Tomter Vel — infrastructure

Everything the vel's own site needs on one host, as containers: Caddy for TLS, NATS with JetStream for the records, Kanidm for who is who, and portal, the site itself. The content (pages, forms, desks) lives in tomtervel/questions and is fetched from there; this repo owns the host.

Internet ──► Caddy (Let's Encrypt)
               ├── PORTAL_HOST ──► portal:3000 ──► NATS (records) + Kanidm (login)
               └── ID_HOST ─────► kanidm:8443 (internal TLS)

Today the vel's draft site runs on Klingenberg Bygg's host as vel.klingenbergbygg.no, sharing that host's Kanidm and NATS. This repo is the same site on a host of the vel's own, with a Kanidm of its own, so nothing about the vel's members or records depends on anyone else.

First start

On a fresh Linux host with podman + podman-compose and openssl (run rootful, i.e. as root, so Caddy can bind 80/443 and Kanidm sees a stable source IP):

git clone https://prosjekt.klingenbergbygg.no/tomtervel/infrastructure /srv/tomtervel/infrastructure
cd /srv/tomtervel/infrastructure
cp .env.example .env      # set PORTAL_HOST and ID_HOST; DNS must point here
sudo sh bootstrap.sh      # renders configs, makes the internal cert, starts, recovers Kanidm admin
sh kanidm-setup.sh        # logs in, creates the portal client and desk groups, restarts portal

bootstrap.sh prints the admin and idm_admin passwords once; write them down. After kanidm-setup.sh, https://PORTAL_HOST serves the site and https://ID_HOST is the login.

Running on kasse (behind the host Caddy)

The vel keeps its own Kanidm and NATS here so it can later lift onto a host of its own unchanged — on kasse it just runs in isolation, fronted by kasse's existing host Caddy (which owns 80/443). So the bundled Caddy stays off (it's behind profiles: [edge]); run the default podman compose up -d --build. The services publish loopback-only ports for the host Caddy to reach:

• portal → 127.0.0.1:3050
• kanidm → 127.0.0.1:8443 (internal self-signed TLS)
• nats → 127.0.0.1:4223 (kasse's shared platform NATS owns 4222)

Add two host-Caddy site blocks in /etc/caddy/conf.d/, using the vel's .env hosts (vel.klingenbergbygg.no → :3050 already exists):

<PORTAL_HOST> {
	reverse_proxy localhost:3050
}
<ID_HOST> {
	reverse_proxy https://localhost:8443 {
		transport http { tls_insecure_skip_verify }
	}
}

Then sudo systemctl reload caddy, point the content repo's lint-and-reload reload step at nats://127.0.0.1:4223, and retire the old systemd portal: sudo systemctl disable --now app@tomtervel-portal.

Accounts, all the way

kanidm-setup.sh sets up everything portal needs from this Kanidm, and is safe to rerun:

• Desk groups, named as the content names them. Read from every qualifies: under questions/ in the content repo, so the list cannot drift from the pages. This Kanidm is the vel's own, so there is no prefix: the board's group is styret. Groups from before are renamed in place (tomtervel_styret → styret), members and all.
• Sign-in. The OAuth2 client's scope map is idm_all_persons: everyone here is the vel's, and what they see is decided by their desk groups in the groups claim.
• Onboarding. The portal-onboarding service account manages every desk group, so an invite from a desk and a grants: can add people to them. Its token goes into portal.env.
• The first person in. SEED_ADMIN_EMAIL in .env: at start, portal invites them into the most privileged group (the board) and mails the link, once. Everyone else they invite from their desk.
• The people who ask the questions. Everyone a page names as responsible gets an account at start, a mail saying they are listed as responsible for asking that question, and a place in RESPONSIBLE_GROUP (redaktor). Signed in, they can suggest changes to their own pages.
• The project Gitea. With GITEA_URL set, redaktor may sign in to it with the same account: its own OAuth2 client, only for that group. Run on the Gitea host, the script adds the sign-in source itself (and requires the redaktor claim there too); elsewhere it leaves the secret in gitea-oauth.secret and prints the one command.

People and desks

Each group in the content (qualifies: under questions/) is a Kanidm group tomtervel_<group>, all members of tomtervel_members. Someone in tomtervel_styret logs in and sees the board's desk.

kanidm -D idm_admin -H https://ID_HOST person create kari "Kari Lien"
kanidm -D idm_admin -H https://ID_HOST person update kari --mail kari@example.no
kanidm -D idm_admin -H https://ID_HOST person credential create-reset-token kari
kanidm -D idm_admin -H https://ID_HOST group add-members tomtervel_styret kari

Membership is read at login; someone added while logged in logs out and in again.

Content changes

A push to tomtervel/questions is linted on push and tells the portal to reload over NATS. That reload reaches the host it runs on: today Klingenberg Bygg's. On this host the runner is a host service (pacman -S gitea-runner, registered against prosjekt.klingenbergbygg.no) — not an in-compose container — so it reaches NATS on the published 127.0.0.1:4222. Give the content repo's lint-and-reload.yml a reload job whose runs-on matches that runner's host label, running nats --server nats://portal:$NATS_PASSWORD@127.0.0.1:4222 pub portal.content.reload "". Until then, podman compose restart portal picks up new content.

Upgrading portal

Bump PORTAL_RELEASE in .env (a tag of uhhm/portal) and podman compose up -d --build portal. Keep IRIS_RELEASE in the content repo's workflow matched to it.

Backups

• Kanidm writes a nightly backup into its volume (/data/backups, seven kept); copy that directory off the host.
• NATS JetStream data is the nats_data volume: every record ever submitted and every state change. Snapshot the volume.
• Caddy's certificates regenerate; nothing to keep.

What is not here

Mail (a person's reset link is a token you hand them), monitoring, and the vel's current website at tomtervel.no, which stays where it is until the vel points the apex at PORTAL_HOST.

Mail

Portal decides what to send - a mail: on a state in the content, or an invite - and publishes it on this stack's NATS. gdo is what hands it to a mail server, and it runs here, in the vel's own stack, so the vel's mail leaves on the vel's own terms.

It has no mail server of its own. It relays through the host's, which on kasse is Klingenberg Bygg's postfix on port 25, reached from the container as host.containers.internal. That is the one thing in this stack the vel borrows, and the one thing that changes if the vel ever moves to a host of its own: point SMTP_HOST at whatever that host runs.

site.yaml in the content repo needs a mail.from, or portal has nothing to send from and says so in its log.

Replies are not read yet: JMAP_URL, JMAP_TOKEN and MAIL_REPLY_DOMAIN turn a reply into a note on the record it answers, and none of them are set here.

How the login page looks

Out of the box Kanidm presents itself: its own name, its own mark. A neighbour arriving at a login page that belongs to nobody in particular is right to hesitate, so kanidm-setup.sh sets the vel's name and mark on both the instance and the portal client:

kanidm -D idm_admin -H https://$ID_HOST system domain set-displayname "Tomter Vel"
kanidm -D idm_admin -H https://$ID_HOST system domain set-image logo.svg svg

The logo is fetched from the content repo - the same images/logo.svg that site.yaml uses as the favicon - so the login page and the site are branded from one place. png, jpg, gif, svg and webp all work.

Two things the setup script deliberately leaves alone, because they are decisions rather than branding:

• system domain set-allow-account-recovery true lets someone who has lost their credentials get a reset link by proving one of their own email addresses, instead of asking the board. Worth having for a vel; it is still a door, so it is turned on knowingly or not at all.
• system domain set-allow-easter-eggs - seasonal icons and birthday surprises. Off in production builds, and this is somebody's vel.

Keep the CLI and the server on the same version

The kanidm CLI on the host and the kanidm/server image in compose.yml must match. They are not merely fussy about it: a 1.11.2 client against a 1.11.1 server looks up the domain entry at a UUID the older server does not have, so system domain set-displayname and set-image fail with "Item not found" - a message that says nothing about versions. The CLI warns on every call; the warning is worth reading.

The same "Item not found" also comes back when the versions do match and the account is idm_admin: the instance's own name and logo are system settings only admin may change. kanidm-setup.sh sets them when an admin session exists and says so when it does not.

The host's CLI comes from pacman and moves on its own, so the image is what to bump: image: kanidm/server:<version> in compose.yml, then podman compose up -d kanidm.

Mail leaving the host

gdo hands the vel's mail to kasse's own postfix over the container network, so postfix has to be willing to relay from it: mynetworks on kasse includes 172.16.0.0/12 and 10.88.0.0/16 (set 2026-09-30, the original is in /etc/postfix/main.cf.before-containers). Without it every address outside the host's own domains is refused with "Relay access denied", and gdo retries a minute apart until it is.