[alsey89]

Closes #28.

switchboard add app 3000 recorded 127.0.0.1:3000. That is one of two addresses meaning this machine, and a server listening on ::1 does not answer on it. Node dev servers commonly bind IPv6 loopback only, so the route was dead the moment it was added.

Measured on a real IPv6-only dev server:

127.0.0.1:3000       FAIL  connection refused
localhost:3000       OK    -> [::1]:3000
[::1]:3000           OK    -> [::1]:3000

The shorthand is now localhost:<port>. Go's dialer tries both families for a name and takes whichever answers (RFC 6555), which is what the browser was doing all along when the user checked localhost:3000, saw their app, and concluded Switchboard was broken.

Why the old failure was expensive

The app plainly worked. ls said down. That is the same word ls uses for a server that is not running at all, so the only diagnostic available pointed away from the cause.

Before and after, against the same live server:

- https://app.test → 127.0.0.1:3000
-   (nothing is listening on 127.0.0.1:3000 yet — the route goes live when your server starts)
+ https://app.test → localhost:3000

Scope

Nothing on disk changes. port = 3000 stays exactly as written. Only the address derived from it moves, so every existing route gets this without anyone editing a config.

ls and doctor come along for free, since both probe through UpstreamAddr.

--upstream is untouched. Naming an address is how you say you meant that address.

Only the outbound dial changes. Every socket Switchboard listens on stays pinned to 127.0.0.1. The security claim is unaffected.

Testing

All packages pass, GOOS=linux and GOOS=windows build clean. Two existing tests asserted the old address and were updated — the behaviour they pinned is the behaviour being fixed.

Worth flagging: go test had these cached as passing after the change. go clean -testcache was needed to see the failures.

No test binds a real IPv6 socket. CI runners cannot be assumed to have IPv6, and a test that silently skips there would give the least coverage exactly where a regression is likeliest to slip through. TestUpstreamAddr pins the returned string instead, which is what a regression would actually change.

Not included

The diagnostic for someone who writes --upstream 127.0.0.1:3000 explicitly against an IPv6-only server. They still get a bare down. Rarer now that the default path works, and worth its own issue if you want it.

Also no CHANGELOG entry, to avoid conflicting with #30, which is still open and rewrites that file. Happy to add one once it merges.