Run the Windows port on Linux

[arlindovs]

The Rust port under windows/ builds and runs on Linux with fairly little
persuasion — every Win32 call was already behind cfg(windows), so what was
left was the handful of places where a Windows mechanism was the only one on
offer, plus a PNG icon and a launcher.

Tested on Ubuntu 26.04 / GNOME 49 (Wayland), against a live Claude Code
install: cargo test --locked passes (121 tests), codenotch doctor reads the
Claude credential and finds the standalone CLI, and the notch draws on the
screen edge and tracks running sessions through the watcher.

What changed

Mechanisms that had no Linux equivalent

Was	Now
HKCU\...\Run value written with reg.exe	an XDG autostart entry — same bargain: per user, no administrator rights, still --silent
GetUserDefaultLocaleName, GetLocaleInfoEx, the taskbar-clock registry value	LC_ALL/LC_MESSAGES/LANG, and the 12/24-hour question from the locale's own t_fmt, translated to the letters the LOCALE_* patterns use so is_24h_pattern keeps reading both
explorer, cmd /C start	xdg-open
claude.exe / claude.cmd only	suffixes chosen per platform, plus ~/.npm-global/bin and ~/.bun/bin; Codex's vendored native binary matched by linux rather than windows
Claude sign-in in a PowerShell window	the same script in a terminal emulator, found from the usual candidates. The CLI path still travels in the environment, so a path with quotes in it is never parsed

Build and packaging

tauri::generate_context! will not build without a PNG icon, so the icon set
grows the four sizes Tauri looks for, generated from the existing icon.ico
and listed beside it. Windows bundling is untouched — NSIS still takes the
.ico. tauri.linux.conf.json is merged only on Linux: deb and appimage, and
an updater with no endpoints, because the feed describes an NSIS installer.

scripts/run-linux.sh

Two things the desktop will not do by itself:

• Wayland does not let a client place its own windows, and the notch has to
  sit on a screen edge — so it runs as an X11 client under XWayland.
• A shell started from a snap (Ubuntu's VS Code, for one) exports that
  snap's library paths, and a binary built against the system glibc then dies
  before main with
  symbol lookup error: … undefined symbol: __libc_pthread_init, version GLIBC_PRIVATE.

The script unsets those and sets GDK_BACKEND=x11. Launched from the desktop
rather than from such a shell, the binary runs on its own.

CI

The Windows workflow grows a Linux job, for the reason that workflow exists at
all: a change that compiles on one of the two says nothing about the other. It
is build-and-test only — there is no runner with a desktop to show a notch on.

What does not work yet

Documented in the README rather than hidden. Each degrades quietly instead of
misbehaving:

• Dragging the pill along its edge — follows the mouse through
  GetAsyncKeyState; needs an X11 pointer query.
• Seen-clears-it, and jumping back to the terminal — focus.rs reads the
  foreground window and the process tree through Toolhelp. /proc plus a
  window-manager call would replace it.
• Antigravity — its credential comes from the Windows Credential Manager;
  libsecret is the equivalent.
• App icons taken from an installed .exe — the built-in provider SVGs
  cover every provider, so little is lost.

Everything else behaves as on Windows: all other providers, the hover card, the
settings window, the tray menu, hooks, and start at sign-in. The tray itself
needs GNOME's AppIndicator Support extension, as every Tauri tray does there.

Two things for you to decide

1. Where this belongs. The README says this tree is developed at
   Im-Midi/codenotch-windows and offered here. I opened it here because the
   README also says a pull request touching the tree is built and tested here.
   Happy to move it, or to mirror it there.
2. The name. A Linux port living under windows/ is odd, and the crate's
   description still says "Codenotch for Windows". I left both alone rather
   than making that call inside a contribution — say the word and I will follow
   whatever you prefer.

I am glad to close the remaining gaps in follow-ups if the direction is welcome.

🤖 Generated with Claude Code

[vinzdg]

Thanks @arlindovs. You should know there is a second Linux port open — #269 by @heliorosadoneto — and the two take opposite approaches, so I'm reviewing them against each other rather than in isolation. Tagging you both so neither is working blind.

Yours is the cheaper shape. It keeps windows/ where it is, adds tauri.linux.conf.json beside the existing config, adds no dependencies, no packaging path, and is based on today's main. It breaks no open PR and commits me to nothing. #269 is better-engineered Linux code — a real 10-function platform facade — but it renames windows/ → desktop/, which invalidates every branch written against those paths, and its own branch predates 1.16.0.

Things to fix here regardless of what I decide about Linux:

• tauri.conf.json:37-43 — the PNGs are prepended to the shared icon list, so generate_context! embeds a PNG window icon on Windows too. Move them into tauri.linux.conf.json; Tauri merges it at build time. (Good news: targets stays ["nsis"] and createUpdaterArtifacts is untouched, so unlike Linux desktop port (Mint / Cinnamon / X11) #269 there's no leak into the Windows installer. That matters — the installer target is now switched on from the workflow when a signing key is present.)
• hooks_install.rs:71 still joins codenotch-hook.exe, so Install hooks fails on Linux with a missing-file error. Both the PR body and your windows/README.md edit say hooks behave as on Windows; they don't yet.
• claude_auth.rs terminal_emulator() passes ["-e","sh","-c",script]. That's right for xterm/konsole/alacritty, but gnome-terminal dropped multi-arg -e in GNOME 42+ and kgx never had it — and on Ubuntu x-terminal-emulator resolves to gnome-terminal first. So Claude sign-in likely fails on the desktop the PR says it was tested on. Use -- with a fallback.
• tauri.linux.conf.json:8 sets updater.endpoints: []. Once a signing key exists, app.updater() will return EmptyEndpoints and Settings will show an error on Linux. Better to gate the check off entirely on non-Windows — see src/updater.rs, which already has a configured() guard you can extend.

Smaller: find_cli() gains ~/.npm-global/bin and ~/.bun/bin un-gated (harmless, but it's a shared path), codex.rs:133 still pushes codex.exe on Linux, and there's ~330 KB of PNGs including a 230 KB icon.png where Tauri needs only 32/128/128@2x.

On the decision itself: I haven't made it, and I won't pretend the holdup is code quality. The code cost of Linux is small and getting smaller. The support cost is permanent — a third packaging path, and bug reports from X11/Wayland desktops I don't run and can't reproduce on. Once a .deb ships it also needs its own update story, because the Tauri feed added in 1.16.0 only describes an NSIS installer.

I'd rather tell you that plainly than leave this open for weeks. Fix the four items above and it's a reversible first step I can say yes to; the facade work in #269 can fold in on top later.