Eric WagonerandClaude Opus 5.5
f0d7237c2b
A start and a finish sent seconds apart can reach Sentry in either order, and a start that lands second is never closed, so a fast rollup was reported as a timeout on nights it succeeded. Hnefatafl's monitor did so nine nights in fifteen. One check-in at the end cannot race; a run that hangs sends nothing and is marked missed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GwFKMuQnPAEHJ5yA1q4orh
Deploying NAME
Every game is a room on a small Node process that keeps it as an
append-only ledger of moves in /var/lib/SLUG/rooms. Production is the
shared Kestrel's Hall host: Caddy terminates TLS with automatic
certificates, serves the static build from /opt/SLUG/build, and proxies
/api and /ws to the game server on port PORT, which runs as the
__SLUG__ user under systemd from /opt/SLUG/app.
The shared host
The games of Kestrel's Hall share one droplet, hall (nyc3, s-1vcpu-1gb,
$6 a month, tag kestrels-hall, IP 159.203.115.252). Each game keeps its
own Caddy site file in /etc/caddy/sites/.caddy, its own port,
service user, /opt/ and /var/lib/ trees, access log
(/var/lib/caddy/-access.log), cron file and ops scripts;
/etc/caddy/Caddyfile only says import sites/*.caddy. A game server is one
small Node process capped at 300 MB by its unit file, and a room is a few
kilobytes of ledger, so the host carries several. Each game needs a port
of its own on the host: wizwar 8787, waving-hands 8788, hnefatafl 8789,
dip 8790.
Current production
- Host: IP
- URLs: https://DOMAIN (A record at Hover, where kestrelsnest.social's DNS lives) and https://SLUG.IP.sslip.io (always works, zero DNS).
- Everyday deploy:
deploy/deploy.sh __IP__ - Players' reports:
deploy/pull-reports.shmirrors /var/lib/SLUG/feedback.jsonl and the screenshots to ~/Desktop/SLUG-reports with a digest;deploy/report-reply.sh __IP__ <id> <status> "text"answers one.
Adding this game to the host
scp deploy/setup-droplet.sh deploy/Caddyfile.tmpl root@<ip>:/root/ && ssh root@<ip> \ "bash /root/setup-droplet.sh '__DOMAIN__, __SLUG__.<ip>.sslip.io'"It installs Caddy and Node if the host lacks them, writes this game's site file, and touches nothing of another game's. Point the A record at the host first, or leave the real name out until it is; running it again with the new line is how a name is added later.deploy/deploy.sh <ip>
A fresh host, when one is needed: doctl compute droplet create hall --region nyc3 \ --size s-1vcpu-1gb --image ubuntu-24-04-x64 --ssh-keys <your-key-ids> \ --tag-name kestrels-hall --wait, then step 1.
Operations scripts on the droplet
deploy.sh installs these to /usr/local/bin, the traffic.py they share to
/usr/local/lib/SLUG, and the cron file to /etc/cron.d/SLUG on every
deploy, so the live copies are the repo copies:
__SLUG__-visitors.sh [day]: who is here now and who came that day.__SLUG__-pulse.sh: the weekly health check (service, errors, rollup trend, backup, box).__SLUG__-rollup.sh [day]: one JSON line per day of counts, run nightly at 00:12 UTC into /var/lib/SLUG/rollup.jsonl.__SLUG__-backup.sh: nightly at 07:23 UTC, mirrors /var/lib/SLUG to thekestrel-wizwar-backupsSpace under__SLUG__/(a current copy less the rollup, and dated snapshots kept 90 days). It needs rclone with the Spaces credentials in /root/.config/rclone/rclone.conf, copied by hand from the wizwar droplet; until then it logs "skipped".
Errors from the game server go to Sentry, project __SLUG__ in the
locallygrownnet organisation; the DSN is in the unit file. The nightly rollup
and backup check in with Sentry Crons when /root/.SLUG-sentry-cron-rollup
and /root/.SLUG-sentry-cron hold their check-in URLs (the ingest
URL with the project's cron path and public key), so a missed night is noticed.
To have every new issue, regression and reappearance posted to Slack the way
wizwar's are, run deploy/sentry-slack-alert.sh [#channel] once from your own
shell with SENTRY_TOKEN (an org auth token with alerts:write) and
SLACK_CHANNEL_ID (the channel's Slack ID) set; the token never leaves the
shell.
Before every deploy, deploy/verify-ledgers.sh <ip> fetches every production
ledger and replays it with the local engine, comparing each room with what the
server shows. A ledger the new engine refuses or replays differently stops the
deploy: the server would rewrite that game on restart.
Everyday deploys
deploy/deploy.sh <ip>
Runs the type-checks, the tests and the build locally, rsyncs build/ to
the droplet keeping the previous week's hashed assets, rsyncs the server and
engine sources, installs dependencies, and restarts the game server, then
waits up to twenty seconds for it to answer HTTP and fails with the tail of
its journal if it does not. A room is replayed from its ledger when the
server comes back, which takes a few seconds; Caddy holds requests that
land in the gap.
Operations
- Logs:
ssh root@<ip> journalctl -u __SLUG__ -ffor the game server,journalctl -u caddy -fand /var/lib/caddy/SLUG-access.log for the web side. - Restart:
ssh root@<ip> systemctl restart __SLUG__ - Rooms:
/var/lib/__SLUG__/rooms/<CODE>.jsonl, one ledger per game; the nightly backup keeps them in the Space.