menu/ansible/readme.md

# The menu deployment

Lives here, beside the app it deploys. One run deploys a Site: it starts the
app's container as a **rootless Podman Quadlet unit** on `emily` and points the
Site's web address at it through the shared **Caddy** unit that
`ansible-playbooks/caddy` owns. There is no older system to keep serving, so
there is no staged rollout and nothing to retire — deploying the image built from
an earlier commit is the rollback.

This is the same shape as the sibling services in `ansible-playbooks` (gitea,
nextcloud, tandoor, bitwarden, valkey, postgres, trillium): rootless Podman, one
Quadlet `.container` / `.network` per service in `~/.config/containers/systemd/`,
state under `~/podman-data/`, and Caddy reaching each app by name on a Podman
network. Nothing here talks to Docker.

## The pieces

```
ansible/
├── hosts                                  # one entry per Site, all on emily
├── playbook.yml                           # connects as tanshu, no become
├── vars/{default,hopsngrains,greatbear,hngmohali}.yml
└── roles/
    ├── register/   # writes menu-<site>.network and the Caddy drop-in
    ├── menu-app/   # the Quadlet unit, start + health + verify
    └── caddy/      # appends the site block, reloads Caddy
```

Two things are worth internalising, because they are the whole design:

**Caddy knows nothing about its upstreams.** A service registers _itself_ by
writing two files into the Quadlet directory: its own `<name>.network` unit and a
drop-in beside `caddy.container` that pulls that network into Caddy. This play
does that in `pre_tasks`, before the app unit starts, because Quadlet fails unit
generation when a unit references a `.network` file that does not exist yet.

**The container's uid is mapped onto `tanshu`.** `UserNS=keep-id:uid=1000,gid=1000`
maps the image's own `node` account onto the user that owns the units, so every
file the app writes is owned by `tanshu` on the host: plain `tar`/`rsync`
backups, no subuids, no `podman unshare`. `scripts/provision-site-env.sh` uses
the same mapping for the same reason.

**A Quadlet `HealthCmd` must not contain quotes.** Quadlet's parser strips a
trailing double quote, so a check written as `node -e "fetch(…)"` reaches podman
as an unterminated shell string: every run fails with `/bin/sh: syntax error:
unterminated quoted string`, and the container serves while being permanently
unhealthy. That is why the check is `wget -q -O /dev/null …` in
`vars/default.yml` — quote-free, and the same reason the sibling apps use a bare
`curl`. Nothing about it is visible in `podman ps` beyond the word `unhealthy`;
`podman inspect <container> --format '{{json .State.Health.Log}}'` shows it.

## What lives where

| Thing                                  | Where                                                                       |
| -------------------------------------- | --------------------------------------------------------------------------- |
| Name, tagline, icon, categories, menus | on the server under `~/podman-data/menu-<site>/data` — the admin edits them |
| Admin password hash, session secret    | `~/podman-data/menu-<site>/site.env` (untracked, mode 600)                  |
| Theme colours                          | `config/<site>.site.json` in the repo, baked into the image                 |
| The unit that runs                     | `~/.config/containers/systemd/menu-<site>.container`                        |
| The network and Caddy drop-in          | `menu-<site>.network`, `caddy.container.d/50-menu-<site>.conf`              |
| Which image tag runs                   | `-e image_tag=<commit>`                                                     |
| Which Site is deployed                 | the inventory entry in `hosts`                                              |
| The web address it serves              | `host:` in `vars/<site>.yml`                                                |

`<site>` is the inventory name (`hopsngrains`, `greatbear`, `hngmohali`), and it
is the single slug for everything: the container, its Podman network and its data
directory are all `menu-<site>`. That is how the sibling `hops` and `greatbear`
apps are deployed, and it keeps `podman logs menu-<site>` and
`systemctl --user status menu-<site>` the same two commands across all three.

## What the host needs

`emily` is already set up — it runs Caddy, Postgres, Tandoor and Valkey the same
way. If you are standing up a fresh host, run `ansible-playbooks/caddy` first: it
installs podman, enables lingering, lowers the privileged-port floor so rootless
Caddy can bind 80/443, creates the Quadlet directory, and starts the Caddy unit.
Those prerequisites deliberately are **not** duplicated here, so a host that has
not run that playbook will fail at the first `systemctl --user` — see
`ansible-playbooks/caddy/readme.md`.

The registry is private, but nothing here needs a manual login: the play logs in
from `registry_username` / `registry_password` in `vars/default.yml`, pulls the
image, and logs out again, so the credential is never left in the target user's
auth file. The unit keeps podman's default `Pull=missing` on purpose, so a
restart or a reboot never depends on the registry being reachable.

`ssh` on some machines complains about
`/etc/ssh/ssh_config.d/20-systemd-ssh-proxy.conf` being owned by `nobody`; that is
a local file-permission bug whose fix is `sudo chown -R root:root
/etc/ssh/ssh_config.d`, not something this playbook works around.

## Deploying

From the repo root:

```sh
./deploy.sh                  # push HEAD, build <commit>, deploy every Site
./deploy.sh --no-push        # skip the push (you already pushed)
./deploy.sh --force          # skip the guards
./deploy.sh --checks         # lint, format, typecheck and test first
SITE=greatbear ./deploy.sh   # one Site instead of every Site
```

`deploy.sh` requires a clean tracked tree and `HEAD == origin/main`, tags the
build with the commit it came from, and refuses a dirty tree for a reason: the
image is built from `git archive HEAD`, so what runs on the host is exactly what
git knows about. It runs `make build-production COMMIT=<commit>`, which builds
`linux/arm64` — the architecture `emily` runs — and pushes both the commit tag
and `latest` to `registry.tanshu.com/menu-admin`.

By hand:

```sh
git push
make build-production COMMIT=$(git rev-parse --short HEAD)
cd ansible
ansible-playbook -u tanshu --inventory=hosts playbook.yml -e image_tag=$(git rev-parse --short HEAD)
```

Every Site runs on the same box, so `hosts` has one entry per Site pointing at
the same `ansible_host` with its own `var_file`. The play carries no `become`: the
Quadlet units belong to `ansible_user` (`tanshu`), and running as that user is
what gives us a real PAM session, which is what makes `systemctl --user` work.
Connecting as anyone else would write units into the wrong home.

Adding a Site is one line in `hosts`, a copy of `vars/hopsngrains.yml`, and a
`config/<site>.site.json`. The image is shared and carries every Site's config,
so a new Site needs a new build — which the commit tag gives it.

## Migrating a Site's data from `monoco`

The menus still serve from Docker containers on `monoco`. A Site moves across
with its data, before it is ever pointed at from DNS. Run this from a machine
that can ssh to both hosts — the `monoco` and `emily` aliases this repo's own
tooling uses:

```sh
scripts/copy-site-data.sh hopsngrains
scripts/copy-site-data.sh greatbear
scripts/copy-site-data.sh hngmohali
```

For one Site it stops the old container, streams `/var/lib/menu.<domain>` across,
starts the old container again so `monoco` keeps serving until DNS moves, and
then proves the copy: the SHA-256 of the whole tree is compared against the
source, and every file is checked to belong to `tanshu`. `--live` skips the stop;
`MONOCO=`, `EMILY=`, `SRC_ROOT=` and `DEST_ROOT=` override the defaults.

Two things make the copy look different from a plain `rsync`:

- **`sudo` on monoco.** `/var/lib/menu.<domain>` belongs to uid 1000 (the
  `ubuntu` account) and the ssh user there is 1001, so it is not readable without
  it. monoco's `tanshu` has passwordless sudo.
- **The workstation relays it.** monoco cannot ssh to emily, so the tar streams
  through the machine running the script; nothing large is written locally.
  Extracting as `tanshu` on emily is what makes the files land owned by the user
  `keep-id` maps the app onto — as root they would arrive owned by uid 1000,
  which on emily is nobody.

By hand, that is:

```sh
SITE=hopsngrains
HOST=menu.hopsngrains.com

ssh monoco "docker stop menu-app.$HOST"
ssh -n monoco "sudo -n tar -C /var/lib/$HOST -cf - ." \
  | ssh emily "mkdir -p ~/podman-data/menu-$SITE && tar -C ~/podman-data/menu-$SITE -xf -"
ssh monoco "docker start menu-app.$HOST"

# Same checks the script runs.
ssh -n monoco "sudo -n sh -c 'cd /var/lib/$HOST && find . -type f -exec sha256sum {} +' | sort | sha256sum"
ssh -n emily  "cd ~/podman-data/menu-$SITE && find . -type f -exec sha256sum {} + | sort | sha256sum"
ssh -n emily  "find ~/podman-data/menu-$SITE ! -user tanshu"
```

The old layout put everything under `/var/lib/menu.<domain>/`, with the content
one level down in `data/`; the new one is `~/podman-data/menu-<site>/` with the
same `data/` inside it and `site.env` beside it. Nothing inside changes.
`site.env` carries over so the admin keeps working; without it the app starts
with the admin disabled and the playbook says so rather than failing.

Then deploy that Site and check it on a `Host:` override before touching DNS.
`monoco` keeps serving until the cutover; only then stop its container — and
leave it in place, not removed, so `docker start` is the rollback.

## The admin password

The app starts with the admin disabled while `~/podman-data/menu-<site>/site.env`
is missing, and the playbook says so rather than failing. Create it once per Site;
run it again with `--force` to reset the password (a fresh session secret signs
everyone out). There is no minimum password length.

```sh
scp scripts/provision-site-env.sh emily:/tmp/
ssh -t emily bash /tmp/provision-site-env.sh ~/podman-data/menu-hopsngrains
systemctl --user restart menu-hopsngrains.service
```

The `-t` matters: the script prompts for the password, and without a terminal on
the remote side the prompt cannot be read, so it would sit there looking frozen.
With no image argument it borrows the one the running container uses. It has to
run as `tanshu`, the user whose rootless Podman owns the container.

## Checking it landed

```sh
npm run check-site -- https://menu.hopsngrains.com
```

Read-only, exits non-zero on failure. Then from a phone: the landing page, a
Section page, the page counter, tap-to-zoom, and signing into `/admin` to publish
something.

## Rolling back

Deploy the image built from an earlier commit — the commit tag is what makes
this a one-liner, since every commit ever pushed to the registry is still there:

```sh
cd ansible
ansible-playbook -u tanshu --inventory=hosts playbook.yml -e image_tag=<older-commit>
```

`podman logs menu-<site>` and `journalctl --user -u menu-<site>` are where a
container that will not start says why. To go all the way back to the Docker
deployment, start the container that was left on `monoco` and point Caddy at it
again.