A host that drops out of the scheduled check now says so
The manual /check has always reported an unreachable host. The nightly run printed it to the container log and stopped there — which is backwards, because the unattended run is the one nobody is watching.
@famewolf added two hosts to one instance, the SSH out of the container could not authenticate, and his nightly check quietly covered one machine out of three. The reason was in a log he had no reason to open, and the only sign was updates that never appeared for two of his boxes.
Said once per outage now, and again when the host comes back. A message every night about the same dead box is a message people learn to ignore, which is the same defect wearing a different coat. It goes to every channel, not only Telegram.
/restore on Discord could not import what it needed
name 'json' is not defined. discord_bot.py has no module-level import of it — every function that needs it imports it itself, which is the style in that file — and the new one did not.
And the fix I shipped in 2.12.2 for this was wrong. The original code caught that NameError in the same try as the download and answered "I could not read that attachment", so the real fault was invisible, and I blamed Discord's CDN — in a changelog, and to the person who reported it. @NotRetarded's second screenshot settled it: once the messages told the truth, the failure moved past the fetch and named itself, which means the download had been working the whole time.
The User-Agent added in 2.12.2 is still correct — every other request this bot makes identifies itself — but it fixed nothing, and saying otherwise was a guess dressed up as a finding.
There is now a check across the whole codebase that every standard-library module a function touches is one it can actually see. A missing import is invisible until the line runs, which on an error path can be weeks.