Commit Graph
2 Commits
Author SHA1 Message Date
bobbanandClaude Sonnet 5 4118062405 Alert when a Semaphore template or Gitea workflow run fails
Every 15 minutes (on the existing health-check timer) the app reads the
latest run of each Semaphore template and each Gitea repo's latest
workflow run. A failed one raises one notification, and another when a
later run succeeds. It is state-based like the server health alerts, so a
job that fails every night alerts on the first failure, not every night.
Gitea alerts include the run's link. Toggle: Settings > Notifications.

What counts:
- Semaphore "error" is a failure, "success" is a pass. A run that is
  waiting, running, stopped by hand or rejected is neither, so it leaves
  the previous state alone: a run in progress must not clear a failure it
  hasn't fixed yet, and a manual stop isn't a failure.
- Gitea failure/success likewise; running, waiting, blocked, cancelled and
  skipped leave things as they were.
- A failing template that gets another failing run does not re-alert.

Not mistaking "couldn't read" for "fixed":
- Semaphore's template listing swallowed per-project errors, so a project
  that failed to load looked like a project with no templates. A new
  checkTemplates adapter method reports which projects failed, and their
  failures are held rather than cleared.
- Gitea reports a run it couldn't fetch as null, the same as "no runs";
  both leave the repo's state alone.
- An unreachable integration holds all of its failures. Nothing is cleared
  or re-announced while it is down.

The first pass only records what is already failing without announcing
it, so upgrading (or adding an integration to a fresh install) doesn't
produce a wall of alerts about months-old failures. That baseline is not
spent while nothing could be read.

Maintenance windows on a Semaphore or Gitea integration silence its
failure alerts with the same rules as the health alerts: a problem that
starts during a window alerts when it ends, and one already announced
stays known. The diff logic is reused from the health monitor rather than
copied. Maintenance page text updated.

Known limit: for Gitea this follows the repo's most recent run on any
workflow or branch, matching what the Gitea page shows; a failure in one
workflow can be masked by a later success of another.

Verified with 53 checks against fake Semaphore and Gitea servers and a
webhook receiver: classification, baseline (including not being consumed
when nothing is readable), single alert per failure, no repeat, in-progress/
stopped/cancelled runs, recovery and re-failure, unreadable project,
unreadable integration, run-fetch errors, maintenance windows (silenced,
then announced after), the toggle, disabled integrations and repos without
Actions. Real dev database mtime untouched.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-09-26 03:11:59 +02:00
bobbanandClaude Sonnet 5 1688de3ea2 Alert when a server goes silent, a disk fills up, or a Synology volume degrades
The data was all being collected (agent last-seen, per-disk usage,
Proxmox storage, Synology volume/disk health) but nothing acted on
it, so a dead server or a full disk was only noticed by opening the
right page.

A new health pass runs every 15 minutes (matching the agent's default
report interval) and raises one notification when a problem starts and
one when it clears: a server's agent silent past a threshold (default
60 min), a server disk / Proxmox storage or root filesystem / Synology
volume at or above a usage threshold (default 90%), and a Synology
volume or disk that isn't "normal", has bad SMART, bad sectors past the
threshold, or life remaining below it. Both thresholds and an on/off
toggle live under Settings -> Notifications.

The parts that make this trustworthy rather than noisy:
- A problem is keyed by identity, so it alerts once and not every run;
  a shared Proxmox storage listed by every node is one problem, not
  one per node.
- Active problems persist across restarts, so a rebuild doesn't
  re-alert everything already known.
- If a source can't be read on a given run (Proxmox/Synology
  unreachable, one node lacking privileges) its existing problems are
  held, not reported "cleared" and then re-alerted when it comes back —
  the integration-failure alert already owns "the integration is down".
- For 20 minutes after startup server-derived problems are held too:
  agents couldn't report while the app was down, so judging them then
  would report every server offline after any restart.
- An offline server's disk figures are stale and are not judged; a
  server that never reported has no agent and raises nothing.
- Tracking continues while the toggle is off (only sending is gated),
  so turning it back on doesn't dump every long-standing problem.

Timestamps without a zone (SQLite's format) are read as UTC; the test
runs on a UTC+2 machine, where reading them as local time gives a
different answer.

Verified with 32 checks: the evaluation rules and the state diff as
pure functions (exact thresholds, the proxmox:1 vs proxmox:10 prefix
trap, the flapping sequence), then a whole pass against a real
Proxmox adapter talking to a fake HTTPS cluster (one node returning
403, the whole API down, a shared storage on two nodes, a node that
recovers), a webhook receiver, the real DB, and the persisted state.

Not exercised end-to-end: the Synology collection path — its rules are
tested on data shaped exactly like the adapter's output types, but I
did not stand up a fake DSM. Real dev database mtime untouched.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-09-26 02:15:19 +02:00