Root cause, confirmed causally rather than inferred: masking systemd-timesyncd (which vmbox does so `-rtc base=` stays pinned, since timesyncd otherwise resets the guest clock to real time) left systemd-time-wait-sync waiting for a sync that can never arrive. That unit sat in "start running" forever and every other job queued behind it as "start waiting" -- sshd included -- so the guest never reached multi-user.target. It booted to a login prompt and looked perfectly healthy while being permanently unreachable. The fix is one line: mask both units together. Four earlier hypotheses were each disproven by measurement. What finally worked was asking the guest instead of reading absence-of-log-lines: `systemctl is-system-running` returned "starting", and `systemctl list-jobs` put systemd-time-wait-sync at the head of the queue. Stopping it made sshd active and bound :22 within seconds, and ssh from the host then succeeded. Rule for next time: a unit that is enabled, has satisfied dependencies, and neither starts nor fails is QUEUED -- check list-jobs first. The smoke gate added earlier did its job here: it caught the unreachable image and refused to seal it, keeping the sandbox so its logs survived. Also fixes serial_exec markers to be plain alphanumerics -- shell metacharacters get echoed and line-wrapped by the tty, which made them unreliable to match and produced spurious timeouts. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VdyMBGCrGVGDWtJ9xe48f3 |
||
|---|---|---|
| .. | ||
| bin | ||
| guest | ||
| lib | ||
| tests | ||
| .gitignore | ||
| .pre-commit-config.yaml | ||
| install.sh | ||
| README.md | ||
vmbox — a disposable Arch VM sandbox
Run destructive system scripts — shutdown installers, chattr +i hosts guards,
pacman hooks — against a throwaway Arch guest, and verify the outcome from the
host. Break it as thoroughly as you like; reset takes under a second.
./install.sh # installs deps (libisoburn), links `vm` onto PATH
vm build # builds the golden image, once (~10 min)
vm new demo # instant: a ~200 KB overlay on the sealed base
vm run demo 'sudo systemctl poweroff'
vm reset demo # back to pristine
Why this exists
Much of testsAndMisc/linux_configuration cannot be tested on the real machine.
The repo's own CI says so — its Ubuntu job runs only the "side-effect-free"
tests and notes the rest "drive pacman, /etc, systemd or adb and cannot run on
an Ubuntu runner without root and an Arch container". Locally they probe the
live system (test_results.log asserts PASS: /etc/hosts is immutable).
Three things make testing them on the host genuinely unsafe:
chattr +iimmutability — 15 files make/etc/hostsand friends immutable. A bad test is not simply revertable;rmfails with EPERM.- Auto-poweroff — the shutdown installers write systemd units that power the machine off on a schedule.
- Pacman hooks —
boot_recovery/install.shinstalls pre/post-transaction hooks and touches mkinitcpio. Blast radius is "your next update".
The hard part: a successful shutdown kills your observer
Running a shutdown script over SSH returns a broken pipe. That is indistinguishable from a crash, a hang, or a network blip — ssh returns 255 for all of them. So "did it actually shut down?" is answered from the host, using two artefacts that outlive the VM:
| Channel | What it proves |
|---|---|
QMP event log (events.jsonl) |
That the machine stopped, and who stopped it |
| Serial console log | How far the shutdown sequence got |
A persistent recorder attaches at launch, before anything runs in the guest. This ordering matters: attach afterwards and QEMU has already exited, the stream is closed, and there is no verdict to read.
Five outcomes, all host-observable
| Observation | Verdict | Exit |
|---|---|---|
SHUTDOWN guest:true + serial reaches reboot: Power down |
clean poweroff | 0 |
SHUTDOWN guest:true, serial truncated |
dirty — stopped mid-sequence | 3 |
SHUTDOWN reason=guest-reset |
rebooted, did not power off | 4 |
GUEST_PANICKED |
crashed | 5 |
| no event, liveness probe fails | hung | 6 |
guest:true alone is not enough for "shuts down completely" — poweroff -f
and a script that kills systemd mid-sequence both produce it. The serial tail is
the discriminator, and it is asserted by grep, not read by eye.
A
POWERDOWNevent is not proof of anything: it means the ACPI request was delivered. With no OS booted, QEMU emitsPOWERDOWNand then runs forever.
Design notes
- Reset is a backing-file trick. The golden image is built once and sealed
(
chmod 444+ a sha256 sidecar verified on every launch). Each sandbox is a thin qcow2 overlay, sovm resetisrm+ recreate. Booting the base itself would silently invalidate every overlay, so the launcher refuses to. - No root anywhere. User-mode networking with a per-VM forwarded SSH port; VM-to-VM traffic rides a socket multicast segment. No bridge, no libvirt, no daemon.
- The guest clock is pinned (
--rtc) and re-applied on every launch. The shutdown installers gate on a 21:00–05:00 window, so an unpinned clock makes their tests pass or fail depending on the time of day. NTP is masked in the image, or it would quietly undo the pin. - The guest gets its own clone of the repos from a read-only mount. Real credentials are never shared; the guest carries sentinel values instead.
Commands
| Command | Effect |
|---|---|
vm build [--force] |
Build/rebuild the golden image |
vm new <name> [--rtc <ts>] |
Create a sandbox |
vm run <name> <cmd...> |
Run a command, print the verdict |
vm ssh <name> |
Interactive shell (starts the VM if needed) |
vm screenshot <name> [out.png] |
Capture the screen (locker/X11 tests) |
vm reset <name> |
Wipe back to pristine |
vm rm <name> / vm list |
Delete / list sandboxes |
Out of scope
A VM cannot cover everything: adb/phone tests (~36 files), GPU monitors, the
i2c RGB in setup_night_lockdown.sh, and the Nextcloud/SearXNG suites (whose
fixtures assume a Debian layout) all stay on the host or stay untested.
Tests
bats tests/test_vmbox.bats — host-side unit tests for name validation, meta
handling, serial rotation and the full verdict table. They do not boot a VM;
booting is covered by the destructive end-to-end demo above.