant: it must
   not meet a container's IDs); the media folder is `deploy`'s, so it can copy files
   in, readable by all. The folders are named in full: `deploy`'s home is closed to
   other users (Ubuntu's default), so a user with sudo can't `cd` into it, but `sudo`
   reaches it by its path, which the shell makes of `~deploy` without entering it.

   ```sh
   sudo groupadd --system --gid 61096 jellyfin
   sudo useradd --system --uid 61096 --gid 61096 --no-create-home --shell /usr/sbin/nologin jellyfin
   sudo mkdir -p ~deploy/halkyon-learn/deploy/jellyfin/config \
     ~deploy/halkyon-learn/deploy/jellyfin/cache ~deploy/halkyon-learn/deploy/jellyfin/media
   sudo chown -R jellyfin:jellyfin ~deploy/halkyon-learn/deploy/jellyfin/config \
     ~deploy/halkyon-learn/deploy/jellyfin/cache
   sudo chown -R deploy:deploy ~deploy/halkyon-learn/deploy/jellyfin/media
   ```

2. **DNS** (the operator, at the domain's DNS): the Jellyfin domain, such as
   `jellyfin.example.org`, pointing at the instance's server (`A`, and `AAAA` if it has
   an IPv6 address).

3. **Switch it on** (the operator, on the instance's server, as `deploy`): in
   `deploy/.env`,

   ```sh
   COMPOSE_FILE=compose.yaml:compose.jellyfin.yaml
   JELLYFIN_DOMAIN=jellyfin.example.org
   ```

4. **The media** (the operator, on their own machine): a folder per film, show or
   album, as Jellyfin expects them, into `deploy/jellyfin/media/`, readable by all;
   through `tar` and SSH, which every server has (rsync would have to be installed on
   it too):

   ```sh
   tar -C "<the folder's parent>" -cf - "<folder>" |
     ssh <deploy's ssh alias> 'tar -C halkyon-learn/deploy/jellyfin/media/movies -xf -'
   ```

5. **Start it** (the operator, on the instance's server, as `deploy`):
   `cd ~/halkyon-learn/deploy && docker compose up -d`. Caddy and the server are
   recreated with it (the site, and the instance's suggestion); the instance is away
   for seconds.

6. **Set Jellyfin up through a tunnel** (the operator, on their own machine):
   `ssh -N -L 8096:172.30.80.10:8096 <deploy's ssh alias>`, then http://localhost:8096.
   - The wizard: an admin of a name nobody guesses, its password in a password
     manager; a library of the type Movies (or Shows, Music) on `/media/…`; remote
     connections allowed.
   - Dashboard › Networking › **Known proxies**: `172.30.80.2` and `fd4a:6c6b:7e80::2`
     (Caddy's addresses on `edge`). Jellyfin then sees each visitor's own address, and the tunnel as local.
   - Dashboard › General: **Quick Connect** off, unless people sign in with it.
   - Dashboard › Users › the admin: **Hide this user from login screens**; **Allow
     remote connections to this server** off.
   - For a demo, Dashboard › Users › **+**: an account without a password, shown on
     login screens, seeing the demo's library alone; off: playback and conversion that
     need transcoding, content downloading, remote control of other users, deleting
     media. On its Profile tab, **Failed login tries before user is locked out** `-1`:
     with Jellyfin's default of 3, anyone could switch the demo off with three wrong
     passwords, and without a password the lockout guards nothing. People sign in to it
     by tapping it and leaving the password empty: Jellyfin doesn't say which accounts
     have none.

7. **Check from outside** (anyone, anywhere): `https://<domain>/System/Info/Public`
   answers, `/Users/Public` lists the demo account, a `POST` to `/Users/Password` gets
   403, and the instance's `instance.info()` names the domain as `jellyfinUrl`.

Switching it off: the compose file out of `COMPOSE_FILE`, then `docker compose up -d
--remove-orphans`; `deploy/jellyfin/` stays until removed by hand.

## Jellyfin at home

The Jellyfin sidecar's site, served by the instance, with Jellyfin on the operator's own
machine rather than beside the instance (ADR 0032): `compose.jellyfin-home.yaml` on the
instance's server, `home-jellyfin/` on that machine, a WireGuard tunnel between them
that the machine dials out on. People reach the instance's server alone; nothing names
the home network's address, and its router opens no port. Caddy's rules are the
sidecar's (`jellyfin.caddy`), and the tunnel reaches Jellyfin from Caddy alone and
nothing of the instance's server (`home-jellyfin/server/wg-jellyfin-rules.sh`; nor can
the CI runner reach it, `IPAddressDeny`). The instance suggests it only while it
answers: `JELLYFIN_URL` stays unset, so the instance looks at `jellyfin.` on its own
domain, which `JELLYFIN_DOMAIN` must therefore be. With the machine off, asleep or
offline, Caddy answers 502, and the instance stops suggesting it at its first look
after that, within about an hour. It takes the place of the sidecar, never runs
beside it.

The tunnel's addresses: 10.61.96.1 the instance's end, 10.61.96.2 the home end, UDP
51820 on the instance's server. Steps, in order; ✱ marks a step as root, 🏠 one on the
home machine.

1. 🏠 **The home end's key** (the operator, in `deploy/home-jellyfin/` of their
   checkout): WireGuard's own tool from the image, so nothing is installed; the
   private key goes into the git-ignored config, which only its owner reads, the
   public key is printed for step 3.

   ```sh
   cd deploy/home-jellyfin
   image=$(sed -n 's/^ *image: \(lscr.io\/linuxserver\/wireguard@.*\)/\1/p' compose.yaml)
   (umask 077 && mkdir -p wireguard/wg_confs &&
     docker run --rm --entrypoint wg "$image" genkey > wireguard/home.key)
   docker run --rm -i --entrypoint wg "$image" pubkey < wireguard/home.key
   ```

2. ✱ **The instance's end** (the operator, on the instance's server, as a user with
   sudo): WireGuard's tools, the key, and the files from the checkout. The key is in
   the password manager too: a new server needs it again, or the home end a new
   public key.

   ```sh
   sudo apt-get install wireguard-tools
   (umask 077 && wg genkey | sudo tee /etc/wireguard/wg-jellyfin.key | wg pubkey)
   from=~deploy/halkyon-learn/deploy/home-jellyfin/server
   sudo install -m 0600 "$from/wg-jellyfin.conf" /etc/wireguard/wg-jellyfin.conf
   sudo install -m 0700 "$from/wg-jellyfin-rules.sh" /etc/wireguard/wg-jellyfin-rules.sh
   sudo install -D -m 0644 "$from/wg-quick-after-docker.conf" \
     /etc/systemd/system/wg-quick@wg-jellyfin.service.d/after-docker.conf
   ```

   The printed public key is the instance's, for step 4.

3. ✱ **Introduce the home end, open the port, start it** (as in step 2):
   `sudo nano /etc/wireguard/wg-jellyfin.conf`, the home end's public key from step 1
   in place of `<the home end's public key>`; where the host has a firewall, UDP 51820
   opened (`sudo ufw allow 51820/udp`); the runner's unit as the checkout has it, so it
   can't reach the tunnel; then the tunnel.

   ```sh
   sudo install -m 0644 ~deploy/halkyon-learn/deploy/runner/forgejo-runner.service \
     /etc/systemd/system/forgejo-runner.service
   sudo systemctl daemon-reload
   sudo systemctl restart forgejo-runner
   sudo systemctl enable --now wg-quick@wg-jellyfin
   ```

4. 🏠 **The home end's config, folders and media** (in `deploy/home-jellyfin/`):
   `wg0.conf.example` as `wireguard/wg_confs/wg0.conf`, with the key from step 1 and
   the instance's public key from step 2; Jellyfin's folders owned by its user ID
   (61096, as on the instance's server: no person and no other container uses it); the
   media folder named in `.env` and readable by all, a folder per film, show or album
   as Jellyfin expects them.

   ```sh
   (umask 077 && sed "s|<the home end's private key>|$(cat wireguard/home.key)|" \
     wg0.conf.example > wireguard/wg_confs/wg0.conf)
   nano wireguard/wg_confs/wg0.conf   # the instance's public key
   sudo install -d -o 61096 -g 61096 jellyfin/config jellyfin/cache
   echo "JELLYFIN_MEDIA=$HOME/Videos" > .env
   chmod -R o+rX "$HOME/Videos"
   docker compose up -d
   ```

   `sudo wg show wg-jellyfin` on the instance's server now shows a handshake.

   Whenever `wg0.conf` fails to come up (a placeholder or a wrong key, an endpoint that
   doesn't resolve: why it names the server by its address, ADR 0032, D11, and a new
   address of the server's goes there too), the image's client mode deletes the namespace's default route
   (`ip route del default` in its `svc-wireguard/run`, ls126), and Jellyfin with it loses
   the internet: TMDb's metadata and images fail with *Network is unreachable*, and a
   library scanned then gets frames of its videos for posters. A tunnel that comes up
   keeps the route, even before its first handshake. Fixing `wg0.conf` alone isn't enough:
   `docker compose up -d --force-recreate`, both containers, since Jellyfin lives in
   WireGuard's namespace and a restart of WireGuard alone leaves it in the old one. The
   folder `wireguard/` is the container's user's (61096) from its first start on, so
   `wg0.conf` is edited with sudo.

5. 🏠 **Set Jellyfin up** at http://localhost:8096 on the home machine, before the site is
   switched on: the sidecar's step 6, but for its networking, which is the tunnel's:
   - Dashboard › Networking › **LAN networks**: `172.29.96.1/32` alone, the gateway
     through which this machine's loopback reaches it (`home-jellyfin/compose.yaml`).
     Left empty, Jellyfin counts every private range as local, the tunnel's too, and an
     admin without remote access could sign in through it.
   - Dashboard › Networking › **Known proxies**: `10.61.96.1`, as which Docker on the
     instance's server masquerades Caddy's requests. Jellyfin then sees each visitor's
     own address.
   - **Allow remote connections** on; the admin hidden from login screens, its remote
     connections off; Quick Connect, the demo account and its lockout as in the
     sidecar's step 6. Every account but those meant for everyone hidden from login
     screens: the instance suggests this server to all its people.

6. **Switch the site on** (the operator, on the instance's server, as `deploy`): in
   `deploy/.env`, `JELLYFIN_URL` left empty,

   ```sh
   COMPOSE_FILE=compose.yaml:compose.jellyfin-home.yaml
   JELLYFIN_DOMAIN=jellyfin.halkyon.eu
   ```

   then `cd ~/halkyon-learn/deploy && docker compose up -d`. Caddy and the server are
   recreated with it; the instance is away for seconds.

7. **Check from outside** (anyone, anywhere): the sidecar's step 7; and on the instance's
   server, as `deploy`, that nothing reaches the home end but Caddy, and the home end
   nothing of the server:

   ```sh
   # Refused (on edge, but not at Caddy's address): times out.
   docker run --rm --network halkyon-learn-edge caddy:2 \
     wget -T 5 -q -O /dev/null http://10.61.96.2:8096/System/Info/Public && echo LEAK
   # Caddy itself gets through.
   docker compose exec -T caddy wget -T 5 -q -O - http://10.61.96.2:8096/System/Info/Public
   ```

   and on the home machine, in `deploy/home-jellyfin/`:

   ```sh
   # Refused (nothing new from the tunnel reaches the host): gives up after 5 s. The
   # image's nc is OpenBSD's, whose own -w ends the try.
   docker compose exec -T wireguard nc -w 5 -z 10.61.96.1 22 && echo LEAK
   ```

   Stopping the home end (`docker compose stop`) makes the site answer 502, and
   `instance.info()` names no `jellyfinUrl` after the instance's next look, within about
   an hour.

Switching it off: on the instance's server the compose file out of `COMPOSE_FILE`, `docker
compose up -d --remove-orphans`, and ✱ `sudo systemctl disable --now
wg-quick@wg-jellyfin`, the port closed again; at home `docker compose down`, with
`deploy/home-jellyfin/jellyfin/` kept until removed by hand.

## LanguageTool

LanguageTool's open-source server beside the instance (ADR 0024), which checks spelling
and grammar for Halkyon Learn: Check in the assistant's field, and the underlines while
writing (the app's Settings › Writing). Only the instance's server reaches it
(`LANGUAGETOOL_URL`, set by the compose file): LanguageTool has no sign-in of its own,
so it gets no site, and the server checks who asks and how often (signed in, 60 checks a
minute each). Nothing written is kept or logged, LanguageTool's own log included: Docker
keeps none for it. It checks without n-gram data (LanguageTool's statistical checks of
words like their and there, tens of GB per language), in about 1.5 GB of memory.

Steps, in order; ✱ marks a step as root.

1. ✱ **Its user and folder** (the operator, on the instance's server, as a user with
   sudo): the user and group ID 61081 are LanguageTool's alone, as the compose file runs
   it (`useradd` warns that the ID lies above the system range, which is meant: it must
   not meet a container's IDs). The folder is named in full, as Jellyfin's.

   ```sh
   sudo groupadd --system --gid 61081 languagetool
   sudo useradd --system --uid 61081 --gid 61081 --no-create-home --shell /usr/sbin/nologin languagetool
   sudo mkdir -p ~deploy/halkyon-learn/deploy/languagetool/fasttext
   sudo chown -R languagetool:languagetool ~deploy/halkyon-learn/deploy/languagetool
   ```

2. **Switch it on** (the operator, on the instance's server, as `deploy`): in
   `deploy/.env`, `COMPOSE_FILE=compose.yaml:compose.languagetool.yaml`.

3. **Start it** (the operator, on the instance's server, as `deploy`):
   `cd ~/halkyon-learn/deploy && docker compose up -d`. The server is recreated with it
   (its `LANGUAGETOOL_URL`); the instance is away for seconds. At its first start
   LanguageTool fetches fastText's model of languages (`lid.176.bin`, 126 MB, from
   Facebook's `dl.fbaipublicfiles.com`) into `deploy/languagetool/fasttext/`, which it
   uses to tell which language a note is in; checks start once it is there.

4. **Check it**: on the instance's server, as `deploy`, `docker compose ps languagetool`
   says `healthy` after a minute; then, from anywhere, the instance names the languages
   it checks:

   ```sh
   curl -s -X POST -H 'Content-Type: application/json' -d '{}' https://<domain>/instance/info
   ```

   answers with `grammarLanguages`, `de-DE` and `en-US` among them.

Switching it off: the compose file out of `COMPOSE_FILE`, then `docker compose up -d
--remove-orphans`; the app's Writing settings then say the instance checks nothing, and
`deploy/languagetool/` stays until removed by hand.

## Train

Halkyon Train's server beside the instance (Halkyon Train's ADR 0030, D46), at its own
domain, a member of the register Learn's server holds (ADR 0016), its database
`halkyon_train` in the instance's PostgreSQL, and BRouter planning its routes (D11, D44)
from the segments of the operator's region (Europe's by default, about 3.3 GB on
28 Oct 2026), refreshed monthly. `redeploy.sh` builds, dumps, migrates and waits for it
as for Learn's server once `.env` names its compose file, and `backup.sh` backs its
database up beside Learn's. The map stays OpenFreeMap's until the instance serves its
own tiles (D22, D23).

Steps, in order; ✱ marks a step as root.

1. **Room on the disk** (the operator, on the instance's server): `df -h ~deploy`.
   BRouter's segments take about 3.3 GB for Europe, and a refresh writes each one beside
   the old before replacing it; Train's dumps go to `~/backups` with Learn's.
2. **Its database** (the operator, on the instance's server, as `deploy`, in
   `~/halkyon-learn/deploy`): a role of Train's own, with the password that goes into
   `.env` as `TRAIN_POSTGRES_PASSWORD` (`openssl rand -hex 32`), and its database:

   ```sh
   docker compose exec -T postgres psql -U halkyon_learn -d postgres -v ON_ERROR_STOP=1 \
     -v password="<TRAIN_POSTGRES_PASSWORD>" <<'SQL'
   CREATE ROLE halkyon_train LOGIN PASSWORD :'password';
   CREATE DATABASE halkyon_train OWNER halkyon_train;
   REVOKE CONNECT ON DATABASE halkyon_train FROM PUBLIC;
   REVOKE CONNECT ON DATABASE halkyon_learn FROM PUBLIC;
   SQL
   ```

   `CREATE ROLE` makes the role Train's server signs in as, which may log in and is
   nothing more (no superuser, no creating roles or databases); `CREATE DATABASE … OWNER`
   gives it its own database, where its migrations make the tables. The two `REVOKE`s
   take away the right every role has by default to connect to a database: only its
   owner (and the instance's superuser, `halkyon_learn`, which Learn's server signs in
   as) connects to `halkyon_train`, and `halkyon_train` no longer connects to Learn's.
   Nothing of Learn's changes. (Learn's server, signing in as the superuser, could still
   read Train's; giving it a role of its own is a step of its own.)
3. ✱ **BRouter's user and folder** (the operator, as a user with sudo): the user and
   group ID 61177 are BRouter's alone, as the compose file runs it; the folder its
   segments go to is theirs.

   ```sh
   sudo groupadd --system --gid 61177 brouter
   sudo useradd --system --uid 61177 --gid 61177 --no-create-home --shell /usr/sbin/nologin brouter
   sudo mkdir -p ~deploy/halkyon-learn/deploy/brouter/segments
   sudo chown -R brouter:brouter ~deploy/halkyon-learn/deploy/brouter/segments
   ```

4. **DNS** (the operator, at the domain's DNS): Train's domain, `train.halkyon.eu`,
   pointing at the instance's server (`A`, and `AAAA` if it has an IPv6 address).
5. **Switch it on** (the operator, as `deploy`): in `deploy/.env`, `compose.train.yaml`
   added to `COMPOSE_FILE`, and Train's settings (`.env.example`, *With Halkyon Train*):

   ```sh
   COMPOSE_FILE=compose.yaml:compose.languagetool.yaml:compose.train.yaml
   TRAIN_DOMAIN=train.halkyon.eu
   TRAIN_POSTGRES_PASSWORD=<as in step 2>
   TRAIN_SESSION_KEY_HASH_PEPPER=<openssl rand -hex 32>
   ```

   `INSTANCE_SECRET` must be set (Learn's register takes no member without it). Where
   `INSTANCE_MEMBERS` is set, add `http://train-server:8082/` to it; unset, the compose
   file names Train.
6. **Start it** (the operator, as `deploy`): `docker compose up -d --build`. Caddy and
   Learn's server are recreated with it (Train's site, and the register's member); the
   instance is away for seconds. Train's server applies its migrations as it starts.
   The segments come in the first hour or so (`docker compose logs brouter-segments`
   says when); until then BRouter finds no route. Where brouter.de doesn't answer or
   the folder isn't BRouter's, the log says so and it tries again an hour later.
7. **Check it**: `https://train.halkyon.eu/` answers; in Train, signed in with a Learn
   account, a route plans between two points of the region.

Workouts' weather (Halkyon Train's ADR 0036) comes from Open-Meteo through Train's
server, on by default, the free API from the server's address (non-commercial use alone,
10,000 requests a day). `WEATHER=off` in `deploy/.env` stops it; an instance with ads or
subscriptions needs one of Open-Meteo's paid plans, its key in `OPEN_METEO_API_KEY`.


## Train's map

The instance's own map for Halkyon Train (Halkyon Train's ADR 0030, D50), beside Train
(`compose.train.yaml` first): Protomaps' basemap cut for Europe from Protomaps' daily
build (the planet's was 139 GB on 29 Oct 2026; Europe's cut is estimated at some 45 GB,
to be measured), served by `pmtiles serve` and drawn by Train's own style, which Train's
server serves. Only people signed in to Train get tiles: `map.source` hands the app a
key, which Caddy has Train's server check before each tile. No tile is logged. The cut
is renewed monthly where Protomaps has a newer build, written beside the old one and
moved over it. Off, Train's maps are OpenFreeMap's (or `MAP_STYLE_URL`'s) again.

Steps, in order; ✱ marks a step as root.

1. **Room on the disk** (the operator, on the instance's server): `df -h ~deploy`.
   A cut takes some 45 GB, and twice that while a new one is cut beside it.
2. ✱ **The tiles' user and folder** (the operator, as a user with sudo): the user and
   group ID 61178 are the tiles' alone, as the compose file runs them.

   ```sh
   sudo groupadd --system --gid 61178 maptiles
   sudo useradd --system --uid 61178 --gid 61178 --no-create-home --shell /usr/sbin/nologin maptiles
   sudo mkdir -p ~deploy/halkyon-learn/deploy/maps/tiles
   sudo chown -R maptiles:maptiles ~deploy/halkyon-learn/deploy/maps/tiles
   ```

3. **Switch it on** (the operator, as `deploy`): in `deploy/.env`, `compose.maps.yaml`
   added to `COMPOSE_FILE` after Train's, and the secret the keys are made with:

   ```sh
   COMPOSE_FILE=compose.yaml:compose.languagetool.yaml:compose.train.yaml:compose.maps.yaml
   MAP_KEY_SECRET=<openssl rand -hex 32>
   ```

   `MAP_STYLE_URL` is then not used. Changing `MAP_KEY_SECRET` refuses every key handed
   out before, until the apps ask for the source again (as each opens).
4. **Start it** (the operator, as `deploy`): `docker compose up -d --build`. Caddy and
   Train's server are recreated with it. The first cut takes hours
   (`docker compose logs -f map-tiles-refresh` says when it is done); until then Train's
   maps stay empty. A failed cut is tried again an hour later.
5. **Check it**: in Train, signed in, a workout's map is drawn and names the instance as
   its source; `curl -I https://train.halkyon.eu/tiles/europe/0/0/0.mvt` is refused
   with 401. Not yet tried on the server (29 Oct 2026): whether `pmtiles serve` reads a
   new cut moved over the old one without a restart (else `docker compose restart
   map-tiles` after each cut), and Caddy's `import` of an empty folder where the map is
   off.

## Train's search

Photon beside Train (Halkyon Train's ADR 0030, D50; `compose.train.yaml` first), which
finds the places Train's route planner searches for, on GraphHopper's weekly database
of Europe (a 32.6 GB download on 29 Oct 2026, more unpacked, to be measured; Germany's
alone 7.5 GB, `PHOTON_REGION=europe/germany`). Only Train's server reaches it, for
people signed in; nothing searched is kept or logged. The database is renewed monthly:
downloaded and unpacked beside the one in use, checked against its MD5, then put in
its place, Photon away for a minute or two.

Steps, in order; ✱ marks a step as root.

1. **Room and memory** (the operator, on the instance's server): `df -h ~deploy` and
   `free -h`. The database takes tens of GB, twice that while a new one comes; Photon
   wants `PHOTON_HEAP` (6g by default) and room beside it (`PHOTON_MEMORY`, 8g).
2. ✱ **Photon's user and folder** (the operator, as a user with sudo): the user and
   group ID 61179 are Photon's alone, as the compose file runs it.

   ```sh
   sudo groupadd --system --gid 61179 photon
   sudo useradd --system --uid 61179 --gid 61179 --no-create-home --shell /usr/sbin/nologin photon
   sudo mkdir -p ~deploy/halkyon-learn/deploy/photon/data
   sudo chown -R photon:photon ~deploy/halkyon-learn/deploy/photon/data
   ```

3. **Switch it on** (the operator, as `deploy`): in `deploy/.env`, `compose.search.yaml`
   added to `COMPOSE_FILE` after Train's:

   ```sh
   COMPOSE_FILE=compose.yaml:compose.languagetool.yaml:compose.train.yaml:compose.maps.yaml:compose.search.yaml
   ```

4. **Start it** (the operator, as `deploy`): `docker compose up -d --build`. Train's
   server is recreated with it. The first database takes an hour or more
   (`docker compose logs -f photon-data` says when it is ready); Photon starts on it
   within a minute, and until then Train's planner offers no search.
5. **Check it**: in Train, signed in, the planner's search finds a town of the region.
   Not yet tried on the server (29 Oct 2026): the database's size unpacked, Photon's
   memory on Europe's, and that 1.3.0 reads GraphHopper's `1.0` build.