Repository navigation
Description
The gap
ADR-0044 records this under its own Consequences and leaves it to a follow-up, which is this:
Record freshness is unmonitored by construction, today. Nothing in the estate asks whether the name still resolves to the WAN address, and the sticky address means a broken updater stays hidden for months. The check is a
dnsprobe from inside compared against the addressmorpheusholds, and it is not built here.
Why it matters now rather than in the abstract
#442 closed on 2026-09-22 and the remote path is live. That record is now the first thing a peer resolves, and ADR-0044 already named the shape of the failure: the WAN address is sticky, tied to the modem's MAC, so a broken updater changes nothing visible until the one day the address moves. On that day the record is stale, the tunnel does not come up, and the person who finds out is the operator trying to get in from somewhere else.
The fallback ADR-0044 names is the bare address on the withheld list, which works exactly until the hardware swap that invalidates it — and restore-the-firewall.md already warns that the spare's different MAC means a different public address. So the fallback is weakest precisely when the record is most likely to be wrong.
What the check is
A comparison, not a reachability probe. Resolve the endpoint hostname against a public resolver and compare it to the WAN address morpheus currently holds. Asking the estate's own resolver proves nothing, because Unbound would answer from cache.
Neither value may be published. The metric should be the comparison alone — something like homelab_ddns_record_matches_wan as 0 or 1 — so the series carries no hostname and no address. make gateway-state already collects the firewall's view of its uplinks (#353) and is the natural place for it, rather than a new mechanism.
The hostname must come from morpheus's own config.xml, which backup-firewall.sh already treats as secret, and not from anything in this repository. security.md's withheld list is why.
Done when
- A collector emits a boolean comparison metric on the schedule
gateway-statealready runs on, publishing neither the hostname nor either address. - An alert fires when it reads
0for long enough to exclude a forced-update window, with a runbook line saying the two causes are a stale record and a changed WAN address, and that Force update on the pfSense client is the first move. docs/observability.mdrecords what the metric does and does not disclose.- ADR-0044's consequence gains a marked note pointing here, per ADR-0001.
Purchases this needs
None.
Activity