All notable changes to the LinuxTask project will be documented in this file.
The [Unreleased] section collects changes since the last release;
tools/release.py moves its body under the new version heading when a
release is cut and resets it to empty, so released notes never linger
as pending. Releases v2.0.0 and v2.2.0 predate the project's git tags,
so their headings carry no compare links (v2.2.0 is also listed with the
same date as v2.4.0, and with no tag that date cannot be checked).
v3.6.0 - 2026-11-15
- The AppImage now says when a newer release exists. Settings → Updates
shows the version you run, a
Check nowbutton andDownload and installwith a progress bar, and the Settings button turns blue when there is something to install. The check runs at most once every 24 hours againstapi.github.com, sends no account or personal data, and can be switched off withCheck automatically. Only the AppImage has it: a git checkout or atools/install.shinstall never touches the network.
v3.5.3 - 2026-11-07
Two bugs, both reported from outside this machine. If you record on KDE
Wayland with more than one monitor, this is the release to take: the
v3.5.2 AppImage does not have the fix, and on Plasma 6.8 it clamps the
cursor to whichever panel kscreen happens to print first. Nothing changed
in the recorder, the file format or the shortcuts.
linuxtask --versionno longer opens a window.main()only handed off to the headless CLI when the first argument wasplay,-hor--help, so an unknown flag — or a subcommand typed wrong,paly macro.json— fell through to the GUI branch and launched the app. Anything on the command line now goes to the CLI and argparse answers with exit 2 and a usage line; a barelinuxtaskstill starts the GUI. Neither launcher passes arguments (Exec=AppRunin the AppImage desktop entry,Exec="$ABS_RUN_PATH"from the installer), so nothing that used to work changes.- KDE Wayland reads the Plasma 6.8 layout again (issue #7). That
kscreen-doctor --jsonputs each output's size on the output itself, next tocurrentModeIdandmodes, and marks the primary withpriority: 1instead of aprimaryflag. The reader only looked insidemode/currentMode, so it dropped every output, detection fell through to thekscreen-doctor -olisting, and the clamp box became the first panel kscreen prints -- on the reporter's desk the 1600x1200 secondary, listed ahead of his 2560x1440 primary. Making the main monitor primary in kscreen could not help: the listing order does not follow priority. The size is now read from the output when no mode object is present, andpriority: 1counts as primary. On his own JSON that givesResolution (kscreen-json): 2560x1440, layout 4160x1440+0+0, the same 4160x1440 kscreen declares asscreen.currentSize. That check runs against the JSON he attached, which is now a test; there is no Plasma session on the machine that builds this, so the confirmation on his desk is still his to give.
- The README points at the download, and counts four CI jobs. Nothing
in it told a reader that
LinuxTask-x86_64.AppImageis on the release page -- the one thing v3.5.0 and v3.5.1 were missing -- and the tests section still said "three jobs" after theappimagejob was added to close that class of failure.
v3.5.2 - 2026-11-06
This is the release that has the download. v3.5.0 and v3.5.1 were both
published with an empty asset list: the AppImage built here and then failed
on GitHub's machine - once at v3.5.0, once at v3.5.1, once more on a push -
over five separate causes: the path the workflow attached from, meson and
patchelf missing from the build environment's PATH, a wheel built at the
wrong version, girepository-2.0 shipped under a different package name
than the apt list used, and the two below. Nothing in the recorder changed
since v3.5.1.
Why a clean machine notices things this one does not: build/appimage/wheels
still holds the output of an older bootstrap, and the build skips the step
that already has an artifact. Deleting that directory is what the runner does
by definition, so from now on CI does it on every push (below).
- The AppImage is attached to this release, built by CI from a clean
checkout:
LinuxTask-x86_64.AppImage. - PyGObject now finds pycairo's C headers from this build, not from the
host. PyGObject's build looks
cairoup by importing it with the interpreter it is compiling for; the loop that source-builds pycairo never installed it into its own scratch environment, so on a runner the import failed, the pkg-config fallback found nothing either, and cc stopped atfatal error: py3cairo.h. Arch hosts hide that: theirpython-cairoanswers the fallback with headers of a pycairo that is not the one that ships in the bundle. Each built wheel now goes back into the build environment, and PyGObject is not attempted unlessimport cairoanswers. - The build names every host dependency its own sdists ask for. The
preflight went from three hand-picked
pkg-confignames to the fulldependency()list of the pinned pycairo, dbus-python and PyGObject, versions included (girepository-2.0 >= 2.80,libffi >= 3.0,cairo-gobject...), and the Ubuntu packages that provide them live in one script that both workflows call instead of a list copy per workflow. A missing or too-old module now costs onepkg-configcall and printsPyGObject -> girepository-2.0 (not installed); before, it surfaced seconds or minutes later inside a meson stack trace that named a subprocess instead of the package to install. - The AppImage build is a CI job, not a tag-time surprise.
Python CIgained anappimagejob that installs the dependencies, runs the whole documented build - including the smoke guard that imports every driver with the bundled interpreter - and checks the artifact exists. It found the pycairo defect on its first run, 39 seconds into a push, with nothing published.
v3.5.1 - 2026-11-05
A patch for the download itself. v3.5.0 produced an AppImage that ran here
and failed on GitHub's clean runner, so that release went out without the
artifact - which is the one thing most readers of the release page came to
get. Nothing about the recorder changed.
The build script had been leaning on state a developer machine has and a
runner does not: build tools already on PATH, and a local wheel cache that
was hiding three separate packaging faults behind one green "build ok". Each
is fixed at its own level, and each has a guard test that fails against the
pre-fix script.
- The AppImage builds on a clean checkout.
meson,meson-python,ninja,patchelf,setuptoolsandwheelare now installed inside the bundled build environment and that environment'sbinis put onPATHbefore wheels are compiled, so the build no longer stops atmeson executable "meson" not foundor atFileNotFoundError: 'patchelf'in the RPATH step. - The wheels that ship are the wheels that are pinned. The build asked
pip for a bare package name and got whatever was newest (
dbus-python1.5.1 while the project pins 1.5.0); it now derives the version from the pinned filename it is about to produce. - A wheel with no compiled extension is an error, not an artifact. The
probe that rejects a platform-tagged wheel containing only Python files ran
before the check that the pinned wheel exists at all, so a version mismatch
reported "contains no compiled extension" out of a
FileNotFoundError. The existence check now comes first and the failure names the exact file to delete and rebuild.
v3.5.0 - 2026-11-05
The point of this release is playing a macro back across more than one
monitor: a recording made on a second screen no longer folds onto the first
one (#6, #7). The second visible change is that v3.5.0 is the first release
that carries an AppImage - the build existed and worked, but nothing published
it, so the download the README pointed at was never there. Everything else is
what those two changes exposed: the installer stopped running pip install on
launch, the Wayland IPC paths started checking who answers, and CI now runs
the suite on every Python the project promises rather than on the one that
happened to be configured.
- Multi-monitor playback (#6, #7). The clamp box is the union of every
enabled output rather than the geometry of the single output the driver
happened to read, so coordinates on a second screen - or one placed left of
or above the origin - are legal space and not an edge. Hyprland enumerates
hyprctl monitors -j, Swayswaymsg -t get_outputs, KDE takes the union fromkscreen-doctor --jsonwith the text listing andxrandras fallbacks (that pattern now accepts a negative offset, which the old+-only regex silently dropped for every monitor above the primary), and X11's root window already spans every CRTC. A detection that finds nothing falls back to 1920x1080 instead of a 0x0 box that pins the whole timeline to one pixel in silence. - A compositor refusal is honoured again.
move_cursor/move_relativereturn a bool across all five drivers, so playback'suinputwarp fallback is reachable when the driver says no - and it warps the clamped coordinates rather than the raw event. - "Humanize" does what the label claimed. Point jitter, delta jitter (up to 2 px per axis) and inter-event timing variance (0-3 % on top of the speed factor) are three primitives now; the old helper jittered nothing and only ever saw the delta path.
- Settings stop going intermittent on long replays. Speed and humanize are read through plain attributes that the Tk variables mirror, because reading a Tkinter variable from the replay thread is not safe.
- The first events of a recording are no longer dropped:
toggle_recordarms the capture thread last, sincedevice_loopstarts as soon as the flag is true. - A missing compositor dependency raises a dialog instead of killing the
window (
python3-dbusabsent on KDE Wayland, an unsupported desktop), and the title bar derives the desktop name from the driver in use. tools/run.shnever installs packages at launch. It checks the three imports it needs, names the script that fixes them and exits; the automaticpip install --userit replaced could half-populate a user site, fail on a managed environment, or reach the network on every start.tools/install.shstopped expanding a username throughevalin favour of agetent passwdlookup retried by uid, with an undeterminable home as a hard error; on GNOME Wayland it now installsydotooland verifies both the client andydotoold, because absolute warps need the daemon.- Python 3.10 installs again. evdev 2.0 declares
Requires-Python >=3.11while the pin carried no marker, so pip could not resolve it there - which madetools/install.shfail on Ubuntu 22.04 and Debian 11 and advertised a support floor that was a claim rather than a fact. The pin is split by interpreter now (2.0.0 on >=3.11, 1.9.3 on 3.10), and CI exercises both - the 1.x branch is also what the distropython3-evdevpackages ship. - The AppImage bundles a working evdev. pip's global wheel cache is keyed
by name and ABI, not by whether the C extensions actually compiled, so a
stale cache entry from an earlier failed build shipped a wheel whose
import evdevdied on the user's machine. evdev is now source-built against the bundled interpreter's ABI with--no-cache-dir, and every wheel is checked for a.sobefore the build continues.
- The Wayland IPC paths verify who answers. On a shared session bus any
process could reply to the KWin scripting probe or emit a portal
Responsesignal, and the GNOME Shell bridge accepted warp requests from anyone. KDE now trusts only replies from the cached unique names oforg.kde.KWinandorg.freedesktop.portal.Desktop, tags every read with its own nonce so a reply cannot be replayed from an earlier one, and uses a random request handle instead of a time-derived one. The GJS extension refuses an invocation whose uid is not the uid it runs as, and the Python side confirms the bridge's owner throughGetConnectionUnixUserbefore believing a coordinate. The rule throughout: fail-open only for the first call, while no name is known yet; fail-closed on a mismatch. SECURITY.mdreplaced a template with this app's actual threat model - what is in scope (command injection throughhyprctl/swaymsg/xdotool/ydotool/gdbus/kscreen-doctor/xrandr, the D-Bus and portal trust decisions, the GNOME bridge), what is not, and that disclosures go to the private advisory form rather than an issue.
- The release carries the AppImage.
release.ymlgained a job that installs the build dependencies, runstools/appimage/build.sh, refuses an empty artifact and attaches the result to the release the previous job created. The build publishes it at the documentedbuild/appimage/dist/LinuxTask-x86_64.AppImage, which is the path the workflow attaches and was not the path the builder produced. - The bundled interpreter is pinned and verified.
tools/appimage/python-build-standalone.lockrecords the release tag, the Python version, the URL and the sha256;build.shrefuses an incomplete pin or a URL that disagrees with its own tag, re-downloads a cached tarball whose digest stopped matching, and publishes the cache only aftersha256sum --check.update_python_lock.pyre-pins from the release API. The ASCII toolbar rewrite moved out of a heredoc intopatch_ui.py, asserted against the realsrc/main.py. - CI is three jobs instead of one hand-listed script: a matrix over
3.10/3.11/3.12/3.13/3.14 running pytest with coverage and then
unittest discover(README documents it as the no-pytest path), adistro-bindingsjob on ubuntu-24.04 where the KDE driver's suite actually runs, and a ruff lint job. The suites that need a distro-only binding skip instead of erroring when it is absent, and the runners got the packages a distribution CPython does not build in (python3-tk,python3-pil.imagetk). - Tooling at the repository root:
pyproject.tomlwith the pins, the Python floor and the lint/test configuration,requirements-dev.txt,tests/__init__.pydoing thesys.pathinsert once instead of ten test files repeating it, and Dependabot for pip and Actions. - Documentation: the README's multi-monitor section describes the layout
box the drivers now read and names the
<driver> resolution: WxH, layout WxH+X+Ystart-up line as the thing to paste into a report; the troubleshooting table gained the wrong-monitor row; the development section describes what actually runs, anddocs/CONTRIBUTING.mdpoints atSECURITY.md. - Test suite: 175 tests at v3.4.1, 294 here, passing on every interpreter CI covers, with the cross-driver contract suite pinning the bool return, the layout union and the log shape.
v3.4.1 - 2026-10-24
- Adwaita-flavored visual polish (chrome only): the dark UI now uses an
explicit neutral palette -
#303030window background,#3a3a3abutton fill - with a discreet 1px outline on every toolbar button, lighter hover overlays on the neutral buttons and the flat settings gear, and the speed drop-down restyled from the CustomTkinter blue theme to flat gray (its height now matches the 32px button rhythm). Fonts are unified on DejaVu Sans (speed menu, settings window, first-recording dialog, tooltip), and "Open"/"Save" got wider buttons so their labels sit inside the new outline. Recording, playback, hotkeys and every driver keep their exact behavior and state colors (red/green/blue/orange); the toolbar is still the same 400x44 window with the same controls, no new options or dependencies.
v3.4.0 - 2026-10-24
- GNOME Wayland absolute mode via a GNOME Shell cursor-bridge
extension (
extensions/gnome-shell/linuxtask-cursorbridge@jadrt22):global.get_pointer()for the real pointer position andseat.warp_pointer()for true absolute replay — the same pattern the KDE driver uses with KWin scripting, and the same one Keymasq/PaperWM use. Install withtools/install-gnome-bridge.sh(one logout/login on first install);tools/uninstall.shremoves it. The GNOME driver prefers the bridge forget_cursor_pos(),move_cursor()and exactGetScreenSize()resolution, degrades to the legacy relative-delta + xdotool/tkinter fallback when it is absent or dead (same escalate/skip rule as the KWin probe), andpython-dbusstays optional for GNOME — calls go in-process when it is importable, short-livedgdbussubprocesses otherwise. New headless suitetests/test_gnome_driver.pyjoins CI. - The one-time "Driver warning" dialog now shows the driver's own
message (
warn_fnno longer discards it) instead of the KDE-portal-specific hard-coded text, so the GNOME delta-fallback and bridge-degrade warnings are described correctly.
v3.3.0 - 2026-10-16
- Headless
playCLI (issue #5): scripts can now replay a saved macro without the GUI -linuxtask play macro.json [--speed N] [--loop] [--humanize] [--start-pos X,Y], also reachable as./tools/run.sh play ...andpython src/cli.py play ....src/cli.pyreusesPlayback.playback_threadverbatim, supplying the window-owned attributes (speed/humanize variables, loop flag, UI queue) as plain stand-ins, so the replay path is the same code the toolbar runs. The macro file is validated with the same_validate_eventasopen_file, and the compositor driver plus the uinput device are only opened after it validates, so a bad path can never pop a portal dialog. The README documents the macro JSON format, so a script can generate a macro instead of recording one in the GUI first. Exit codes:2invalid macro,1unsupported desktop,130on Ctrl+C.main.pyroutesplay/-h/--helpto the CLI and everything else to the GUI; the AppImage ships the module and its smoke guard imports it.
src/cli.pyexited with SIGSEGV once in a 42-run batch (exit code-11instead of0). Not reproduced in 30 later runs under-X faulthandler, which produced no traceback, so whether the crash is during replay or at teardown is unknown.
v3.2.1 - 2026-10-10
- KDE Wayland absolute moves landed at the wrong coordinates: a live
cursor read never fed the driver's dead-reckoning origin.
_curstarts at(0, 0)and was only ever updated bysync_for_playbackormove_relative, somove_cursorcomputed every delta from the corner of the screen — asking for(400, 300)moved the pointer by(+400, +300)from wherever it happened to be instead of to that coordinate.get_cursor_posnow writes the position it reads into_cur, for the KWin source and for the xdotool fallback alike, so a single read before an absolute move is enough to anchor it. Relative moves, scroll and buttons were unaffected. Verified live against the running portal: 4/4 warps exact within 3 px, with and without a priorsync_for_playback. - Absolute moves no longer vanish when a driver refuses them: a
posevent calledmove_cursorand, onFalse, only queued the one-shot driver warning — the cursor went nowhere and the macro replayed as nothing. Every other event type (rel,scroll,key) already fell back to the virtualuinputdevice; the absolute path was the only one without a fallback. The refused move now decomposes into a delta from the driver's last known position and is written asEV_REL, which the kernel accepts regardless of driver state. The same fallback covers the start-position warp at the top of each playback pass, where a dead driver used to leave the pointer wherever the previous pass ended and loop repeats drifted. - The first absolute move of a session now lands on target: nothing
read the cursor position before the first
move_cursor, so the dead-reckoning origin was still cold(0, 0)and the pointer ended up at(real + target). This broke both halves of the app — recording (the first click of a macro) and replay (the firstposevent) — and survived the fix above, because that one only anchors the origin once a read has already happened.move_cursornow anchors a cold origin with one live read before computing its delta, so a driver that never went throughsync_for_playbackstill lands correctly. - A failed cursor read no longer freezes the tracked origin: the lazy
anchor marked the position initialised before it knew the read had
succeeded, so a single unreadable position (no KWin scripting, no
xdotool, no XWayland) left the origin stuck at
(0, 0)for the rest of the driver's life. The flag is now set only on a successful read, and the next absolute move tries to anchor again instead of dead-reckoning from the screen corner forever.
v3.2.0 - 2026-10-10
- Sway driver:
src/drivers/sway.pydrives Sway (and swayfx) through the seat IPC —cursor movefor relative motion,cursor setto warp absolute,cursor press/releasefor buttons (button1/button2/button3) and scroll viabutton4/button5. Detected first throughSWAYSOCKand thenXDG_CURRENT_DESKTOP, before the generic X11 fallback, so the warning from issue #4 no longer fires on Sway. Because the sway IPC has no way to read the cursor position back, the driver setssupports_absolute_positioning = Falseand records relative deltas like the GNOME driver;cursor setis still used at replay forstart_posandposevents. Resolution comes fromswaymsg -t get_outputs(outputrect, withcurrent_modeas fallback).tools/install.shchecks forswaymsgon Sway and fails with install guidance instead of installing the compositor. - Sway driver verified against a live compositor: the author ran sway 1.12
headless on a private
SWAYSOCK(leaving the desktop session untouched) and exercisedSwayDriverend to end — seat resolved toseat0, resolution read as 1024×768 fromget_outputs,cursor move 10 -5,cursor set 100 200, clamping to the last pixel, button press/release andbutton4/button5scroll all answeredsuccess: true. That run exposed two bugs the mocked tests could not: swaymsg answers commands with a JSON array ([{"success": true}]), not a bare object, and itsgetopteats negative coordinates (cursor move -30 40arrives ascursor move 40). Both are fixed and covered by regression tests, and the README status table now lists Sway as author-tested instead of "implemented, not tested".
- Sway seat IPC replies are parsed correctly:
_seat_cmdrequired a JSON object withsuccess, but swaymsg answers commands with a list of result objects ([{"success": true}]), so every cursor move, button and scroll reported failure against a real sway while passing the mocked tests. - Sway negative coordinates survive swaymsg's getopt:
cursor move -30 40reached sway ascursor move 40. A--separator is inserted before a negative number, and stays off for-t <type>queries where the leading-tis the option itself. - Sway IPC commands are no longer passed as a single argv entry: the
command string was handed to
swaymsgas one argument, so-t get_seatsarrived as a literal message type and sway answeredUnknown message type get_seats. The command is nowshlex.splitinto argv, which also fixes the seat and resolution detection that silently fell back toseat0and 1920×1080. - Driver-factory tests no longer break CI by importing every driver: they
patched
drivers.kde_wayland..., which force-imports the module. That module needsdbus-pythonandPyGObjectat import time and CI installs onlyrequirements.txt, so the suite failed withmodule 'drivers' has no attribute 'kde_wayland'. Stub modules are injected intosys.modulesinstead, and no real driver module is imported. - The AppImage smoke guard imports every driver: it listed the split
modules, the D-Bus stack and the UI stack, but no driver module — the Sway
driver could have shipped (and did ship) with a broken import. The guard now
imports
drivers.sway,x11,gnome,hyprlandandkde_waylandplusXlibwith the bundled interpreter, andtests/test_appimage_smoke_coverage.pyreads the guard out ofbuild.shand fails when a module undersrc/drivers/is missing from it, so the gap cannot reopen. Verified by a full AppImage rebuild with the guard expanded.
- README status table matches the capability matrix: Hyprland and the X11 desktops are marked as author-tested (record, replay and clicks), on par with the KDE Plasma Wayland column that was already listed as verified. Before this the table said only KDE was tested while the matrix below it claimed three columns were verified. GNOME Wayland stays marked as implemented but not tested by the author.
- Sway column added to the capability matrix, with a note explaining why absolute recording is impossible there (no cursor query in the sway IPC) and that macros still replay through the uinput fallback.
v3.1.2 - 2026-10-07
- Installer confirms before sudo:
tools/install.shasks for confirmation before escalating and accepts--yesto skip the prompt for unattended installs.
- KDE Wayland buttons reach the portal as real mouse buttons: clicks, right-clicks and middle-clicks are translated to the evdev codes the RemoteDesktop portal expects (272/273/274) instead of X11 button numbers 1/2/3, which portal versions ignored or misread.
- A dead portal is reported instead of silently swallowing input: when
the portal session fails to start or a portal call fails, the driver
raises a one-time warning in the UI (repeats stay in the log) and the
event falls back to
uinput, instead of dropping it with no signal. - KDE Wayland cursor position is read from the compositor, not from
frozen
xdotool:xdotool getmouselocationonly tracks the pointer over XWayland surfaces, so on a Wayland desktop it froze on one stale coordinate and every recorded macro carried the same position — replays then clicked wherever the pointer happened to be. The driver now reads the true position from KWin's scripting D-Bus API (workspace.cursorPos, thekdotoolmechanism, ~7 ms per read), falls back toxdotooland then to the last known position, and warns once if KWin is unavailable so a stale-coordinate macro can never be recorded silently again.
v3.1.1 - 2026-10-04
- Vision/capture helpers moved out of
src/:vision.pyandcapture.pynow live inexperimental/with the add-ons that use them, keeping them out of the main package and the AppImage. - X11 driver verified: record/replay tested by the author on a KDE Plasma X11 session; the README status table and capability matrix now mark the X11 column as tested.
evdevpinned end to end:requirements.txt, the AppImagepyprojectand the.venvcreated byinstall.shall run the pinned 2.0.0, so CI, the venv and the packaged app stop testing three different versions.
- Coordinate suite cannot touch a live
ydotoold: the test mocks the daemon spawn and the socket cleanup, registers every patcher withaddCleanup, and ships a regression test proving a stale socket file survives the suite (reviewer-reproduced before the fix). - AppImage guard is total:
build.shselects an interpreter that actually hasappimage(existing env, then system python) and, when none exists, fails with pip/uv guidance instead ofuv: command not found. - Release script keeps the README in sync:
update_source_version()now rewrites theThis README describes vX.Y.Zline (v3.1.0 shipped while the README still said 3.0.3). color_spinsys.path inserts normalized: one spelling, membership guard, same pattern as the test suite.
v3.1.0 - 2026-10-04
- Uninstall script:
tools/uninstall.shmirrors the installer (udev rule, desktop entry, optionalinputgroup removal, pip packages behind an explicit purge flag) with a root./uninstall.shwrapper; the README keeps the manual procedure as a fallback.
main.pysplit: the 967-lineLinuxTaskAppgod object was reduced to UI and wiring (~507 lines); device capture, deduplication and timeline building moved tosrc/recorder.py, and replay, humanize jitter and macro validation tosrc/playback.py. No behavior change.- Wider CI: the workflow now runs the X11, Hyprland and input-device suites alongside the core three — 6 suites, 68 tests, all headless-safe.
v3.0.3 - 2026-10-04
- First-recording password warning: the app now warns before the first
recording that everything typed (including passwords) is captured, with
a "Don't show again" opt-out persisted to
~/.config/linuxtask/config.json; recording can be cancelled from the dialog, and stale hotkey actions are drained while it is open.
- English-first installer:
tools/install.shandtools/run.shmessages standardized in English, and the README documentstools/as the canonical install path. - Honest status table: the README lists KDE Plasma (Wayland) as tested by the author and Hyprland/GNOME/X11 as implemented but untested, with known limitations, manual uninstall steps, and an AI-assistance note.
- Readable-only input devices on evdev 2.x: device enumeration now
calls
evdev.list_devices(writable=False)(read-only access, matching the0440udev rule) with a no-arg fallback for evdev 1.x, instead of returning an empty list; an actionable warning dialog appears when no devices are accessible. - Arch-proof, PEP 668-safe installer:
tools/install.shinstalls the distro evdev/xlib/Tk packages, creates a project.venv(--system-site-packages) so pip runs outside the externally-managed system Python (PEP 668, never--break-system-packages), fails fast with a logout notice, guards the real user under sudo, and smoke-tests the app;tools/run.shprefers the venv. - Own virtual device excluded from listeners: input listeners skip the
app's
LinuxTask-Virtualuinput device by name, stopping replay echo and hotkey re-triggering during playback.
v3.0.2 - 2026-10-04
- English-first README: Portuguese survives only as a short footnote with a cross-link to the English guide.
- Animated demo GIF: shows keyboard-only record/replay on KDE Wayland instead of a static screenshot.
- Quickstart heading style: uses the plain style, matching the rest of the docs.
- Prototypes quarantined: the unfinished libei prototypes moved to
.quarantine/, out of the AppImage payload. - Font patch retargeted: the AppImage font patch targets the actual toolbar font size instead of a stale pattern that never matched.
- Add-on-only dependencies:
requirements.txtmarks numpy/Pillow as add-on-only rather than app dependencies. - Narrowed JSON ignore: the global
*.jsonignore is narrowed so config files stay visible. - House-style rewrite: changelog rewritten in one house style (plain Keep a Changelog, no emoji) from v3.0.1 down to v2.0.0.
- Compare-link footer: compare-link footer added, and dead links to the untagged v2.0.0/v2.2.0 removed.
- Unreleased draining: the release script now drains
[Unreleased]into the new version heading and resets it, instead of leaving released notes on top as still pending. - Version-based base tag: the script selects the base tag by version rather than commit distance, and keeps the footer links current.
- Faithful dry run:
--dry-runpreviews the entry that will actually be published instead of a generated draft.
- Per-event failure tally: playback counts per-event failures and reports a summary afterward, so a replay with skipped events is visible.
- Isolated bad timestamps: a malformed event timestamp now fails that event only instead of aborting the whole macro.
- Hyprland read fallback:
get_cursor_pos()returnsNoneon read failure instead of(0, 0), so a transienthyprctlfailure falls back to uinput instead of jumping the cursor to the corner. - Shared portal screen size: the KDE portal driver uses the shared screen-size attributes instead of private copies stuck at zero.
- Missing hyprctl contained: a missing
hyprctlno longer escapes playback mid-macro via an unhandledFileNotFoundError. - No corner jump on failure: X11/GNOME cursor reads return
Noneon failure instead of(0, 0), so a failed read aborts the move rather than jumping to the corner. - Immediate Stop response: playback sleeps in slices that check the stop flag, so Stop responds immediately instead of after the full gap.
- Bounded event dedupe: event dedupe evicts partially through a deque instead of clearing wholesale, closing the double-fire window in long sessions.
- Missing ydotool handled: the GNOME driver returns early when the
ydotoolpath is missing instead of raisingTypeError. - Lock-guarded recorder state: shared recorder state is now guarded by a lock across device threads and record/playback.
- Backend call timeouts: the automated portal steps time out after 15s
instead of freezing the UI at startup (the
Startcall that waits for the user to approve the dialog keeps its 120s), andxrandr/gdbuscalls pass a timeout like the other backends. - No shell invocation: the release script passes an argument list
with
shell=False, closing tag-based shell injection. - Validated macro I/O: macro save/load uses UTF-8 encoding and validates its input.
v3.0.1 - 2026-10-01
- Image-click add-on:
experimental/image_click/finds an image on screen and clicks it, using grim capture with a pure-numpy NCC matcher. - Color-spin add-on:
experimental/color_spin/clicks in a loop until a target color appears, with a stability guard against animated false positives and a compact GUI with live log and start/stop buttons. - Independent add-on installer:
experimental/install.shadds only numpy/Pillow, grim/slurp and desktop shortcuts;--removecleans them up, and the main install neither includes nor requires the add-on. - Vision and capture helpers: reusable
src/vision.py(template matching,color_fraction) andsrc/capture.py(grim/slurp backend) used only by the add-on, so the app itself gains no new dependency. - Add-on tests:
tests/test_experimental_addons.pycovers the add-on headless with fake backends; run manually, not in CI.
- Compact toolbar: window reduced from 420x50 to 400x44 with 32px buttons and tighter spacing (TinyTask-inspired density); all controls keep their places, with smaller Open/Save labels, the 0.5x-10x speed selector, the settings gear and hover tooltips intact.
- README and CONTRIBUTING rewrite: both files rewritten for accuracy and professional tone, with a one-line quickstart in the README.
- Release script placement: new entries are inserted below the
hand-written
[Unreleased]section instead of prepended above it (the behavior at the time; the script now drains[Unreleased]under the new heading).
v3.0.0 - 2026-09-15
- KDE Wayland support: new driver using the Portal RemoteDesktop interface with dead-reckoning position tracking, re-synced once per playback.
- Reproducible AppImage build:
tools/appimage/build.shstages and patches the source for python-build-standalone (ASCII toolbar labels for the bundled Tk). - Core flow test suite: new
tests/test_main_flow.pywith 38 headless unit tests for the heart ofmain.py(no display or hardware needed): event deduplication across devices, recording of keys/scroll/relative and absolute motion withEV_SYNflush, hotkey mapping and exclusion, humanize jitter bounds, playback dispatch with UInput fallback, speed scaling and stop interruption, macro validation and save/load roundtrip, and the thread-safe hotkey queue. Added to the CI workflow (python-app.yml).
- X11 driver rewrite:
X11Drivernow usespython-xlib(XTest) instead of shelling out toxdotool. - Thread-safe hotkeys: global hotkeys are dispatched to the UI thread via a queue; hover tooltips added for emoji-only toolbar buttons.
- Monotonic timing: macro recording and playback schedules use
time.monotonic()instead oftime.time(), making event deltas immune to NTP/DST clock jumps. Saved.jsonmacros stay compatible (they store deltas, not absolute timestamps).
- Missing dependency crashes:
factory.pynow checks forXlib(X11) anddbus/gi(KDE Wayland) before importing each driver and raises a clearRuntimeErrorwith exact install instructions (apt/pacman/dnf) instead of a rawImportError.main.pyshows the message in a visible dialog rather than dying with a traceback. - KDE Wayland uninstallable out of the box:
tools/install.shnow installs the distro D-Bus bindings (python3-dbus/python3-gi) the portal driver needs;run.shwarns with manual commands when they are absent. - X11 fallback missing python-xlib:
install.shandrun.shpip fallbacks now includepython-xlib(previously onlycustomtkinter evdev), so the generic X11 driver no longer fails to import on manual installs. - Implicit state:
_rel_dx/_rel_dyare now initialized in the app constructor instead of being created lazily insidetoggle_record(); evdev listener threads no longer depend on record having run first. - Hyprland 0.55+: cursor moves use the new Lua dispatcher
(
hl.dsp.cursor.move) with fallback to the legacymovecursorform. - Release script version regex:
tools/release.pynow reads theAPP_VERSIONconstant (the old pattern parsed a hardcoded version out of the window title and always failed) and mirrors the bump totools/appimage/pyproject.toml.
- Read-only input devices: udev rule for
event*devices changed from0660to0440— recording only needs read access, and write access on real input devices is a full injection primitive for any process in theinputgroup./dev/uinputstays0660(replay requires it). - Matching ACLs:
install.shandfix_linuxtask_perms.shnow grantu:USER:r(notrw) on/dev/input/event*, so the immediate ACL grant cannot reopen the write access the udev rule closes. - Experiment quarantined: the unfinished libei PoCs moved out of
src/drivers/intoexperimental/libei/(with a README), and the compiledei_sendbinary was removed —build.shcopiessrc/drivers/wholesale, so the AppImage no longer bundles stray artifacts.
v2.6.0 - 2026-04-07
- Full Cinnamon/X11 support: new driver using
xdotoolfor absolute and relative movement on X11 desktops (Cinnamon, MATE, XFCE, and others). - Mouse scroll recording: hardware-level capture of
REL_WHEELevents with playback support across all drivers. - Improved hotkey configuration: pressing Esc cancels hotkey remapping, and the UI confirms newly assigned keys immediately.
- GNOME driver fallback: mouse button handling via
ydotoolfor better reliability when UInput is unavailable. - Enhanced documentation:
README.mdrewritten with architecture diagrams, a support matrix, and an updated demo screenshot. - Improved installer:
install.shrefined so the desktop shortcut appears immediately and uses absolute paths. - CI automation: GitHub Actions workflows added for automated testing and tagged GitHub Releases.
- UI refresh: window width increased to
420pxto prevent text overlap, with the internal component structure refactored. - Cross-distro installation:
install.shandfix_linuxtask_perms.shdetect and supportapt,pacman, anddnfautomatically. - Enhanced core precision: relative movement accumulation logic refactored to eliminate coordinate drift and stuttering during playback.
- Legacy cleanup: the obsolete virtual-coordinates system removed in favor of the native compositor drivers.
- Double-movement bug: driver return values corrected so events no longer fire twice through UInput on Hyprland and GNOME.
- Permission management: input-device detection improved and
permission grants automated via
setfacl. - Theme consistency: settings window background fixed to respect the user theme instead of a hardcoded color.
- Release script regex: version lookup supports dynamic window titles instead of matching only a hardcoded title.
v2.4.0 - 2026-03-01
- Release automation script:
tools/release.pyadded to cut releases from conventional commits and update the changelog. - Professionalized documentation: README overhauled and
CONTRIBUTING.mdadded, giving new contributors a single entry point. - Standardized code headers: license headers and PEP 8 compliance applied across all Python files for consistent tooling output.
- Structural reorganization: repository laid out into
src/,tests/,docs/, andtools/so code, tests, docs, and helpers live apart.
- Hyprland driver validation: driver covered with mocked unit tests, locking its behavior against regressions.
- GNOME cursor research: absolute cursor position on GNOME investigated with a PoC script kept as research, not shipped as a feature.
- Dynamic screen resolution: GNOME/Wayland resolution detected at
runtime via
gdbusandxrandrinstead of assuming a fixed size.
- GNOME driver crash: missing screen-resolution attributes added to
GnomeDriver, resolving theAttributeErroron startup.
- Hardware access automation:
fix_linuxtask_perms.shadded to configure ACL and udev rules without manual steps. - Pure relative-movement engine: relative movement logic implemented for GNOME Wayland users without absolute cursor reads.
- Ydotool integration:
ydotoolddaemon management optimized for Arch/CachyOS.
- Delta-based tracking: cursor movement now uses delta-based tracking so it no longer drifts.
- Strict coordinate clamping: coordinates are strictly clamped so the cursor no longer snaps to the corner.
- Humanize mode: playback adds ±2px jitter and 0-3% timing delays so macros mimic human input more closely.
- Settings black-screen fix: the "black screen" bug on Wayland/Hyprland is fixed.
- Better settings-window contrast: contrast improved across the settings window.
- Stable desktop shortcut: shortcut icon pinned to
input-mousefor a consistent launcher appearance. - Movement-precision tests:
test_jitter.pyadded to cover movement precision with unit tests.