ghostty-rs/docs/rewrite/notes/verification-VR.md

Verification VR (IME, dialogs, quick terminal, flatpak host command, hscroll)

Replay hook additions (crates/ghostty-apprt-gtk/src/debug_input.rs, class/surface.rs)

• IME: ime-preedit <text>, ime-preedit-end, ime-commit <text>, ime-state, ime-key <preedit|commit|commit-simple|handled-commit|pass> <key> [text]. The steps call the surface's real template callbacks (im_preedit_start/changed/end, im_commit, now pub(super) and taking &gtk::IMContext) with a mock_im::MockIm (a real GtkIMContext subclass whose preedit_string() returns text + underline attrs + cursor) as emitter. ime-key runs the real in_keyevent/im_composing bookkeeping: the IM half of Surface::key_event was extracted into Surface::im_filter(filter) (behaviour unchanged, key_event passes filter_keypress). Orders covered: ibus (commit then preedit-end), simple (preedit-end then commit), dead key (handled while composing -> not encoded), plain translation commit (handled, encoded with the key), pass. Zig has no retrieve-surrounding/delete-surrounding handlers (checked surface.zig), so there is nothing to port.
• Dialogs: dialog-info, dialog-remember, dialog-click <label> (clicks the real GtkButton in the AdwAlertDialog, so the AdwAlertDialog response -> confirm/deny signals run). animations off|on.
• Quick terminal: windows-dump now logs per GhosttyWindow: quick_terminal, visible, mapped, active, size and the layer-shell state (layer, keyboard mode, namespace, anchors, exclusive zone); quick-focus-out emits notify::is-active on the quick window (the window is never compositor-focused here) to run the autohide handler.
• hscroll <dx> (the handler body was moved to Surface::scroll_horizontal_impl; the controller keeps the unit check), tabs-dump. Replay target window: if GTK's active window is hidden (autohidden quick terminal) the first visible Ghostty window is used.

Results

Area	Result
IME preedit	PASS. ime-preedit かn renders the underlined preedit at the cursor (screenshot); core preedit cleared on commit/end.
IME commit / dead keys	PASS. pty receives こんáxyé (cat -v also shows the exact UTF-8 bytes); preedit text never reaches the pty; composing/consumed state sequence matches Zig's logic.
Unsafe paste dialog	PASS. Dialog (heading, body, Deny/Allow, contents) verified structurally; Deny writes nothing, Allow writes the text once.
OSC 52 write=ask	PASS. Deny keeps the clipboard, Allow writes it, Allow+remember (switch) makes later writes silent (clipboard_write -> allow).
OSC 52 read=ask	PASS. Allow: pty gets ESC]52;c;c2VjcmV0...; Deny: nothing sent.
Quick terminal	PASS after the link-order fix below: layer surface (visible in niri msg layers: Top/Overlay, namespace ghostty-quick-terminal, keyboard interactivity), anchors/size for top/bottom/left/right/center match Zig's quick-terminal-size and anchor logic, toggle hides and re-shows, autohide on focus loss (default false on Linux, as Zig) hides, autohide=false stays.
hscroll tabs	PASS (threshold 120, direction, no wrap at the ends, 500 ms accumulator reset).
Flatpak host command	PASS against a mock service (see below).

Not verifiable here: real keyboard focus on the layer surface (quick-terminal-keyboard-interactivity=exclusive is applied, niri msg layers shows it, but GtkWindow::is_active stays false because the replayed windows are never compositor-focused; the autohide path was driven with quick-focus-out); KDE slide animation (needs the KDE protocol, niri has none); rendering of AdwDialog in screenshots (window-screenshot does not capture the dialog layer; use dialog-info); a real input method (ibus/fcitx are not available).

Bug found and fixed: gtk4-layer-shell load order

libgtk4-layer-shell must be loaded before libwayland-client. The Rust binary's DT_NEEDED order depends on how rustc orders native libs, which flipped between builds (adding an unrelated gtk4_layer_shell:: path in a file was enough); with the wrong order is_supported() is false and the quick terminal is silently disabled ("compositor does not support wlr-layer-shell"). Fix: wayland-sys with features client,dlopen is a direct dependency of apprt-gtk, so libwayland-client is no longer a direct DT_NEEDED of the binary (it is only pulled in transitively by GTK, which is loaded after all direct libraries, same effect as Zig's explicit link order). Check: readelf -d target/*/ghostty | grep NEEDED has no libwayland-client. Regression: quick_terminal_layer_surface_toggle_and_autohide.

Flatpak host command

• Seam: ghostty_os::flatpak::set_flatpak_info_path(path) (OnceLock, production code, unused by the app) makes is_flatpak() check that path instead of /.flatpak-info.
• Mock service: crates/ghostty-os/tests/common/flatpak_mock.rs (zbus, claims org.freedesktop.Flatpak, implements Development.HostCommand/HostCommandSignal, really runs the command locally with the received stdio fds in its own process group, emits HostCommandExited with the raw wait status, records spawns and signals). Tests re-exec themselves under dbus-run-session (skipped if it is missing).
• crates/ghostty-os/tests/flatpak_mock.rs: spawn returns host pid, cwd/env/fds forwarded, stdout via passed fd, exit status propagated (7), wait state reset, wait_callback, host spawn failure -> FlatpakSpawnFail, signal_handle().signal(1, true) after the command moved into the watcher -> recorded (pid, 1, true) and the host process exits.
• crates/ghostty-termio/tests/flatpak_exec.rs (dev-deps: zbus, rustix): termio Exec flatpak branch end to end (output on the pty, exit code 5 propagated as ChildExited, input echoed through the pty, stop() sends SIGHUP to the host process group and the host process dies). One test per binary: the override is process-global.

Regression tests

• tests/gtk_replay.rs (root package, runs the real binary on the current Wayland/X11 session; skips without a display or with GHOSTTY_SKIP_GTK_REPLAY): IME flow, paste deny/confirm, OSC 52 write/read ask (+remember), quick terminal layer surface/toggle/autohide/no-autohide, horizontal scroll tab switching. ~45 s total, serialized (they share the session clipboard).
• Parser unit tests for the new steps in debug_input.rs.

Notes

• screenshot can be one frame stale right after input (wait >= 1 s); text after key escape in bash is meta-prefixed by readline (not a bug).
• Child programs reading the tty from sh -c need timeout --foreground (job-control SIGTTIN otherwise).
• quick-terminal-autohide defaults to false on Linux (as Zig), so the quick terminal does not hide on focus loss by default.