halkyon/deploy/host/first-start.md

halkyon.li's first start

The instance on balthasar, once (ADR 0039), after the host's steps 1 to 6 (README.md): Learn, Train, Todo and Meditate with Jellyfin, LanguageTool, BRouter, Train's map and place search, all from the start (the developer's choice of 8 Nov 2026). Each step leaves what came before running. Commands run on balthasar as the developer (ssh balthasar from melchior) unless a step says otherwise; those for deploy after sudo -iu deploy. Later deploys are redeploy.sh's (D13).

1. The sidecars' users, each ID the sidecar's alone, as its compose file runs it (useradd warns that the IDs lie above the system range, which is meant: they must never be a person's, ../sidecars/README.md):

   for each in jellyfin:61096 languagetool:61081 brouter:61177 maptiles:61178 photon:61179; do
     name=${each%:*} id=${each#*:}
     sudo groupadd --system --gid "$id" "$name"
     sudo useradd --system --uid "$id" --gid "$id" --no-create-home --shell /usr/sbin/nologin "$name"
   done
   

2. deploy, which owns the checkout and runs the instance, in Docker's group (so as good as root on this host: its key may run redeploy.sh alone, step 3). Its home is on the HDD, as /home is:

   sudo useradd --create-home --shell /bin/bash --groups docker deploy
   sudo -iu deploy
   ssh-keygen -t ed25519 -N '' -C 'deploy@balthasar: halkyon, read-only' -f ~/.ssh/id_ed25519
   cat ~/.ssh/id_ed25519.pub
   

   In Forgejo (code.halkyon.app, the repository's Settings › Deploy keys), that key as a deploy key, without write access. Then, still as deploy, the clone; SSH first shows the forge's host key, which must be SHA256:98mCW/wdTIK7Jl6vwDE4qQqLECa8wkgDWye7+nQONn4 (ED25519):

   git clone ssh://git@code.halkyon.app/halkyon-days/halkyon.git ~/halkyon
   

3. The developer's keys: the GPG key redeploy.sh checks every commit's signature with (deploy/allowed-signers), and the SSH key it alone may be run with. On melchior:

   gpg --export --armor B172B9ED76A32EBCABA24A9F07F6FCF405488B64 >developer.asc
   scp developer.asc ~/.ssh/raisondetredev@ramiel.pub balthasar:/tmp/
   

   On balthasar, as deploy:

   gpg --import /tmp/developer.asc
   git -C ~/halkyon log -1 --format='%G? %GP'   # U B172B9ED…B64: good, vouched for by the list
   printf 'restrict,command="/home/deploy/halkyon/deploy/redeploy.sh" %s\n' \
     "$(cat /tmp/raisondetredev@ramiel.pub)" >>~/.ssh/authorized_keys
   chmod 600 ~/.ssh/authorized_keys
   ssh-keygen -lf ~/.ssh/authorized_keys   # SHA256:eSJ3RtDT0oJxCUq33B50snmkt1jSYw19Z0U6X+W9+hA
   

   (rm /tmp/developer.asc /tmp/raisondetredev@ramiel.pub as the developer afterwards.) On melchior, in ~/.ssh/config, through the tunnel, so it works at home and away alike:

   Host balthasar-deploy
       HostName 10.61.97.2
       User deploy
       IdentityFile ~/.ssh/raisondetredev@ramiel
       IdentitiesOnly yes
       ProxyJump ramiel
   

4. .env, as deploy: every secret made here, the developer's settings, never over an existing file (halkyon-li-env.sh):

   cd ~/halkyon
   sh deploy/host/halkyon-li-env.sh deploy/.env
   cat deploy/.env   # the whole file into Proton Pass: Infra, a note "halkyon.li .env"
   

5. SimpleLogin's app, the only way in (D4): in SimpleLogin, Settings › Developer › New website, named Halkyon (halkyon.li), its URL https://learn.halkyon.li, its redirect URL (callback) https://learn.halkyon.li/simplelogin-callback. Its client ID and secret into deploy/.env (SIMPLELOGIN_CLIENT_ID, SIMPLELOGIN_CLIENT_SECRET, as deploy with nano ~/halkyon/deploy/.env) and into the note in Proton Pass.

6. The sidecars' folders, and Jellyfin moved in. As the developer:

   d=/home/deploy/halkyon/deploy
   sudo install -d -o languagetool -g languagetool $d/languagetool $d/languagetool/fasttext
   sudo install -d -o brouter -g brouter $d/brouter/segments
   sudo install -d -o maptiles -g maptiles $d/maps/tiles
   sudo install -d -o photon -g photon $d/photon/data
   

   Then ../sidecars/README.md, Jellyfin on halkyon.li, steps 2 to 5: the library to /data/media, the configuration and cache into the deploy, network.xml for Caddy, ADR 0032's folder removed. .env names the library already.

7. The databases. PostgreSQL first, alone, which makes the instance's superuser halkyon_learn with .env's password; then a role and a database for each of the other three, as ../sidecars/README.md explains for Train's (Train, step 2: each role logs in and is nothing more, owns its database alone and connects to no other). As deploy, the passwords read from .env rather than typed:

   cd ~/halkyon/deploy
   docker compose up -d --wait postgres
   set -a; . ./.env; set +a
   docker compose exec -T postgres psql -U halkyon_learn -d postgres -v ON_ERROR_STOP=1 \
     -v train="$TRAIN_POSTGRES_PASSWORD" -v todo="$TODO_POSTGRES_PASSWORD" \
     -v meditate="$MEDITATE_POSTGRES_PASSWORD" <<'SQL'
   CREATE ROLE halkyon_train LOGIN PASSWORD :'train';
   CREATE DATABASE halkyon_train OWNER halkyon_train;
   CREATE ROLE halkyon_todo LOGIN PASSWORD :'todo';
   CREATE DATABASE halkyon_todo OWNER halkyon_todo;
   CREATE ROLE halkyon_meditate LOGIN PASSWORD :'meditate';
   CREATE DATABASE halkyon_meditate OWNER halkyon_meditate;
   REVOKE CONNECT ON DATABASE halkyon_learn FROM PUBLIC;
   REVOKE CONNECT ON DATABASE halkyon_train FROM PUBLIC;
   REVOKE CONNECT ON DATABASE halkyon_todo FROM PUBLIC;
   REVOKE CONNECT ON DATABASE halkyon_meditate FROM PUBLIC;
   SQL
   exit   # leaves deploy's shell, with .env's secrets in it
   

8. The first start, as deploy in ~/halkyon/deploy. It builds every image (the four servers, the dictionaries, BRouter, the map's and Photon's), half an hour or more, then starts everything; each server applies its migrations as it starts:

   docker compose up -d --build
   docker compose ps   # every service up; the servers healthy within a minute or two
   

   What builds on afterwards, by itself: the dictionaries (some 3 GB downloaded within the first hour), BRouter's segments for southern Bavaria (minutes), the map's cut (an hour or more) and Photon's Germany (7.5 GB, an hour or more). Until then Train has no routes, an empty map and no search. From melchior, anywhere:

   for name in learn todo train meditate jellyfin; do
     curl -s -o /dev/null -w "$name.halkyon.li %{http_code}\n" https://$name.halkyon.li/
   done
   

   Every name answers (200, or a redirect), over a certificate Caddy got at home, as ramiel passes TLS through unopened.

9. The first operator (ADR 0021, D20, D32): in Halkyon Learn, another instance, https://learn.halkyon.li, Sign in with SimpleLogin, a handle chosen; the account then waits. As deploy, in ~/halkyon/deploy, the developer's handle in lower case in place of their_handle:

   docker compose exec -T postgres psql -U halkyon_learn -d halkyon_learn -v ON_ERROR_STOP=1 <<'SQL'
   BEGIN;
   UPDATE "serverpod_auth_core_user"
   SET "scopeNames" = ("scopeNames"::jsonb || '["operator"]')::json
   WHERE "id" = (SELECT "authUserId" FROM "public_profile" WHERE "handleKey" = 'their_handle')
     AND NOT ("scopeNames"::jsonb ? 'operator');
   UPDATE "serverpod_auth_core_session"
   SET "scopeNames" = ("scopeNames"::jsonb || '["operator"]')::json
   WHERE "authUserId" = (SELECT "authUserId" FROM "public_profile" WHERE "handleKey" = 'their_handle')
     AND NOT ("scopeNames"::jsonb ? 'operator');
   COMMIT;
   SQL
   

   UPDATE 1 twice, or the first once and the second UPDATE 0 while there's no session yet. Back in the app, Try again on the wait screen: an operator never waits, and Settings shows This instance. Everyone else now signs up and waits for the operator, who approves them there, the four words read out first.

10. Backups (D18): ../backups.md, Setting it up, for halkyon.li: on the same Storage Box a sub-account of its own (halkyon-li), a key and a restic repository of their own, the repository's password into Proton Pass (Infra, "halkyon.li restic"), then the timer.

Then the CI runner, README.md, step 7, and the Instance console (ADR 0040).