chore: 1.5.1

Every AppImage came out as "cannot tell" because the payload is a
compressed filesystem and none of the Chromium markers is on disk to be
found. The names are reachable, though: squashfs keeps the name of
everything it holds in one table of a few kilobytes, and a name is all
this question needs. That table is now read and inflated in place -
nothing is unpacked, and the image is never run, which a scan started by
the path unit has no business doing anyway.

So a Tauri AppImage such as Modrinth is recognised as WebKitGTK and
dropped from the list instead of being offered as something that could be
switched on - there is no autoscroll in WebKit to ask for - and an
Electron one is covered like any other application.

Left as "cannot tell": lzo and lz4, which squashfs stores as bare blocks
no tool will read without their own framing, and the original ISO9660
layout.

Alongside:

- The detect cache carries a format line. Its entries are keyed on size
  and mtime, so an AppImage that has not changed would otherwise keep
  answering the way an older version decided it did, forever.

- A marker list for whole-tree searches, which is what a Flatpak, a snap
  and an AppImage need. icudtl.dat, snapshot_blob.bin and resources.pak
  are out of it: Flutter ships an icudtl.dat in data/, and next to a
  binary those names are evidence while four directories down they are
  not. The Flatpak and snap searches now share that list.

.github/FUNDING.yml

@@ -0,0 +1,15 @@
+# These are supported funding model platforms
+
+github: Felitendo
+patreon: # Replace with a single Patreon username
+open_collective: # Replace with a single Open Collective username
+ko_fi: felitendo
+tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
+community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
+liberapay: # Replace with a single Liberapay username
+issuehunt: # Replace with a single IssueHunt username
+lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
+polar: # Replace with a single Polar username
+buy_me_a_coffee: # Replace with a single Buy Me a Coffee username
+thanks_dev: # Replace with a single thanks.dev username
+custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

Makefile

@@ -8,7 +8,7 @@
 # Overridable so a packager can pass the version it is actually building
 # (`make VERSION=$pkgver`). The literal below is the fallback for builds
 # straight from a checkout, and is what a release tag has to carry.
-VERSION      ?= 1.4.0
+VERSION      ?= 1.5.1
 
 PREFIX       ?= /usr
 DESTDIR      ?=
@@ -23,13 +23,17 @@ MANDIR       ?= $(DATADIR)/man
 # distribution that still keeps /lib separate from /usr/lib says so here - and
 # only guessed at when there is no systemd installed to ask.
 #
-# A build with a prefix of its own keeps the units under that prefix instead.
+# A build with a prefix of its own keeps the units under that prefix instead,
-# systemd searches $(PREFIX)/lib/systemd/user as well, and a file outside the
+# and under share rather than lib. systemd searches both for a system prefix,
-# prefix it was asked for is not this build's to place.
+# but only share for a home one: $XDG_DATA_HOME/systemd/user is a search path
+# and ~/.local/lib/systemd/user is not, so an install into ~/.local that put
+# the units in lib would leave the watcher impossible to enable. That prefix is
+# the one an atomic distribution leaves a user - there is no writing to /usr on
+# Bazzite or Silverblue without layering a package and rebooting.
 ifeq ($(PREFIX),/usr)
 USERUNITDIR  ?= $(shell pkg-config --variable=systemduserunitdir systemd 2>/dev/null || echo /usr/lib/systemd/user)
 else
-USERUNITDIR  ?= $(PREFIX)/lib/systemd/user
+USERUNITDIR  ?= $(DATADIR)/systemd/user
 endif
 
 LINGUAS      := de

README.md

@@ -1,9 +1,55 @@
 # middleclick-autoscroll
 
-Middle-click autoscroll is a CLI for Linux that enables autoscroll in every application that supports it.
+middleclick-autoscroll (what a great name, wow) is a CLI for linux that enables autoscroll in every application that supports it.
-Works with browsers, Electron apps like Discord and Spotify, Steam, and anything else that runs on Chromium under the hood.
+Works with browsers, electron apps like discord and spotify, steam, and anything else that runs on chromium.
 
-There is no app list. My tool just looks at what uses chromium under the hood and applies the necessary steps to get it working.
+This tool looks at which of your apps run chromium under the hood and applies the necessary steps to get autoscrolling working (often just a feature flag). One install, one command and autoscroll _✨just works✨_ (like on windows).
+
+## How to use
+
+Just run `middleclick-autoscroll`. This will open the configuration TUI that looks like this:
+
+```
+  Middle-Click Autoscroll
+
+  Autoscroll                     ON
+
+  Applications covered           11 of 13
+  Not identified                 1 - see the applications list
+  Steam                          ON
+  New applications               ON
+  Middle-click paste             off
+  Last applied                   2 minutes ago
+
+  Applications pick this up the next time they are started.
+
+  [1] Turn autoscroll on or off
+  [2] Re-apply everything
+  [3] Applications
+  [4] Settings
+  [q] Quit
+
+  >
+```
+
+Normally you just need to press `[1]` and the magic is done.
+Pressing `[3]` lets you see every app that was found and toggle each one individually. 
+
+```
+  Applications
+
+  ▸ Steam                              on (Steam)
+    Chromium                           on (launcher)
+    Discord                            on (flag file)
+    Obsidian                           on (launcher)
+    Spotify                            on (launcher)
+    Slack                              off
+    Cursor                             cannot tell
+
+  Up/Down select - Space turns one on or off - q goes back
+```
+
+You can also press `[4]` for more settings.
 
 ## How to install
 
@@ -21,6 +67,15 @@ sudo curl -fsSL -o /etc/yum.repos.d/middleclick-autoscroll.repo \
 sudo dnf install middleclick-autoscroll
 ```
 
+**Bazzite**
+
+```bash
+sudo curl -fsSL -o /etc/yum.repos.d/middleclick-autoscroll.repo \
+  https://felitendo.github.io/middleclick-autoscroll/middleclick-autoscroll.repo
+sudo rpm-ostree install middleclick-autoscroll
+systemctl reboot
+```
+
 **Debian**
 
 ```bash
@@ -41,99 +96,37 @@ sudo zypper addrepo --gpgcheck --refresh \
 sudo zypper install middleclick-autoscroll
 ```
 
-## How to use
-
-Just run `middleclick-autoscroll`. This will open the configuration TUI that looks like this:
-
-```
-  Middle-Click Autoscroll
-
-  Autoscroll                     ON
-
-  Applications covered           11 of 13
-  Not identified                 1 - see the applications list
-  Steam                          ON
-  New applications               ON
-  Last applied                   2 minutes ago
-
-  Applications pick this up the next time they are started.
-
-  [1] Turn autoscroll on or off
-  [2] Re-apply everything
-  [3] Applications
-  [4] Settings
-  [q] Quit
-
-  >
-```
-
-`[1]` is all you need for the normal case. `[3]` lists every app that was found
-and what is being done with it, so you can leave a single one out or switch on
-an AppImage that couldn't be identified:
-
-```
-  Applications
-
-  ▸ Steam                              on (Steam)
-    Chromium                           on (launcher)
-    Discord                            on (flag file)
-    Obsidian                           on (launcher)
-    Spotify                            on (launcher)
-    Slack                              off
-    Cursor                             cannot tell
-
-  Anything not identified is left alone until it is turned on here.
-  Up/Down select - Space turns one on or off - q goes back
-```
-
-`[4]` has the same switches per category - browsers, Electron and CEF apps,
-Flatpaks, snaps, autostart entries, Steam, Spotify - plus a field for extra
-Chromium arguments.
-
 ## How it works
 
 Blink (the engine in Chromium, Electron, and CEF) already has autoscroll, but
-it's off on Linux because middle click does primary-selection paste there. This
+it's off on Linux by default because middle click normally pastes the clipboard (??).
-flag turns it on:
+This flag turns it on:
 
 --enable-blink-features=MiddleClickAutoscroll
 
 But doing that for every app is kinda bothersome and it also might break with updates.
 That's why I created this small tool to automate that.
 
-Browsers get the same thing asked for differently:
+Browsers get the same thing but this time without "blink":
 
 --enable-features=MiddleClickAutoscroll
 
-Both mean the same feature - Blink generates a matching feature name for each of
+But both do the same ¯\_(ツ)_/¯
-its runtime flags - but the first one is on Chromium's list of flags worth
+
-warning about, so a browser started with it shows a yellow "unsupported
+Except in Helium and Brave, which have their own names for it
-command-line flag" bar over every page. The second one is not on that list, so
+(`HeliumMiddleClickAutoscroll`, and Brave's is actually spelled
-there's no bar. It only works from Chromium 124 on, which is why apps that
+`MiddelButtonClickAutoscroll`), so browsers get those as well.
-embed something older (Steam's CEF, older Electron) keep the first one - they
-have no such bar to begin with. Helium, which carries the feature under its own
-name, is asked for `HeliumMiddleClickAutoscroll` alongside it.
 
 New apps are picked up by a systemd path unit that watches the
-relevant directories. Without systemd, `middleclick-autoscroll apply` does the
+relevant directories. If you hate systemd; `middleclick-autoscroll apply` does the
 same thing manually.
 
-## Steam
+AppImages get looked into as well. The payload is a squashfs image glued onto
-
+the runtime, and squashfs keeps a small table with the name of every file in
-Steam's web UI supports autoscroll but has no way to pass extra arguments to its
+there - so that table gets read and unpacked (a few kb, nothing is extracted
-helper, so the program patches the script that starts it. Steam checks its own
+and the thing is never run) and the Chromium files are looked for in it. An
-files against its manifest - by size and timestamp - and reinstalls the whole
+Electron AppImage is covered like any other app; a Tauri one is left alone,
-client package over anything that differs, so the patch is written to look
+because WebKitGTK simply has no autoscroll to switch on.
-unchanged: the bytes the argument costs are taken back out of the script's
-comments and the timestamp is put back.
-
-If that can't be done - a script with no comments left to pay for the argument,
-or one this version doesn't recognise - nothing is written and Steam is left
-alone. A file of the wrong length costs the client package downloaded and
-installed again, and a Steam that quits at the end of it instead of coming up;
-no autoscroll in Steam is the smaller problem. Nothing is added to Steam's
-launcher entry or to the shortcuts it writes for games, and Steam keeps
-repairing its own installation.
 
 ## Commands
 

doc/middleclick-autoscroll.1.scd

@@ -26,10 +26,11 @@ runtime flags - but it is not on the list of flags Chromium warns about, so the
 browser does not put a bar reading "You are using an unsupported command-line
 flag" above every page. That spelling only works from Chromium 124 onwards,
 which is why everything else keeps the first one: an application that embeds an
-older Chromium, such as Steam's CEF, has no such bar to show anyway. Helium
+older Chromium, such as Steam's CEF, has no such bar to show anyway. Two
-knows the feature under a name of its own and is asked for
+browsers know the feature under a name of their own and ignore Chromium's:
-*HeliumMiddleClickAutoscroll* as well; a name a browser does not know is
+Helium, which is asked for *HeliumMiddleClickAutoscroll* as well, and Brave,
-ignored. It works on any distribution:
+which is asked for *MiddelButtonClickAutoscroll* - spelled the way Brave spells
+it. A name a browser does not know is ignored. It works on any distribution:
 which of the routes below an application takes is read off its launcher, not
 assumed from where the launcher came from.
 
@@ -172,11 +173,66 @@ and has a configuration file with a slot for extra arguments; that slot is
 where the flag goes. Installed as a plain package, a Flatpak or a snap, it is
 an ordinary desktop entry and needs nothing special.
 
-# WHAT CANNOT BE DETECTED
+# MIDDLE-CLICK PASTE
 
-An AppImage keeps its payload in a compressed filesystem, so there is no way to
+Inside a Chromium application the argument settles this on its own: with
-tell from the outside whether it contains Chromium. Those are listed as *cannot
+autoscroll on, Blink stops pasting the primary selection on middle click,
-tell* and left alone until they are switched on from the applications screen.
+because the button is doing something else now. Everywhere else on the desktop
+middle click goes on pasting, so the desktop's own switch for it is turned off
+too.
+
+KDE has such a switch, and only for a Wayland session, where the paste is a
+protocol the compositor either offers or does not:
+
+	~/.config/kwinrc, [Wayland], EnablePrimarySelection=false
+
+It is written with *kwriteconfig6*, the same tool System Settings uses, because
+kwinrc has a cascade behind it and entries a distribution can mark immutable;
+an edit that reaches past all that writes something KDE goes on to ignore. KWin
+reads the key once, while it starts, so the change is true from the next login
+on rather than straight away.
+
+What the key said before is recorded, and *disable* puts that back. A key that
+already said *false* before any of this is left alone and not recorded either,
+so a paste the user switched off themselves is never switched back on.
+
+On X11 there is nothing to switch. The primary selection is part of X itself
+and every toolkit reaches for it on its own; no one place can say no. The
+status screen leaves the row out where there is nothing to report.
+
+Turn it off under *Settings* to keep the desktop's middle-click paste.
+
+# APPIMAGES
+
+An AppImage is the AppImage runtime, an ordinary executable, with a squashfs
+image appended to it. The payload is compressed, so none of the files that
+identify Chromium is on disk where it could be found.
+
+The names are, though. squashfs keeps the name of everything it holds in one
+table of its own, a few kilobytes of it, and a name is all this question needs -
+so that table is read and decompressed in place. Nothing is unpacked, and the
+image is never run: this can happen from the watcher, and starting a program
+because it was installed is not something a scan gets to do.
+
+Two kinds of image are still out of reach. One is packed with *lzo* or *lz4*,
+which squashfs stores as bare blocks that neither tool will read without the
+framing their own file formats add. The other is the original AppImage layout,
+which is an ISO9660 filesystem rather than a squashfs one. Both are listed as
+*cannot tell* and left alone until they are switched on from the applications
+screen.
+
+# WHAT HAS NO FLAG AT ALL
+
+Autoscroll is a Blink feature, so only an application drawn by Blink can be
+given it. An application built on WebKitGTK - which is what Tauri uses on
+Linux, and GNOME Web, and anything else linked against libwebkit2gtk - is not
+one of those. WebKit has no equivalent feature to ask for, on any command line
+or in any configuration file.
+
+Such an application is identified as what it is and then left out of the list
+entirely, rather than being offered as something that could be switched on. An
+application that is absent from *list* is absent because there is nothing this
+tool could do for it.
 
 # FILES
 
@@ -190,9 +246,17 @@ _~/.local/state/middleclick-autoscroll/ledger_
 _~/.local/state/middleclick-autoscroll/backup/_
 	Copies of the files that are edited in place rather than shadowed.
 
+_~/.config/kwinrc_
+	KDE's own configuration. One key in it, *EnablePrimarySelection*, is
+	written when the desktop's middle-click paste is turned off, and put
+	back by *disable*.
+
 _~/.cache/middleclick-autoscroll/detect_
 	Which programs were found to be Chromium, keyed by size and modification
-	time. Safe to delete.
+	time. Its first line says which version of the program wrote it, and a
+	file from an older one is ignored rather than trusted - an entry for a
+	file that has not changed would otherwise never be looked at again.
+	Safe to delete.
 
 # ENVIRONMENT
 
@@ -205,9 +269,14 @@ Bash 4.2 or newer and GNU coreutils. The watcher needs a systemd user session;
 without one everything else works and applications installed later are picked
 up at the next *apply* rather than on their own.
 
+*gzip* is what reads the name table of an AppImage, and *xz* or *zstd* the ones
+packed with those instead. None of the three is required: an image whose
+compressor has no tool installed is listed as *cannot tell*, exactly like the
+two formats there is no reader for at all.
+
 # SEE ALSO
 
-*systemctl*(1), *flatpak*(1), *snap*(8)
+*systemctl*(1), *flatpak*(1), *snap*(8), *kwriteconfig6*(1)
 
 # AUTHORS
 

po/de.po

@@ -63,6 +63,9 @@ msgstr "%d – siehe Anwendungsliste"
 msgid "Steam"
 msgstr "Steam"
 
+msgid "Middle-click paste"
+msgstr "Mittelklick-Einfügen"
+
 msgid "New applications"
 msgstr "Neue Anwendungen"
 
@@ -112,6 +115,9 @@ msgstr "Programme, die sich beim Anmelden selbst starten"
 msgid "Spotify"
 msgstr "Spotify"
 
+msgid "Turn off KDE's middle-click paste"
+msgstr "KDEs Mittelklick-Einfügen abschalten"
+
 msgid "Apply to newly installed applications"
 msgstr "Auf neu installierte Anwendungen anwenden"
 
@@ -179,6 +185,9 @@ msgstr "Steam startet seine Oberfläche auf eine Art, die diese Version nicht ä
 msgid "Steam's interface cannot be given the flag without leaving a file Steam would reinstall; the change has been taken back."
 msgstr "Steams Oberfläche lässt sich das Flag nicht geben, ohne eine Datei zu hinterlassen, die Steam neu installieren würde; die Änderung wurde zurückgenommen."
 
+msgid "Middle click stops pasting at the next login."
+msgstr "Mittelklick fügt ab der nächsten Anmeldung nichts mehr ein."
+
 
 #. Help
 msgid "Usage: middleclick-autoscroll [command]"

po/middleclick-autoscroll.pot

@@ -63,6 +63,9 @@ msgstr ""
 msgid "Steam"
 msgstr ""
 
+msgid "Middle-click paste"
+msgstr ""
+
 msgid "New applications"
 msgstr ""
 
@@ -112,6 +115,9 @@ msgstr ""
 msgid "Spotify"
 msgstr ""
 
+msgid "Turn off KDE's middle-click paste"
+msgstr ""
+
 msgid "Apply to newly installed applications"
 msgstr ""
 
@@ -179,6 +185,9 @@ msgstr ""
 msgid "Steam's interface cannot be given the flag without leaving a file Steam would reinstall; the change has been taken back."
 msgstr ""
 
+msgid "Middle click stops pasting at the next login."
+msgstr ""
+
 
 #. Help
 msgid "Usage: middleclick-autoscroll [command]"

src/lib/apply.sh

@@ -134,6 +134,11 @@ mca_apply() {
 
 	[[ $CFG_SPOTIFY == yes ]] && mca_spotify_apply
 
+	# Not about any one application: middle click goes on pasting everywhere
+	# else on the desktop, and KDE is the one desktop that can be told not to.
+	# See kde.sh.
+	[[ $CFG_PASTE == yes ]] && mca_kde_paste_apply
+
 	mca_prune_orphans
 
 	if (( MCA_CHANGES )) && mca_have update-desktop-database; then

src/lib/common.sh

@@ -35,15 +35,19 @@ MCA_PRETTY="Middle-Click Autoscroll"
 # the one that always works.
 MCA_FLAG="--enable-blink-features=MiddleClickAutoscroll"
 
-# Helium ships the feature under its own name and does not answer to the
+# Two browsers do not answer to the Chromium name. Helium ships the feature
-# Chromium one, so browsers are asked for both. A name a browser does not know
+# under its own name and sets the runtime flag from that alone. Brave takes
-# is ignored, which is what makes one list safe to hand to all of them.
+# the generated feature away and switches the flag on from one of its own,
-MCA_BROWSER_FLAG="--enable-features=MiddleClickAutoscroll,HeliumMiddleClickAutoscroll"
+# spelled MiddelButtonClickAutoscroll - the typo is in Brave's source, and the
+# correct spelling does nothing. So browsers are asked for all three. A name a
+# browser does not know is ignored, which is what makes one list safe to hand
+# to all of them.
+MCA_BROWSER_FLAG="--enable-features=MiddleClickAutoscroll,HeliumMiddleClickAutoscroll,MiddelButtonClickAutoscroll"
 
 # The feature names on their own, for merging into a list that an application
 # (or the user) already carries, and for taking them back out again.
 MCA_FEATURE="MiddleClickAutoscroll"
-MCA_BROWSER_FEATURES="MiddleClickAutoscroll,HeliumMiddleClickAutoscroll"
+MCA_BROWSER_FEATURES="MiddleClickAutoscroll,HeliumMiddleClickAutoscroll,MiddelButtonClickAutoscroll"
 
 # Bumped when the flags above change, and when what gets written where does.
 # An installation that was set up by a version with a different answer is taken
@@ -52,7 +56,10 @@ MCA_BROWSER_FEATURES="MiddleClickAutoscroll,HeliumMiddleClickAutoscroll"
 #
 # 3 is what takes -noverifyfiles back out of the Steam entries an older version
 # put it in.
-MCA_FLAG_SCHEME=3
+#
+# 4 adds Brave's name to the browser list, which an entry edited in place would
+# otherwise never get.
+MCA_FLAG_SCHEME=4
 
 MCA_LIBDIR="${MCA_LIBDIR:-@LIBDIR@}"
 MCA_LOCALEDIR="${MCA_LOCALEDIR:-@LOCALEDIR@}"
@@ -278,9 +285,12 @@ mca_state_write() {
 #   shadow  <generated desktop entry>   <source entry it was generated from>
 #   inplace <edited file>               <basename of its backup copy>
 #   flags   <flag file>                 created | appended | merged
+#   kwin    <kwinrc>                    what EnablePrimarySelection said before
 #
-# `inplace` and `flags` differ in how they are undone: a backup is restored
+# The kinds differ in how they are undone: a backup is restored wholesale, a
-# wholesale, a flag file only loses the one line that was added to it.
+# flag file only loses the one line that was added to it, and kwinrc gets one
+# key written back - it is a file KDE writes to itself, and putting a whole
+# copy of it back would take everything else settled since with it.
 
 mca_ledger_add() {
 	local kind="$1" path="$2" detail="${3:-}"
@@ -303,6 +313,15 @@ mca_ledger_has() {
 	awk -F'\t' -v p="$1" '$2 == p { found = 1 } END { exit !found }' "$MCA_LEDGER"
 }
 
+# mca_ledger_detail <path>
+# The third column of that file's line, empty when the ledger has none. What it
+# means is the kind's business; for `kwin` it is the value to put back.
+mca_ledger_detail() {
+	[[ -f $MCA_LEDGER ]] || return 1
+	awk -F'\t' -v p="$1" '$2 == p { d = $3 } END { if (!length(d)) exit 1; print d }' \
+		"$MCA_LEDGER"
+}
+
 mca_ledger_forget() {
 	local path="$1" tmp
 	[[ -f $MCA_LEDGER ]] || return 0

src/lib/config.sh

@@ -144,6 +144,7 @@ mca_config_load() {
 	CFG_STEAM=no;      mca_config_bool PatchSteam       yes && CFG_STEAM=yes
 	CFG_SPOTIFY=no;    mca_config_bool PatchSpotify     yes && CFG_SPOTIFY=yes
 	CFG_WATCH=no;      mca_config_bool WatchNewApps     yes && CFG_WATCH=yes
+	CFG_PASTE=no;      mca_config_bool DisablePaste     yes && CFG_PASTE=yes
 
 	CFG_EXTRA_FLAGS="$(mca_config_get ExtraFlags '')"
 

src/lib/detect.sh

@@ -21,6 +21,22 @@ MCA_MARKERS=(
 	libvk_swiftshader.so LICENSES.chromium.html libcef.so
 )
 
+# The markers that still mean something when the whole of a tree is searched
+# rather than just the directory the binary sits in - inside a Flatpak, a snap
+# or an AppImage, where there is no "next to the binary" to look at.
+#
+# Three of the names above are missing here on purpose. icudtl.dat is ICU's
+# data file and not Chromium's: a Flutter application ships one in data/, and
+# a whole-tree search would find it. snapshot_blob.bin and resources.pak are
+# the same kind of shared name. Next to a binary they are still evidence,
+# because nothing but Chromium unpacks its payload there - a few directories
+# deeper they are not.
+MCA_MARKERS_STRICT=(
+	chrome_crashpad_handler chrome-sandbox chrome_100_percent.pak
+	v8_context_snapshot.bin libvk_swiftshader.so LICENSES.chromium.html
+	libcef.so app.asar
+)
+
 # Shared directories, where a marker belongs to the system rather than to the
 # program that happens to live there. An application ships its payload in a
 # directory of its own; nothing unpacks Chromium straight into /usr/lib.
@@ -309,14 +325,27 @@ _mca_stat_batch() {
 	return 0
 }
 
+# The first line of the cache file, and the reason it is there: the verdicts
+# below it are keyed on size and mtime, so an entry for a file that has not
+# changed is never looked at again. A cache written when a verdict meant
+# something else - before an AppImage could come out as anything but "no" -
+# would therefore keep answering the old way forever. Bumping this is how such
+# a cache gets dropped instead.
+MCA_CACHE_FORMAT='# middleclick-autoscroll detect 2'
+
 _mca_cache_load() {
-	local path stamp verdict
+	local path stamp verdict first=1
 
 	(( MCA_CACHE_LOADED )) && return 0
 	MCA_CACHE_LOADED=1
 
 	[[ -r "$MCA_CACHEDIR/detect" ]] || return 0
 	while IFS=$'\t' read -r path stamp verdict; do
+		if (( first )); then
+			first=0
+			[[ $path == "$MCA_CACHE_FORMAT" ]] || return 0
+			continue
+		fi
 		[[ -n $path && -n $stamp ]] || continue
 		MCA_DETECT_CACHE["$path"]="$stamp"$'\t'"$verdict"
 	done < "$MCA_CACHEDIR/detect"
@@ -331,6 +360,7 @@ mca_cache_flush() {
 	mkdir -p "$MCA_CACHEDIR" 2>/dev/null || return 0
 
 	tmp="$(mktemp "$MCA_CACHEDIR/detect.XXXXXX")" || return 0
+	printf '%s\n' "$MCA_CACHE_FORMAT" > "$tmp"
 	for path in "${!MCA_DETECT_CACHE[@]}"; do
 		entry="${MCA_DETECT_CACHE[$path]}"
 		printf '%s\t%s\n' "$path" "$entry" >> "$tmp"
@@ -362,6 +392,26 @@ _mca_has_markers() {
 	return 1
 }
 
+# _mca_find_markers <directory> <depth>
+# The same question for a whole tree, which is the shape a Flatpak, a snap and
+# an unpacked AppImage come in: everything the application ships is somewhere
+# under one root and there is no single directory that is "next to the binary".
+# Hence the narrower list - see MCA_MARKERS_STRICT.
+_mca_find_markers() {
+	local root="$1" depth="$2" m
+	local -a names=()
+
+	[[ -d $root ]] || return 1
+
+	for m in "${MCA_MARKERS_STRICT[@]}"; do
+		(( ${#names[@]} )) && names+=(-o)
+		names+=(-name "$m")
+	done
+
+	[[ -n "$(find "$root" -maxdepth "$depth" \
+		\( "${names[@]}" \) -print -quit 2>/dev/null)" ]]
+}
+
 # The plain assignments the script made before it handed over, for
 # _mca_script_subst to read. A variable of its own rather than something passed
 # around: every caller of _mca_script_target reads it through a command
@@ -521,18 +571,295 @@ mca_flags_candidates() {
 	fi
 }
 
-# mca_is_chromium <program>
+# ---------------------------------------------------------------------------
-# Succeeds when the program is a Chromium, CEF or Electron process.
+# Looking inside an AppImage
-mca_is_chromium() {
+# ---------------------------------------------------------------------------
-	local prog="$1" verdict real stamp cached
+#
+# An AppImage is the AppImage runtime - an ordinary ELF executable - with a
+# squashfs image appended to it. The payload is compressed, so nothing the
+# application ships is on disk where the marker check could find it, and for a
+# long time that made every AppImage a "cannot tell".
+#
+# It does not have to be. squashfs keeps the names of everything it holds in
+# one table of its own, the table is a few kilobytes, and the names are the
+# whole of what this question needs. So that table is read and inflated here -
+# rather than unpacking a hundred megabytes, and rather than running the image
+# to ask it, which is the one thing a scan started by a path unit must never
+# do.
+#
+# Everything below fails closed. The moment a field is not where it should be,
+# or the compressor is one there is no tool for, the answer goes back to
+# "cannot tell" and the application is left to the applications screen.
 
-	[[ -n $prog && -e $prog ]] || return 1
+# _mca_bytes <file> <offset> <count>
+# Count bytes from an offset, as numbers, in MCA_BYTES. od rather than a shell
+# read because a variable cannot hold a NUL byte and these headers are full of
+# them - and a whole header at a time rather than a field at a time, because
+# the first question below is asked about every program on the system and one
+# od per field would be four forks per program.
+MCA_BYTES=()
+
+_mca_bytes() {
+	local file="$1" off="$2" count="$3"
+
+	MCA_BYTES=()
+
+	# read stops on end of input rather than on the delimiter it was given, so
+	# it reports failure for input that is perfectly complete. The count is
+	# the check.
+	read -r -d '' -a MCA_BYTES \
+		< <(od -An -tu1 -j"$off" -N"$count" -v -- "$file" 2>/dev/null)
+
+	(( ${#MCA_BYTES[@]} == count ))
+}
+
+# _mca_slice <file> <offset> <length>
+# Length bytes from an offset, on stdout.
+_mca_slice() {
+	dd if="$1" bs=65536 iflag=skip_bytes,count_bytes \
+		skip="$2" count="$3" status=none 2>/dev/null
+}
+
+# _mca_word <index>
+# One 64-bit little-endian field out of MCA_BYTES, in MCA_WORD. Assigns rather
+# than prints for the same reason _mca_bytes reads a whole header at a time:
+# printing means a command substitution, and that is a fork per field.
+#
+# Fails for the all-ones value squashfs writes down for a table that is not in
+# the image, which is a case the callers below have to tell from a real offset
+# and bash arithmetic - signed, 64-bit - cannot represent.
+MCA_WORD=0
+
+_mca_word() {
+	local at="$1"
+
+	(( ${#MCA_BYTES[@]} >= at + 8 )) || return 1
+	(( MCA_BYTES[at+7] == 255 && MCA_BYTES[at+6] == 255 )) && return 1
+
+	MCA_WORD=$(( (MCA_BYTES[at] | MCA_BYTES[at+1] << 8 \
+		| MCA_BYTES[at+2] << 16 | MCA_BYTES[at+3] << 24) \
+		| ((MCA_BYTES[at+4] | MCA_BYTES[at+5] << 8 \
+		| MCA_BYTES[at+6] << 16 | MCA_BYTES[at+7] << 24) << 32) ))
+	return 0
+}
+
+# _mca_appimage_offset <file>
+# Where the appended image starts, in MCA_APPIMAGE_AT: directly after the ELF,
+# which is where its section header table ends. This is how the runtime works
+# it out for --appimage-offset, and it is exact.
+#
+# Searching for the squashfs magic instead would not be: the runtime carries a
+# copy of it in its own code. Neither is the AppImage magic at byte 8 of the
+# ELF header a way in - it is there to be zeroed, and a build pipeline that
+# does not want its image picked up by a desktop integration daemon does
+# exactly that. So the offset is computed for any ELF at all and the caller
+# finds out whether there is an image at it.
+MCA_APPIMAGE_AT=0
+
+_mca_appimage_offset() {
+	local shoff shentsize shnum
+
+	MCA_APPIMAGE_AT=0
+	_mca_bytes "$1" 0 64 || return 1
+
+	# \x7fELF, and then the class byte: 2 for the 64-bit header, 1 for the
+	# 32-bit one, which puts every field after it somewhere else.
+	(( MCA_BYTES[0] == 127 && MCA_BYTES[1] == 69 \
+		&& MCA_BYTES[2] == 76 && MCA_BYTES[3] == 70 )) || return 1
+
+	if (( MCA_BYTES[4] == 2 )); then
+		_mca_word 40 || return 1
+		shoff="$MCA_WORD"
+		shentsize=$(( MCA_BYTES[58] | MCA_BYTES[59] << 8 ))
+		shnum=$(( MCA_BYTES[60] | MCA_BYTES[61] << 8 ))
+	elif (( MCA_BYTES[4] == 1 )); then
+		shoff=$(( MCA_BYTES[32] | MCA_BYTES[33] << 8 \
+			| MCA_BYTES[34] << 16 | MCA_BYTES[35] << 24 ))
+		shentsize=$(( MCA_BYTES[46] | MCA_BYTES[47] << 8 ))
+		shnum=$(( MCA_BYTES[48] | MCA_BYTES[49] << 8 ))
+	else
+		return 1
+	fi
+
+	(( shoff > 0 && shentsize > 0 && shnum > 0 )) || return 1
+	MCA_APPIMAGE_AT=$(( shoff + shentsize * shnum ))
+	return 0
+}
+
+# _mca_inflate <compressor> <file> <offset> <length>
+# One squashfs metadata block, decompressed onto stdout.
+#
+# lzo and lz4 are missing because squashfs stores them as bare blocks and
+# neither lzop nor the lz4 tool will read one without the framing their own
+# file format puts around it. An image compressed with either comes out as
+# "cannot tell", which is where it started.
+_mca_inflate() {
+	local comp="$1" file="$2" off="$3" len="$4"
+
+	case "$comp" in
+		1)
+			# squashfs stores a zlib stream and gzip only reads its own
+			# container, but underneath both are the same deflate data with a
+			# different wrapper around it - so the wrapper is swapped: zlib's
+			# two header bytes are dropped and a minimal gzip header put in
+			# front. gzip then writes every byte of the block and complains
+			# about the trailer it did not get, which is why its status is
+			# thrown away here and the check is on the output instead.
+			(( len > 2 )) || return 1
+			{
+				printf '\037\213\010\000\000\000\000\000\000\003'
+				_mca_slice "$file" $(( off + 2 )) $(( len - 2 ))
+			} | { gzip -dc 2>/dev/null || true; }
+			;;
+		2)
+			mca_have xz || return 1
+			_mca_slice "$file" "$off" "$len" \
+				| { xz -dc --format=lzma 2>/dev/null || true; }
+			;;
+		4)
+			mca_have xz || return 1
+			_mca_slice "$file" "$off" "$len" | { xz -dc 2>/dev/null || true; }
+			;;
+		6)
+			mca_have zstd || return 1
+			_mca_slice "$file" "$off" "$len" | { zstd -dc 2>/dev/null || true; }
+			;;
+		*)  return 1 ;;
+	esac
+}
+
+# _mca_squashfs_names <file> <offset>
+# The directory table of the image at that offset, inflated onto stdout. It is
+# not parsed: the names sit in it as plain text between the records that
+# describe them, and a name is all the caller is looking for.
+#
+# Fails unless the walk lands exactly on the end of the table. That is the
+# integrity check - the block sizes adding up to the table's own length is
+# what says the fields were read from a real superblock and that the output is
+# the whole of the names rather than some of them - and it is why the caller
+# has to collect this before trusting it, never pipe it.
+_mca_squashfs_names() {
+	local file="$1" base="$2"
+	local comp dir_start end field header size pos
+
+	_mca_bytes "$file" "$base" 96 || return 1
+	comp=$(( MCA_BYTES[20] | MCA_BYTES[21] << 8 ))
+
+	_mca_word 72 || return 1
+	dir_start="$MCA_WORD"
+	(( dir_start > 0 )) || return 1
+
+	# Where the names stop: the first table squashfs writes after them. Which
+	# one that is depends on the image - there is no fragment table when
+	# nothing was packed into a fragment, and no export table unless it was
+	# asked for - so they are tried in the order they are written and the
+	# first one that is actually there wins. bytes_used closes the list for an
+	# image that has none of them.
+	end=0
+	for field in 80 88 48 40; do
+		_mca_word "$field" || continue
+		(( MCA_WORD > dir_start )) || continue
+		end="$MCA_WORD"
+		break
+	done
+	(( end > dir_start )) || return 1
+
+	# Names for a hundred thousand files would still fit in a fraction of
+	# this. A table that claims more than it is a table that was misread.
+	(( end - dir_start > 8388608 )) && return 1
+
+	pos=$(( base + dir_start ))
+	end=$(( base + end ))
+
+	while (( pos < end )); do
+		_mca_bytes "$file" "$pos" 2 || return 1
+		header=$(( MCA_BYTES[0] | MCA_BYTES[1] << 8 ))
+		size=$(( header & 0x7fff ))
+		(( size > 0 && pos + 2 + size <= end )) || return 1
+
+		# The top bit says the block was stored as it is, which squashfs does
+		# for the ones compression made no smaller.
+		if (( header & 0x8000 )); then
+			_mca_slice "$file" $(( pos + 2 )) "$size" || return 1
+		else
+			_mca_inflate "$comp" "$file" $(( pos + 2 )) "$size" || return 1
+		fi
+
+		pos=$(( pos + 2 + size ))
+	done
+
+	(( pos == end ))
+}
+
+# _mca_appimage_verdict <file>
+# What the image appended to this file contains, left in MCA_APPIMAGE:
+#   yes      - Chromium, CEF or Electron
+#   no       - the names of everything inside were read and none of them is
+#   unknown  - there is an image, but its contents could not be read
+#   none     - not an AppImage; there is nothing appended to this ELF
+MCA_APPIMAGE=none
+
+_mca_appimage_verdict() {
+	local file="$1" base names m
+	local -a args=()
+
+	MCA_APPIMAGE=none
+
+	_mca_appimage_offset "$file" || return 0
+	base="$MCA_APPIMAGE_AT"
+
+	# hsqs, the squashfs magic. Nothing there means nothing was appended, so
+	# this is an ordinary executable and not an AppImage at all.
+	_mca_bytes "$file" "$base" 4 || return 0
+	(( MCA_BYTES[0] == 104 && MCA_BYTES[1] == 115 \
+		&& MCA_BYTES[2] == 113 && MCA_BYTES[3] == 115 )) || return 0
+
+	# From here on there is an image, so the worst this can end on is "cannot
+	# tell" - never "no", which would be an answer about contents that were
+	# never read.
+	MCA_APPIMAGE=unknown
+
+	names="$(mktemp "${TMPDIR:-/tmp}/mca-names.XXXXXX")" || return 0
+
+	# Nothing at all in the table means every block failed to inflate - a
+	# compressor whose tool is not installed - which is a "cannot tell" too.
+	if _mca_squashfs_names "$file" "$base" > "$names" 2>/dev/null \
+		&& [[ -s $names ]]
+	then
+		for m in "${MCA_MARKERS_STRICT[@]}"; do args+=(-e "$m"); done
+		if grep -qaF "${args[@]}" -- "$names"; then
+			MCA_APPIMAGE=yes
+		else
+			MCA_APPIMAGE=no
+		fi
+	fi
+
+	rm -f -- "$names"
+	return 0
+}
+
+# mca_detect_verdict <program>
+# What running this program starts, left in MCA_VERDICT:
+#   yes      - Chromium, CEF or Electron
+#   no       - something else
+#   unknown  - an image whose payload could not be read, so neither answer has
+#              been earned and the applications screen offers it as a choice
+#
+# Assigns rather than returns three states through an exit code, and is the one
+# place the memo and the on-disk cache are consulted.
+MCA_VERDICT=''
+
+mca_detect_verdict() {
+	local prog="$1" real stamp cached
+
+	MCA_VERDICT=no
+	[[ -n $prog && -e $prog ]] || return 0
 
 	# Memoized under the path as given, so the same launcher named twice in a
 	# scan costs nothing at all the second time.
 	if [[ -n ${MCA_DETECT_MEMO[$prog]+set} ]]; then
-		[[ ${MCA_DETECT_MEMO[$prog]} == yes ]]
+		MCA_VERDICT="${MCA_DETECT_MEMO[$prog]}"
-		return $?
+		return 0
 	fi
 
 	_mca_cache_load
@@ -546,26 +873,45 @@ mca_is_chromium() {
 	if [[ -n $stamp && -n ${MCA_DETECT_CACHE[$prog]+set} ]]; then
 		cached="${MCA_DETECT_CACHE[$prog]}"
 		if [[ "${cached%%$'\t'*}" == "$stamp" ]]; then
-			verdict="${cached#*$'\t'}"
+			MCA_VERDICT="${cached#*$'\t'}"
-			MCA_DETECT_MEMO[$prog]="$verdict"
+			MCA_DETECT_MEMO[$prog]="$MCA_VERDICT"
-			[[ $verdict == yes ]]
+			return 0
-			return $?
 		fi
 	fi
 
 	real="$(readlink -f -- "$prog" 2>/dev/null)" || real="$prog"
 
-	verdict=no
+	MCA_DETECT_UNSURE=0
-	if _mca_detect_uncached "$real"; then verdict=yes; fi
+	if _mca_detect_uncached "$real"; then
+		MCA_VERDICT=yes
+	elif (( MCA_DETECT_UNSURE )); then
+		MCA_VERDICT=unknown
+	else
+		MCA_VERDICT=no
+	fi
 
-	MCA_DETECT_MEMO[$prog]="$verdict"
+	MCA_DETECT_MEMO[$prog]="$MCA_VERDICT"
 	if [[ -n $stamp ]]; then
-		MCA_DETECT_CACHE["$prog"]="$stamp"$'\t'"$verdict"
+		MCA_DETECT_CACHE["$prog"]="$stamp"$'\t'"$MCA_VERDICT"
 		MCA_CACHE_DIRTY=1
 	fi
-	[[ $verdict == yes ]]
+	return 0
 }
 
+# mca_is_chromium <program>
+# Succeeds when the program is a Chromium, CEF or Electron process. "Cannot
+# tell" is not that, so it fails here - anything that has to treat the two
+# differently asks mca_detect_verdict instead.
+mca_is_chromium() {
+	mca_detect_verdict "$1"
+	[[ $MCA_VERDICT == yes ]]
+}
+
+# Set by _mca_detect_uncached when it reaches the end without finding anything
+# and the reason is that something could not be read, rather than that there
+# was nothing there. Only mca_detect_verdict reads it, straight after the call.
+MCA_DETECT_UNSURE=0
+
 _mca_detect_uncached() {
 	local real="$1" depth="${2:-0}" dir target
 
@@ -590,15 +936,28 @@ _mca_detect_uncached() {
 	_mca_has_markers "$dir" && return 0
 	[[ ${dir##*/} == bin ]] && _mca_has_markers "${dir%/*}" && return 0
 
+	# An AppImage keeps all of that inside a filesystem appended to itself, so
+	# there is nothing next to the binary to find - but the names of everything
+	# in there can be read, and that settles it either way.
+	_mca_appimage_verdict "$real"
+	case "$MCA_APPIMAGE" in
+		yes) return 0 ;;
+		no)  return 1 ;;
+	esac
+
 	# Last resort: Chromium's own argument table is in the binary. -m1 stops at
 	# the first hit, so this reads far less than the file size suggests.
-	#
-	# AppImages are the one thing this cannot see through - their payload is a
-	# compressed filesystem - which is why they come out as unknown and are
-	# left to the applications screen.
 	grep -qaFm1 -- 'enable-blink-features' "$real" 2>/dev/null && return 0
 	grep -qaFm1 -- 'CHROME_VERSION_EXTRA' "$real" 2>/dev/null && return 0
 
+	# What is left is an image that could not be read - an unsupported
+	# compressor - or the older AppImage layout, which is an ISO9660 filesystem
+	# and has no name table of this shape at all. Either way the answer is not
+	# "no", it is "nobody looked".
+	if [[ $MCA_APPIMAGE == unknown || $real == *.AppImage || $real == *.appimage ]]; then
+		MCA_DETECT_UNSURE=1
+	fi
+
 	return 1
 }
 
@@ -717,10 +1076,12 @@ mca_scan() {
 			# Steam is Chromium inside, but nothing about it can be changed
 			# from a command line argument; it has its own module.
 			kind=steam
-		elif mca_is_chromium "$prog"; then
+		else
-			(( c_browser[i] )) && kind=browser || kind=app
+			mca_detect_verdict "$prog"
-		elif [[ $prog == *.AppImage || $prog == *.appimage ]]; then
+			case "$MCA_VERDICT" in
-			kind=unknown
+				yes)     (( c_browser[i] )) && kind=browser || kind=app ;;
+				unknown) kind=unknown ;;
+			esac
 		fi
 
 		[[ $kind == no ]] && continue
@@ -813,13 +1174,7 @@ mca_snap_is_chromium() {
 
 	for d in "${MCA_SNAP_DIRS[@]}"; do
 		root="$d/$name/current"
-		[[ -d $root ]] || continue
+		if _mca_find_markers "$root" 5; then
-		if [[ -n "$(find "$root" -maxdepth 5 \
-			\( -name 'chrome_crashpad_handler' -o -name 'app.asar' \
-			   -o -name 'libcef.so' -o -name 'v8_context_snapshot.bin' \
-			   -o -name 'chrome-sandbox' \) \
-			-print -quit 2>/dev/null)" ]]
-		then
 			MCA_DETECT_MEMO[snap:$name]=yes
 			return 0
 		fi
@@ -842,12 +1197,7 @@ mca_flatpak_is_chromium() {
 	fi
 
 	loc="$(flatpak info --show-location "$id" 2>/dev/null)"
-	if [[ -n $loc && -d "$loc/files" ]] \
+	if [[ -n $loc ]] && _mca_find_markers "$loc/files" 4; then
-		&& [[ -n "$(find "$loc/files" -maxdepth 4 \
-			\( -name 'chrome_crashpad_handler' -o -name 'app.asar' \
-			   -o -name 'libcef.so' -o -name 'v8_context_snapshot.bin' \) \
-			-print -quit 2>/dev/null)" ]]
-	then
 		MCA_DETECT_MEMO[flatpak:$id]=yes
 		return 0
 	fi

src/lib/kde.sh

@@ -0,0 +1,167 @@
+# shellcheck shell=bash
+#
+# KDE's middle-click paste.
+#
+# Inside a Chromium application the flag settles this on its own: with
+# autoscroll on, Blink stops pasting the primary selection on middle click,
+# because the button is doing something else now. Everywhere else on the
+# desktop middle click goes on pasting, which is the behaviour this whole
+# program exists to get away from - so the desktop's own switch for it is
+# turned off too.
+#
+# KWin has such a switch, and only for a Wayland session, where the paste is a
+# protocol the compositor either offers or does not:
+#
+#   ~/.config/kwinrc, [Wayland], EnablePrimarySelection=false
+#
+# On X11 there is nothing to switch. The primary selection is part of X itself
+# and every toolkit reaches for it on its own; no one place can say no.
+#
+# KWin reads the key once, while it starts, so what is written here is true
+# from the next login on rather than straight away.
+
+MCA_KWINRC="${MCA_XDG_CONFIG}/kwinrc"
+MCA_KWIN_GROUP="Wayland"
+MCA_KWIN_KEY="EnablePrimarySelection"
+
+# The kconfig command line tools, whichever generation this session has. They
+# are what System Settings writes kwinrc with, and worth the fork: the file has
+# a cascade behind it - /etc/xdg, the kdedefaults directory - and entries a
+# distribution can mark immutable, and an edit that reaches past all that with
+# sed writes something KDE goes on to ignore.
+MCA_KDE_TOOL=''
+MCA_KDE_TOOL_LOOKED=0
+
+_mca_kde_tool() {
+	local gen
+
+	if (( ! MCA_KDE_TOOL_LOOKED )); then
+		MCA_KDE_TOOL_LOOKED=1
+		for gen in 6 5; do
+			if mca_have "kreadconfig$gen" && mca_have "kwriteconfig$gen"; then
+				MCA_KDE_TOOL="$gen"
+				break
+			fi
+		done
+	fi
+
+	[[ -n $MCA_KDE_TOOL ]] || return 1
+	printf '%s\n' "$MCA_KDE_TOOL"
+}
+
+# mca_kde_session
+# Whether this is a KWin Wayland session.
+mca_kde_session() {
+	local desktop="${XDG_CURRENT_DESKTOP:-}"
+
+	# The watcher is a systemd service, and one started before the session
+	# imported its environment sees none of these variables. A running
+	# kwin_wayland answers both halves of the question at once, and is what
+	# such a run ends up going by.
+	if [[ -z $desktop && -z ${KDE_FULL_SESSION:-} ]]; then
+		mca_have pgrep || return 1
+		pgrep -u "$(id -u)" -x kwin_wayland > /dev/null 2>&1
+		return
+	fi
+
+	# XDG_CURRENT_DESKTOP is the one every desktop sets, and the one to go by
+	# wherever it is set; KDE_FULL_SESSION is what answers when it is not.
+	[[ -z $desktop || ${desktop^^} == *KDE* ]] || return 1
+
+	[[ ${XDG_SESSION_TYPE:-} == wayland || -n ${WAYLAND_DISPLAY:-} ]]
+}
+
+mca_kde_available() {
+	mca_kde_session && _mca_kde_tool > /dev/null
+}
+
+# mca_kde_read [file]
+# What the key says now, lowercased, empty when it is not set at all.
+mca_kde_read() {
+	local file="${1:-$MCA_KWINRC}" gen val
+	gen="$(_mca_kde_tool)" || return 1
+
+	# kreadconfig creates the file it is pointed at, and the status screen -
+	# which asks on every redraw - has no business writing anything. A file
+	# that is not there has no value in it to read anyway.
+	[[ -f $file ]] || return 0
+
+	val="$("kreadconfig$gen" --file "$file" --group "$MCA_KWIN_GROUP" \
+		--key "$MCA_KWIN_KEY" 2>/dev/null)"
+	printf '%s\n' "${val,,}"
+}
+
+# mca_kde_write [file] <value>
+# An empty value takes the key back out rather than writing nothing into it -
+# a key that is missing and a key that says nothing are not the same thing to
+# KDE, and only the first one is what an untouched kwinrc looks like.
+mca_kde_write() {
+	local file="$1" value="$2" gen
+	gen="$(_mca_kde_tool)" || return 1
+
+	if [[ -z $value ]]; then
+		"kwriteconfig$gen" --file "$file" --group "$MCA_KWIN_GROUP" \
+			--key "$MCA_KWIN_KEY" --delete 2>/dev/null
+	else
+		"kwriteconfig$gen" --file "$file" --group "$MCA_KWIN_GROUP" \
+			--key "$MCA_KWIN_KEY" "$value" 2>/dev/null
+	fi
+}
+
+# mca_kde_paste_apply
+# Turns the paste off, and remembers what the key said before so that undoing
+# puts that back rather than a guess.
+#
+# A key that already says false without the ledger knowing about it was set by
+# the user or by the distribution. That one is left alone and nothing is
+# recorded, so `disable` cannot switch a paste back on that this program never
+# switched off.
+mca_kde_paste_apply() {
+	local current previous changed=0
+
+	mca_kde_available || return 1
+	current="$(mca_kde_read)" || return 1
+
+	if mca_ledger_has "$MCA_KWINRC"; then
+		# Ours already. What goes back is the value recorded the first time
+		# round, not what reading the file now would say - that is false, and
+		# false is what is being undone.
+		previous="$(mca_ledger_detail "$MCA_KWINRC")" || previous=unset
+	elif [[ $current == false ]]; then
+		return 0
+	else
+		previous="${current:-unset}"
+	fi
+
+	if [[ $current != false ]]; then
+		mca_kde_write "$MCA_KWINRC" false || return 1
+		MCA_CHANGES=$(( MCA_CHANGES + 1 ))
+		changed=1
+	fi
+
+	mca_ledger_add kwin "$MCA_KWINRC" "$previous"
+
+	(( changed )) && mca_note "$(mca_msg "Middle click stops pasting at the next login.")"
+	return 0
+}
+
+# mca_kde_paste_revert <file> <previous value>
+mca_kde_paste_revert() {
+	local file="$1" previous="$2"
+
+	_mca_kde_tool > /dev/null || return 1
+	[[ -f $file ]] || return 1
+
+	# Only what is still recognisably ours. A key somebody has since put a
+	# value of their own into is theirs now.
+	[[ "$(mca_kde_read "$file")" == false ]] || return 1
+
+	[[ $previous == unset ]] && previous=''
+	mca_kde_write "$file" "$previous" || return 1
+
+	# Nothing was in the file but the key, which means there was no file before
+	# this program wrote one. An empty kwinrc says nothing that its absence
+	# does not say better.
+	[[ -s $file ]] || rm -f -- "$file"
+	return 0
+}

src/lib/menu.sh

@@ -194,6 +194,18 @@ mca_ui_status() {
 		fi
 	fi
 
+	# Not an ON/OFF like the rows around it: what is being reported is the state
+	# of the paste, not of a switch, and off is the state this program is after.
+	if mca_kde_available; then
+		if [[ "$(mca_kde_read)" == false ]]; then
+			_mca_row "$(mca_msg "Middle-click paste")" \
+				"${MCA_C_GREEN}$(mca_msg "off")${MCA_C_RESET}"
+		else
+			_mca_row "$(mca_msg "Middle-click paste")" \
+				"${MCA_C_DIM}$(mca_msg "on")${MCA_C_RESET}"
+		fi
+	fi
+
 	_mca_row "$(mca_msg "Last applied")" "$(mca_time_ago "$last")"
 
 	if [[ $CFG_ENABLED == yes ]]; then
@@ -216,6 +228,7 @@ MCA_SETTINGS=(
 	"PatchAutostart|bool|yes|Programs that start themselves at login"
 	"PatchSteam|bool|yes|Steam"
 	"PatchSpotify|bool|yes|Spotify"
+	"DisablePaste|bool|yes|Turn off KDE's middle-click paste"
 	"WatchNewApps|bool|yes|Apply to newly installed applications"
 	"ExtraFlags|text||Additional Chromium arguments"
 )

src/lib/patch.sh

@@ -646,6 +646,10 @@ mca_revert_all() {
 				mca_steam_revert "$path" "$detail" \
 					&& MCA_CHANGES=$(( MCA_CHANGES + 1 ))
 				;;
+			kwin)
+				mca_kde_paste_revert "$path" "$detail" \
+					&& MCA_CHANGES=$(( MCA_CHANGES + 1 ))
+				;;
 		esac
 	done < <(tac "$MCA_LEDGER" 2>/dev/null)
 

src/middleclick-autoscroll

@@ -17,7 +17,7 @@ set -uo pipefail
 
 MCA_LIBDIR="${MCA_LIBDIR:-@LIBDIR@}"
 
-for _mod in common config detect patch steam apply menu; do
+for _mod in common config detect patch steam kde apply menu; do
 	# shellcheck source=/dev/null
 	if ! source "$MCA_LIBDIR/$_mod.sh"; then
 		printf 'middleclick-autoscroll: cannot load %s/%s.sh\n' "$MCA_LIBDIR" "$_mod" >&2