Nextcloud: its own "Cron last run" setup check (GET /settings/setupcheck,
Nextcloud 28+, found by class name) raises an alert when background jobs have
stopped; other checks that report an error are one alert, warnings are shown
on the page. The core version Nextcloud says is available is shown too. Before
28, cron and the checks are unknown, never "fine".
Alerts page: bulk Close... (one reason, written on each incident, undoable) and
Open again. Bulk Wake now leaves closed alerts alone.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The Close drawer can also close the alert's osTicket ticket. osTicket's API
cannot, so this writes to its database with a separate, optional login: the
ticket is set to Closed, a note in its thread says who closed it and why, and
the closed event is recorded, in one transaction. Tables and columns are
checked first, only an open ticket is closed, and a ticket closed by somebody
else at the same moment is left alone. Summary tickets are not closed from here.
"Test closing" (Alert tickets settings) tries every write and rolls it back.
If the ticket cannot be closed the alert is still closed and the reason is
said, noted on the incident and audited.
Not run against a real osTicket: tested against a pretend database only.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Three read-only integrations, each with a monitoring page, alerts, maintenance
silencing and held-when-unreadable handling.
Nextcloud: database needing upgrade (critical), maintenance mode, and the data
disk running low; users, files and app updates with an admin app password.
NetBox: no background worker running, counts, and a list of the servers this
app manages that have no address in NetBox (a list, not an alert).
n8n: an active workflow whose last run failed, clearing when it next succeeds.
Disk messages for Immich, Paperless and Nextcloud now use the same units as
the pages that show them.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Three read-only integrations, each with a monitoring page, alerts, maintenance
silencing and held-when-unreadable handling.
Prometheus: scrape targets that stop answering (5 minutes, one alert per job)
and alerts that Prometheus's own rules fire (critical when labelled so).
Immich: disk under the library filling, failed background jobs, counts.
Paperless-ngx: worker/database/Redis in error, documents that failed to
import, index and sanity check, disk. Immich and Paperless also in /metrics.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
New Authentik integration (read-only, every 5 minutes): version, workers,
outposts and their health, certificates and API tokens that expire, failed
sign-ins and configuration errors. Monitoring -> Authentik page.
Alerts: an outpost that serves apps has gone quiet (critical), certificate or
token about to run out (critical close to it), no worker, failed sign-ins,
configuration errors. Held by maintenance, held when unreadable, in /metrics.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
New Jellyfin integration (read-only, every 2 minutes): version, restart wanted,
streams and transcoding, failed scheduled tasks, item counts, failed sign-ins.
Monitoring -> Jellyfin page with a peak-streams history, three alerts, /metrics.
When Jellyfin runs on a managed server, an update run asks first if people are
streaming (409 "watching" unless startAnyway), and a patch plan waits up to two
hours for them before updating that server.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Duplicati reports each operation to a secret address (its own send-http-url option), so
nothing calls Duplicati and no password for it is kept. Each machine is a source; each job
is learned by name and shows its last result, duration, size, changes and a history.
A failed run is a warning (critical from the second in a row); one that has not succeeded
for longer than expected (per job, 2 days by default) is a warning, critical at three
times that. Maintenance on the machine's server silences them; /metrics series added.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
With Dockhand connected, an image the app recognises (about 20 with a health endpoint of
their own, plus a few dozen web UIs) is offered as an endpoint check on that app's own
health address, with what a good answer looks like set up as a value check: Immich's ping
says pong, Nextcloud's status.php is not in maintenance, Gitea's health says pass, and so
on. Asked on the Docker host's address at the port the container publishes. Container
ports are now read from Dockhand's list.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
From Dockhand, on the health pass: a container that keeps restarting (critical, with
its exit code), one whose own health check fails, one that stopped unexpectedly, one
created and never started for an hour, and a Docker host Dockhand cannot reach. A stop
is only news when it was seen happen, so stacks switched off long ago stay quiet.
A Leave alone button on the Docker page and a list in Settings; on by default.
Also fixes a date-dependent storage-forecast test that pinned its clock while the code
under test reads the real one.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
New UPS integration (NUT or SNMP v2c, read only). Monitoring -> UPS shows charge,
runtime, load and mains voltage with history and the times mains power was lost.
Mains lost and a nearly empty battery are critical alerts; battery to be replaced
and a heavy load are warnings. The UPS is a node in the dependency map and counts
as down while on battery, so servers that go offline with it are folded under its
alert; the page lists what is on it in the order to shut it down. /metrics series
and docs added.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- New docs/ folder: SSO, API, Prometheus and Grafana, webhooks, heartbeats and calendar, how monitoring works, and an index; the old root API.md is split across them
- Read API: alerts carry snoozedUntil
- DATABASE.md: table and migration counts, indexes, relations, integration config, settings and internal flags, delete rules, retention
- ROLES.md: the sidebar table rebuilt from the menu
- NOTIFICATIONS.md, INTEGRATIONS.md, BACKUP.md, AGENT_UPDATE.md: what had drifted
- README: links to the guides; accounts cannot be deleted in the app
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- Racks page and a network diagram (UniFi tree with servers, printable)
- Notes on servers and devices; To do list shown in Coming up and My work
- Agents report the operating system (Linux 7, Windows 6); end-of-life badge and alert
- Suggested dependencies from where a service is reached
- Nginx Proxy Manager integration: proxy hosts page, certificate and host alerts, suggested checks
- Home Assistant integration: watch chosen entities, history chart, limit and unavailable alerts
- TimeChart gets a "value" scale that fits negatives; docs updated
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
UniFi: the controller's client list is matched to servers by IP address, so a
server is connected through a switch or access point (and port). That is a
dependency made by itself, shown on the server page, the Network devices page
and the Connections map. IP Addresses can sync the devices' addresses.
A server's page is split into Overview, Health, Network, History, Tasks and
Settings, by ?view=; alerts link to the section they are about.
My work lists what has your name on it: alerts you have, maintenance you
started (extend, end with undo) and update runs you started, with a badge.
Settings has a search box (and the command palette lists every setting).
Removing a dependency, admin link, runbook, hand-made IP address, port opening
or restore test is undone from the message instead of asked about first.
Two backup tests used a fixed date and failed as days passed; they follow the
real clock now.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A UniFi controller (API key on Network 9+, or a read-only local login on any
version) is read every five minutes. Its switches, access points and gateways
are kept with state, IP, uptime, firmware and uplink; its networks and Wi-Fi
names beside them (passphrases are never read). Nothing is written to the
controller.
Network > Network devices lists them. Devices are a new node type in the
dependency map: uplinks become dependencies by themselves, others can be
written by hand. A device offline past a limit is an offline alert, folds what
depends on it, can be silenced with maintenance and have runbooks.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Notifications gains Alert tickets, Display gains Badge colours, and a new Data tab holds Backup,
Logs and Cache. Domains is the 'Domain names' tab of DNS. Old addresses redirect; the docs, the
Privacy page and the links from alerts and the agenda name the new places.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- Servers can say what they run on (or it is worked out from the Proxmox node); when the host is
down the guests' offline alerts fold under it: held back, not counted, not notified, still recorded.
- Recurring maintenance schedules made into ordinary windows a week ahead (maintenance_schedules).
- Run a Semaphore template limited to a server from its page or an alert, and watch the run.
- Agents report pending updates (apt, dnf, zypper, Windows Update); an Updates page lists them.
- Migrations 0024 and 0025.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Uptime Kuma often watches a host by another address than the one it was added
with here (a Proxmox host monitored over Tailscale, say). Address aliases on
the Maintenance page say that an address belongs to a server, integration or
DNS provider, so the sync puts the window there; matching stays exact.
Migration 0018 adds maintenance_aliases.
Also fixes monitors whose empty hostname/URL/port arrive from Kuma's metrics as
the text "null": they were taken as an address, hiding the real URL.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A monitor whose address is the host of an integration's or DNS provider's own
address now gets a maintenance window there, alongside the existing match
against servers. Matching stays exact (no inference through DNS), and monitors
that match nothing are listed with their address in the sync result.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Uptime Kuma shares its maintenance windows only over its dashboard Socket.IO
connection, which takes a login. With a login username and password on the
integration, "Sync schedule" on the Maintenance page (and a 15-minute
background pass when switched on) reads each window's running and next time
slot and the monitors it covers, and makes a maintenance window on every
server those monitors match. Windows that haven't started are listed as
Scheduled and silence nothing until they begin; ones ended or cancelled by
hand stay ended, and ones removed in Kuma are removed here.
maintenance_windows gets a source column (migration 0017) and active windows
now require their start to have passed. Adds socket.io-client, and tests
against a stand-in Uptime Kuma speaking the same protocol.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The ticket API was first verified with a test ticket (key and IP restriction, creation, message text arriving
as written), then with a ticket opened for a real alert. A summary ticket (many problems at once) has not been
seen on a real osTicket yet. Also records the two mistakes the real osTicket showed that the stand-in server
had hidden: percent-encoded text, then base64.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Percent-encoding put "%20" in every space of a real ticket, and the base64 that replaced it arrived
as base64. osTicket (Format::parseRfc2397, via ApiJsonDataParser::fixup) takes the message's type from the
"data:" prefix and uses everything after the first comma exactly as it is, decoding nothing; its own
documented example is plain text after the comma. The text is now sent raw (JSON takes care of quotes and
line breaks; later commas are fine, since only the first one splits).
The test now mimics that parsing -- split at the first comma, decode nothing -- and requires the text to
arrive exactly as written; it fails if percent-encoding or base64 comes back. The stand-in server in the
earlier test had decoded the encoding, which is why neither mistake was caught. INTEGRATIONS.md says what
happened and that it hasn't been re-checked against a real osTicket.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A test ticket sent to a real osTicket arrived with every space shown as %20. osTicket takes the
message's type from an RFC 2397 data URL and strips the "data:" part, but leaves percent-encoded text
as it is. The message is now sent as base64 (data:text/plain;charset=utf-8;base64,...), which osTicket
decodes (it is how it takes attachments), and which works whether or not a given osTicket version
decodes percent-encoding.
The stand-in osTicket in the test had decoded the percent-encoding, which is why this wasn't caught:
the test now requires base64 and fails if any percent-encoding is in the message. INTEGRATIONS.md says
the ticket API is only partly verified against a real osTicket, and what that first test showed.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
README, INTEGRATIONS (the osTicket ticket API: fields, making the key, what is sent, what it can't do,
what wasn't verified), NOTIFICATIONS (the rules, in one table), ROLES, DATABASE (the new table, the
settings key, the internal flags, what deleting an integration does), and the Privacy page (the one new
thing that leaves the app: an alert's text, to your osTicket).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
osTicket's own REST API only supports creating tickets, not listing or
reading them, so this reads osTicket's MySQL/MariaDB database directly
with a read-only user instead - the only integration in this app that
isn't a REST API. Joins the ticket, status, priority, department,
staff, team, and user tables, filtered to tickets in the "open" state
(status names are customizable per install, but that state flag isn't).
Surfaces per-ticket subject, priority, department, assignee, requester,
and osTicket's own overdue/awaiting-reply flags, plus a page at
/osticket and a Dashboard widget with open/overdue/awaiting-reply
counts.
Not verified against a live instance: unlike the HTTP-based
integrations, there was no way to fake a MySQL server to test against
in this environment, so the query is built from osTicket's published
schema but has never actually run against a real database. See
INTEGRATIONS.md for the read-only grant needed and further caveats.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Proxmox VE already shows whether the last vzdump push to PBS succeeded, but
has no visibility into PBS's own backup verification, GC/prune health, or
host status. This adds PBS as its own integration (own adapter, page, nav
entry, and Dashboard widget) that reads datastore usage and, for every
stored snapshot, its verification state directly from PBS.
A new daily check (mirroring the existing Proxmox backup-failure check)
notifies when a snapshot has failed verification or a datastore couldn't be
read, with its own toggle in Settings -> Notifications and its own
maintenance-window silencing.
Not verified against a live PBS instance — built from PBS's published API
docs and a scratch test against a mocked PBS server exercising the adapter's
parsing and auth-header format (PBSAPIToken uses a colon separator, unlike
PVE's PVEAPIToken which uses =). See INTEGRATIONS.md for details and the
"not verified" caveat.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New "Sync from phpIPAM" action on IP Addresses, alongside the existing
"Sync from Tailscale" / "Sync from Proxmox" ones and built the same way:
a new integration type (config in-app, URL + API app ID + app token,
credentials encrypted at rest) that this page pulls from on demand.
Deliberately import-only, not a full integration -- no dedicated page,
dashboard widget, or nav entry, since that's all this was asked for.
Auth is phpIPAM's static "App token" method: create an API app under
Administration -> API with its security set to "SSL with App token", and
its one-time code goes straight in as the `token` header (also sent as
`phpipam-token`, in case a given version expects that name instead) --
no login call, no token to renew. The user/password "User token" method
isn't implemented.
Addresses are read the standard way: GET /subnets/, then GET
/subnets/{id}/addresses/ for each, rather than assuming a single
"all addresses" endpoint exists on every version. phpIPAM wraps every
response as {code, success, data} -- including an empty result: a subnet
with nothing in it answers success:false, message:"No addresses found"
rather than success:true, data:[]. That's read as "nothing here", not a
failure; anything else with success:false throws with phpIPAM's own
message. One subnet failing outright (e.g. the app lacks permission on
it) is skipped with a note rather than aborting the whole sync. Every
address field is read defensively -- optional, independently
type-checked -- so a field phpIPAM renames or drops in some version
leaves that value blank instead of breaking the import.
Imported entries: label from hostname or description, "phpIPAM" as
vendor, the subnet's own description (or its CIDR, if it has none) as
location, and description/note/MAC folded into notes. Existing sync
plumbing (upsertSyncedEntry) gained a location parameter so this and any
future sync can set it; the two existing syncs pass null, unchanged.
Endpoints, the token header, and the address/subnet field names are
cross-checked against phpIPAM's own published API documentation. Not
verified against a live instance -- there wasn't one available while
building this, so if a real sync comes back empty or with the wrong
fields, that's the next thing to check.
Verified with 23 backend checks against a fake phpIPAM server matching
that documented shape (the empty-subnet quirk, a subnet that fails
outright, malformed/missing fields, a non-JSON response) and the real
route (added/updated/skipped counts, a manually-entered IP never
overwritten, roles, no-enabled-integration, upstream failure surfaced
per-integration rather than as a 500, audit entries) plus a browser check
of the real IP Addresses page against the real routers: the sync button,
its result message, re-syncing (updates rather than duplicates), the
manual entry staying untouched, and the viewer view.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New "Import from Uptime Kuma" action on the Maintenance page (operator):
pick an Uptime Kuma integration and a duration, and it starts (or
extends) a maintenance window here for every server whose address
matches a monitor Uptime Kuma currently reports as being in maintenance,
reusing the existing monitor-to-server matching from the Uptime Kuma
integration itself.
Uptime Kuma's metrics endpoint only exposes a monitor's *current* status,
not its scheduled start/end time (there's no API for that), so this
deliberately doesn't try to mirror Uptime Kuma's own schedule -- it starts
a window for the duration you choose, the same bounded/required-end
window this feature has always used. Running it again while Kuma is
still in maintenance extends the same window rather than stacking a
second one; when Kuma later shows nothing in maintenance, already-active
windows are left alone rather than force-ended, since ending them isn't
something only Uptime Kuma's state should decide. Two Kuma monitors that
match the same server are deduped to one window. Monitors with no
matching server are reported back by name so nothing is silently missed,
and monitors that aren't in maintenance are ignored entirely.
The manual "Start maintenance" endpoint's start-or-extend logic (dedupe,
pruning old rows, the response shape) is now a shared
services/maintenance.ts function instead of living only in that route
handler, so the import path can't drift from how a manual window behaves.
Likewise the server-matching helper gained a small toMatchableServers()
so the existing Uptime Kuma monitors route and this new one build the
same match input the same way instead of each parsing server rows on
their own.
No schema change -- imported windows are ordinary maintenance windows;
their Uptime-Kuma origin is only in the reason text ("Imported from
Uptime Kuma (<integration>): <monitor>"), visible in the Active table and
the audit log like any other window.
Verified with 28 backend checks (the refactored manual start/extend
flow as a regression check; roles; unknown/wrong-type/disabled
integration; validation; nothing-in-maintenance; matching including
same-server dedup and unmatched monitors; re-running extends rather than
duplicating; windows left alone once Kuma exits maintenance; upstream
failure; audit entries) and by driving the real Maintenance page against
the real routers in a browser: import, re-import (extends), the
nothing-in-maintenance state, and the viewer view (no edit card, active
windows still visible). Real dev database mtime untouched.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New integration, following the existing pattern: config in-app (URL +
API key, credentials encrypted at rest), its own Uptime Kuma page, an
Integrations list entry, a Dashboard widget, and diagnostic-log/
integration-down-alert coverage for free via the shared withDiagLogging
wrapper. Read-only -- no start/stop equivalent exists for a monitor.
Uptime Kuma has no conventional REST API (the dashboard talks to it over
Socket.IO); researched before writing any code, since guessing wrong here
would have cost real time. The one machine-readable, authenticated
endpoint that lists every monitor is its Prometheus exporter at
GET /metrics, gated by HTTP Basic auth -- an API key as the password with
the username left blank on current installs, or the real dashboard
login on installs from before the API-key feature existed. This adapter
authenticates the same way and parses that endpoint's text-exposition
format itself (metrics: monitor_status, monitor_response_time,
monitor_cert_days_remaining, monitor_uptime_ratio; labels: monitor_id,
monitor_name, monitor_type, monitor_url, monitor_hostname, monitor_port),
verified against the documented metric/label set and the actual upstream
source (server/prometheus.js). A malformed line is skipped rather than
failing the whole scrape.
"What server is being monitored for what": each monitor's target (an IP
for TCP checks, or the hostname out of the URL for HTTP/keyword checks)
is matched against your servers' own IPs and hostnames -- reusing the
same kind of match already used in the consistency report -- and linked
to that server's page. Monitors with no single network target (groups,
push monitors, DNS/keyword checks with a complex URL) are left unmatched
rather than guessed at. Uptime Kuma's tags aren't read, since the
Prometheus endpoint doesn't reliably distinguish a tag label from any
other label it might add later.
The username field is the first genuinely optional integration config
field this app has had; IntegrationField gained an `optional` flag
(server validation and both the add/edit web forms honor it) rather than
special-casing Uptime Kuma.
Verified with 48 backend checks (Prometheus text parsing including
escaped quotes, decimals, negative numbers, and malformed lines; TCP vs.
HTTP target/port extraction; every documented status code; server
matching by IP, hostname, and short name, including no-match cases; the
route's real HTTP round trip against a fake Uptime Kuma server, wrong
credentials, upstream failures, roles, wrong/disabled/missing
integration, diagnostic-log entries; the optional-field validation rule)
and by driving the real page and the real Dashboard widget in a browser
against the real routers, including CSV export and column sorting. Real
dev database mtime untouched.
Not verified: a real Uptime Kuma instance. Everything here was checked
against Uptime Kuma's documented metric format, its actual upstream
source, and a fake server built to match both -- not against a live
installation. If your instance's /metrics output differs from what's
documented (older version, unusual monitor types), the parser should
degrade to an empty or partial monitor list rather than error, but that
degradation itself hasn't been observed against the real thing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Proxmox already runs vzdump backups, but nothing in the app said
whether they were actually succeeding — a silent backup failure is
one of the more dangerous blind spots a homelab admin can have. Adds
a "Backups" card to the Proxmox page: configured backup job
schedules (storage target, which guests, enabled/disabled) from
GET /cluster/backup, and recent vzdump task history per node from
GET /nodes/{node}/tasks?typefilter=vzdump, with a banner at the top
if the most recent run didn't succeed.
New "Proxmox backup failed" notification toggle under Settings ->
Notifications, on the same daily schedule as the other checks. The
scheduler checks each node's own most-recent vzdump run independently
(not just the single most recent task overall) so one node's healthy
backup can't mask another node's failing one in a multi-node cluster.
Known limitation, documented in the adapter's own header comment:
Proxmox's task list doesn't reliably expose which specific guest
failed within an "all guests" job — only the task's own log text has
that — so this surfaces job- and task-level status rather than
guessing at per-guest outcomes.
Verified against a fake Proxmox server (real self-signed HTTPS, since
the adapter's node:https usage can't be monkey-patched under ESM)
reproducing the documented /cluster/backup and task-list response
shapes: job parsing (all-guests+exclude vs specific-vmids+disabled)
correct, task OK/failure parsing correct, and the critical multi-node
scenario confirmed — one node's failing latest run flagged, the
other's healthy latest run correctly left alone, with exactly one
notification of the right content. This reproduces Proxmox's
documented API shape rather than a live-verified one; flag if the
real cluster's response differs in some way this didn't anticipate.
Confirmed the real dev database's mtime was untouched throughout.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Each adapter performs write actions, not just reads (start/stop a
guest, edit a DNS record, rerun a CI job, etc.), so a read-only
credential silently works for the dashboard views but fails the
moment you use an action. Lists the exact endpoints/permissions
needed per target system, drawn from each adapter's own auth code and
header comments (e.g. Proxmox's Sys.Audit/Datastore.Audit split,
Azure's DNS Zone Contributor role, Loopia/Pi-hole/cPanel having no
scoped-credential option at all).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>