This is a living record of concrete gaps encountered while using WinVM Testbed. Keep observed behavior, effect, workaround, and a likely improvement direction together so later work can reproduce the problem.
Status: substitution fixed; occasional loss remains open. Typing into Windows 11 Notepad (11.2607, ARM64 appliance, spell check on) sometimes produced the wrong characters. "the quick brown fox 0123" became "the iiick nnnnn 11111123". The resident reported the text as delivered, with an unverifiable effect.
- Cause. The failures followed word boundaries and depended on Notepad's
spell check, not on window activation:
- With spell check off, worded text was correct 60 of 60 times.
- Text without spaces was correct 30 of 30 times, even with background activation.
- Turning autocorrect off alone did not help.
- A WinForms fixture never failed.
- Keys that arrive while Notepad's text pipeline is behind are processed
late. A Unicode packet (
VK_PACKET) does not carry its character in the key message, so the late keys came out as the most recent character or were lost.
- Fix.
typenow sends each character the receiving keyboard layout produces with at most Shift as a layout keystroke. The result reports this aswindows.native/send_input_layout_keysand gives the counts indata.textInput. Unicode packets remain for other characters, and for all text when a modifier is held, Caps Lock is on, or an input method is open. - Results. Before the fix, targeted typing from the background failed 4 of 9 times. Afterwards, every Notepad and fixture scenario showed no substituted characters in 10 attempts each.
- Still open. Notepad with spell check still drops a few layout
keystrokes near word boundaries: about 3% of short strings in the guest,
and 2 of 30 resident attempts. The effect stays
unverifiable. When the exact text matters, read the result back, or useset_valuewhere the control supports it.
Status: open. desktop input text "é" arrived at the resident as two
UTF-16 units (mojibake) through the common CLI and through winvm control
with a raw UTF-8 JSON argument. The same request with the character escaped
as \u00e9 in ASCII JSON arrived intact. The controller-to-guest argument
path needs an explicit UTF-8 contract.
Status: resolved by policy and tooling. The retained UTM appliance booted
to the Windows sign-in screen, but this controller held no credential for it;
the factory setup password had been rotated and the new value was not stored.
One sign-in attempt with the stale setup password was rejected and the VM was
destroyed and rebuilt. The adapter now keeps the current password in
WINVM_LOGIN_SECRET_FILE (a per-UUID default under the controller's secret
store), winvm credential rotate never changes Windows without first writing
the new value, winvm login signs in from it after a cold boot, and doctor
warns when none is stored.
The broker reported that rejected sign-in as delivery: confirmed, effect: no_effect while Windows displayed "The password is incorrect". A typed
credential-rejected result remains open.
Status: resolved. The UTM factory's in-place EFI repair wrote inside UTM's
container, which macOS denies to agent sessions; it now normalizes through
export and import. The UTM provider lacked the advertised factory-status and
trust-ssh-host-key commands; both now use the UTM guest agent. On a fresh
appliance reached only over SSH, winget had no community source
(0x8A15000F); the development bootstrap now installs Microsoft's signed
source package and retries once. The bootstrap's in-session helper needs a
signed-in desktop, so sign in with winvm login before bootstrapping after a
cold boot. The first semantic snapshot after installation timed out once and
then passed on repeated conformance runs.
Status: mitigated 2026-08-23. A cold UTM 4.7.5 process reported zero registered VMs even though its persisted registry and the exact pinned UTM bundle were both present. The bundle's embedded UUID still matched private inventory. Opening that bundle caused UTM to repopulate its runtime inventory; the Windows target then started and passed the full common doctor.
Effect: status returned no state, doctor collapsed the power and resident
dimensions to unavailable, and callers reasonably but incorrectly proposed
re-pinning the target. Changing the pin would have weakened the safety guard
without repairing UTM's missing runtime registration.
Mitigation: doctor now reports exact identity separately and directs this
specific failure to winvm repair-registration. The explicit repair verifies
the selected bundle's name and embedded UUID against private inventory before
asking UTM to register it. It neither starts the VM nor rewrites the pin. A
bundle absence or metadata mismatch still fails closed and requires deliberate
private-inventory repair.
Status: resolved 2026-08-04. Window discovery now considers UTM windows from every macOS Space, prefers an on-screen match when more than one exists, and falls back to the largest matching off-Space console. The normalized 1399×985 capture succeeded while the console remained on another Space.
The VM, SSH, guest relay, and UTM input channel all remained healthy, and
macOS accessibility still identified the standard UTM Windows window.
Nevertheless, winvm screenshot failed with No visible UTM window because
Core Graphics discovery used optionOnScreenOnly.
Effect: provider-level visual recovery disappeared merely because macOS moved the UTM window between Spaces. That is especially hazardous when screenshot plus raw input is the fallback for a system-modal prompt.
Status: open; correlation observed, cause not yet isolated. The first LAN listen by an unsigned candidate displayed the Windows Defender Firewall consent dialog. While that system-modal dialog was present, both guest-agent IP discovery and SSH became unavailable. Semantic UI could not dismiss the dialog because its session-1 relay is reached through SSH; provider screenshot and raw click remained the only control path. A raw Cancel dismissed the dialog and Windows created public-profile block rules for the candidate, but the transport did not recover. A normal restart from the secure screen then wedged on a black console and required a bounded force-stop/start recovery.
The firewall dialog and transport loss happened together, but this observation does not prove the dialog caused the SSH or guest-agent failure. The guest's OpenSSH firewall rules included the active profile after recovery.
Effect: a production-style first-LAN-listen test can lose both high-level control channels at the exact point where a system prompt needs attention.
Workaround: take a provider screenshot before the first listen, keep provider raw input ready, and use it for system-modal dialogs rather than waiting on the SSH-backed semantic relay. After recovery, remove any candidate-specific block rules and, when appropriate for the test, add a narrowly scoped temporary inbound rule for the exact executable and port. Remove that rule during cleanup.
Possible direction: teach the production-test helper to recognize the firewall dialog from provider capture, report that semantic automation is unavailable, and offer a bounded raw-input recovery path. Add bounded failure times and diagnostics around IP discovery, SSH, and restart so a transport loss cannot look like an indefinite controller hang.
Status: open; observed 2026-08-20. After the accelerated Windows display initialized, Core Graphics still enumerated the UTM console but macOS refused both its window-ID capture and a tiny screen-region permission probe. The Windows-resident GDI capture remained healthy and returned the complete 1280×1024 desktop with no host interference.
Effect: outer visual recovery is unavailable in that host session even though ordinary target-native control is healthy. The provider now fails closed with a diagnostic that points to host lock/display state and Screen Recording permission rather than emitting a black or unrelated host image.
Open: repeat the provider capture after the controller desktop is unlocked and confirm whether host session state, Screen Recording authorization, or the accelerated surface is the deciding condition.
The clean ARM64 Windows guest could drive installed applications but could not
build a Rust/Tauri candidate. A successful native ARM64 build required the
official Visual Studio 2022 Build Tools with the C++ workload, the ARM64 C++
tools component, LLVM/Clang for crates such as ring, and the ARM64 Rustup
toolchain. Visual Studio component modification had to be launched with
Start-Process ... -Wait; invoking the bootstrapper directly with --wait
returned before component installation finished.
Effect: controller health is not the same as release-build readiness, and a candidate run can otherwise discover a multi-gigabyte toolchain gap midway through the platform matrix.
Possible direction: add a separate, opt-in doctor-build check or a documented
release-builder bootstrap. Keep the ordinary testbed requirements small: the
controller should not silently install a compiler toolchain merely to drive a
prebuilt application.
Status: re-resolved 2026-08-20. The original fixed-resolution correction worked only while the UTM console retained the configured size. Dynamic resolution later changed the live Windows display between 1024×768 and 1280×1024. The stale 1399×985 configuration then either caused normalization to fail or produced the wrong click coordinate space. Provider capture now removes the configured UTM title-bar height and derives output dimensions from the live viewport. Explicit width and height remain an opt-in for a guest that does not follow console size.
winvm screenshot captures the full macOS UTM window with
screencapture -l. On the observed Retina host the PNG was 2798×2050 and
included UTM chrome/title-bar pixels, while the guest display and UTM mouse
API used roughly half-scale guest coordinates.
Effect: a point read from the screenshot cannot be passed directly to
winvm click; it needs an undocumented scale and title-bar transform. This
made WebView-only controls need trial-and-error clicks.
The display dimensions remain explicit configuration because screenshot recovery must work even when SSH and PowerShell display discovery are unavailable. Update them when the guest logical display mode changes.
WinApp's InvokePattern reported that the NSIS installer's Finish button
was invoked, but the installer remained on the same page and the application
did not launch. The semantic click operation on the same named button did
advance the installer and launch the app.
Effect: a successful ui invoke response is not proof that the target handled
the action. Workaround: re-inspect the window after every material action and
retry with semantic click when the transition did not occur.
Possible direction: add this warning and the click fallback to the UI guide,
and consider an optional action helper that waits for a window/tree change.
The 200 OK Tauri frame was discoverable, but its WebView2 controls were absent from UIA. This limit is already noted in the README, but it is consequential: settings, switches, server controls, and inline messages required screenshots, raw clicks, and keyboard focus traversal. That path is substantially less reliable than WinApp's native-control automation and made a visually clipped dialog difficult to inspect.
Possible direction: add a WebView2/CDP-capable driver or a documented application opt-in that exposes embedded web content to accessibility tooling.
Current: A ScriptingBridge failure can abort utmctl before guest
dispatch. Shell provider and workspace calls now retain private, bounded
metadata even when probes discard stderr. See
UTM diagnostics for log locations,
correlation with Apple crash reports, a read-only reproduction, and coverage
limits. A read-only health guard now refuses the observed broken automation
endpoint before CLI launch, and library recovery no longer repeats arbitrary
status failures. Restoring the underlying UTM endpoint remains unresolved.