How alerts work
lanowl's promise is one message per incident, not one per blip. This page is how it keeps it: what counts as down, what reaches your phone at once, what waits for a digest, and what never reaches you at all.
Down is decided by code
Every minute (cadence.sweep_interval_s), every device in the inventory is probed. A device
is down after two misses in a row (cadence.debounce_fails), and up again at the first
answer. The internet has its own loop every 15 seconds. The model takes no part in any of
this: switch it off and detection is exactly the same.
A device is not an incident when its silence is explained:
| Why | How lanowl knows |
|---|---|
| Asleep on schedule | expect_offline: sun or day, with a grace period at the edges |
| Paused by you | /pause or its sheet on the dashboard (Pausing a device) |
| Rebooting on your say-so | An approved reboot holds its alerts until it is back |
| Behind something that is down | depends_on: the parent's alert names it instead |
| lanowl itself lost the network | The router and most of the network are unreachable at once: one alert says so |
Two doors: now, or in the digest
Each problem has a severity, from the device's criticality:
| Severity | From | Door |
|---|---|---|
| ๐ด critical | criticality: critical; a group's majority dark; no internet; on the backup line |
At once, one message per incident |
| ๐ high | criticality: high |
The next digest, once |
| ๐ก warning | criticality: warning; a named service failing on a device that answers |
The next digest, once |
| โช info | criticality: low or info |
The next digest, once. Never makes the network "degraded". |
Everything raised in the same minute leaves as one message.
An alert, as it arrives
๐ด CRITICAL 07:13:00
Home Assistant (192.168.88.12)
unreachable (icmp, Home Assistant)
๐ฆ Hoo. Likely cause (the owl, 07:13): Went quiet at 07:12 right after installing 2026.10.1;
nothing else dropped. The update's own restart.
โ Nothing to do: it should answer again within ten minutes.
The ๐ฆ part is not a second message. On a new critical incident the owl investigates at once
(cadence.llm_on_incident), and its diagnosis is edited into the alert you already have,
which does not notify you again. Under it, a ๐ชถ line for each check the owl ran by itself to
find the cause, โ for one that failed. Internet incidents are explained by lanowl's own evidence
instead, so they never wait for the model.
The recovery
๐ข RECOVERED
Group 'cameras' is back online
back at 02:19:00 ยท unstable for 5m ยท 2 outages
A recovery is sent only after the problem has stayed away for 15 minutes
(alerts.recovery_confirm_s), and it says when it really came back. A device that drops,
returns and drops again inside that window is one incident: the recovery counts the
outages instead of each one paging you.
After a recovery, the same problem stays quiet for an hour (alerts.cooldown_s): a new
occurrence is folded into the count. If it is still down when the hour is over, it pages
once more, so a real problem cannot go silent.
The digest
On a network where nothing has changed, the scheduled audit is skipped (the log says "audit
skipped: nothing changed since 14:05", and the Now card says so); the digest rules below apply
all the same. At least one audit a day runs, and "Audit now" always does
(cadence.llm_when_unchanged).
The digest is sent by the owl's hourly audit (cadence.llm_interval_s), and only when
there is something new to say:
- a problem that is not critical opened since the last digest and is still open;
- the owl's scheduled looks found something new (a configuration change, something slowly getting worse);
- once a day, a problem that is still open (
alerts.digest_repeat_s), so it is not forgotten.
โช lanowl digest โ DEGRADED
26/27 up ยท WAN OK
๐ฆ Hoo. Cam Garage (192.168.88.23) is down and needs you.
๐ Cam Garage (192.168.88.23): unreachable (icmp) โ since 13:35:00
โณ Stopped answering at 13:35; the four other cameras on the same PoE switch answer, so
not the switch. Its DHCP lease is still bound: it lost power or hung. โ Power-cycle it
with its smart plug; if it stays dark, check the cable at the garage end.
When the audit ran checks of its own, the digest ends with them: "The audit looked for itself", then a ๐ชถ line for each, โ for one that failed.
A healthy network gets no digest (cadence.digest_when_ok: false). A problem you were told
about in a digest gets a ๐ข line when it goes away (alerts.notify_recovery,
alerts.recovery_min_severity); one that opened and healed between two digests was never
mentioned, so its recovery is not news either.
Without a model, digests still go out: the monitor's own report, without the ๐ฆ lines.
The internet
| What happened | What you get |
|---|---|
A blip shorter than 90 seconds (wan.watch.min_outage_s) |
Nothing at the time. The digest counts it ("1 brief drop in 24h, longest 45s"), and the dashboard says where it broke. |
| No internet for longer | One ๐ด alert, queued and delivered when the line is back, then a recovery with how long it lasted. |
| On the backup line | One ๐ด alert and one ๐ข when the main link is back (A backup internet line). |
A port slower, a device's own log
Read every hour on each device lanowl has a login for, every minute on the main router over its
API connection (devwatch, Device kinds).
A port comes at once. A log takes the two doors above: at once from a criticality: critical
device, or when the owl reads someone getting in or a device about to fail; from any other
device, the next digest.
| What happened | What you get |
|---|---|
| A port runs slower than usual, read twice a minute apart | ๐ก LINK SLOWER: the port, its speed, its usual one, and on a MikroTik what the other end offers |
| It is back at its usual speed | ๐ข LINK BACK, with how long it was slower โ only if you were told it dropped. A port slower for a week runs at its new speed: that is its usual from then on |
| The same port drops and comes back twice in a week | The second LINK BACK says the device behind it changes the link by itself; then nothing more about that port until a month passes with no change |
| Its log shows something that is not normal, on a critical device | ๐ก or ๐ด DEVICE LOG: one ๐ฆ text from the owl, with the device's own words. Nothing for the first quiet_days (7): kept on the Timeline instead |
| The same, on any other device | A line under Device logs in the next digest: the owl's newest note on it, and how many since the last digest. It never sends a digest by itself: one still waiting at 08:00 goes with the first digest after it |
One event is one block in a digest: when a device's open issue and its log note are about the same thing, the digest shows the owl's line once, under the issue, and the owl's opening summary does not retell what has a line of its own below it.
Beyond Telegram
Every message above can also be sent as a JSON POST to an address of your own (Apprise, ntfy, n8n, a gateway), with or without Telegram: Webhook. It gets what Telegram gets, and can only be narrowed down.
What never reaches you
- A single missed ping.
- A device asleep on schedule, paused, or rebooting because you approved it.
- Twelve cameras behind one switch: that is one incident.
- A device's log when it is normal: Wi-Fi clients, DHCP, a port going down and up once, your own logins from home, lanowl's own.
- Latency alone, a recovery of something you were never told about, a problem already over by the next digest.
- "All is well".
Tuning
| Key | Default | Turn it when |
|---|---|---|
cadence.debounce_fails |
2 | Flaky gear pages on a single bad minute: raise it, or per device |
alerts.recovery_confirm_s |
900 | You want recoveries sooner (and more flaps as separate incidents) |
alerts.cooldown_s |
3600 | A problem that comes back pages too soon, or too late |
alerts.renotify_after_s |
0 | You want a reminder page while a critical problem stays open |
alerts.digest_repeat_s |
86400 | The daily reminder of an open problem is too often (0: never) |
alerts.recovery_min_severity |
info | You want "back online" lines only from warning up, say |
Edit this page on GitHub.