A sortable ledger of every box in the fleet: purchase year, hardware notes, last known-good kernel, and which machine pays for itself and which is kept purely for sentimental reasons. Plain TOML underneath; the reporting layer is a forty-line shell script.
A tiny runner that walks a directory of config templates, renders them against a host's facts and prints a unified diff against reality. No agent, no daemon — it runs over an existing SSH session and exits. The point is to catch drift before it becomes an outage.
A growing archive of post-mortems written while memories were still warm. Each entry answers three questions: what broke, what made it visible, what to change in monitoring so the next report is not about the same thing.