Repository navigation
Conversation
This was referenced Oct 3, 2026
The driver only ever met a real kernel on a developer's VMware VM. This adds a Windows guest that any Linux host with /dev/kvm can run unattended, and a workflow that loads every built package into it. - ci/guest/image: builds a Server Core golden image from the evaluation ISO, with OpenSSH, the QEMU guest agent, test signing on and boot recovery off, and checks it over SSH before publishing it. - ci/guest/host/guestctl: one CLI to boot, ssh, push/pull, reboot, snapshot/revert, screenshot and guest-agent-exec the guest, for CI and for agents acting in the guest. - ci/guest/in-guest: prepare (verifier, dumps), run (package integrity, corpus, server-rs, install, listing, Test-BlorgCorrectness, extra suites), and collect diagnostics with a bugcheck verdict. - ci/guest/run-guest-tests.sh: one end-to-end run with an exit code. - guest-runtime.yml: runs it on ubuntu-latest after each successful build, against the blorg-package-windows-x64 artifact. - ci/guest/infra/azure: optional long-lived nested-virtualization host.
Copies the whole tests/guest-suites directory (only its top-level *.ps1 run), and third_party/schemas/conformance into suites/contract, where the Contract suite from the API alignment work looks for it. The workflow fetches the schemas submodule without credentials, since it runs PR code.
…r's image cache server-rs now runs on the KVM host as the Linux build and the guest reaches it at 10.0.2.2, so the driver's traffic crosses a real NIC and the backend is case-sensitive, as deployed. The corpus is generated on the host by make-corpus.py (one generator for both topologies); suites' fixtures ship as suites/<name>.corpus/. The in-guest Windows server remains as a fallback when the package has no Linux binary. A workflow_run triggered by a pull request runs in master's cache scope, so it now uses the default branch's rig scripts and never saves the golden image.
guest-runtime.yml takes a session input: bash run against the live guest after the tests, with guestctl on PATH, its transcript and files returned in the guest-results artifact. PRs touching the rig run session-probe.sh, which exercises every channel once (ssh, push/pull, the driver's volume, screenshot, guest agent, snapshot/revert).
…r into the image The package now carries server/linux-x64/server-rs, so the in-guest Windows server fallback is gone along with its corpus push and -CorpusDir. Driver Verifier is set in the golden image, so a default run needs no reboot before its tests, and the image is compressed with zstd for faster first reads.
Chuccle
force-pushed
the
claude/windows-guest-test-rig-vl5f6l
branch
from
October 3, 2026 22:12
5387da9 to
a044fb3
Compare
The first image build ran 67 minutes with no way to see why from outside the artifact. Changed screens now go to the log as base64 every two minutes, as do a session's screenshots, so an agent that can read logs but not download artifacts can still see the guest.
Agents behind an egress proxy can read the checks API but not job logs or artifacts; the verdict (blorg guest-verdict) and the agent session's transcript tail (blorg guest-session) now reach them as annotations.
guest-runtime.yml no longer takes commands: the package from build.yml, an image from pinned inputs and the suites in the commit, nothing else. Acting in the guest from a remote session moves to guest-session.yml, manual only and gating nothing. The host setup both share is one composite action, ci/guest/action.yml. Image inputs (install ISO, virtio-win, OpenSSH zip) are pinned by SHA-256; an unpinned one is reported with its hash so it can be pinned. Runs queue instead of cancelling, so an image build is never thrown away, and the image builds even when no package is found.
…host The guest section moves out of AGENTS.md into ci/guest/README.md, written for a person running the scripts on any Linux host with KVM. Comments, the workflows and the session script no longer assume who is driving.
A '--' inside an XML comment made autounattend.xml malformed, so Setup ignored it and sat at the language screen until the build timed out. build-image.sh now checks the answer file is well-formed before booting.
- Install-BlorgFS.ps1 gets its .inf and .cer paths explicitly: under powershell -File, Windows PowerShell 5.1 leaves $PSScriptRoot empty in param() defaults. - A child script's stderr no longer aborts the step under 'Stop'; its exit code decides. - A step that returned $true was recorded as a skip ($true -eq 'skip' is true in PowerShell). - The guestctl probe runs even when the tests fail.
Versioned URLs and SHA-256 pins for the Windows evaluation ISO and virtio-win 0.1.302, and a SHA-256 pin for the Win32-OpenSSH zip, taken from the first full image build.
The first full run installed the driver on Server 2022 (build 20348) and sc start failed with 127, procedure not found: BlorgFS.vcxproj targets NTDDI 0x0A000010 (24H2, build 26100), so it imports kernel routines 2022 does not have. The golden image is now Server 2025 evaluation (26100). Its ISO hash is pinned after this build reports it.
The probe failed on 'sc query BlorgFS' after the driver bugchecked the guest. Whether the driver is healthy is the tests' verdict; the probe only checks that every way into the guest works.
When a guest run brings minidumps back, a windows-latest job (which has cdb) runs tools/Get-CrashVerdict.ps1 over them with the driver's PDB from the same Build BlorgFS run, and reports the bugcheck, faulting line and stack in its log and as a 'blorg guest-crash' check annotation. Show-CrashAnalysis.ps1 does the same by hand on any Windows machine. Get-CrashVerdict.ps1 gains -SymbolDir for the driver's own PDB.
A minidump's !analyze reports BUGCHECK_CODE rather than the full dump's 'BugCheck' banner, so the verdict left Bugcheck empty for the guest's 0xD1.
Owner
Author
|
Windows guest (KVM) is red because of a driver bug, not because of this PR. The package under test is #63's build (run 37157505659). Its driver bugchecks with 0xD1 in A fix exists as a patch that moves the address copy before the spinlock. It is not ported into this PR, because this PR does not build the package. The fix has to land where the package is built (#63, or the consolidated #68), and this check will turn green once the guest tests a package that includes it. |
Owner
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Before: the driver only ever met a real Windows kernel on a developer's VMware VM, driven by hand through
vmrun. CI built and unit-tested it but never loaded it.After: any Linux host with
/dev/kvm(a stock GitHububuntu-latestrunner, a cloud VM with nested virtualization, a Linux dev box) can boot a scripted Windows guest, install the packaged driver in it, test it against the package's Linux server-rs on the host, and return a pass/fail verdict with logs and crash dumps. CI does that as a fixed test for every successful build, and anyone can work inside the guest by hand through one CLI.The input is the
blorg-package-windows-x64artifact from #63, so the rig tests exactly what ships and builds nothing itself.How
ci/guest/image/build-image.sh: unattended Windows Server 2022 Server Core install from Microsoft's evaluation ISO, every input pinned by SHA-256.Setup-GoldenImage.ps1installs OpenSSH (key-only, PowerShell shell) and the QEMU guest agent, turns test signing on and boot recovery off, sets Driver Verifier onBlorgFS.sys, keeps kernel dumps, and disables Windows Update. The result is booted once and checked over SSH before it is published.ci/guest/host/guestctl:up,down,ssh,ps,push/pull,reboot,snapshot/revert,screenshot,qga-exec,qmp. SSH is forwarded on host loopback only.ci/guest/run-guest-tests.sh: one end-to-end run. server-rs runs on the host and the guest reaches it at10.0.2.2; a deterministic corpus comes fromhost/make-corpus.py. A bugcheck or unexplained reboot fails the run even if every step passed. The verdict is also ablorg guest-verdictcheck annotation..github/workflows/guest-runtime.yml: a fixed test that takes no commands. Runs onworkflow_runof Build BlorgFS, on PRs touching the rig, and on demand. Image cached by recipe hash and quarter; runs queue rather than cancel; no secrets..github/workflows/guest-session.yml: manual only, gates nothing. Runs a script ofguestctlcommands against a live guest, for working in it without a KVM host of your own.ci/guest/README.md: how to run all of it by hand.ci/guest/infra/azure/main.bicep: optional long-lived KVM host.Temporary: rig PR runs take #63's package build until #63 is merged; that commit comes out afterwards.