Metric history: the per-minute sampler (clusev:sample-metrics) now reads
cpu/mem/disk from the `servers` DB row — the same source the live gauges read,
shared via MariaDB — instead of a cross-container Redis cache the poller may
not share in prod. It gates on last_seen_at freshness (skip offline/unpolled)
and keeps `load` as a best-effort cache read. History now fills whenever the
gauges have data, independent of the cache driver or container split — the
prod detail page no longer shows "Noch keine Verlaufsdaten" after hours.
Service-action failures are persisted: a failed systemctl start/stop/restart
writes a red service.<op>.failed AuditEvent with the FULL systemctl reason in
meta['output'] (mb_scrub-ed, 4000-cap), shown expandable in the audit log — a
truncated, ephemeral toast is no longer the only trace. The failure toast is a
short pointer to the audit log, level=error, and error toasts now linger 9s.
Alert delivery: a warning banner appears on the alerts page when NO channel
would reach anyone (no SMTP + no safe webhook + no Gotify). A new Gotify push
channel (server URL + encrypted app token) posts title/message/priority with
the token in X-Gotify-Key; a self-hosted LAN Gotify is allowed (scheme-checked,
not private-range-blocked like the SSRF-guarded webhooks) — a real push when a
service goes down, no e-mail needed.
Secret-retarget hardening (Codex-found class): changing a target while leaving
its secret blank could send the OLD secret to the NEW endpoint. Closed for the
Gotify token, the SMTP password (save + test-send + the runtime mailer), and
the host-terminal SSH credential (username/port change) — each drops/forbids
the stale secret before the new target is stored. Setting::uncached() reads the
security-sensitive keys DB-direct so a cache-repopulation race can't re-pair an
old secret with a new target. Three Codex rounds, clean; 778 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ROOT CAUSE of "alert fires but no e-mail", third and final layer: the
metrics poller (clusev:poll-metrics, a long-lived supervisor process) is
what queues alert mail — and Laravel bakes the DEFAULT MAILER NAME into a
mailable at Mail::to()->queue() time. A poller that booted before SMTP
was configured bakes `log` into every alert e-mail forever; the worker
then faithfully "sends" them into laravel.log no matter how correct its
own config is. This is why manual test sends (fresh processes → smtp)
arrived while real alert mails (poller → log) never did.
Fix: extract the SMTP-settings application into App\Support\MailSettings
(::apply(), now with Mail::purge on change so long-lived processes
rebuild the transport) and call it from all three process shapes:
- web requests → AppServiceProvider::boot()
- queue worker → Queue::before
- poller/loops → AlertNotifier::notify() right before queueing,
so the CURRENT mailer name is baked in.
Also: enterprise HTML alert e-mail (frontend-design). Dark Tactical
Terminal card — signal rail + status pill (red ALARM / green BEHOBEN),
rule headline, mono terminal readout (server/metric/value/threshold),
bulletproof dashboard CTA, CID-embedded logo (no external assets; renders
in Gmail/Outlook with images blocked), full de/en, plain-text part kept
as fallback. Offline rules render state text (offline/back online)
instead of a bare 0/1.
Verified: app restarted (poller reloaded), synchronous + queued sends to
the admin both deliver over real SMTP (476ms send, no new log-mailer
entries). 750 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
boban