steamify-cachyos-dev/.claude/skills/cachyos-vm-testing/SKILL.md

name, description

name	description
cachyos-vm-testing	Use when testing Steamify CachyOS (repo steamify-cachyos, steamify.sh) changes in the CachyOS QEMU test VM - starting/stopping the VM, restoring snapshots, running the wizard over SSH with scripted menu input, checking each component's state, reboot checks and screenshots.

Testing Steamify CachyOS in the CachyOS VM

The wizard (steamify.sh + lib/*.sh) opens a menu that detects which components are on and toggles them to match the user's choice. Menu order:

1. SteamOS conversion (boot into gaming mode via autologin, Return to Gaming Mode shortcut, Steam desktop autostart)
2. └ Boot into: gamescope / desktop (sub-option of 1, shown while 1 is ticked; Left/Right or the number switches it)
3. SteamOS theme (installs cachyos-vapor and applies the Vapor global theme with its desktop and window layout; needs a running Plasma session)
4. Steam Deck/Machine icons (STEAM_GAMEPADUI_ARGS -steamos3)
5. Single user mode (SDDM, no lock screen/user switching/log out, Valve's empty KDE wallet; enabling it enables 1, disabling 1 disables it)
6. Steamify shortcut (desktop icon + launcher entry that curl | bash the newest release; its gear icon is a release asset)
7. HDMI-CEC (any PC; pre-ticked on a first run only on Fremont: Valve's cecd/cec-audio-control/inputattach-cec-units from the holo repo, the steam-launcher drop-in that overrides STEAM_ENABLE_CEC=0; see "Fake HDMI-CEC")
8. Steam Machine support (only on DMI Valve/Fremont: leds-valve-dkms-git built for every kernel via /etc/dkms/leds-valve-dkms.conf, ensure-kernel-headers.service, udev rule, steamos-manager, powerdevilrc)
9. └ Pin the kernel to 7.1.6-1 (sub-option of 8, ticked along with it; packages in /var/cache/steamify/kernel)
10. Update BIOS (only on Fremont, an opt-in action, never pre-ticked; tickable only when Valve has a newer BIOS)

Without --fremont the menu has only 1-7. Numbers shift when a parent is unticked (its sub-option is hidden), so read the ticks with 'q\n' first.

Undo journals live in ~/.local/state/cachyos-gamescope-boot/ in the guest.

Run everything from the root of this repo on the host. Defaults: user theupriser, SSH port 2222 (the scripts take VM_USER / VM_PORT / VM_HOST).

The VM's disk, vars*.fd and run.sh copy may live outside the repo (~/vms/cachyos-test on the main dev machine). Find it with readlink /proc/$(pgrep -f '^qemu-system')/cwd when the VM runs; the snapshot commands below run in that directory. scripts/vmreset.sh takes VM_DIR and defaults to the repo if it holds disk.qcow2, else ~/vms/cachyos-test.

Quick start (the usual loop)

scripts/vmreset.sh --fremont        # restore ssh-ready, boot, mount repo, autologin, wait for Plasma
                                    # (also skips the broken krfoss mirror, installs shellcheck)
scripts/cmp.sh save                 # baseline of the KDE configs
scripts/vmwatch.sh 'q\n' "Menu"   # wizard in a visible Konsole in the VM (read the ticks first)
scripts/vmwatch.sh --release '\ny\nm\nq\n' "Full run"   # the newest GitHub release instead of /mnt
scripts/vmstate.sh                  # component state
scripts/vmshot.sh --clean /path/shot.png    # screenshot (--clean: close Steam/Hello/Konsole first)

Always finish on a fresh ssh-ready snapshot, ideally with --release once published. Re-applying or toggling on an already set-up VM hides first-install bugs: the DKMS override silently failed on a fresh system (/etc/dkms doesn't exist before dkms is installed), which no re-apply showed because dkms was there already.

Long runs (full install: packages, yay, DKMS) take 10+ minutes: run them with run_in_background and wait with an until-loop, not chained sleeps.

Prefer vmwatch.sh over vmrun.sh when the user is watching the QEMU window: vmrun.sh runs invisibly over SSH, so they see nothing happen.

VM lifecycle

Start in the background:

./run.sh --fremont > /tmp/vm.log 2>&1 &     # flags: [install] [--fremont] [--nvidia] [--vulkan [--amd]]

Wait for SSH:

until ssh -p 2222 -o BatchMode=yes -o ConnectTimeout=3 theupriser@localhost true 2>/dev/null; do sleep 3; done

Shut down and wait until QEMU is gone. Use '^qemu-system': a plain pgrep -f qemu-system also matches the shell running it.

ssh -p 2222 -o BatchMode=yes theupriser@localhost sudo systemctl poweroff
while pgrep -f '^qemu-system' >/dev/null; do sleep 2; done

Snapshots

Only while the VM is off. Keep vars.fd together with the disk.

qemu-img snapshot -l disk.qcow2                                               # list
qemu-img snapshot -a ssh-ready disk.qcow2 && cp vars.ssh-ready.fd vars.fd     # restore
qemu-img snapshot -c my-state disk.qcow2 && cp vars.fd vars.my-state.fd       # create

Existing snapshots: clean (fresh install) and ssh-ready (sshd + host key + passwordless sudo). Reset to ssh-ready before each full test run.

First-time guest setup

Only when building from clean. In the guest console:

sudo mount -t 9p -o trans=virtio,version=9p2000.L vmtools /media && /media/guest-ssh-setup.sh

This installs and enables sshd, opens the firewall (ufw/firewalld if active), adds the host keys (share/host-keys.pub, written by run.sh) and a NOPASSWD sudoers rule for the test user.

Mounting the project repo

The mount is lost on every reboot (unless the fstab line below was added). Without it the wizard fails with /mnt/steamify.sh: No such file or directory, which is easy to miss in filtered output. Remount after each reboot.

The repo (REPO, default ~/projects/cachyos-gamescope-boot) is shared read-write as 9p tag repo. The ssh-ready snapshot does not mount it:

ssh -p 2222 -o BatchMode=yes theupriser@localhost bash -s << 'EOF'
sudo mount -t 9p -o trans=virtio,version=9p2000.L repo /mnt
# optional, survives reboots:
grep -q ' /mnt 9p ' /etc/fstab || echo 'repo /mnt 9p trans=virtio,version=9p2000.L,nofail 0 0' | sudo tee -a /etc/fstab
EOF

Remote commands: the guest shell is fish

Pass remote commands as bash heredocs, never as inline strings containing $:

ssh -p 2222 -o BatchMode=yes theupriser@localhost bash -s << 'EOF'
echo "$HOME"
EOF

Getting a Plasma session

On a fresh snapshot the plasmalogin greeter is shown. Enable test autologin:

ssh -p 2222 -o BatchMode=yes theupriser@localhost bash -s << 'EOF'
sudo mkdir -p /etc/plasmalogin.conf.d
printf '[Autologin]\nUser=theupriser\nSession=plasma.desktop\n' | sudo tee /etc/plasmalogin.conf.d/00-test-autologin.conf
sudo systemctl restart plasmalogin
EOF

Remove /etc/plasmalogin.conf.d/00-test-autologin.conf before testing the "normal login screen" case.

Running the wizard non-interactively

scripts/vmrun.sh '<input>' does this (log also at /tmp/wizard.log in the guest). Manually, in the guest:

export XDG_RUNTIME_DIR=/run/user/1000 WAYLAND_DISPLAY=wayland-0 DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/1000/bus
printf '\ny\nm\nq\n' | /mnt/steamify.sh

Menu input: a number toggles a component, an empty line continues, a re-applies what is on, q quits; then y answers "Go ahead?". The wizard loops: after a run it asks for Enter (back to the menu), or [m]/[r] (menu / restart now) when a restart is needed. On q with a restart pending it asks "Restart now? [Y/n]" (default yes): n goes back to the menu and the next q asks again. Input that runs out quits without restarting, so end scripted input right after the last q (or with n), never with an empty line at the restart question.

Input	Meaning
'\ny\nm\nq\n'	first run, accept what's pre-ticked (not Boot into desktop, not BIOS; HDMI-CEC only on Fremont), back to the menu, quit
'3\n\ny\n'	toggle the theme (while 1 is ticked)
'7\n\ny\nm\nq\nn\n'	toggle HDMI-CEC (while 1 is ticked)
'a\ny\nm\nq\n'	re-apply what is on (includes the conversion, so a restart is pending)
'q\n'	just show the menu
'1\n\ny\n'	from a state where 1 is off: turn the conversion on

Know the starting ticks before choosing input. When everything is off (fresh snapshot, or after turning the last component off), the menu treats it as a first run and pre-ticks the components (not Boot into desktop, BIOS, or HDMI-CEC off Fremont). Otherwise the ticks show what is on now. Unticking 1 also unticks single user mode (5) and hides the Boot into row (2), so every later number moves up one; the same for Steam Machine support (8) and its kernel pin row (9). Ticking 5 ticks 1 again. The numbers apply to the menu as it is when you type them, one line at a time. When unsure, run 'q\n' first and read the ticks.

a only re-applies components that are already on: it does not retry one that failed. Re-applying the conversion also resets the autologin session to gamescope, so set it back to plasma before the next reboot (below).

Gamescope does not render in this VM

Everything was tried (2026-09-25); don't spend time on it again:

• virgl (default): no Vulkan, black.
• --vulkan --amd (venus on the host 780M, RADV): gamescope and Steam run, but virtio-gpu rejects every framebuffer gamescope makes ("Cannot import FB to DRM ... not supported for scan-out", AR24 and XR24, also with --force-composition --disable-layers): black. KWin crashes on venus ("Illegal command buffer"), so Plasma freezes too.
• nested gamescope with lavapipe (vulkan-swrast): refuses to start, lavapipe lacks VK_KHR_present_id/present_wait.
• Other hypervisors (VirtualBox, VMware) have no Vulkan at all. Check gaming mode (Steam's settings, HDMI-CEC section, refresh rates) on the real Steam Machine. The VM covers everything else.

(Also after any re-apply that includes the conversion.)

No suitable Vulkan (venus is unstable with the host NVIDIA driver). After enabling the conversion, before rebooting:

sudo /usr/lib/steamos/steam-set-session plasma.desktop

so autologin lands in Plasma. Stuck in a gamescope relogin loop: over SSH set the session to plasma and sudo systemctl restart display-manager (or stop the DM, set the session, start it).

Checking component state

scripts/vmstate.sh prints:

• display manager, SDDM autologin file, plasmalogin [Autologin] section
• session sync bridge units (sync-steamos-session.path) and sudoers rule (gamescope-session-switch)
• Return to Gaming Mode shortcut Exec= line
• steam-desktop-autostart user unit; glyphs env file (99-gamescope-steam-glyphs.conf)
• LookAndFeelPackage, ColorScheme, GTK theme, font
• KDE Action Restrictions (lock_screen), kscreenlockerrc Autolock, Lock Session shortcut
• panel floating/thickness in plasmashellrc
• kickoff primaryActions / systemFavorites / icon
• leds-valve-dkms-git package, udev rule, leds_valve module, /sys/class/leds/valve-leds* count and owner (user must be able to write)
• steamos-manager package and whether it is active
• undo journals present, and duplicate keys (cut -f1-3 journal | sort | uniq -d)

scripts/cmp.sh save stores a baseline of the KDE/GTK configs in the guest (~/.cache/vm-baseline, not /tmp); scripts/cmp.sh diffs against it.

After changing the theme, cmp.sh should show only spectacle's kglobalshortcutsrc entries (from screenshots) and keys whose value equals ~/.config/kdedefaults (e.g. ColorScheme=BreezeDark, widgetStyle=Breeze): KDE drops or writes those on its own.

Testing the newest release

scripts/vmwatch.sh --release '<input>' downloads releases/latest/download/steamify.sh in the guest and runs it with the piped input (WIZARD_KEEP_STDIN=1; the bundle otherwise reattaches the terminal). Release assets only exist from the version that added them: before v0.9.0 was published, steamify.sh and the shortcut's steam-gaming-settings.svg gave 404 and the shortcut fell back to Steam's icon. Check a release with curl -sIL -o /dev/null -w '%{http_code}' .../releases/latest/download/<asset>.

Reboot checks

• Single user mode on: SDDM autologin without greeter: pgrep sddm-greeter empty, plasmashell running.
• Everything off: plasmalogin greeter shown: loginctl list-sessions shows a greeter session, no user plasmashell.

Full test matrix

Reset to ssh-ready, start with --fremont, mount the repo, enable test autologin, then (all passed last run; scripts/vmstate.sh after every step):

1. Fresh run turning everything on ('\ny\nm\nq\n'), set session to plasma, reboot: SDDM logs straight in, Vapor desktop, three desktop icons (Return to Gaming Mode, Steam, Steamify CachyOS with the gear icon), LEDs loaded (17 nodes), uname -r 7.1.6-1-cachyos and pacman -Qu shows linux-cachyos [ignored], ~/.local/share/kwalletd/kdewallet.kwl.bak-steamify next to Valve's empty wallet, steamos-manager-configure-cecd enabled and ~/.config/cecd/config.d/00-steamos-manager.toml written.
2. Rerun: all shown on, "Everything is already the way you want it".
3. a re-apply: no duplicate journal entries.
4. Theme on: Steam Deck wallpaper, full-width 46px panel, distributor-logo-steamdeck launcher icon, dark-lnf=com.valve.vapor.desktop (Brightness & Color's Dark Mode toggle is on, hint "Switch to Breeze"); with single user on, kickoff keeps primaryActions=3. Theme off: CachyOS wallpaper, floating 30px panel, CachyOS launcher icon, BreezeDark colors (not light), cachyos-vapor removed, cmp.sh clean.
5. Single user off: switches to plasmalogin + sync bridge + sudoers; shortcut Exec uses sudo -n.
6. Single user on: back to SDDM.
7. Icons + Steam Machine support off: driver, udev rule, modules-load, steamos-manager removed; yay kept; kernel pin removed (IgnorePkg empty, pacman -Syu back to the current kernel), files kept in /var/cache/steamify/kernel. On again (even with the CachyOS servers blocked in /etc/hosts): installs 7.1.6 from that directory; DKMS builds the LED driver once per kernel. 7c. HDMI-CEC: load vivid, then scripts/vmcec.sh (all PASS), off/on via the menu. 7b. DKMS: vmstate.sh shows installed for every kernel. Headers at boot: sudo pacman -R --noconfirm linux-cachyos-lts-headers (DKMS drops the LTS build), set the session to plasma, reboot, then journalctl -b -u ensure-kernel-headers shows the install and DKMS lists LTS again.
8. Conversion off: single user auto-unticked, plasmalogin [Autologin] back to CachyOS's Session=plasma, journals empty.
9. Remove the test autologin file, reboot: normal login screen.

BIOS update item

Only shown with --fremont, and only tickable when the guest's BIOS version differs from Valve's newest. Fake an older one with BIOS_VERSION=F7F0107 ./run.sh --fremont (SMBIOS type 0; vmreset.sh passes the environment through). fwupd correctly refuses the firmware in the VM ("not for this machine's hardware"), so walk the whole flow with WIZARD_BIOS_DRY_RUN=1 (skips only that check, never flashes). Scripted: printf '10\n\ny\ny\nUPDATE\nm\nq\nn\n' | WIZARD_BIOS_DRY_RUN=1 /mnt/steamify.sh (BIOS is item 10 with --fremont; the final n answers the restart question).

A successful dry run counts as staged, so [m]/[r] and the restart question on q show up too; in dry-run mode "restart" only prints.

Fake HDMI-CEC (vivid): scripts/vmcec.sh

The VM has no CEC hardware; the kernel's vivid test driver emulates it: a virtual HDMI input (/dev/cec0, plays the TV via cec-ctl/cec-follower) and output (/dev/cec1, the PC, where cecd runs). Needs the HDMI-CEC item on; the Steam-settings checks need Steam Machine support (steamos-manager).

scripts/vmcec.sh              # all checks + a real suspend/resume (~5 min)
scripts/vmcec.sh --no-sleep   # without suspending the VM
scripts/vmcec.sh --sleep-only # only the suspend/resume check
scripts/vmwake.sh             # wake a sleeping VM (QMP socket from run.sh)

It prints PASS/FAIL/SKIP for: identity (address 1.0.0.0, name "Steam Machine", Valve vendor ID), every TV remote key and the key it becomes, the Steam settings over D-Bus (remote control, WakeTv, SuspendTv, SuspendDevice and cecd's config), PC to TV (make active, wake, volume, audio status, standby), TV standby putting the PC to sleep (blocked by an inhibitor), and sleep/wake turning the TV off and on.

Gotchas it handles, learned the hard way:

• The cecd service (-e) grabs every CEC device, the fake TV too: the test uses a temporary drop-in (~/.config/systemd/user/cecd.service.d/vmcec.conf) with -d /dev/cec1, removed at the end.
• The TV's power button (KEY_POWER) and TV standby with SuspendDevice on really suspend the VM (Steam Machine support makes the power button sleep): block with sudo systemd-inhibit --what=sleep:handle-power-key (as user it's refused). If the VM sleeps anyway: scripts/vmwake.sh (the QEMU window's keyboard doesn't wake it; a VM started before run.sh had -qmp needs a hard restart).
• rtcwake skips logind, so cecd never sees the sleep: suspend with systemctl suspend and wake over QMP.
• steamosctl get/set-hdmi-cec-suspend-tv|suspend-device don't work (steamos-manager 26.4.1); Steam uses the D-Bus properties of com.steampowered.SteamOSManager1.HdmiCec2, and so does the test.
• The kernel answers "give OSD name" itself, even without cecd: it's no proof cecd runs. cecd logs "Putting TV in standby" only at RUST_LOG=debug.
• CecDevice1 methods take arguments (VolumeUp y 0, Daemon1 Standby b true).
• After the test (or a hard restart) CachyOS may put autologin back on gamescope: set the session to plasma and restart the display manager.

The menu item is 7 (8 = Steam Machine support, 9 = kernel pin, 10 = BIOS with --fremont).

Steamify shortcut

Test a double-click with systemd-run --user kioclient exec ~/Desktop/cachyos-gamescope-boot-wizard.desktop. If Plasma opens the file in Kate instead, it saw the file before it was executable. After the wizard the window counts down 10 seconds and closes; after an error (e.g. 404) it waits for Enter.

Visual checks

scripts/vmshot.sh [--clean] <out.png> does the below. After a reboot, Steam's sign-in window and CachyOS Hello cover the desktop: --clean shuts Steam down (closing only steamwebhelper leaves a black window) and closes Hello and the wizard's Konsole. Steam comes back at the next login. Starting spectacle straight from SSH core-dumps; it has to run as a user unit (systemd-run --user --wait).

ssh -p 2222 -o BatchMode=yes theupriser@localhost bash -s << 'EOF'
export XDG_RUNTIME_DIR=/run/user/1000 WAYLAND_DISPLAY=wayland-0 DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/1000/bus
qdbus6 org.kde.plasmashell /PlasmaShell org.kde.PlasmaShell.activateLauncherMenu   # optional, toggles
sleep 1; systemd-run --user --wait -q spectacle -b -n -f -o /tmp/x.png
EOF
scp -P 2222 theupriser@localhost:/tmp/x.png /tmp/x.png    # then view it

The launcher call toggles; retake if the screenshot misses it. Apps started directly over SSH lack the session's Qt platform theme and look light; start them with systemd-run --user <app>.

Pitfalls

• After a reboot or hard restart the autologin may be back on gamescope (black screen): sudo /usr/lib/steamos/steam-set-session plasma.desktop && sudo systemctl restart display-manager.

• vmreset.sh sometimes stops at "Plasma did not start" (the greeter stays up although the test autologin file is there): restart the display manager.

• pkill -f '<pattern>' from the host shell also matches the command running it and kills it (exit 144): match something narrower, or kill by PID.

• Counting sudo password prompts (the VM has NOPASSWD): move /etc/sudoers.d/99-test-vm aside, give the user a password, run the wizard in a Python pty that answers [sudo] password and logs each prompt, then put the file back. makepkg -i runs sudo -k and always asks again.

• Broken mirror: mirror5.krfoss.org served a bad .sig ("Maximum file size exceeded"), failing the conversion's package install. vmreset.sh comments it out; by hand: sudo sed -i '/krfoss/s/^Server/#Server/' /etc/pacman.d/*mirrorlist*.

• The snapshot has no shellcheck; vmreset.sh installs it. Check the bundle like CI: .github/tools/bundle.sh on the host, then in the guest cd /mnt && shellcheck -S warning -e SC2034,SC2154 dist/steamify.sh.

• The host shell is zsh: $var holding several file names isn't split (sed -i ... $files gets one "name"); use xargs.

• Windows open in the snapshot (System Settings, CachyOS Hello) show stale data (e.g. a theme list from before cachyos-vapor) and cover screenshots; vmreset.sh closes them.

• The guest's /tmp is cleared on reboot; keep baselines elsewhere.

• The fake Fremont DMI makes leds-valve load 17 LED nodes, but there is no real hardware.

• Per-file shellcheck (shellcheck -S warning -x steamify.sh lib/*.sh in /mnt) shows SC2154/SC2034 cross-file false positives; the bundle check above is the one CI runs.

• This repo's .gitignore is an allowlist: it ignores everything and un-ignores only the tracked scripts/docs. Add any new tracked file to it explicitly, or git will silently ignore it.