[pyjeebz]

Closes #23, with a changed design (below). Follow-up for what it leaves out: #29.

The most useful fact about a dev server is where to open it. The table now shows it:

PORT   PROCESS      PROJECT   UPTIME   URL
3458   Next.js      ports     2s       http://localhost:3458
9239   Next.js      ports     2s
8791   python3.13   shop      3s

• The URL column is last, and appears only when some row has a URL. A machine running only databases looks exactly as before.
• Clickable (an OSC 8 link) in a terminal that renders links. Plain text when piped or with NO_COLOR. On Windows, links appear only in Windows Terminal or a terminal that sets TERM_PROGRAM (like VS Code's), because the legacy console prints the escape codes raw.
• --json gains url. It's additive and omitted when empty.

The design change: no request is ever sent

#23 asked to probe every listener with an HTTP request. I didn't, because the listing has to stay side-effect free:

• a HEAD makes a Next.js dev server compile a page;
• a database port never answers, so it stalls until the request times out;
• --watch would repeat all of that every second.

Instead the URL comes from framework detection. There's a curated list of dev servers that serve plain HTTP unless told otherwise: Next.js, Vite, Astro, Remix, Nuxt, SvelteKit, Storybook, Angular, NestJS, FastAPI, Django, Flask, Rails, Laravel, Phoenix, Spring Boot, Gin and their kin.

• Left out, rather than guessed: nginx and Apache (may be HTTPS), Mailpit (also SMTP), databases, and Wrangler, Echo and Chi (see below).
• On demand, ports doctor still sends a real request, and ports open (open: open a running server in the browser #24) will too.
• Servers outside the list (Express, python -m http.server, Bun, Axum…) get no URL yet. table: URLs for HTTP servers outside the framework list #29 weighs an opt-in probe for them.

One URL per app, on its own port

A framework label belongs to the process, and one process can hold several ports, so:

• A --port/-p on the command line names the app port, and only that port gets the URL. A process that names a port it doesn't hold gets none. Next.js runs as two processes: the next dev -p 3458 parent holds only its debugger port, and its server child holds the app port.
• Otherwise, the lowest port that isn't a known side port gets it. Side ports are Node --inspect 9229–9230, JDWP 5005, debugpy 5678 and LiveReload 35729.
• HTTPS chosen on the command line gets no http:// URL. That covers --ssl, --ssl-keyfile, --cert, --certfile, --experimental-https and similar. HTTPS set in a config file can't be seen; the docs say so.

How it was verified

Gate, locally with CI's versions: unit ✓, build ✓, Windows build ✓, integration ✓ (uncached), installers ✓, gofmt ✓, vet ✓, lint ✓ (v2.13.2). Site: npm run lint ✓, npm run build ✓.

Live, against a real next dev (the site in web/) plus a python -m http.server:

• Piped: the Next.js row has the URL; the python row has none, and no trailing spaces.
• --json: "url" appears on the Next.js row only.
• In a terminal: an OSC 8 link. With NO_COLOR: plain text.
• The dev server's own log shows 0 requests from ports. curl of the URL returns 200.
• NODE_OPTIONS=--inspect next dev -p 3458: only 3458 has a URL. The parent's inspector (9239) and the child's (9240) have none. My first version gave the parent's inspector a URL, and this live run is what caught it.

New tests:

• TestLocalURL: binds, IPv6 including zone ids, nginx not guessed, Windows rows, dropped frameworks.
• TestTableURLColumn: the column only when present, last, no trailing spaces, links only when enabled.
• TestFillURLs: the inspector, -p/--port, the parent/child Next.js case, JDWP, HTTPS flags, two apps.

Every README, docs and --help table containing a web server was re-rendered exactly as the binary prints it. The commands page documents the URL column and the url field. The landing page's styled tables are left for the Phase 4 site refresh.

Risk

Tier: risky (internal/inspect/output.go, --json).
Public surface: a new table column, and a new url JSON field (additive).
Rollback: revert the squash commit.

A fresh-context review blocked the first commit. On the Windows legacy console, the default ports would print raw escape codes. That's fixed, and so are its should-fixes:

• a URL was going to debugger ports, HTTPS servers and every port a process held;
• IPv6 zone ids weren't escaped;
• the weak echo/chi matches and multi-process Wrangler were on the list;
• the narrowed scope needed recording (table: URLs for HTTP servers outside the framework list #29).

Notes for the release

The table shows http://localhost:<port> for known web dev servers (Next.js, Vite, FastAPI, Rails…), clickable in a terminal. --json rows gain url.

---

Generated by Claude Code