windows-vm.sh: stop the input helper before giving up on a click

On a timeout or a setup error the request now stops the scheduled task and
waits for it to end before releasing the mutex, so a stalled helper can't
act later inside another request.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

README.md

@@ -222,7 +222,7 @@ Frame's software fits together, all checked against a real headset and labelled
 | [Frame Control for iPhone](docs/iphone.md) | The iPhone and iPad app, how it runs the server on the Frame, pairing |
 | [Recovery and OS images](docs/recovery-and-images.md) | Where to download the Frame's OS, what's inside, testing without the headset |
 | [AI agents and assistant](docs/agents.md) | Key-free MCP tools, human approvals, and an opt-in assistant panel |
-| [Testing](docs/testing.md) | Unit tests, end-to-end tests against a fake Frame in Docker, and the headset smoke test |
+| [Testing](docs/testing.md) | Unit tests, end-to-end tests against a fake Frame in Docker, the headset smoke test and a Windows test VM |
 | [Open questions](docs/open-questions.md) | What's still unchecked |
 
 <details>

docs/evidence/linux-vr/2026-09-29.txt

@@ -0,0 +1,29 @@
+Locked real-Frame repeat, 2026-09-29
+SteamOS 0.4.1, BUILD_ID 20260925.6191901; aarch64.
+mkdir /tmp/frame-test.lock succeeded before installs/launches; rmdir issued after cleanup.
+Preflight battery 44%, charging; before WiVRn 46%, before ALVR 47%, cleanup 47%.
+Original Steam PID 49823 and vrserver PID 49571 present after cleanup.
+Same unmodified upstream release APKs and SHA-256s as 2026-09-28.txt.
+Installer functions loaded from a613735 before checkout was fast-forwarded to current main.
+No compatibility layer injected. Per-app immersive Lepton instances, Steam shortcut launches.
+Headset unworn. No Linux gaming host. No pairing or streaming session reached.
+
+WIVRN: selected journal lines (local +1000 prefix, Android timestamps UTC).
+Sep 29 11:03:15 frame lepton-steamlaunch-2817846116[1967]: 09-29 01:03:15.202  1153  1181 E OpenXR-Loader: Error [GENERAL | xrCreateInstance | OpenXR-Loader] : LoaderInstance::CreateInstance, no support found for requested extension: XR_KHR_convert_timespec_time
+Sep 29 11:03:15 frame lepton-steamlaunch-2817846116[1967]: 09-29 01:03:15.210  1153  1181 I WiVRn   : [2026-09-29 01:03:15.210] [WiVRn] [info] Failed to create OpenXR instance version 1.1.58: XR_ERROR_EXTENSION_NOT_PRESENT
+Sep 29 11:03:15 frame lepton-steamlaunch-2817846116[1967]: 09-29 01:03:15.248  1153  1181 E OpenXR-Loader: Error [GENERAL | xrCreateInstance | OpenXR-Loader] : LoaderInstance::CreateInstance, no support found for requested extension: XR_KHR_convert_timespec_time
+Sep 29 11:03:15 frame lepton-steamlaunch-2817846116[1967]: 09-29 01:03:15.256  1153  1181 I WiVRn   : [2026-09-29 01:03:15.256] [WiVRn] [info] Failed to create OpenXR instance version 1.0.58: XR_ERROR_EXTENSION_NOT_PRESENT
+Sep 29 11:03:15 frame lepton-steamlaunch-2817846116[1967]: 09-29 01:03:15.257  1153  1181 E WiVRn   : [2026-09-29 01:03:15.257] [WiVRn] [error] Error during initialization: Failed to create OpenXR instance: XR_ERROR_EXTENSION_NOT_PRESENT
+Screenshot API exit 0; 1920x1080 uniformly dark image; no client scene visible.
+
+ALVR: selected journal lines (local +1000 prefix, Android timestamps UTC).
+Sep 29 11:03:36 frame lepton-steamlaunch-2831623938[1967]: 09-29 01:03:35.553  1139  1167 E OpenXR-Loader: Error [GENERAL | xrCreateInstance | OpenXR-Loader] : LoaderInstance::CreateInstance, no support found for requested extension: XR_KHR_convert_timespec_time
+Sep 29 11:03:36 frame lepton-steamlaunch-2831623938[1967]: 09-29 01:03:35.553  1139  1165 I RustStdoutStderr: Error [GENERAL | xrCreateInstance | OpenXR-Loader] : LoaderInstance::CreateInstance, no support found for requested extension: XR_KHR_convert_timespec_time
+Sep 29 11:03:36 frame lepton-steamlaunch-2831623938[1967]: 09-29 01:03:35.611  1139  1167 E [ALVR NATIVE-RUST]: panicked at alvr/client_openxr/src/lib.rs:220:10:
+Sep 29 11:03:36 frame lepton-steamlaunch-2831623938[1967]: 09-29 01:03:35.611  1139  1167 E [ALVR NATIVE-RUST]: called `Result::unwrap()` on an `Err` value: ERROR_EXTENSION_NOT_PRESENT
+Screenshot API exit 0; 1920x1080 uniformly dark image; no client scene visible.
+
+Cleanup: both test app directories, compatdata, shadercache, containers and shortcuts absent.
+Capture output directory removed. No global settings changed; Steam/SteamVR not stopped.
+The shared lock was subsequently acquired by another thread (new directory timestamp 11:03:49 +1000).
+Native clients and Valve host streaming not exercised: no native build or Linux gaming host available.

docs/linux-vr-streaming.md

@@ -38,6 +38,33 @@ after its container exited. No headset was worn and no host was connected.
 All test app files, compatdata, shortcuts and containers were removed afterwards.
 SteamVR's original process remained running. No global settings changed.
 
+### Locked repeat, 2026-09-29 (verified)
+
+Acquired `/tmp/frame-test.lock` before installing or launching anything and
+released it after cleanup. Battery was 44% and charging at preflight, 46–47%
+during the launches, and 47% at cleanup. The SteamOS version/build was unchanged.
+
+Reinstalled and launched both original APKs in immersive Lepton instances.
+WiVRn again failed at OpenXR 1.1 and 1.0 with the missing timespec extension;
+ALVR again panicked on `ERROR_EXTENSION_NOT_PRESENT`. This repeats the
+unmodified-client test, not the newer installer's automatic compatibility-layer
+path. Neither reached a session that could be paired or exercised further.
+Fresh [journal excerpts and cleanup evidence](evidence/linux-vr/2026-09-29.txt)
+record the failures.
+
+SteamVR's screenshot API returned a 1920×1080 headset capture after each
+launch. Both are uniformly dark: [WiVRn](evidence/linux-vr/2026-09-29-wivrn.png)
+and [ALVR](evidence/linux-vr/2026-09-29-alvr.png). These images do **not** prove
+rendering or a working client. The headset was unworn; visibility, controllers,
+frame rate and motion-to-photon latency could not be judged. The explicit
+OpenXR errors, rather than the dark captures, establish the client blocker.
+
+Both test installs, app data, shader caches, shortcuts, containers and temporary
+capture files were removed. The original Steam and SteamVR process IDs were
+unchanged. No reboot, power action or global setting change was used. Native
+clients remain untested, and no Linux gaming host was available for Valve's
+streaming path. The recommendation below is unchanged.
+
 ### Relation to the VR APK branch
 
 **Documented from source:** [PR #20](https://github.com/saphid/frame-control/pull/20)

docs/streaming.md

@@ -21,6 +21,20 @@ The confidence labels are the same as in [ssh.md](ssh.md).
 | **ADB + scrcpy (Lepton only)** | A mirror of the Android container | **Guess** | `brew install scrcpy android-platform-tools`, then `adb connect frame.local:5555` while Lepton Development is running ([adb_lepton](https://partner.steamgames.com/doc/steamhardware/steamframe/adb_lepton)), then `scrcpy`. This only shows Android apps, not SteamOS. |
 | VNC server on the Frame (krfb / wayvnc) | A mirror of the Plasma desktop | **Inferred (SteamOS)** | Deck users run krfb in Desktop Mode ([one.vg](https://one.vg/blog/remote-control-your-steam-deck)). On the Frame, the in-headset desktop is a virtual screen, and krfb isn't known to be preinstalled. RDP and Steam Link cover this case, so it's not recommended. |
 
+**RDP from Windows, verified 2026-09-30:** Windows 11 25H2's Remote Desktop
+(`mstsc`) against BUILD_ID 20260925.6191901. xrdp picks TLS, not NLA, so
+Remote Desktop never asks for a user or password. It warns that the certificate
+(`www.xrdp.org`) can't be verified. After **Yes**, xrdp shows its own "Login to
+frame" box with the username blank. Any user but `steamos` gets "User does not
+exist, or could not be authenticated". Signing in as `steamos` with the
+Developer Mode password opens a Plasma (X11) desktop within about 6 seconds.
+It's a new session on display `:10`, separate from what the headset shows,
+and it uses about 1.3 GB of the Frame's memory. Closing Remote Desktop leaves
+it running, and the next sign-in reconnects to it. To end it, find its
+session with `loginctl list-sessions` over SSH (its leader is `xrdp-sesexec`)
+and run `loginctl terminate-session <id>`; the headset's own session keeps
+running.
+
 **Recommendation for A:** start with Steam Link for macOS, because Valve
 documents it. Use Windows App (RDP) when you want a proper Linux desktop on the
 Mac with keyboard, mouse, and clipboard.

docs/testing.md

@@ -8,6 +8,7 @@ A fourth, a SteamOS VM, may come later ([issue #6](https://github.com/saphid/ste
 | Unit tests (`tests/*.py`) | `python3 -m unittest discover -s tests` | Nothing | Parsing, validation, request guards; SSH and HTTP are mocked |
 | Fake Frame (`tests/e2e`) | `scripts/e2e.sh` | Linux with Docker | The real server and scripts against a container that behaves like a Frame |
 | Headset smoke test | `scripts/frame-smoke.sh` | A Frame on the `frame` alias | Install, launch and remove on the real device, recorded with its BUILD_ID |
+| [Windows test VM](#windows-test-vm) | `scripts/windows-vm.sh` | A Linux machine with KVM and Docker | The Windows build on a real Windows desktop, against a real Frame when needed |
 
 ## Unit tests
 
@@ -163,6 +164,86 @@ refuses ids with a hyphen (`missing/invalid arguments`), which the fake had
 accepted. The fake now refuses them the same way, and Frame Control makes ids
 Steam accepts.
 
+## Windows test VM
+
+The unit tests run on Windows in CI, but the app itself doesn't. Anything that
+depends on the Windows desktop (Remote Desktop, the installer, the bundled
+Python, file dialogs) needs a real Windows machine. This is a Windows 11 VM in a
+[dockur/windows](https://github.com/dockur/windows) container on a Linux machine
+with KVM, driven from the Mac with `scripts/windows-vm.sh`.
+
+Setting it up, once, on the Linux machine:
+
+- **Windows comes from Microsoft.** The container downloads the Windows 11
+  image from Microsoft on first start. Windows runs unactivated, which is fine
+  for testing. Don't use activation workarounds or third-party Windows images.
+- **Publish its ports on loopback only.** Map `127.0.0.1:2222:22` (SSH),
+  `127.0.0.1:8006:8006` (the web console) and, if you need it,
+  `127.0.0.1:13389:3389`. The Mac reaches them through `ssh -J`. Use
+  `restart: "no"` and `stop_grace_period: 2m` so it only runs when someone is
+  testing, and a `docker stop` shuts Windows down cleanly.
+- **Give it SSH on first sign-in.** The container runs `/oem/install.bat`
+  once. Have it add the OpenSSH Server capability, start `sshd`, set PowerShell
+  as its default shell, and put a dedicated public key (for example
+  `~/.ssh/id_ed25519_winvm` on the Mac) in
+  `C:\ProgramData\ssh\administrators_authorized_keys`.
+- Give it 4 cores, 8 GB of memory and a 32 GB disk. That's enough for the app
+  and the tests.
+
+Using it, with `WINVM_HOST` set to the Linux machine's ssh alias:
+
+```sh
+scripts/windows-vm.sh up                        # start it and wait for SSH (1-2 minutes)
+scripts/windows-vm.sh put Frame-Control-Setup-x64.exe
+scripts/windows-vm.sh ps 'Start-Process "$env:USERPROFILE\Downloads\Frame-Control-Setup-x64.exe" /S -Wait'
+scripts/windows-vm.sh shot screen.png            # what's on its screen
+scripts/windows-vm.sh click 723 359              # screen pixels, as in the screenshot
+scripts/windows-vm.sh keys s t e a m o s ret     # QEMU key names
+scripts/windows-vm.sh down                       # shut Windows down
+```
+
+- **Clicks go through a scheduled task.** Commands over SSH run in a
+  session with no desktop, so `click` and `scroll` write the position to a
+  file, and a scheduled task running as the signed-in user replays it. The
+  VM's screen must be signed in; it is after `up`. Clicks and scrolls sent at
+  the same time run one after another. QEMU's own `mouse_move` is relative
+  and drifts, so the script doesn't use it.
+- **Screenshots may not show the pointer.** Check the result of a click (a
+  menu that opens, a button that changes) rather than the pointer's position.
+- **Windows' `ssh` waits for stdin.** The script closes it for every command.
+  Do the same if you run `ssh` in the VM by hand.
+- **Starting the app.** Run Frame Control in the signed-in session, not over
+  SSH. Use its Start-menu shortcut via `click`, or a scheduled task like the
+  one `click` uses.
+
+**Testing against a real Frame.** The VM reaches the headset on the LAN like
+any other computer. Run **Set Up Connection** in the VM's Frame Control once;
+that makes the VM's keys. So the VM doesn't keep access between test runs,
+afterwards take the lines it added out of the headset's
+`~/.ssh/authorized_keys`: the ones matching the VM's
+`~/.ssh/id_ed25519_frame.pub` and, if pairing made one,
+`~/.ssh/id_rsa_frame_devkit.pub`. Check that `ssh -o BatchMode=yes frame true`
+in the VM now fails. Then, around each run:
+
+```sh
+scripts/windows-vm.sh frame-key add       # let the VM's key into the headset
+# ... test ...
+scripts/windows-vm.sh frame-key remove    # and take it out again
+```
+
+`add` appends the VM's key tagged `windows-vm-test`, unless the headset
+already trusts that key through another line. `remove` deletes only the exact
+line `add` wrote, so it doesn't undo what Set Up Connection did, and it leaves
+the file alone if it can't rewrite it. Follow the shared-device
+procedure in [Headset smoke test](#headset-smoke-test) before installing or
+launching anything on the headset.
+
+**Verified 2026-09-30:** Windows 11 Pro 25H2 (build 26200) against a Frame on
+BUILD_ID 20260925.6191901. The script was used to install Frame Control 0.4.0,
+connect it to the Frame and follow Remote Desktop through to the Frame's
+desktop ([what it does](streaming.md#a-see-and-control-the-frame-from-the-mac)).
+`frame-key` left `authorized_keys` byte-for-byte as it was.
+
 ## Owned media player
 
 `tests/test_media.py` covers layout evidence and overrides, OU eye ordering,

scripts/windows-vm.sh

@@ -0,0 +1,161 @@
+#!/usr/bin/env zsh
+# Mac or Linux: drive a Windows 11 test VM for Frame Control's Windows build.
+# The VM is a dockur/windows container on another machine; this reaches it over
+# SSH through that machine. See docs/testing.md#windows-test-vm.
+#
+# Usage: scripts/windows-vm.sh <command> [args]
+#   up | down | status      start it (waits for SSH), shut Windows down cleanly, show state
+#   ps '<PowerShell>'       run PowerShell as the VM's user
+#   put <file> [<dir>]      copy a file in (default: the user's Downloads)
+#   shot <out.png>          save the VM's screen
+#   click <x> <y>           click screen pixel x,y in the signed-in session
+#   scroll <x> <y> <n>      turn the wheel n notches at x,y (negative scrolls down)
+#   keys <key>...           type QEMU key names: a, shift-a, ret, tab, esc, spc ...
+#   frame-key add|remove    let the VM's Frame Control key into the headset (`frame`
+#                           alias) for a test run, then take it out again
+#
+# Set WINVM_HOST to the ssh alias of the machine running the container. Optional:
+# WINVM_CONTAINER (frame-winvm), WINVM_USER (frame), WINVM_PORT (2222: the VM's
+# sshd, published on that machine's loopback), WINVM_KEY (~/.ssh/id_ed25519_winvm).
+set -euo pipefail
+setopt extendedglob
+
+host=${WINVM_HOST:?set WINVM_HOST to the ssh alias of the machine running the VM}
+ctr=${WINVM_CONTAINER:-frame-winvm}
+user=${WINVM_USER:-frame}
+port=${WINVM_PORT:-2222}
+key=${WINVM_KEY:-$HOME/.ssh/id_ed25519_winvm}
+opts=(-o ConnectTimeout=20 -o StrictHostKeyChecking=accept-new
+      -o UserKnownHostsFile=${TMPDIR:-/tmp}/windows-vm-known_hosts -i $key -J $host)
+TAG=windows-vm-test  # comment on the VM's key in the headset's authorized_keys
+
+die() { print -u2 "windows-vm: $*"; exit 1 }
+int() { [[ $1 == (-|)<-99999> ]] || die "not a whole number (up to 99999): $1" }
+# These go into commands run by a shell on the other machine, so keep them plain.
+for v in $host $ctr $user; do [[ $v == [A-Za-z0-9_.]##[A-Za-z0-9_.-]# ]] || die "not a plain name: $v"; done
+[[ $port == <1-65535> ]] || die "WINVM_PORT isn't a port: $port"
+
+# Windows' OpenSSH waits for stdin to close, so it always gets /dev/null. The
+# script travels UTF-16 base64-encoded, so no quoting survives two shells.
+vm_ps() {
+  local b64=$(print -rn -- "\$ProgressPreference = 'SilentlyContinue'"$'\n'"$1" |
+              iconv -f UTF-8 -t UTF-16LE | base64 | tr -d '\n')
+  ssh $opts -p $port $user@127.0.0.1 "powershell -NoProfile -NonInteractive -OutputFormat Text -EncodedCommand $b64" </dev/null
+}
+
+# QEMU's monitor inside the container: one command per line on stdin.
+monitor() { ssh $host "docker exec -i $ctr nc -q 1 -U /run/shm/monitor.sock" >/dev/null }
+
+# Input has to come from the signed-in desktop session, not SSH's session 0, so
+# a scheduled task running as the user replays one click or wheel turn written
+# to input.txt, then deletes the file to say it's done. Any error stops it
+# before that, so the caller times out instead of reporting a click.
+INPUT_PS1='$ErrorActionPreference = "Stop"
+$a = (Get-Content "$PSScriptRoot\input.txt").Trim() -split "\s+"
+Add-Type -Namespace WinVm -Name Input -MemberDefinition @"
+[DllImport("user32.dll")] public static extern bool SetProcessDPIAware();
+[DllImport("user32.dll")] public static extern bool SetCursorPos(int x, int y);
+[DllImport("user32.dll")] public static extern void mouse_event(uint flags, int dx, int dy, int data, System.IntPtr extra);
+"@
+[WinVm.Input]::SetProcessDPIAware() | Out-Null
+if (-not [WinVm.Input]::SetCursorPos([int]$a[0], [int]$a[1])) { throw "SetCursorPos failed" }
+Start-Sleep -Milliseconds 150
+if ($a[2] -eq "click") {
+  [WinVm.Input]::mouse_event(0x2, 0, 0, 0, [IntPtr]::Zero); Start-Sleep -Milliseconds 60
+  [WinVm.Input]::mouse_event(0x4, 0, 0, 0, [IntPtr]::Zero)
+} else { [WinVm.Input]::mouse_event(0x800, 0, 0, 120 * [int]$a[3], [IntPtr]::Zero) }
+Remove-Item "$PSScriptRoot\input.txt"'
+
+pointer() {  # x y click|wheel [notches]
+  local b64=$(print -rn -- $INPUT_PS1 | base64 | tr -d '\n')
+  vm_ps '$ErrorActionPreference = "Stop"
+$d = Join-Path $env:LOCALAPPDATA "windows-vm"; $req = "$d\input.txt"; $mine = $false
+# Giving up: end the helper first, or it could wake up and act in a later request.
+function Abandon {
+  Stop-ScheduledTask -TaskName WindowsVmInput -ErrorAction SilentlyContinue
+  foreach ($i in 1..25) {
+    if ((Get-ScheduledTask -TaskName WindowsVmInput -ErrorAction SilentlyContinue).State -ne "Running") { break }
+    Start-Sleep -Milliseconds 200
+  }
+  Remove-Item $req -ErrorAction SilentlyContinue
+}
+trap { [Console]::Error.WriteLine("windows-vm: $_"); if ($mine) { Abandon }; exit 1 }
+# One request at a time: the helper, the task and input.txt are shared. Windows
+# frees the mutex when this process ends, however it ends.
+$lock = New-Object Threading.Mutex($false, "windows-vm-input")
+try { $mine = $lock.WaitOne(15000) } catch [Threading.AbandonedMutexException] { $mine = $true }
+if (-not $mine) { [Console]::Error.WriteLine("windows-vm: another click or scroll is still in progress"); exit 1 }
+New-Item -ItemType Directory -Force $d | Out-Null
+[IO.File]::WriteAllText($req, "'"$*"'")
+[IO.File]::WriteAllText("$d\input.ps1", [Text.Encoding]::UTF8.GetString([Convert]::FromBase64String("'$b64'")))
+$act = New-ScheduledTaskAction -Execute powershell.exe -Argument "-NoProfile -WindowStyle Hidden -ExecutionPolicy Bypass -File `"$d\input.ps1`""
+$who = New-ScheduledTaskPrincipal -UserId $env:USERNAME -LogonType Interactive
+Register-ScheduledTask -TaskName WindowsVmInput -Action $act -Principal $who -Force | Out-Null
+Start-ScheduledTask -TaskName WindowsVmInput
+foreach ($i in 1..50) { if (-not (Test-Path $req)) { exit 0 }; Start-Sleep -Milliseconds 200 }
+Abandon
+[Console]::Error.WriteLine("windows-vm: no input after 10 s: is $env:USERNAME signed in on the VM screen, and is x,y on it?"); exit 1'
+}
+
+(( $# )) || die "usage: see the top of $0"
+cmd=$1; shift
+case $cmd in
+  up)
+    ssh $host "docker start $ctr" >/dev/null
+    for i in {1..60}; do
+      vm_ps 'exit 0' 2>/dev/null && { print "up"; exit 0 }
+      sleep 5
+    done
+    die "Windows didn't answer on SSH within 5 minutes" ;;
+  down)   # the container turns SIGTERM into an ACPI shutdown and waits for Windows
+    ssh $host "docker stop -t 150 $ctr" >/dev/null && print "down" ;;
+  status) ssh $host "docker ps -a --filter 'name=^$ctr\$' --format '{{.Names}}: {{.Status}}'" ;;
+  ps)     (( $# == 1 )) || die "usage: ps '<PowerShell>'"; vm_ps "$1" ;;
+  put)
+    [[ -f ${1:-} ]] || die "usage: put <file> [<dir>]"
+    scp -q $opts -P $port $1 "$user@127.0.0.1:${2:-C:/Users/$user/Downloads}/${1:t}" </dev/null ;;
+  shot)
+    (( $# == 1 )) || die "usage: shot <out.png>"
+    print "screendump /tmp/windows-vm-shot.ppm" | monitor
+    sleep 1
+    ssh $host "docker exec $ctr sh -c 'cat /tmp/windows-vm-shot.ppm && rm /tmp/windows-vm-shot.ppm'" | python3 -c '
+import re, struct, sys, zlib
+d = sys.stdin.buffer.read()
+m = re.match(rb"P6\s+(\d+)\s+(\d+)\s+255\s", d) or sys.exit("windows-vm: no screen dump")
+w, h = int(m[1]), int(m[2]); px = d[m.end():]
+raw = b"".join(b"\0" + px[y * w * 3:(y + 1) * w * 3] for y in range(h))
+chunk = lambda t, b: struct.pack(">I", len(b)) + t + b + struct.pack(">I", zlib.crc32(t + b))
+open(sys.argv[1], "wb").write(b"\x89PNG\r\n\x1a\n" + chunk(b"IHDR", struct.pack(">IIBBBBB", w, h, 8, 2, 0, 0, 0))
+                              + chunk(b"IDAT", zlib.compress(raw)) + chunk(b"IEND", b""))
+' $1
+    print $1 ;;
+  click)  (( $# == 2 )) || die "usage: click <x> <y>"; int $1; int $2; pointer $1 $2 click ;;
+  scroll) (( $# == 3 )) || die "usage: scroll <x> <y> <n>"; int $1; int $2; int $3; pointer $1 $2 wheel $3 ;;
+  keys)
+    (( $# )) || die "usage: keys <key>..."
+    for k; do [[ $k == [a-z0-9_.,/=-]## ]] || die "not a QEMU key name: $k"; done
+    for k; do print "sendkey $k"; done | monitor ;;
+  frame-key)
+    [[ ${1:-} == (add|remove) ]] || die "usage: frame-key add|remove"
+    pub=(${=$(vm_ps 'Get-Content (Join-Path $env:USERPROFILE ".ssh\id_ed25519_frame.pub")' | tr -d '\r')})
+    [[ ${pub[1]:-} == ssh-ed25519 && ${pub[2]:-} == [A-Za-z0-9+/=]## ]] ||
+      die "the VM has no Frame Control key yet: run Set Up Connection in the app first"
+    # $1: add|remove, $2: the key's base64, $3: the exact line this script owns.
+    ssh frame "sh -s -- $1 '$pub[2]' '$pub[1] $pub[2] $TAG'" <<'EOF'
+f=~/.ssh/authorized_keys
+if [ "$1" = add ]; then
+  if grep -qxF "$3" "$f"; then exit 0; fi
+  if grep -qF "$2" "$f"; then echo "the headset already trusts this key through another entry; left as it is"; exit 0; fi
+  [ -z "$(tail -c 1 "$f")" ] || echo >> "$f"   # a last line without a newline would swallow ours
+  echo "$3" >> "$f"
+else
+  t=$(mktemp "$f.XXXXXX") || exit 1
+  grep -vxF "$3" "$f" > "$t"   # 0: lines left, 1: none left, more: couldn't read or write
+  if [ $? -gt 1 ] || ! chmod 600 "$t" || ! mv "$t" "$f"; then
+    rm -f "$t"; echo "couldn't rewrite $f; it's unchanged" >&2; exit 1
+  fi
+fi
+EOF
+    print "frame-key $1: done" ;;
+  *) die "unknown command: $cmd (see the top of $0)" ;;
+esac