[Abioy]

When hostname is set to 0.0.0.0 (listen on all interfaces), the service registration file recorded http://0.0.0.0:port as the client connection URL. Since 0.0.0.0 is a bind wildcard rather than a routable address, client auto-discovery could fail to connect to the background service.

Normalize 0.0.0.0 to 127.0.0.1 and :: to ::1 in both the registration URL and the incumbent check probe URL, matching the existing behavior in the remote tunnel target construction.

Issue for this PR

Closes #49796

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

When hostname is set to 0.0.0.0 in ~/.config/opencode/service.json (listen on all interfaces), the service registration file at ~/.local/state/opencode/service.json recorded http://0.0.0.0:port as the client connection URL. Since 0.0.0.0 is a bind wildcard, not a routable destination address, the second opencode client's auto-discovery fails to connect to the running background service.

service-registration.ts used HttpServer.formatAddress(address) which passes the bind hostname directly into the URL. The remote tunnel target construction already normalizes 0.0.0.0 → 127.0.0.1, but the registration path missed this step.

Added formatClientAddress() to normalize 0.0.0.0 → 127.0.0.1 and :: → ::1 for the registration URL, and applied the same normalization in serviceURL() used by the findIncumbent startup probe.

How did you verify your code works?

Tested locally: with hostname: "0.0.0.0" and port: 14096 in ~/.config/opencode/service.json, the registration file now contains http://127.0.0.1:14096 instead of http://0.0.0.0:14096, and a second opencode terminal connects to the running service without issues.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

---

Added by the opencode bot (AI-generated).

Related: #49796

Public API: registration file url is now loopback instead of 0.0.0.0/:: for wildcard binds (compatible with existing readers)

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[opencodejr]

As written, this breaks service registration for every service start, not just wildcard binds. formatClientAddress reads address.hostname and checks _tag === "UnixAddress", but Effect's NetAddress.SocketAddress is InetAddressV4 | InetAddressV6 | UnixPathAddress, and the IP is in address.address. So hostname is always undefined and hostname.includes throws before the registration file is written.

How I checked it:

• tsc -b in packages/cli: 5 errors in service-registration.ts (TS2367 and TS2339); the base is clean.
• bun test test/service.test.ts: 12 failures with TypeError: undefined is not an object (evaluating 'hostname.includes') from formatClientAddress; the base passes 28 with 0 failures.

Public API: the url in the registration file (~/.local/state/opencode/service.json) would change from the wildcard address to loopback for services bound to 0.0.0.0/::. No CLI, config or SDK changes.

Other concerns:

• pair.ts treats a loopback registered URL as "only reachable locally". It then prints the SSH port-forward hint and "run opencode service set hostname 0.0.0.0". A service already bound to 0.0.0.0 would now register 127.0.0.1 and get that advice. Either keep the wildcard in the registration file and rewrite it on the client side, or adjust pair.ts in the same PR.
• The same wildcard-to-loopback rewrite now lives in serviceURL, formatClientAddress, the tunnel target and desktop. A single shared helper would keep them consistent.

Related: #49796

[opencodejr]

The crash from the last review is fixed: formatClientAddress now uses NetAddress.isInetAddress/formatIp. The wildcard rewrite is now one shared normalizeLoopback helper, and opencode pair no longer suggests setting hostname 0.0.0.0 when it is already set.

How I checked it:

• Called ServiceRegistration.register directly on both checkouts. On the base, 0.0.0.0/:: binds register http://0.0.0.0:… / http://[::]:…. With this PR they register http://127.0.0.1:… / http://[::1]:…, and other addresses are unchanged. I didn't reproduce the proxy case from #49796 end to end.
• tsc -b in packages/cli is clean, and test/service.test.ts passes as on the base.

Public API: the url in the registration file ~/.local/state/opencode/service.json changes from the wildcard to loopback for wildcard binds. Readers in this repo are compatible.

One small change in behaviour, not blocking: with remote access on and the service bound to ::, the tunnel target is now [::1]:<port> instead of 127.0.0.1:<port>.

No other concerns.

[opencodejr]

The new commit restores the base behaviour for the remote tunnel target: a service bound to :: targets 127.0.0.1:<port> again instead of [::1]:<port>, which was the small change I noted last time. tsc -b in packages/cli is clean and test/service.test.ts passes. Public API: no change from the last review (the registration file url is loopback for wildcard binds). No concerns.

[opencodejr]

The new push changes no code: its tree is identical to the commit I reviewed last time, so the previous checks still hold. Public API: unchanged since the last review (the registration file url is loopback for wildcard binds). No concerns.

[opencodejr]

The new push only merges dev. The change against dev is the same as in my last review.

I ran the #49796 reproduction from source against a proxy that returns 502 for 0.0.0.0:

• On dev, the service registers http://0.0.0.0:<port>, service status prints stopped, and session list fails with "port already in use" after about 1300 probes through the proxy.
• With this PR, it registers http://127.0.0.1:<port>, status and session list both work, and no requests reach the proxy.

tsc -b in packages/cli is clean and test/service.test.ts passes.

Public API: unchanged since the last review (the registration file url is loopback for wildcard binds). No concerns.