[algorithmx]

Closes part of the gap between the Mac app (app/) and the Linux app (desktop/). docs/platform-parity.md (new) compares the two row by row, names the code each row rests on, and ranks what's still missing. Five commits, each reviewable on its own:

1. Test the Linux daemon against a simulated box. tools/test-linux-host.sh runs the real vibebuddyd against the firmware core's simulator through a PTY: no box, desktop, user config or network. The simulator gains a test-only control socket (buttons, clock, reboot). Desktop helpers (hyprctl, swaymsg, wmctrl, notify-send, systemctl, …) are stubbed and recorded. CI runs it on Ubuntu.
2. Tell the Linux user when the daemon fails, as the Mac does. The Mac notifies when its supervisor gives up; Linux said nothing. Now the app sends one notify-send after 30 s without the daemon, again only after a recovery. A unit stopped on purpose stays silent but is rechecked, so a failed restart after it still notifies.
   • Restart daemon on Advanced works while the daemon is down (through systemd, like the tray).
   • The Linux status stream has header (10 s) and inactivity (45 s) deadlines, so a stalled daemon counts as down.
   • Mac: the stream uses the same 45 s inactivity timeout, and the failure notification follows the last reported "Notify me" setting instead of falling back to on once status is cleared.
3. Focus K2 windows on Sway and X11, not only Hyprland (daemon/src/window_focus.rs: swaymsg by container id, wmctrl by window id). The window-system variables are read from systemctl --user show-environment at each press, so a daemon systemd started before the desktop imported them still finds the session. The AUR package lists wmctrl as optional. GNOME/KDE under Wayland still get only the tmux pane switch.
4. Show the daemon log in the desktop's own terminal. It falls back from xdg-terminal-exec (missing on Ubuntu 22.04) to GNOME Terminal, Konsole, Xfce Terminal and xterm. With none installed, Advanced says which command to run. The Mac has no such button; its daemon logs to a file.
5. Docs: platform-parity.md, linux-host-testing.md, README (both languages), app.md, architecture.md, two LESSONS entries.

UI strings have zh-Hans translations (286 keys).

Validation

On Linux (Ubuntu), from 574b700:

• cargo test --workspace --locked: 525 passed, 1 ignored (the native-USB ROM test).
• cargo clippy --workspace --all-targets --locked: only the four warnings main already has.
• tools/check-localization.py and the Python tool tests pass.
• tools/test-linux-host.sh: 55 scenarios pass (28 device/host, 11 daemon outage, 9 Sway, 7 X11).

Not run: the Swift changes are not compiled here (the mac-app job will be their first build); no real Sway or X11 login; no real notification server or systemd unit; no tray on Fedora/upstream GNOME/KDE; no box attached. Those stay listed as pending in platform-parity.md.

🤖 Generated with Claude Code

[longzhi]

Review notes on 6cfd1a3. Local checks: cargo test -p vibebuddyd window_focus (22 passed) and tools/test-linux-host.sh (55 scenarios passed).

1. "Restart daemon" now needs systemd — desktop/src/client.rs:162

• Old behaviour: while the daemon ran, the app sent POST /v1/daemon/restart.
• New behaviour: on Linux, the app always runs systemctl --user restart vibebuddyd.
• Good result: the button now works when the daemon is down.
• Bad result: if systemd did not start the daemon, the button fails. Examples: cargo run during development, or a system without systemd.
• Suggestion: when systemctl fails, fall back to POST /v1/daemon/restart. Or state in the docs that this button needs the systemd unit.

2. Hyprland detection does not check that the session is live — daemon/src/window_focus.rs:52

• compositor() accepts SWAYSOCK only when the socket exists.
• HYPRLAND_INSTANCE_SIGNATURE wins when it is set. Nothing checks that it is still valid.
• session() takes values from systemctl --user show-environment first.
• Failure case:
  1. An earlier Hyprland session imported HYPRLAND_INSTANCE_SIGNATURE into the user manager (for example, with linger on).
  2. The user logs in to Sway or X11.
  3. The daemon still selects Hyprland.
  4. Each K2 press fails in hyprctl.
• Suggestion: accept the signature only when $XDG_RUNTIME_DIR/hypr/<signature> exists, the same way the code handles Sway.

3. Comment refers to a missing document — daemon/src/source_opener.rs:526

• The comment says "(G4 review 3.4)".
• The repository has no document with that name.
• Suggestion: remove the reference, or write the reason in the comment.

---

Generated by Claude Code

[algorithmx]

@longzhi Thanks for the review. All three points were real, and each turned out to be one case of a wider problem, so 887ec20 fixes the cause rather than the symptom.

1. Restart without systemd

I didn't add the POST /v1/daemon/restart fallback. That endpoint only exits, so without a supervisor it would stop the daemon instead of restarting it. The same hidden assumption, that systemd runs the daemon, also broke two other things. Show daemon log opened an empty journal. Worse, the failure alert read the unit's inactive as a deliberate stop, so a crash of a hand-started daemon never notified.

• The daemon now reports daemon.supervisor (systemd or other) in its status, from its own cgroup. It doesn't use INVOCATION_ID, because a terminal that runs as a systemd service passes it down to everything started from it.
• The app keeps the last answer while the daemon is down. For a daemon outside the service, Restart and Show daemon log say why they can't act on it, and the failure alert goes out without asking systemd.
• POST /v1/daemon/restart refuses with 409 when nothing would start the daemon again.
• No Mac change was needed: the Mac app already knows who supervises its daemon (managesDaemon).

2. Stale Hyprland signature

Fixed as you suggested, and more broadly:

• K2 now reads the window system from the environment of the processes it is about to focus (/proc/<pid>/environ), taken whole from the nearest one with a graphical session.
• Hyprland and Sway count only if their socket accepts a connection: $XDG_RUNTIME_DIR/hypr/<sig>/.socket.sock, or /tmp/hypr/<sig> before Hyprland 0.40.
• With no evidence, K2 focuses nothing and says so in the log, instead of assuming Hyprland.
• systemd's user manager is only a fallback, for processes whose environment can't be read.

This also covers stale X11 variables in the user manager and a Sway socket left behind by a crashed compositor.

3. "(G4 review 3.4)"

Removed.

Tests

The host suite now starts a real process for each placeholder pid, with the session environment each scenario sets up, plus real listening sockets for Hyprland and Sway. New scenarios cover stale Hyprland and X11 variables, the user-manager fallback, a socket nobody listens on, and a daemon outside the service.

• Host suite: 61 scenarios pass, up from 55.
• cargo test --workspace: 532 passed, 1 ignored.
• Clippy: only the four warnings main already has.
• Localization: passes, 3 new zh-Hans strings.
• Breaking each new rule on purpose made the scenarios meant to catch it fail.

Nothing has been run on a real Sway, X11 or GNOME session yet.

Parity scenario notes

7acc6a9 adds docs/parity-scenarios/: 70 notes, one per thing the Mac app does, read from its code. They cover lifecycle and daemon supervision, the menu bar, notifications, every Settings tab, the first-run guide and errors.

• Each note gives the initial state, the trigger, the app's steps (with the exact UI strings and daemon calls) and the final state.
• Each ends with a checklist of questions for the Linux app and a blank result: same, differs by design, gap or not applicable.
• The notes describe the Mac app only. Where platform-parity.md already records a difference as by design, the note says so; everything else is an open question, not a claim about Linux.

The next commit will be based on these notes: I'll go through them against the Linux app, fill in the results, and fix the gaps they turn up.