Repository navigation
Description
Filed so the design is written down before it is wanted, not because anything
currently needs publishing. The estate exposes nothing today, and that is
correct.
The design, when the time comes
Never expose the lab directly. Instead: a cheap VPS running WireGuard and a
reverse proxy, tunnelling to a host on an isolated home DMZ segment.
Public DNS points at the relay; the relay terminates TLS and proxies over the
tunnel; the internal host has no inbound path of its own. A compromise of the
exposed service reaches the DMZ host and stops there.
The relay's own firewall scoped tightly: the tunnel port and 80/443 open to the
world, SSH restricted to the operator's current public address, and
everything else closed.
What it would cost
$5–24/month for the VPS — Hetzner at the low end, DigitalOcean at the high —
plus $10–20/year for a domain. Avoid .xyz, .biz, .top, .zip and
.ru; they attract blocklists regardless of what you put on them.
The two decisions this trips, and they are the reason to file it early
- ADR-0022.
External reachability is one of three triggers that end the SSO deferral.
Publishing anything starts that clock, and the ADR is explicit that
re-accepting is allowed but arriving there by never looking is not. - ADR-0011. The wiki is
deliberately internal, and the reasoning — a published thing depends on more
components that can fail, and an expired internal certificate fails quietly
and looks like a browser problem — generalises to anything else considered for
publication.
What would make this worth doing
A writeup site or a portfolio page. Given the career direction that is a real
possibility rather than a hypothetical, and it is the one use that justifies the
trigger — because the whole point of publishing would be that people can see it.
What it waits on
#442, which builds the WireGuard
understanding this depends on, and a reason.
Purchases this needs
A VPS and a domain, $5–24/mo + $10–20/yr. Nothing until there is something
to publish.
Corrected 2026-09-19
"#442, which builds the WireGuard understanding this depends on" — the understanding is written now, in ADR-0042 and open-the-remote-path.md (the AllowedIPs asymmetry, the no-MASQUERADE argument), even though #442 is unbuilt. Read those rather than waiting on the build.
One overlap worth knowing before pricing a relay: ADR-0042 §5's unsolved prerequisite is a reachable endpoint, and a VPS with a static address is one of the answers. That is #530. If either this or #530 is taken, the two should be read together — a relay answers the endpoint question for free, and the endpoint decision may make a relay unnecessary.
Corrected 2026-11-08
The 2026-09-19 correction is itself stale.
- WireGuard to the lab jumpbox, terminating on VLAN 30 and not on Winterfell #442 is built (2026-09-22). The WireGuard understanding is no longer only on paper:
open-the-remote-path.md's status block has the measured result. That also means "the estate exposes nothing" now reads publishes nothing — there is one inbound path, the tunnel's UDP forward, and it reaches the lab only. - The WireGuard endpoint: a static WAN address or dynamic DNS — the decision ADR-0042 §5 said would have its own issue #530 is decided, and not by this. ADR-0044 answered the endpoint with dynamic DNS from
morpheusand, in §5, declined the relay for that purpose: a relay bought for an address is this issue's cost without this issue's reason. Its What would reopen this is the contract with this issue — the day the relay exists, peers'Endpointmoves to it and the DDNS client is removed; and if CGNAT arrives on the WAN, this (or an outbound-only overlay) becomes the endpoint answer regardless. - ADR-0022 now reads with ADR-0075. Trigger 1 is spent; trigger 2, external reachability, has not fired for the tier and is exactly what publishing would fire.
What it waits on is now one thing: a reason. Nothing else blocks it.
Noted 2026-11-09 — what the operator's existing Cloudflare account changes
The operator already runs a portfolio site on Cloudflare, on a domain they pay for. That settles one thing now and changes the shape of the design if it is ever taken.
- The portfolio is not this issue's reason. It is already published, off-estate, with no path to the house — the shape this issue exists to protect. Writeups and project pages are static and belong there too. The reason this issue still waits on is something that must be served live from the estate and cannot be exported to the existing host on a schedule.
- The domain line in Purchases is already paid. A relay would take a subdomain of the existing domain; the remaining cost is the VPS alone. The existing domain must not be pointed at the house to stand in for ADR-0044's dynamic DNS record — ADR-0044's What would reopen this names that case as this issue's design, not a DNS change.
- Cloudflare Tunnel is the alternative to weigh against the VPS.
cloudflaredon the DMZ host dials out; Cloudflare serves the hostname and carries requests back down that connection. It is the outbound-only overlay ADR-0044 already names: no monthly cost, no port forward onmorpheus, no relay host to patch, and the DMZ host has no inbound path at all — closer to this issue's own goal than the VPS design. The cost is trust: Cloudflare terminates TLS and sees every request in plaintext, and becomes part of the estate's trust boundary, where a VPS keeps that in the operator's hands. Choosing between them is the ADR this issue would produce; neither avoids ADR-0022's trigger 2.
Activity