Local VM UI testing

Run PowerToys .Next UI tests in a persistent, interactive Windows VM while keeping product and test execution off the host. Use this skill as the execution complement to ui-tests-migration. Restore the baseline checkpoint or recreate the guest when clean-profile behavior must be validated.

The guest is a Hyper-V virtual machine. Nothing runs nested, so the same scaffold works on x64 and on Windows on ARM, where nested virtualization is unavailable to any Linux-hosted emulator.

The local correctness gate passes when the full suite is green on Windows 10 and on Windows 11, in two separate VMs. Run Windows 10 Enterprise LTSC 2021 first for the fastest feedback, then run the same unfiltered suite on Windows 11. Differences in the shell, compositor, theming, and timing break tests that contain nothing Windows 11-specific, and those are exactly the failures worth catching locally instead of in CI. On a Windows on ARM host, Windows 11 ARM64 is the only practical guest; run it with -Platform ARM64 and get the Windows 10 half from an x64 host.

Start guests with the default resource profile: 4 vCPUs and 8 GB RAM. Get the target suite fully green before lowering resources with the Constrained profile (1 vCPU and 4 GB RAM).

For create, migrate, or stabilize tasks, complete both profiles and then automatically invoke ui-tests-pipeline-ci: pass its access preflight, commit and push the scoped changes, queue CI, and remain attached until terminal validation. Do not mark the task complete at local green or wait for another user request to publish/run CI. Respect explicit local-only/no-push/no-CI scope and the Microsoft FTE/access gates. Local execution or VM setup alone does not authorize publication. See the pipeline skill for the shared completion and blocker rules.

How the host reaches the guest

When to use this skill

Use it when asked to:

• Create or migrate UI tests and iterate repeatedly without reprovisioning Windows each run.
• Validate Explorer, hotkey, WebView2, shell-extension, foreground, or composed-visual behavior in a real interactive desktop.
• Run tests as a true standard user while retaining a separate administrator control channel.
• Reuse unchanged PowerToys, winappcli, and .NET payloads and refresh only changed archives.
• Keep a reproducible local VM baseline with optional .NET 10, WebView2, diagnostics, or msvsmon.
• Collect durable status.json, TRX, transcripts, logs, screenshots, and failure attachments.
• Compare revisions in one stable VM before confirming the result from a restored checkpoint.

Do not treat a persistent VM as proof of clean-profile behavior. Caches, registrations, settings, and first-run state survive between runs. Restore the baseline checkpoint or recreate the VM when those are the behavior under test.

Relationship to other UI-test skills

Do not modify stabilized tests merely to make the local VM green. First prove that the suite executes, produces assertion-bearing TRX, and has a useful success rate. Classify environment-specific failures separately unless the task explicitly asks for stabilization.

Guest OS policy

• Two guests, two full suites. Windows 10 Enterprise LTSC 2021 (build 19044/21H2, newer than the Windows 10 20H2 baseline) and Windows 11. Both must be fully green before the local gate passes. LTSC is not available through Fido; -Source Fido -Windows 10 automates Microsoft's official mobile-user-agent ISO page and is the practical public default. The public ISO/MCT images are too old for .NET 10 CET, so Setup Dynamic Update must bring Win10 to 1904x.5007+ before the baseline. Use licensed Microsoft subscription media for LTSC when available, and always record the edition, ISO hash, and installed full build (references/setup.md §3).
• Windows 10 runs first: it is the faster loop and surfaces most defects. Windows 11 then runs the same suite, unfiltered - not only the tests that look Windows 11-specific. A test with no Windows 11 content can still fail there, which is the whole reason for the second pass.
• Narrow filters belong to iteration and diagnosis. They are never the evidence for the Windows 11 pass.
• Give each OS its own VM name, vm.config.psd1, VHDX, checkpoints, and exchange. Never upgrade or repurpose the Windows 10 guest into the Windows 11 guest; the two baselines must stay independent.
• Surfaces that exist only on Windows 11, such as the tier-1 Explorer context menu, need no special handling: the full Windows 11 run already covers them.
• On a Windows on ARM host, use a Windows 11 ARM64 guest with -Platform ARM64. Hyper-V does not emulate a foreign architecture, so that host cannot supply the Windows 10 x64 pass - run it on an x64 host and report both.
• A Windows 11 host requirement does not make Windows 11 the first guest target.
• -Platform is not cosmetic: it flows to the guest as the platform environment variable, names visual baselines, and marks the run as pipeline-like. Use only x64Win10, x64Win11, or ARM64.

Required reads

Read only what the task needs:

1. references/setup.md - host requirements, scaffolding, media acquisition, unattended guest creation, credentials, standard-user desktop, and checkpoint baselines.
2. references/agentic-loop.md - payload contract, controller usage, focused-to-suite iteration, evidence, verdicts, and reset strategy.
3. references/customization.md - persistent image customization, .NET 10, WebView2, msvsmon, resource profiles, and golden-baseline guidance.
4. references/troubleshooting.md - guest creation, PowerShell Direct, interactive session, scheduled-task, focus, timeout, and export failures.
5. references/shell-extensions-and-signing.md - read for any shell-extension module (context menu, preview/thumbnail handler). Why unsigned CI PR builds cannot register a sparse MSIX (0% on CI), classic (registry-COM, signing-free) vs modern (sparse-MSIX) surfaces, Debug vs Release/NDEBUG gating, runtime detection, and reproducing CI's classic scenario on a local signed VM.
6. ui-tests-migration - required whenever test code or framework behavior is being created, migrated, or stabilized.

Default agentic cycle

flowchart LR
    A[Design or edit test] --> B[Host build]
    B --> C[Package changed payload]
    C --> D[Start or reuse VM]
    D --> E[Probe standard-user desktop]
    E --> F[Run focused test]
    F --> G[Export status TRX evidence]
    G --> H{Need test change?}
    H -- Yes --> A
    H -- No --> I[Run full suite on Win10]
    I --> J[Run same full suite on Win11]
    J --> K{Both fully green?}
    K -- No --> A
    K -- Yes --> L[Full constrained suites on both OSes]
    L --> M{Local matrix green?}
    M -- No --> A
    M -- Yes --> N[Clean-profile confirmation when required]
    N --> O[ui-tests-pipeline-ci: preflight, commit, push]
    O --> P[Scoped CI and synchronous terminal validation]

The CI continuation applies to implementation tasks; explicit local-only and setup-only tasks stop at their requested local outcome.

Create and maintain this task list:

- [ ] 0. Verify host setup FIRST: `Initialize-LocalVmHost.ps1 -VmRoot <root> -CheckOnly`. If it reports
        IsReady=false, STOP and ask the user to run the elevated command it prints - Hyper-V group
        membership, the DPAPI credential, and guest creation all need a human. Never autopilot past it
- [ ] 1. Read ui-tests-migration guidance for the target test surface
- [ ] 1a. Read the target module's dev docs — `doc/devdocs/modules/<module>.md` (search `doc/devdocs/`,
        including `common/`, if the exact file is missing) — for development-cycle gotchas such as
        Release/`NDEBUG` registration gating, signed sparse-MSIX context menus, and Explorer restarts,
        so a module's registration/deployment requirements do not surface as opaque test failures.
        For shell-extension modules also read references/shell-extensions-and-signing.md.
- [ ] 2. Scaffold or verify the local VM - references/setup.md
- [ ] 3. Build product and test projects on the host to exit code 0
- [ ] 4. Package a lean exchange and verify archive hashes
- [ ] 5. Run the controller with -PlanOnly and inspect its request/plan
- [ ] 6. Probe the non-admin interactive desktop before test execution
- [ ] 7. Run one focused test synchronously; keep the agent turn attached, read the controller
  result's `.Failed` array (non-passed tests + first error line) instead of re-parsing TRX, and
  use `scripts/Invoke-GuestScript.ps1` for guest-state inspection
- [ ] 8. Diagnose the first controlling failure without weakening assertions
- [ ] 9. Rebuild and rerun with -ReuseStagedPayload
- [ ] 10. Widen to the full module suite on Windows 10 and report pass rate/root-cause groups
- [ ] 11. Run the same full suite in the Windows 11 VM; both must be green before the local gate passes
- [ ] 12. Restore the baseline checkpoint for clean-profile confirmation when required
- [ ] 13. Run the full Constrained suite on both OSes; preserve default and constrained evidence
- [ ] 14. For implementation tasks, invoke ui-tests-pipeline-ci without another user prompt, pass
        access preflight, commit/push task-owned changes, and record the exact branch/SHA
- [ ] 15. Queue scoped CI and wait synchronously for verified terminal results; local green alone
        cannot complete the task unless the user explicitly requested local-only scope

Quick start

Host setup is a one-time, human-only step: Hyper-V group membership, the DPAPI guest credential, and guest creation all need elevation or a password. Check it before anything else - this needs no elevation and changes nothing:

pwsh .github\skills\ui-tests-local-vm\scripts\Initialize-LocalVmHost.ps1 -VmRoot X:\PowerToysUiTestVm -CheckOnly

If it reports IsReady=false, stop and ask the user to run the elevated command it prints (see references/setup.md §0). Otherwise scaffold the VM directory outside the repository:

pwsh .github\skills\ui-tests-local-vm\scripts\Initialize-LocalVm.ps1 `
  -DestinationRoot X:\PowerToysUiTestVm

The scaffold and its shared exchange can live on any volume, including a Dev Drive. Only the VHDX and VM configuration paths, set separately in vm.config.psd1, should point at NTFS.

Follow references/setup.md to write the untracked vm.config.psd1, save the administrator credential with Windows DPAPI, obtain install media, and create the guest with New-UiTestVm.ps1. Then stage the archives described in references/agentic-loop.md and run:

pwsh .github\skills\ui-tests-local-vm\scripts\Invoke-LocalVmUiTest.ps1 `
  -VmName PowerToysUiTest-Win10 `
  -VmRoot X:\PowerToysUiTestVm `
  -ExchangeRoot X:\PowerToysUiTestVm\shared\PowerToysUiTests\MyModule `
  -TestExecutable MyModule.UITests.Next.exe `
  -Filter 'Name=MyModule.FocusedTest' `
  -Platform x64Win10 `
  -BuildLabel (git rev-parse HEAD) `
  -SuiteTimeout 15m `
  -TimeoutMinutes 25 `
  -ReuseStagedPayload

The controller starts the VM if needed, requests automatic host sleep prevention, verifies the interactive standard-user desktop, dispatches the guest runner, and waits synchronously for matching status.json. It reports whether sleep prevention succeeded, fails early if the guest task never starts or exits without status, summarizes TRX, and leaves the persistent VM running by default.

Non-negotiable rules

• Complete host setup before anything else and never autopilot around it. Hyper-V group membership, the DPAPI guest credential, and guest creation are human-only: two need elevation that no tool call can approve, one needs a password that must never reach a model. Initialize-LocalVmHost.ps1 performs all three; agents run it -CheckOnly, and on IsReady=false report BLOCKED, print the elevated command it emits, and wait. Do not ask for a password, do not substitute a weaker channel, do not proceed on a partial setup. Invoke-LocalVmUiTest.ps1 enforces the same check.
• Keep the guest's VHDX and VM configuration on NTFS. On this project's host, keeping them on a Dev Drive wedged the VM management service twice - vmms at 0% CPU, even Get-VM hanging, host reboot to recover - and moving to NTFS fixed it. This is an observation on one host, not a property of ReFS: Hyper-V on plain ReFS is supported, so -AllowReFsVolume overrides the default refusal. The scaffold and the exchange are unaffected and run fine on a Dev Drive.
• Build on the host; run PowerToys and tests only in the VM when host execution is prohibited.
• Invoke every focused, full-suite, and constrained controller run synchronously. Keep the active agent turn attached until the controller returns matching status/TRX evidence; never background the controller or end the turn while it runs.
• Complete the local correctness gate with two green full suites: Windows 10 and Windows 11, in separate VMs. Windows 10 runs first for speed; Windows 11 runs the same unfiltered suite, never a Win11-only subset. Narrow filters are for iteration, not for sign-off. On a Windows on ARM host the ARM64 Windows 11 guest covers the Windows 11 half and the Windows 10 half needs an x64 host.
• Establish a fully green correctness baseline with the default (4 vCPU / 8 GB) resources before running the same tests under Constrained (1 vCPU / 4 GB) resources.
• For implementation tasks, require both resource profiles and the subsequent pipeline handoff. Report unavailable publication/CI prerequisites as blocked, not as full validation success.
• Keep VM files and writable exchange folders outside the repository.
• Keep the guest's default inbound network posture. The control channel does not need connectivity, so do not enable remoting, open ports, or attach the guest to a routable network for test dispatch.
• Use a DPAPI-protected credential file. Never put credentials in prompts, scripts, request JSON, unattend files kept after install, source control, or command-line arguments.
• Run UI tests in an already logged-on standard-user desktop, never in session 0 or as SYSTEM.
• Keep a separate administrator account only for VM control and scheduled-task registration.
• Verify user, token integrity, Explorer presence, session ID, and display size before tests.
• Run product/tests from guest-local storage under C:\PowerToysUiTestRun.
• Use PowerShell 7 for interactive probe/test scheduled tasks. PowerShell Direct and OEM bootstrap remain PS5.1-compatible by design; do not enable a remoting endpoint merely to use PS7.
• Reuse payloads by per-component hashes; refresh only changed tests/product/tools.
• Preserve assertions and visual thresholds. Classify VM-specific failures from evidence.
• Always parse TRX and require total > 0 plus executed == total. A process exit code alone cannot distinguish assertions, skipped/inconclusive tests, zero tests, timeout, or infrastructure failure.
• Keep the VM after normal runs for iteration. Stop it explicitly when idle; delete its VHDX only for an intentional baseline reset.
• The controller requests automatic host sleep prevention and reports HostSleepPrevented; failure is a warning. It cannot override manual sleep, lid-close policy, reboot, shutdown, or power loss.
• Final clean-profile claims require a restored baseline checkpoint or a recreated guest. Restore with Reset-LocalVm.ps1 -Restore.