Run the dependency-free adapter tests without live targets:
python3 -m unittest discover -s tests/client -vWith an accepted target already running, run the shared guarded workflow:
tests/client/live-desktop-conformance.sh windows
tests/client/live-desktop-conformance.sh macos
tests/client/live-desktop-conformance.sh linuxThe live script uses only bin/machine-control. Platform cases provide fixture
setup, independent-effect reads, and cleanup through the explicit testbed or
os escape hatches. The ordinary status, capabilities, local/outside parity,
snapshot, action, capture, and artifact operations all use the common client.
It acquires one two-hour exclusive target-use claim after the read-only doctor,
carries that claim through every operation, and releases it from its cleanup
trap even when conformance fails.
Every run forces the relevant outer-UI prohibition and emits one minimized
machine-control-common-desktop-conformance/v0 result.
Dependency-free tests also cover native common projections. Android, iOS, and
Quest expose device-shaped target status, target doctor, and
target capabilities without claiming a desktop/resident process. ChromeOS
exposes a desktop-shaped native doctor with current-boot SSH persistence,
profile lock, and an empty common lifecycle set. A lifecycle mutation is
dispatched only when that doctor declares the exact operation.
The fake adapter also covers the explicitly iOS common family: platform
gating, allowlisted request construction, standard-input transport that keeps
fill text out of adapter arguments, strict machine-control/v0 result
validation, and logical-target/client metadata. Provider-native passthrough
remains behind testbed --; Android and Quest do not inherit iOS operations.
The fake VM adapter covers the additive workspace surface on macOS, Linux, and Windows: strict capabilities, configured and explicit intent, opaque selection of later adapter calls, normalized mechanisms, typed refusal, private-field rejection, receipt-bound release, and dry-run-only garbage collection. Shared provider tests separately exercise private receipt modes/redaction and UTM persistent, disposable, and policy-gated full-copy behavior.
The same fixtures cover coordinator-neutral target-use claims: policy discovery, bounded attribution, acquisition/status/renewal/release, global claim selection, pre-dispatch refusal, and atomic workspace composition. Shared provider tests add exact-resource contention, fencing, expiry, private state permissions, adapter enforcement, and claim-checked workspace cleanup.
test_scoped_run.py exercises run with a stateful short-lease adapter:
inherited selection, task exit status, renewal, exact workspace release,
refused/malformed management responses, and interrupted descendant cleanup.
It checks held-resource files and delayed child effects independently of the
runner's audit output. These tests use no live targets. See
scoped tasks for the public workflow and its
process/lifecycle boundaries.
For a read-only live check, run python3 tests/client/live-scoped-run.py ALIAS.
It requires an available claim, performs doctor, runs two claimed status
operations across renewal, and independently verifies release and unchanged
power. It never starts the target or takes host focus. Native Windows process
cleanup is exercised when the portable suite runs on Windows; a Mac run does
not establish Windows acceptance.
The same fake adapter proves that target ensure-ready is a no-op when ready,
starts an off target only through its declared up operation, and refuses to
invent a running-target repair. A reported start failure still receives an
independent final doctor observation without exposing adapter diagnostics.
Candidate tests require a fresh running-ready observation and prove that only
a cleanly stopped second identity assertion is eligible for private promotion.
Linux additionally proves credential verification before shutdown and a matching
receipt afterward; either failure prevents eligibility. Ordinary candidate
validation remains usable during incomplete provisioning. The Linux platform
suite covers private receipt invalidation, explicit password-free profiles,
secret-file permissions, pinned-host refusal before secret consumption, and
both factory routes. These fixtures do not establish live SSH/guest acceptance.
The maintenance fixtures prove that capability discovery never invokes an
adapter, Windows/macOS/Linux audit and repair use the same common operation
names, and ChromeOS can expose runtime audit/repair while refusing image
certification before dispatch. Reboot remains repair-only and explicit,
certification selects an available exact-source platform composition, and
valid unhealthy JSON preserves a nonzero result. The common projection removes
private boot epochs, endpoints, staging, and platform-only detail.
ensure-ready only recommends the read-only audit; it still never performs
maintenance itself.