# Silver Server Ops

Server-wide maintenance scripts and operational notes for the Silver Cloud Hetzner host.

This repo is for host-level improvements that are not tied to one specific app or
service. SSL expiry alerts are the first check in the repo; more server-wide
maintenance checks can be added here over time.

## Implemented Checks

- SSL certificate expiry alerts

## SSL Certificate Expiry Alerts

`scripts/check-cert-expiry.sh` checks every Let's Encrypt certificate under
`/etc/letsencrypt/live` and sends an ntfy alert when any certificate expires in
less than 25 days.

The threshold sits just below certbot's own renewal window. Certbot renews at 30
days remaining, so a healthy certificate never drops below that. A 25 day
threshold fires roughly five days after renewal first fails and still leaves 25
days to fix it. A threshold below the renewal window means renewal has to stay
broken for weeks before the alert trips, which is too late to be useful.

### Install

```bash
sudo install -m 0755 scripts/check-cert-expiry.sh /opt/scripts/check-cert-expiry.sh
```

### Root cron

```cron
15 8 * * * /opt/scripts/check-cert-expiry.sh >>/var/log/cert-expiry-check.log 2>&1
```

### Defaults

```bash
CERT_DIR=/etc/letsencrypt/live
EXPIRY_DAYS=25
NTFY_URL=http://127.0.0.1:2586/certbot
ALERT_ON_NO_CERTS=true
CHECK_SERVED=true
SERVED_ADDR=127.0.0.1:443
SERVED_TIMEOUT=10
HEALTHCHECK_URL=
```

### Served certificate check

Certbot writes a renewed certificate to disk, but nginx keeps serving the old
one from memory until it is reloaded. A check that only reads
`/etc/letsencrypt/live` reports everything as healthy while browsers are being
handed an expired certificate.

With `CHECK_SERVED=true` the script also opens a TLS connection to
`SERVED_ADDR` using each lineage's name as the SNI hostname and compares the
served expiry against the file on disk. It warns only when the served
certificate expires earlier than the one on disk, which is the signature of a
renewal hook that stopped firing.

nginx completes a handshake with a fallback vhost certificate when SNI matches
no server block, so the comparison is skipped unless the served certificate's
SAN list actually covers that domain. A lineage that nginx no longer serves is
therefore skipped rather than compared against an unrelated certificate. Set
`CHECK_SERVED=false` to disable the check.

### Overrides

```bash
EXPIRY_DAYS=40 /opt/scripts/check-cert-expiry.sh
```

If the ntfy topic is protected with an access token:

```bash
NTFY_TOKEN=your-token /opt/scripts/check-cert-expiry.sh
```

### Manual test

```bash
sudo /opt/scripts/check-cert-expiry.sh
```

The script exits `0` when all certificates are healthy and `1` when it sends an
alert or cannot run the check correctly.

## Notification Notes

ntfy is a good default for this server because it is already self-hosted, simple
to call from shell scripts, and supports useful alert metadata such as title,
priority, and tags.

Alerts go to ntfy over `http://127.0.0.1:2586` rather than the public
`https://ntfy.silverwal.com` URL. The public URL is served by the same nginx
whose certificates this script watches, so an expired or broken certificate on
this host would fail curl's TLS verification and silently drop the very alert
that reports it. The loopback address removes DNS, nginx, and TLS from the
alerting path.

For jobs where silence is also a failure, pair ntfy with a dead man's switch such
as Healthchecks. ntfy tells you what the script found; Healthchecks tells you when
the script did not run at all.

Set `HEALTHCHECK_URL` to enable it. The script pings that URL after a clean run
and `$HEALTHCHECK_URL/fail` when it alerts or cannot complete the check, so a
dead cron, a bad chmod or a host that never came back up stops looking like a
healthy fleet:

```cron
15 8 * * * HEALTHCHECK_URL=https://hc-ping.com/<uuid> /opt/scripts/check-cert-expiry.sh >>/var/log/cert-expiry-check.log 2>&1
```

Leaving `HEALTHCHECK_URL` empty disables the pings entirely.

This matters more than it looks for a certificate check specifically. The script
is designed to stay silent on a healthy host: certbot renews at 30 days and the
threshold is 25, so a correctly working fleet produces no notifications, ever.
Without a heartbeat, "no alert" and "the check has not run since March" are the
same observation.