fy59/Dotfiles_ArchAP
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Dotfiles_ArchAP/docs/hyprwhspr-swayfx-focus-and-paste-fix-v1.md

8.6 KiB
Raw Permalink Blame History

HYPRWHSPR_SWAYFX_FOCUS_AND_PASTE_FIX_V1

Implemented and verified on 2026-11-07 against the locally installed Arch hyprwhspr 1.52.1-1. No real dictation was started and no real application received test text. Manual audio-to-editor acceptance remains pending.

Evidence and diagnosis

The installed _get_active_window_info probes Niri, Hyprland, X11/XWayland, then AT-SPI. It has no Sway detector. Consequently native Kitty and Firefox report not detected, while the XWayland ChatGPT reports class Chatgpt. Kitty then resolves the fallback Ctrl+V instead of its terminal Ctrl+Shift+V. This explains the observed terminal control character; no global paste_mode setting is needed.

Sway IPC was tested with Kitty, ChatGPT and Firefox individually focused, restoring the preceding focus after the audit. It correctly returns:

Application app_id window_properties.class Transport
Kitty kitty absent native Wayland
Firefox firefox absent native Wayland
ChatGPT absent Chatgpt XWayland

The service's own /proc/<pid>/environ contains:

  • SWAYSOCK=/run/user/1000/sway-ipc.1000.2363.sock
  • WAYLAND_DISPLAY=wayland-1
  • DISPLAY=:0
  • XDG_RUNTIME_DIR=/run/user/1000

A swaymsg get_tree and corrected config focused-window executed with this service environment succeed for all three applications. No global systemd environment import or socket workaround was required. The socket's PID-based path is session-specific and is not hard-coded by the patch.

For ChatGPT, source inspection shows wtype is the normal Wayland injection transport; ydotool is a fallback. A disposable XWayland Tk widget reproduced the following with actual synthetic keys:

Transport Return Key event received
Installed wtype chord success / exit 0 keysym 1, Control mask 4
Patched XWayland dispatch success Control_L, then keysym v, Control mask 4

Adding 50 ms delays to wtype did not correct the wrong keysym. This establishes a reproducible XWayland keymap/keysym transport failure despite a successful wtype exit. It is a concrete explanation for the ChatGPT symptom, not proof that ChatGPT itself has now accepted a real transcript. A manual ChatGPT test is still required. Clipboard contents were not inspected or logged. The retained probe blocks all default key bindings and mocks clipboard copy/save/restore.

The OSD source already sets Gtk4LayerShell.KeyboardMode.NONE. No focus-stealing OSD window appeared in the audit; focus loss remains guarded at delivery time. Clipboard writes and delayed restoration already use the same clipboard lock; no demonstrated concurrency defect required changing that mechanism. Firefox already worked according to the user and retains native wtype + Ctrl+V.

Patch and installation

hyprwhspr/integration/sway-focus-paste.patch adds an authoritative Sway backend using the i3 IPC GET_TREE protocol directly: little-endian fixed header, 500 ms socket deadline, 8 MiB response bound, iterative nodes/floating_nodes traversal. It uses focused, app_id, window_properties.class, name and X11 window ID. It does not invoke a shell, create a poller, or log titles/transcripts.

When SWAYSOCK is set, unavailable/malformed IPC fails closed rather than falling through to stale XWayland focus. Other compositor detection is retained when SWAYSOCK is absent. Existing per-application auto_paste rules and automatic terminal detection are preserved.

Delivery snapshots the current Sway identity, then checks it again after clipboard publication/settling, immediately before dispatch. Focus change or IPC failure aborts the paste. No previous window is activated. For an identified XWayland target it additionally verifies xdotool getactivewindow against Sway's X11 window ID, then uses the existing XTest Ctrl+V implementation. It never retries another transport after a possibly dispatched X11 chord. Native Kitty keeps Ctrl+Shift+V; native Firefox keeps Ctrl+V through wtype. The focus check and keyboard dispatch are separate operations, not an atomic compositor transaction.

No pacman-owned file was edited. The installer verifies exact SHA-256 baselines:

  • text_injector.py: e6bb5d4291d83cbd189d33bee1dc062b466d59b6a0fb1e1da170dba9b648ca29
  • /usr/bin/hyprwhspr: bfe5516566ac12b2f9958109f395d6504ef17dc3ac200eea8779f502a9896f22

It copies the installed lib source into ~/.local/share/hyprwhspr-swayfx/1.52.1-sway-v1/, applies the patch with zero fuzz, and links other package resources. Its launcher retains the packaged interpreter/environment routing. A SHA guard refuses execution if the original injector changes after a package update; rebase/review is then required.

The user CLI symlink ~/.local/bin/hyprwhspr selects this same launcher. The new drop-in ~/.config/systemd/user/hyprwhspr.service.d/50-sway-focus-paste.conf overrides only ExecStart. Existing config and the pre-existing modified service unit were preserved. One preparation snapshot was retained as 1.52.1-sway-v1-before-validation; it is not active.

Reproduction:

python ~/.dotfiles/hyprwhspr/integration/install-sway-fix.py
# When dictation and transcription are idle:
systemctl --user daemon-reload
systemctl --user restart hyprwhspr.service

Installer rerun/idempotency was verified. The controlled restart completed; the daemon reports Vulkan (AMD/Intel), large-v3-turbo and READY. NRestarts=0. Copilot SUPER+SHIFT+F23, push_to_talk, language fr, audio settings, backend and model remain unchanged. QuickShell was not edited; its existing IPC reports idle/Dictée inactive after the restart.

soundfile

The warning originates in FeedbackMixin._report_missing_dependencies running inside ~/.local/share/hyprwhspr/venv/bin/python, Python 3.14, with system site packages enabled. The runtime executable resolves to /usr/bin/python3.14, but sys.prefix correctly identifies this user venv. soundfile was absent there.

Installed only in that venv, without sudo or dependency upgrades:

~/.local/share/hyprwhspr/venv/bin/python -m pip install --no-deps soundfile==0.13.1

Import now succeeds (0.13.1); missing_imports('vulkan') returns an empty tuple. The restarted service no longer emits the missing-module error. The backend and model were not replaced.

Validation and remaining acceptance

  • 16 automated production-method tests pass. They first apply the actual patch to installed upstream in a temporary directory, then cover identities/chords, fragmented/invalid/oversized/truncated IPC, timeout, socket unavailability, stale X focus, delivery-time focus change/loss, XWayland/native routing, no retry after X dispatch failure, per-app override and non-Sway fallback.
  • Safe real XWayland key probe passes; retained at tools/hyprwhspr-sway-key-probe.py. It restores focus and terminates its own fixture, with no clipboard access or real text delivery.
  • Corrected live CLI detection succeeds for all three apps using the service environment. User window titles were omitted from the report.
  • Python compilation, launcher bash syntax and git diff --check pass.
  • Single daemon and original QuickShell instance remain; no fixture or probe processes remain. No new scheduler, persistent worker or poller was added.
  • No automated tests remain failing. Initial pre-fix focus and XWayland key failures are the reproduced defects described above.

Manual checks: use Copilot to dictate a harmless sentence into an empty Kitty prompt (do not press Enter), a Firefox test field and ChatGPT's input without submitting. Verify correct accented text, normal transcription/insertion, QuickShell synchronization and repeated dictation. Actual audio-to-application insertion remains unclaimed until these tests are performed.

Rollback

When dictation/transcription are idle, remove only the two newly generated selectors, then restart the original service:

unlink ~/.config/systemd/user/hyprwhspr.service.d/50-sway-focus-paste.conf
unlink ~/.local/bin/hyprwhspr
systemctl --user daemon-reload
systemctl --user restart hyprwhspr.service

The original unit's ExecStart=/usr/lib/hyprwhspr/bin/hyprwhspr takes effect again. The user-owned source snapshots can remain for diagnosis; no package restore, Git restore or cleanup is required. soundfile can safely remain. If its removal is desired separately, use the same venv's python -m pip uninstall soundfile; it was absent before this task. Preserve all other config/unit modifications.

Upstream opportunity

A focused upstream contribution is appropriate: Sway identity detection, fail-closed delivery-time focus checks and the independently reproduced wtype XWayland keysym failure with targeted X11 dispatch. Keep the local installer and systemd paths out of the upstream code patch. No issue, PR or publication was created.