v0.4.0

jcloud can now watch a server, not only check a tree. A new command gathers the observations on the host, jcloud.yml declares what the machine itself must be, and survey tells a new failure from one already known. Templates under contrib/ run all of it on a schedule. jcloud still reaches no network, keeps no state and alerts nobody: the transport, the last run and the alert stay the platform's.

observe: the collector, shipped with jcloud

• jcloud observe [--root <dir>] runs on the host, in the deployed tree, and prints the lines survey reads: each container docker has up and how it stands, the repo digest it runs, each service's pin from pin_file, and each compose profile a manifest declares (on when COMPOSE_PROFILES names it or one of its services is up). Names come from jcloud.yml, so a platform no longer needs a collector of its own. A docker that cannot be asked is an error, never an empty answer.

host: in jcloud.yml

Absent, nothing about the host is asserted. A platform opts in by declaring:

host:
  timers:
    - japode-healthwatch                  # must be active, last run succeeded
    - {name: japode-backup, max_age: 7h}  # …and finished within 7h
  paths:
    - {path: ., mode: "750", owner: runner:runner}
  firewall: ufw                           # installed and active
  disks: [/, /var/lib/docker]
  disk_free: 15%
  certs: [caddy/certificates/*/*.crt]     # a glob is allowed
  cert_days: 14

• Timers are read with systemctl: active, how the unit they fire last ended, and when. An active timer whose job fails every night is reported, as is a backup older than its max_age.
• Disks are held to a free-space floor (df), and certificates to a number of days left, read from the file with crypto/x509, never by connecting to the host name.
• A fact that could not be read (ufw status without root, a failed stat) is reported as unread, never as a pass.

survey

• --since <previous observations> marks each finding new or standing and lists the ones gone. Findings are matched by a stable code and subject, so a reworded sentence, or a disk going from 9% to 8% free, is still the same finding. The exit code is still this run's verdict.
• --json prints {"findings": [...]}, each with code (container.missing, digest.moved, timer.failed, backup.stale, disk.low, cert.expiring, …), subject, says and, with --since, state. Nothing else goes to stdout in this mode.

contrib/: a watch that runs itself

Templates a platform copies; jcloud never runs them.

• contrib/systemd/: a script, service and timer (every 15 minutes) that observe, survey against the last run and leave report.json and report.txt. The unit fails when the server diverges, so it shows in systemctl --failed.
• contrib/github/jcloud-watch.yml: a workflow for a self-hosted runner on the server that keeps one GitHub issue per finding: opened when new, updated only when its wording changes, closed when gone.

agree

• A compose environment value that is a number or boolean (PORT: 2700) no longer stops agree --compose at the decode. It is rendered as compose renders it and compared.
• A file it could not read is named once, with slashes, on every OS. The message no longer repeats the absolute path from the OS error, and the Windows build no longer mixes services\paste\… and services/paste/… in one report.

Not yet

• A manifest's backup implying its own timer waits on Viglet Cloud's manifests (JCT23). Until then, list each timer under host.timers.

Install

Download a binary below, or go install github.com/Japode/jcloud@v0.4.0.