fix(serve): surface the real bind error instead of "Unexpected error / ServeError"

[dzianisv]

Issue for this PR

Closes #38738
Closes #38739

Type of change

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

What does this PR do?

Every bind failure in opencode serve printed the same two useless lines:

$ opencode serve --hostname 100.68.120.26 --port 4096
Error: Unexpected error

ServeError

No port, no hostname, no errno. A busy port, an address that isn't on the machine, and a privileged port were indistinguishable.

There are two separate reasons for that.

The errno is discarded. Effect's HttpServerError.ServeError is a Data.TaggedError, so its message is "" and the real failure sits on .cause. errorMessage() checked message then fell straight through to name — printing the literal string ServeError. Fixed by checking cause in between. That helps every tagged error, not just this one.

Bun reports the wrong errno anyway. Under Bun, node:http listen() returns EADDRINUSE for everything. Same script, two runtimes:

                        bun 1.3.14      node
100.68.120.26:45041     EADDRINUSE      EADDRNOTAVAIL
127.0.0.1:80            EADDRINUSE      EACCES
127.0.0.1:4096          EADDRINUSE      EADDRINUSE

Bun.serve() does the same, so it isn't the node:http shim specifically. This is why unwrapping the cause alone isn't enough — the errno you unwrap is wrong. So the new assertBindable() checks a literal-IP --hostname against os.networkInterfaces() before calling listen, where the answer is unambiguous. Hostnames are skipped, since listen resolves those.

bindError() handles what's left: it walks the cause chain and maps the errno. Sub-1024 ports are reported as a privilege problem rather than a conflict, because Bun folds EACCES into EADDRINUSE and a privileged port is the far likelier explanation down there.

Finally serve raises a CliError, so the message prints on its own instead of behind Unexpected error.

bind.ts throws plain Errors and has no CLI dependency on purpose — Server.listen is also called by the Desktop/WSL sidecar and by tests.

Worth noting the Bun errno behaviour is a Bun bug and should be reported there too. This PR doesn't depend on that being fixed.

How did you verify your code works?

Drove a real serve process through every case, on dev @ f516651:

command	before	after
--hostname 100.68.120.26 --port 4096 (interface down)	Unexpected error / ServeError	Cannot bind to 100.68.120.26: it is not an address on any local network interface. Available: 127.0.0.1, ::1, 192.168.0.40, … Use 0.0.0.0 to listen on every interface, or bring the interface up first (e.g. start Tailscale/VPN).
--hostname 127.0.0.1 --port 4096 (busy)	Unexpected error / ServeError	Address 127.0.0.1:4096 is already in use. Choose a different --port, use --port 0 to pick a free one, or stop the process holding it.
--hostname 127.0.0.1 --port 80	Unexpected error / ServeError	Cannot bind 127.0.0.1:80: ports below 1024 are privileged. Use a port >= 1024, or run with elevated privileges.
--hostname 192.168.0.40 --port N	listens	listens (unchanged)
--hostname 0.0.0.0 --port N	listens	listens (unchanged)
--hostname localhost --port 0	listens	listens (unchanged)

Also reproduced the original bug on the released 1.17.9, not just on dev.

Tests:

• bun test test/server/ → 306 pass, 2 skip, 0 fail
• bun test test/cli/serve/ → 2 pass, 0 fail
• New test/server/bind.test.ts, 13 cases — including a fixture mimicking ServeError's empty-message shape, the sub-1024 case Bun mislabels, IPv6 zone ids, and a self-referential cause chain
• 3 new cases in packages/tui/test/util/error.test.ts for cause unwrapping, and that an explicit message still wins over cause
• tsc --noEmit clean, oxlint 0 warnings / 0 errors on all changed files

Screenshots / recordings

Not a UI change. CLI output is in the table above.

Checklist

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

[github-actions]

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

[dzianisv]

test / typecheck / nix-eval show action_required — they need a maintainer to approve workflow runs for a first-time contributor.

Ran the same commands locally in the meantime, on dev @ f516651:

• bun typecheck → 30/30 packages pass
• bun turbo test → passes for opencode, tui, core, and the rest

One caveat so it isn't mistaken for fallout from this PR: @opencode-ai/app fails i18n parity > non-English locales have every English key. It fails identically on unmodified dev (671 pass, 1 fail), and this PR touches no files under packages/app.