- Endpoints: HTTP, TCP and TLS checks with answer time, uptime and certificate expiry; failures and expiring certificates become alerts and notifications; linked to a server for maintenance and tags - Restore tests: record when a backup was restored and how it went; failed tests always alert, overdue ones with a limit set; shown on the Backup coverage page - Change log: one filterable timeline made from the audit log and the incident record Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
22 KiB
Notifications
Every notification in Homelab Manager is sent through the same pipe
(server/src/services/notify.ts) to whichever channels are enabled under
Settings → Notifications: Gotify, ntfy, SMTP (email), a generic
JSON webhook, and push to phones and browsers. All five fire for every notification below unless a
routing rule (see Who hears what below) says otherwise; each event also has its on/off toggle and, for the
daily ones, one shared time/timezone.
The settings page has four tabs, with one Save button for all of them (it
shows "Unsaved changes" until you save, and the browser asks before the page
is closed with changes unsaved): Channels (the five channels, each a row
you open to fill in and send a test), What to send (the event toggles
grouped by topic, each group's thresholds beside its switches), Who hears what (routing rules and escalation) and When
(the time zone, the daily reminder time, quiet hours and the weekly digest).
The address keeps the tab, e.g. /settings/notifications?tab=when.
Two settings apply to every notification regardless of what triggered it:
- Quiet hours (Settings → Notifications → Quiet hours): while enabled and inside the configured window, a notification is held instead of sent immediately, then delivered as a single digest at the window's end time. This matters most for the real-time alerts below — the daily reminders already fire once at a time you pick, usually outside the window anyway.
- Maintenance windows (the Maintenance page): opening one for a server, integration, or DNS provider silences the alerts that target it specifically (offline/disk-full for a server; storage, backup, and automation-failure alerts for an integration; "integration down" for that whole service type) — see the Maintenance page's own "What gets silenced" panel for the exact list.
Where a trigger has a configurable threshold, that's noted in its row below; several are fixed and can't be changed from the UI.
Who hears what
Under Settings → Notifications → Who hears what.
Routing rules. By default every notification goes to every channel that is switched on. A rule narrows that for what it matches, on any of: the kind of problem (server down, disk and storage, failing services, backups, updates, expiry, automation, integration failing …), how bad it is (warning or worse, critical only) and the tags of the server it is about (any of them). The first rule that matches decides which channels get it — put the specific rules first — and a rule with no channels tells nobody (the Alerts page still shows the problem, and webhooks for events still fire). A notification no rule matches goes to every channel that is on, as before. Rules can be switched off and reordered.
- A message that holds problems for different routes is split: a disk warning can go to ntfy while the server outage in the same check goes to everything, as two messages. When all of them take the same route it is one message, as ever.
- A recovery follows the route its alert took (the health check remembers how bad each problem was).
- What a rule can see: health and automation problems carry a kind, a severity and (for servers) the server; secret, domain and Tailscale expiry, backup checks, Docker image updates, update runs and integration failures carry a kind and a severity. DNS changes, the weekly digest and similar carry neither, so only rules that ask about nothing in particular match them.
- Notifications held for quiet hours keep who they were for: the digest at the end goes only to the channels their rule picked.
Each browser's own wishes. Under Phone and browser, every subscribed browser can say it only wants warnings or worse, or only critical alerts, and only some kinds of alert — a phone for the critical things, a desktop for everything. It applies after routing: a rule that sends something to Phone and browser still doesn't reach a browser that turned that kind down. A notification that says nothing about its kind or severity reaches all.
Escalation. When an alert has been raised for a while and nobody has acknowledged it, it is told again, in up to three steps — for example after 30 minutes: push; after 2 hours: email the neighbour who holds the key. Each step names its time, the least severity it applies to, the channels to use (chosen freely — routing rules don't apply to escalations), an optional extra email address to tell as well (sent through the email channel's server) and whether to send during quiet hours (otherwise it waits for them to end like any other). Every step is sent once per alert; an acknowledgement stops the rest, and the alert starts over if it goes away and comes back. Problems under maintenance, held back because something they depend on is down, or only informational never escalate. It is looked at every five minutes, so a step can arrive a few minutes after its time. Every escalation is written in the audit log (category alert, action escalate).
Endpoint checks and restore tests
An endpoint check that fails several times in a row (the check's own number, 2 by default) is announced like any other server problem — under the server offline,
disk full … switch (healthAlerts) — once when it starts, with a recovery message when it answers again; a certificate that is within its warning time is announced
the same way. For routing they count as Monitoring (a check that fails) and Expiry (a certificate), and carry the tags of the server the check is linked to; maintenance on that
server holds them back. A restore test that fails, or one that is overdue (when a limit is set), is on the Alerts page and in the incident record but is not sent as a
notification of its own — it is something to look at when you are looking at backups.
Push to a phone or browser
The fifth channel, Phone and browser, sends the same messages as the others to browsers that have asked for them — a phone with this page added to its home screen, a desktop browser — with no app, account or extra service. An administrator switches it on under Settings → Notifications → Channels (the optional contact is given to the browsers' push services, as the standard asks), and then each person opens the same row and presses Notify this browser; the browser asks permission and the row lists their browsers, with a button to send a test and one to remove each.
- It needs https. Browsers only allow push for pages served over https (or from
localhost), so a copy opened over plain http on a LAN address shows why it can't be used instead of the button. - How it travels. The message is encrypted for each browser (RFC 8291) and signed with a key pair this server makes the first time it is needed (RFC 8292, "VAPID"), then handed to the browser's own push service — Google's, Mozilla's or Apple's — which delivers it. The push service sees that a message is being sent and its size, not what it says. The address of each browser's push service is checked before use: https only, and never an address on a private network.
- Browsers' own wishes (a minimum severity, which kinds) are set per browser under the channel's row.
- Quiet hours hold push messages like every other channel, and each event's on/off toggle applies. A tap on a notification opens the Alerts page.
- A browser that has gone (the site's permission was removed, the app uninstalled) is dropped automatically when its push service says so. If the server's key pair is ever lost (the credentials key changed), everyone has to press the button again.
Daily reminders
These six checks share one schedule: Settings → Notifications → Daily reminder time (default 08:00) and Timezone (default UTC). Unlike the state-based alerts further down, these re-send every day the condition is still true — there's no "only once" de-duplication, so an expired secret you haven't renewed yet will be mentioned again at the next day's check, and the day after that.
Each one also runs once at server startup if it hasn't already run today (e.g. after an upgrade or a period offline), so you're not waiting until the next scheduled time to catch up.
| Check | "Notify on" toggle | Fires when | Configured at |
|---|---|---|---|
| Secret expiry | Secret expiry reminder | Any tracked secret (API token, SSL cert, password, generic) is expired or within its own configured warning window. SSL certificates with a host:port are re-checked live first, so this reflects the real current expiry, not a stale saved date. A certificate that couldn't be re-checked live is mentioned separately, so the expiry shown may be stale. | Per-secret warning threshold, set when adding/editing that secret |
| Domain expiry | Domain registration expiring or expired (daily reminder) | Any tracked domain registration is expired or within the warning window; every domain is re-looked-up against its registry first. A domain whose registry lookup has been failing for 3+ days is also mentioned, separately, as "may be stale." Registries that don't publish an expiry (e.g. .de, .eu) are tracked but never trigger this. |
Settings → Notifications → Health checks → Domain expiry warning (default 30 days) |
| Tailscale key expiry | Tailscale key expiry reminder | Any device's node key (across every enabled Tailscale integration) is within 30 days of expiring, or already expired. Devices with key expiry disabled are skipped. | Fixed at 30 days |
| Docker image updates | Docker image update available | Any container (across every enabled Dockhand integration) has an image update available, per Dockhand's own cached update-check results — this doesn't trigger a fresh registry lookup, just reads what the Docker page itself would show. | Not configurable (reflects Dockhand's own check interval) |
| Proxmox backups — gone stale | (same toggle as the Proxmox backup check) | Once a day with the other backup messages: a guest whose newest backup (on Proxmox or a Backup Server) is older than the limit. Guests under maintenance (on the server or the Proxmox integration) are left out. | Settings → Notifications → A guest's backup is too old after (default 48 h; 0 off); a server can have its own limit on its page |
| Update run finished | An update run you started finished, or didn't | An update run started from the Updates page ended: what is still waiting and whether a restart is due, or why it failed. Sent at once (not held for the daily check). | Settings → Notifications → What to send |
| Proxmox backups | Proxmox backup failed or a guest has no coverage | Two independent conditions, both under this one toggle: (1) a node's most recent vzdump backup task (across every enabled Proxmox integration) didn't succeed; (2) a VM/LXC isn't covered by any enabled backup job at all. |
Not configurable |
| Proxmox Backup Server verification | Proxmox Backup Server snapshot failed verification | Across every enabled PBS integration: any datastore has at least one stored snapshot whose verification state is "failed", or a datastore couldn't be read at all (e.g. a permissions problem). This is distinct from the Proxmox check above — PVE only knows a backup ran, PBS is the only place that knows whether the stored data still verifies. | Not configurable |
State-based alerts
These two run on a fixed 15-minute interval (matching the agent's own default report interval) — not the daily reminder time above, and not user-configurable. Unlike the daily reminders, these are state-based: a problem is announced once when it first appears, and once more when it clears — not repeated on every 15-minute pass while it continues. A condition that can't currently be read (an integration that's down, a server whose agent hasn't reported) is held exactly as it was rather than cleared or re-announced, so a temporary read failure can't fake a recovery. On first-ever run (or right after upgrading to this feature), whatever's already failing is recorded silently as the starting baseline rather than announced all at once.
| Check | "Notify on" toggle | Fires when | Configured at |
|---|---|---|---|
| Health — server offline | Server offline, disk nearly full, or Synology volume/disk problem | A server's agent hasn't reported for longer than the offline threshold. Held for the first 20 minutes after this app restarts, since agents haven't had a chance to report yet. | Settings → Notifications → Health checks → Server offline after (default 60 min) |
| Health — host health | (same toggle as the others in this group) | A Linux agent says a service has failed, a disk's SMART check fails (or it has used 90% or more of its rated life), or a ZFS pool is not ONLINE. A server that is currently offline isn't also judged on these (its figures are stale). Silenced by a maintenance window on the server like the others. | The same health-check settings; there is no separate threshold. |
| Health — disk usage | (same toggle) | A server disk, a Proxmox node's root filesystem or any of its storages, or a Synology volume, is at or above the usage threshold. A server that's currently offline isn't also judged on disk usage (its figures are stale). | Settings → Notifications → Health checks → Disk usage alert at (default 90%) |
| Health — Synology status | (same toggle) | A Synology volume's status isn't "normal", or a disk's status/SMART result isn't normal, or it's over the bad-sector or under the remaining-life threshold. | Not configurable |
| Automation — failed run | Semaphore template or Gitea workflow run failed | A Semaphore template's, or a Gitea repo's, most recent run status is a clear failure. A run that's still going, was cancelled, or was stopped by hand doesn't count as failed or as a recovery — it's left exactly as it was. For Gitea this follows the repo's most recent run on any workflow or branch. | Not configurable |
Real-time alerts
These fire immediately, as the triggering action happens — not on any schedule.
| Event | "Notify on" toggle | Fires when |
|---|---|---|
| DNS record added | DNS record added | A DNS record is created against any DNS provider through this app (manually, or via any automated action that writes one). |
| DNS record updated | DNS record updated | An existing DNS record is edited. |
| DNS record deleted | DNS record deleted | A DNS record is deleted. |
| Integration/DNS provider down | Integration/DNS provider failing repeatedly | Any integration or DNS provider adapter's calls fail a configurable number of times in a row. Tracked per adapter type (e.g. "proxmox"), not per individual integration row — with two Proxmox integrations, a streak of failures on either one counts toward the same total. Only fires once per failure streak (not on every failure past the threshold), and is skipped entirely if that type is currently in a maintenance window, or if it is explained by something it needs that is down (see Dependencies in the README): a service that starts failing while the server it runs on is offline is not announced, and if it is still failing once that server is back, the next failed call alerts. |
| Integration/DNS provider recovered | (same toggle) | The next call for that source succeeds, after a "down" alert was already sent for the current streak. |
Digest
| Notification | Fires when |
|---|---|
| "Notifications from quiet hours" | Once, at quiet hours' configured end time, only if enabled and only if at least one notification was held during the window. Bundles every held notification's title and message into one message, then clears the queue. |
Backups of Homelab Manager itself
A scheduled backup of the app's own database that fails sends one message through the channels ("Homelab Manager — its own backup failed", with the reason), under the same toggle as server and disk alerts (Server offline, disk nearly full…). The same problem is also an alert on the Alerts page of the Backups kind — "the newest backup of Homelab Manager's own database is … old" once it is more than two days old, critical after a week, and "no backup of its own yet" with the reason when the very first one fails. No alert is raised for a new install that hasn't tried yet, or when daily backups are switched off. See BACKUP.md.
Weekly digest
| Notification | Fires when |
|---|---|
| "Homelab weekly digest" | Once a week, on the day and at the time chosen under Settings → Notifications (in the notification time zone), only if switched on — it is off until then. One plain-text message through every channel that is on; quiet hours hold it like any other notification. |
It gathers what the app already knows into a few lines per topic, and leaves out any topic with nothing to say: what is wrong right now; uptime over the last seven days (what was down and for how long, with the percentage); what happened (servers that stopped reporting and came back, services that stopped answering, maintenance); incidents (how many began and ended, still open, the longest); disks heading for full; secrets and domains that are expiring or have expired (and hardware whose warranty ends within 60 days); guests that are not backed up, have a failed or stale backup; updates waiting (how many servers, how many security fixes, who needs a restart, the longest wait); and planned maintenance in the next seven days. A week with nothing in it says "A quiet week: nothing needs attention." On the settings page, Preview shows what it would say right now and Send it now sends it at once (through the channels that are on, whatever the time).
The Alerts page
Operations → Alerts shows what these notifications are about — the problems that exist right now — as a list you can look at, filter, and export. It uses the same checks as the notifications above, so a problem appears there for exactly the reason it would be notified, but it differs in three ways:
- It ignores the "Notify on" toggles. Turning a notification off doesn't hide the problem from the page.
- Problems under a maintenance window are kept on the list, marked silenced and counted separately, instead of being dropped.
- It also lists things nothing notifies about: Uptime Kuma monitors that are down, osTicket tickets that are overdue, and any integration that's failing its last few calls (before the threshold that triggers an "integration down" notification).
It runs the checks live when opened (a recent result is reused for a minute), and shows what it couldn't read at the top, so a missing section means "couldn't check" and not "all clear". An alert that has an osTicket ticket (see the next section) shows its ticket number.
Tickets in osTicket
Not a notification, but it acts on the same alerts: with Settings → Alert tickets switched on (off by default), a problem on the Alerts page opens a ticket in osTicket. It runs every 15 minutes, starting about a minute and a half after the app starts.
| Rule | What it means |
|---|---|
| Which problems | The severities and kinds you choose — critical and warning, every kind by default. Info (updates waiting and the like) is off. |
| Must last | A problem must be there in 2 checks in a row (15–30 minutes) before it opens a ticket, so a blip that fixes itself — one failed call to an integration, a monitor that flickers — never does. Can be set from 1 (at once) to 4. |
| One ticket per problem | However many checks see it, once. After a ticket is open, the problem isn't ticketed again while it lasts — even if someone closes the ticket in osTicket. |
| Comes back | A problem that clears and returns within 60 minutes is the same problem (no new ticket). After longer, it is a new one and gets a new ticket. |
| Many at once | More than 5 new problems in one check become a single summary ticket listing them all, so one outage doesn't fill the queue. Up to 5 get a ticket each. |
| Maintenance windows | A problem under one never opens a ticket. If it is still there after the window ends, it does. |
| Never about osTicket | Overdue tickets, and osTicket's own calls failing, never open a ticket — it would go to the thing that is down, or add to the pile. |
| Already there at switch-on | Left alone by default: the first check after switching on records what is wrong as the starting line, and only problems that start after it get a ticket. Can be set to ticket those too. Switching it off and on again starts a new line. |
| If osTicket can't be reached | Nothing is recorded as opened; the next check tries again. A refused key, a wrong address and similar are shown on the settings page and in the Diagnostic Log. |
| If something can't be read | When a check couldn't read everything (an integration was down), nothing is treated as gone, so a problem isn't called cleared just because its source couldn't be looked at. |
The ticket says what is wrong, the severity, the kind, where it comes from, when it was first seen, and links back to the problem in this app. osTicket's
API can't update or close a ticket, so it isn't touched again: close it in osTicket when the problem is dealt with. Every ticket opened (and every test
ticket) is in the Audit Log as alert_ticket, and changes to the settings as settings / update_alert_tickets. "Send a test ticket" and "Check now"
on the settings page use the saved settings.
What does not send a notification
Worth calling out explicitly, since it's easy to assume everything in the app alerts on something:
- osTicket's own tickets — the integration lists open tickets, but there is no notification for it (e.g. nothing pings you about a newly-overdue ticket); the page and the Alerts page show them. (The other direction — an alert opening a ticket — is the section above.)
- phpIPAM sync, Uptime Kuma monitor/maintenance import, and any other manual "Sync from X" action — these run when you click the button and report their result on the page itself, not via a notification.
- Test buttons (Settings → Notifications → each channel's "Send test") send a one-off message through that one channel only, to confirm it's wired up correctly — not a real event and not affected by quiet hours.
- Consistency reports, Diagnostic Log, and Audit Log are all read-only views you check yourself; none of them push a notification on their own (the Diagnostic Log's failures are what feed the integration-down alert above, but reading the log itself never triggers anything).