Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Adding extension-state checks for smoke or bazzite suites
Debugging AT-SPI accessibility node failures in headless QEMU
When NOT to Use
SSH-based system checks → docs/skills/test-authoring/behave/SKILL.md shared SSH steps
CI workflow or runner container setup → docs/skills/ci-ops/ops/SKILL.md
Suite scaffolding or step hygiene → docs/skills/test-authoring/behave/SKILL.md
Core Process
Identify whether the scenario needs AT-SPI interaction, Shell.Eval, or only subprocess/CLI checks.
Reuse existing smoke helpers first (launch_background(), _run_host(), _eval_bool(), _wait_eval_bool()).
Prefer desktop-file launch targets before direct commands for GUI apps so D-Bus activation and AT-SPI registration work in CI.
Poll for visible widgets or windows; avoid unconditional sleeps when a retry loop can prove readiness.
Validate locally with python3 -m py_compile tests/<suite>/features/steps/*.py, duplicate-step detection, ruff, and behave --dry-run.
tests.shared.wait_for_shell.wait_for_shell() is the GNOME Shell startup gate. Its contract is: retry Shell.Eval failures, retry when AT-SPI exposes no panel yet, retry on transient exceptions, retry when the session bus socket disappears during a GDM restart (re-resolving the bus address each attempt), require consecutive stable checks, then fail with a per-error-class breakdown once the wall-clock budget is exhausted.
Stack
Layer
Component
Install
BDD runner
behave
pip
Session bridge
qecore-headless
pip
GUI automation
dogtail (AT-SPI)
pip
Wayland coord bridge
gnome-ponytail-daemon
sudo dnf install gnome-ponytail-daemon inside VM
Shell bridge
org.gnome.Shell.Eval
built-in (requires unsafe_mode=true)
Screen Lock/Unlock D-Bus calls (GNOME 50)
In GNOME 50, Main.screenShield.lock(true) via Shell.Eval is deprecated and fails. Use the stable D-Bus interface org.gnome.ScreenSaver.Lock to lock the session, and org.gnome.ScreenSaver.SetActive false to unlock it:
Commands that access the GNOME user session, including gsettings, gdbus --session, and Mutter DisplayConfig helpers, must source the session
environment on the VM before running:
source /tmp/session.env 2>/dev/null; gsettings get org.gnome.mutter experimental-features
The SSH connection itself does not inherit DBUS_SESSION_BUS_ADDRESS or
WAYLAND_DISPLAY; without this prefix, remote session calls can target no bus
or the wrong user session and produce misleading test failures.
GNOME Shell extensions and AT-SPI health in smoke
Use the public org.gnome.Shell.Extensions.GetExtensionInfo D-Bus method to
assert an extension is enabled (state1). If the scenario promises visible
product behavior, enabled state is only the first diagnostic and must not
replace a rendering assertion. For Dash-to-Dock v106, recursively traverse the
public Clutter actor tree for its source-defined dashtodockContainer name and
require the actor to be mapped, visible, allocated, opaque, and slid open. Do
not inspect the extension's private stateObj or dockManager object graph.
Bluefin's welcome modal is not GNOME Initial Setup. Poll for its visible Skip
button through AT-SPI after the sandbox is ready and click it. Do not create a
system-wide gnome-initial-setup-done marker or kill GNOME first-run processes;
those do not target the Bluefin-specific dialog.
For AT-SPI health, ask org.a11y.Bus.GetAddress through the smoke suite's
_run_host() helper after sourcing /tmp/session.env. A bare subprocess (or
pgrep) can inspect the Fedora runner container rather than the VM GNOME
session and therefore does not prove the accessibility bus is usable.
Overview search entry
Do not call Main.overview._onSearchChanged() — it was removed in GNOME 47.
Use clutter_text.set_text() which emits the text-changed signal and
triggers the search controller via the public signal path:
_shell_eval('Main.overview.searchEntry.clutter_text.get_text()')
# returns: (true, 'Files') — parse with regex on the second element
Activities overview (GNOME 50 QEMU)
Main.overview.visible.toString() consistently returns false in QEMU on GNOME 50
even after Main.overview.show() is called. Do NOT assert Main.overview.visible or
switch to Main.overview._shown without confirming on a live GNOME 50 QEMU run —
the behavior is not reproducible locally without a full VM boot. Scenarios that depend
on overview visibility must be quarantined (@quarantine) until the correct GNOME 50
API is confirmed.
Screenshot on failure
Hook in after_scenario, before sandbox cleanup:
from tests.shared.screenshot import take_screenshot
defafter_scenario(context, scenario):
if scenario.status == "failed":
take_screenshot("failed", context)
take_screenshot() calls the native org.gnome.Shell.Screenshot D-Bus API.
Do not call context.sandbox.shell.eval_js(...) for screenshots — in qecore
4.16 sandbox.shell is an accessibility object and has no eval_js method.
GNOME Extensions CLI (subprocess)
Smoke-suite extension steps run inside the VM via subprocess, not AT-SPI:
import subprocess
# List installed extensions
result = subprocess.run(["gnome-extensions", "list"], capture_output=True, text=True)
extensions = [e.strip() for e in result.stdout.splitlines() if e.strip()]
# List enabled extensions
result = subprocess.run(["gnome-extensions", "list", "--enabled"], capture_output=True, text=True)
enabled = [e.strip() for e in result.stdout.splitlines() if e.strip()]
Note: gnome-extensions requires the GNOME session to be running. These steps run inside the qecore VM (local subprocess), not over SSH.
Extension state via D-Bus (bazzite / GNOME 50)
For suites that need to poll an extension's activation state (e.g. the bazzite suite which runs over SSH), do not use Shell.Eval + Main.extensionManager.lookup(uuid)?.state. On GNOME 50 this API consistently returns state=6 (INITIALIZED) regardless of actual activation.
Use org.gnome.Shell.Extensions.GetExtensionInfo instead:
import subprocess, re
def_extension_state(uuid: str) -> str:
"""Return extension state as a string integer. 99 = unknown / uninstalled."""
result = subprocess.run(
['gdbus', 'call', '--session',
'--dest', 'org.gnome.Shell',
'--object-path', '/org/gnome/Shell/Extensions',
'--method', 'org.gnome.Shell.Extensions.GetExtensionInfo',
f"'{uuid}'"], # ← single-quotes required; see GVariant note below
capture_output=True, text=True, timeout=10,
)
if result.returncode != 0:
return"99"
m = re.search(r"'state':\s*<uint32\s+(\d+)>", result.stdout)
return m.group(1) if m else"99"
GVariant quoting (critical): Extension UUIDs contain @ and . which are invalid in a bare GVariant token. Always wrap the UUID in single quotes inside the Python string: f"'{uuid}'" → produces 'logomenu@aryan_k' on the command line.
Poll through 6 and 8 with a deadline (Bazzite: use 90s — 11 extensions need time post-boot).
Extension state in smoke suite (local subprocess / SSH bridge)
The smoke suite's UUID-specific extension checks use the same
GetExtensionInfo D-Bus call as bazzite, but route through the suite-local
_run_host(...) helper so they work both inside the VM and from the Fedora
runner container over SSH.
Bluefin's 9 bundled extensions each get a named scenario in
tests/smoke/features/bluefin_extensions.feature. Tag the
search-light@icedman.github.com scenario with @bluefin so dakota smoke runs
skip it via environment.py.
Use the distinct step phrase GNOME extension "{uuid}" is enabled (not
bazzite's Extension "{uuid}" is enabled) to avoid cross-suite step collisions.
Bazaar on Bluefin: wait out the Refreshing spinner
Bluefin ships Bazaar (io.github.kolunmi.Bazaar), not GNOME Software's old
Explore/Installed toggle-button layout. For Bazaar UI tests:
wait for a visible window named Bazaar
then poll until any visible tab named Curated, Explore,
Library, or Search appears
accept both page tab and toggle button roles for those tabs
The first launch often shows a Refreshing spinner page before the
AdwViewStack content is ready. On GNOME 50, AT-SPI cache drops can also make
nodes disappear mid-query, so wrap Bazaar window/tab lookups in retry loops
with short sleeps and re-query the tree each attempt.
Desktop notifications via gdbus (smoke suite)
Send a test notification from inside the VM:
gdbus call --session \
--dest org.freedesktop.Notifications \
--object-path /org/freedesktop/Notifications \
--method org.freedesktop.Notifications.Notify \
'' 0 '''Title''Body''[]''{}' 3000
# Returns: (uint32 N,) — N is the notification ID (>0 on success)
Parse the ID from context.notify_output with re.search(r'\(uint32 (\d+),\)', output). An ID of 0 means failure.
Smoke desktop-identity checks: use _run_host + session env
For smoke steps that need session-scoped shell state (XDG_SESSION_TYPE,
DISPLAY, WAYLAND_DISPLAY) or VM-installed tools like glxinfo, prefer the
suite-local _run_host(...) helper over plain subprocess.run(...).
Why: local smoke scenarios execute inside the VM during ad-hoc runs, but CI can
run them from the Fedora runner container. _run_host(...) transparently hops
to the VM over SSH in that case, and source /tmp/session.env 2>/dev/null; ...
preserves the GNOME user-session environment before probing Wayland or renderer
state.
Unit-testing smoke step modules
Smoke step modules drive AT-SPI, dogtail and live GNOME state, so most of their
surface is not unit-testable. What is testable is the pure logic they wrap:
command construction, output parsing, polling loops and assertion branches.
Import them in tests/unit/ with behave, qecore, dogtail and
app_support stubbed via sys.modules, then patch the shell helper
(_run_host, _run_in_vm) with unittest.mock.patch.object.
Notes for the three a11y/input/XWayland modules:
orca_steps.py — wraps _run_host from steps.steps (the smoke steps
directory is only importable during a behave run, so unit tests must register
a steps package stub with a __path__ before importing). Unit-testable:
the Run command on VM context bookkeeping, return-code/substring assertion
messages, the gsettings set …screen-reader-enabled command string,
_orca_is_running() (rc and non-empty stdout), _wait_for_orca() polling
and its start/stop timeout wording, and the toggle step's guarantee that the
screen-reader key is restored to false even when the start assertion fails.
Not unit-testable: whether Orca actually starts.
input_methods_steps.py — _run_in_vm() always prefixes
source /tmp/session.env 2>/dev/null; and dispatches to _ssh_run when
_IN_CONTAINER, else subprocess.run(shell=True). _restore_input_sources()
is idempotent via a _restored flag so the explicit restore step and the
registered context.add_cleanup do not double-apply; the flag is latched only
when every gsettings set returned 0, so a failed restore raises and the
cleanup hook can retry instead of leaking state. Saved gsettings values
contain single quotes and are re-applied through shlex.quote. Not
unit-testable: whether IBus owns the bus name or a layout actually switches.
xwayland_steps.py — _xwayland_display_env() parses pgrep -a -x Xwayland
output: it takes the first line only, reads -auth <file> into XAUTHORITY
(omitted when absent or dangling), and picks the first :<digits> token as
DISPLAY, defaulting to :0. Not unit-testable: xprop -root against a real
X root window, or glxgears rendering.
Per-app accessibility launch environment
Session-wide gsettings set org.gnome.desktop.interface toolkit-accessibility true
(set by e2e.yml) only enables the GTK atk-bridge. Applications that render
their own chrome build their AT-SPI tree from their own environment and must be
launched with explicit accessibility variables:
launch_background(targets, env=...) in tests/smoke/features/steps/app_support.py
applies env on every launch path: exported before the command in SSH mode,
merged into os.environ for local Popen, and forwarded across the Flatpak
sandbox boundary with --env=. Environment set outside a Flatpak sandbox is
not visible inside it — always use the env= parameter, never a shell prefix.
An exported Flatpak desktop entry is a trap: an app ID under
/var/lib/flatpak/exports/share/applications/ looks like an ordinary desktop
target, but gio launch / gtk-launch starts it inside the sandbox and silently
drops the launch environment. launch_background() detects the export path and
reroutes it through flatpak run --env=. Keep an explicit
("flatpak", app_id) target before the exported desktop entry when both are
listed, so the intended launch path is clear at the call site.
Symptom when this is missing: the application appears in the AT-SPI tree but its
window node has no descendants — no entry, tool bar, or page tab list.
Steps then fail late with confusing messages such as "address bar not found".
Window-role checks must prove the subtree is populated
Since GNOME 50 some apps expose their toplevel as filler rather than frame,
so smoke helpers accept {"frame", "filler"}. Accepting a bare filler node on
its own is a false pass: it is exactly what an app exposes when its
accessibility engine never started. Require evidence of a real subtree:
def_has_populated_a11y_tree(node) -> bool:
returnbool(node.findChildren(lambda n: n.roleName in CHROME_ROLES))
Prefer a frame with a populated subtree, fall back to a populated filler, and
otherwise fail with a message that names the likely cause
(GNOME_ACCESSIBILITY / toolkit-accessibility). Keep a
require_a11y_tree=False escape hatch for pure liveness checks such as
"the app is no longer running".
Sleep discipline in step definitions
Unconditional sleep(N) calls inflate suite time — avoid them. Rules:
launch_background() from qecore — do NOT add sleep(1) after calling qecore's built-in launch_background(). The immediately-following "window is accessible" step has its own AT-SPI polling loop; the launch sleep is redundant there.
EXCEPTION — _launch_app() custom launcher: The suite-local _launch_app() in gnome_apps_steps.py uses D-Bus app activation (gio open or gtk-launch), which is asynchronous. sleep(1) after a successful _launch_app() return is required as a D-Bus activation settle time — without it, _wait_for_window() starts polling before the process has registered with the AT-SPI accessibility bus, exhausts all retries (~10s), and fails on slower images (e.g. Dakota testing). Do not remove this sleep. The regression in PR #465 was caused by removing it.
Polling loop intervals — use 0.2s intervals in retry loops (for _ in range(N): sleep(0.2)). 0.5s is the old default; the loops already exit-early on success so tighter intervals help.
GNOME Shell open/close animations — use 0.2s after _shell_eval() open/close commands before checking state. The _wait_eval_bool() helper handles the real confirmation wait.
Screenshot fastfetch — terminal keep-open is fastfetch; sleep 3 (not 10). Pre-screenshot delay time.sleep(2) (not 4). Both are already on QEMU where timing is slow.
Never remove — small sleeps after user-visible actions (sidebar clicks, key combos, focus transitions) that have no async poll to catch up: sleep(0.2) is the minimum. Do not go below 0.1s.
The pattern for _ in range(N): ... sleep(X) that returns early already IS exit-early. The gains come from removing the PRECEDING unconditional sleep, not from changing the loop.
Red Flags
Using 'true' in out to check a Shell.Eval result (success_bool is always true)
Calling Shell.Eval without first setting global.context.unsafe_mode = true
Using requireResult=False with findChild (removed in dogtail 4.16)
Importing tests.shared.ssh_steps into the smoke suite
Using AT-SPI coordinate clicks on the top-bar (unreliable on GNOME 50+)
Polling Main.extensionManager.lookup(uuid)?.state via Shell.Eval (returns 6 on GNOME 50)
Using the SSH-based _extension_state() pattern in the smoke suite (smoke uses local subprocess)
_<app>_app() helper does a single-pass lookup with no retry loop — will flake on GNOME 50 QEMU
A _<app>_window() helper accepts roleName "filler" without checking that the node has descendants (false pass)
A GUI app that renders its own chrome is launched without GNOME_ACCESSIBILITY=1
A Flatpak-exported .desktop entry is launched with gio launch/gtk-launch
while carrying an env= payload (the env never reaches the sandbox)
A session-readiness loop caches the D-Bus session address or treats a missing bus socket as fatal (breaks across the qecore-headless GDM restart)
Verification
Shell.Eval results extracted with _eval_bool() / regex on second tuple element, never 'true' in out
unsafe_mode=true set before any Shell.Eval that reads protected state
Extension state checked via org.gnome.Shell.Extensions.GetExtensionInfo, not Shell.Eval
UUID wrapped in single quotes for GVariant: f"'{uuid}'" not uuid
Smoke suite steps use subprocess.run, not SSH helpers
behave --dry-run tests/smoke/features/ passes before pushing
Window-role helpers that accept filler also assert the node has a populated subtree
Launch targets for Flatpak-packaged apps resolve to flatpak run --env=, not an exported desktop entry
Common Rationalizations
"A direct command launch is simpler." → For GUI apps, desktop-file activation is usually more reliable for AT-SPI registration.
"I'll just sleep after launch." → Poll for the visible window instead; fixed sleeps bloat the suite and still flake.
"This title match is good enough." → Prefer app-level AT-SPI lookup first, then use title fallback only when the app name is unstable.
Red Flags
New smoke app steps hardcode /usr/share/applications/... for Flatpak-only apps
Step code uses findChild(..., requireResult=...)
New GNOME steps duplicate existing step phrases in the suite
New launch steps add unconditional post-launch sleeps instead of relying on accessibility polling
Verification
Reused existing GNOME/smoke helpers before adding new ones
Launch targets prefer desktop files, with Flatpak or command fallback only when needed
AT-SPI polling or Shell.Eval assertions replace fixed waits where possible
behave --dry-run tests/<suite>/features/ passes for the touched suite
Session readiness across a GDM restart
qecore-headless restarts GDM, which destroys the session D-Bus socket and
brings up a fresh autologin session. tests/shared/wait_for_shell.py is the
canonical readiness helper and encodes the resulting contract:
ServiceUnknown (bus up, org.gnome.Shell unowned) and
Could not connect: No such file or directory (socket gone, GDM restarting)
are both retryable, never terminal.
The session bus address is re-resolved on every attempt
(resolve_session_bus_env()); an address or connection cached before the
restart points at a destroyed socket and can never recover. The address is
never unset — an empty DBUS_SESSION_BUS_ADDRESS sends gdbus down the
dbus-launch --autolaunch path instead of the real session socket.
Readiness must hold for two consecutive checks so a check does not latch onto
the outgoing session moments before GDM tears it down.
The loop is bounded by a 300s wall-clock deadline, and the timeout message
reports a per-error-class attempt breakdown plus the last error.
When the socket file is absent the probe short-circuits instead of spawning
gdbus, because an unreachable/empty address sends GIO down the
dbus-launch --autolaunch path, which cannot work in the test container.
collect_session_diagnostics() snapshots socket presence, loginctl list-sessions and systemctl status gdm on the first failure, every 15th
failure, and at timeout. If the socket never returns and no user session is
listed, the fault is lane-side GDM provisioning, not this helper.
Reuse this helper rather than writing a new gdbus-poll loop. See
docs/skills/ci-ops/ops/references/qecore-headless-restarts-gdm-bus-socket-churn.md.
