ci(android): make the emulator smoke survive a slow emulator, not app failures (#68)

Evidence from runs 39866280662, 39870244857 and 40091145294 shows the red
smoke runs were the emulator, not TGLock: on the software GPU the WebView
paints about one frame per second while the page re-renders every second,
so a tap's down/up can straddle a re-render and Chromium drops the click
(no setEnabled call at all in the failed Stop run). Play services and Pixel
Launcher post-boot work also starved system_server into foreign ANR dialogs.

- AOSP image, 4 cores, 4 GB RAM, no snapshot save; wait for CPU to settle
  after boot before launching.
- Answer foreign "isn't responding" dialogs with Wait and record them; a
  TGLock ANR/error dialog, crash or missing process still fails.
- Re-tap only when the button is unchanged 10s after a tap (at most 3 taps);
  foreground-service and SOCKS listener effects stay strictly asserted with
  a 45s budget sized for this emulator.
- Timed step markers in stdout and logcat, failure screenshot and window
  dump, one evidence artifact per run attempt.

Co-authored-by: babin <Thartewerner536e@engineer.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

.github/workflows/android.yml

@@ -79,7 +79,7 @@ jobs:
     name: Android 15 emulator smoke
     needs: apk
     runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 25
     steps:
       - uses: actions/checkout@v6
       - uses: actions/download-artifact@v4
@@ -91,20 +91,25 @@ jobs:
           echo 'KERNEL=="kvm", GROUP="kvm", MODE="0666", OPTIONS+="static_node=kvm"' | sudo tee /etc/udev/rules.d/99-kvm4all.rules
           sudo udevadm control --reload-rules
           sudo udevadm trigger --name-match=kvm
+      # The AOSP image has no Play services or Pixel Launcher. Their post-boot
+      # work starved system_server on the 4-vCPU runner and produced foreign ANR
+      # dialogs over the app. The app needs neither; WebView is still present.
       - name: Launch and exercise the native app
         uses: reactivecircus/android-emulator-runner@v2
         with:
           api-level: 35
           arch: x86_64
-          target: google_apis
+          target: default
           profile: pixel_2
+          cores: 4
+          ram-size: 4096M
           disable-animations: true
-          emulator-options: -no-window -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none
+          emulator-options: -no-window -no-snapshot -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none
-          script: timeout 180 python3 scripts/android_smoke.py emulator-apk
+          script: timeout 420 python3 scripts/android_smoke.py emulator-apk
       - uses: actions/upload-artifact@v4
         if: always()
         with:
-          name: android-emulator-smoke-evidence
+          name: android-emulator-smoke-evidence-attempt-${{ github.run_attempt }}
           path: android-smoke-evidence/
           if-no-files-found: warn
           retention-days: 14

docs/ANDROID.md

@@ -57,8 +57,13 @@ restart, and explicit force-stop/relaunch. It verifies the foreground service an
 performs a real SOCKS5 greeting through ADB port forwarding to the Rust listener.
 If buttons are inaccessible after a bounded wait, CI explicitly reports the
 lifecycle checks as skipped; a launch-only pass is not lifecycle evidence.
-`android-emulator-smoke-evidence` retains the exact result, UI dumps, service
+`android-emulator-smoke-evidence-attempt-N` (one per run attempt) retains the
-state and logcat. This does not test automatic low-memory eviction, battery
+exact result, timed steps, UI dumps, service state, logcat and, on failure, a
+screenshot. The emulator is the AOSP image without Play services; foreign
+system ANR dialogs are answered with Wait and recorded, while a crash, ANR or
+error dialog of TGLock itself fails the job. A tap the WebView dropped (the
+button is unchanged ten seconds later) is repeated at most twice; the service
+and listener effects are then asserted with the same strictness. This does not test automatic low-memory eviction, battery
 behavior, Telegram connectivity, or a physical phone.
 
 Checked-in `gen/android` contains the native source and Gradle wrapper. Tauri's

scripts/android_smoke.py

@@ -14,23 +14,103 @@ PACKAGE = "com.bysonic.tglock"
 COMPONENT = f"{PACKAGE}/.MainActivity"
 EVIDENCE = Path("android-smoke-evidence")
 EVIDENCE.mkdir(exist_ok=True)
-RESULT = {"launch": "not_run", "lifecycle": "not_run"}
+RESULT = {"launch": "not_run", "lifecycle": "not_run", "foreign_dialogs": [], "taps": {}}
+STARTED = time.monotonic()
+
+# Budgets measured on the GitHub-hosted emulator, not on a phone. With the
+# software GPU the WebView paints about one frame per second, and dumpsys
+# itself can take seconds while system_server catches up after boot.
+SETTLE_SECONDS = 120
+VISIBLE_SECONDS = 60
+EFFECT_SECONDS = 45
+# A click that reaches the handler re-renders the button as "Подождите…" or
+# the opposite action within a frame. Only a button left unchanged this long
+# was never clicked.
+RETAP_AFTER_SECONDS = 10
+MAX_TAPS = 3
+SYSTEM_DIALOG = re.compile(r"(Application Not Responding|Application Error): ([\w.:]+)")
 
 
-def adb(*args, check=True, timeout=20):
+def adb(*args, check=True, timeout=30):
     return subprocess.run(
         ["adb", *args], check=check, capture_output=True, text=True,
         encoding="utf-8", errors="replace", timeout=timeout,
     ).stdout.strip()
 
 
-def wait_for(description, predicate, seconds=20):
+def step(message):
+    # The same marker in stdout and logcat lines up the job log with evidence.
+    line = f"[{time.monotonic() - STARTED:6.1f}s] {message}"
+    print(line, flush=True)
+    with open(EVIDENCE / "steps.txt", "a", encoding="utf-8") as steps:
+        steps.write(line + "\n")
+    adb("shell", "log", "-t", "tglock-smoke", json.dumps(message), check=False)
+
+
+def system_dialog():
+    try:
+        windows = adb("shell", "dumpsys", "window", "windows", check=False)
+    except subprocess.SubprocessError:
+        return None
+    match = SYSTEM_DIALOG.search(windows)
+    return (match.group(1), match.group(2)) if match else None
+
+
+def clear_foreign_dialogs():
+    # A loaded emulator regularly declares its own launcher or system apps
+    # unresponsive, and that dialog covers the WebView. Ours is a failure.
+    dialog = system_dialog()
+    if dialog is None:
+        return
+    kind, process = dialog
+    if process.split(":")[0] == PACKAGE:
+        raise AssertionError(f"{kind} dialog for {process}")
+    button = "android:id/aerr_wait" if kind == "Application Not Responding" else "android:id/aerr_close"
+    try:
+        point = node_center(ui_dump(f"dialog-{len(RESULT['foreign_dialogs'])}"),
+                            lambda node: node.get("resource-id") == button)
+    except (subprocess.SubprocessError, ET.ParseError):
+        point = None
+    if point is not None:
+        adb("shell", "input", "tap", str(point[0]), str(point[1]))
+    outcome = f"tapped {button}" if point else "button not dumped yet"
+    RESULT["foreign_dialogs"].append(f"{kind}: {process}, {outcome}")
+    step(f"foreign dialog {kind}: {process}, {outcome}")
+
+
+def wait_for(description, predicate, seconds=EFFECT_SECONDS):
     deadline = time.monotonic() + seconds
     while time.monotonic() < deadline:
+        clear_foreign_dialogs()
         if predicate():
+            step(f"ok: {description}")
             return
         time.sleep(1)
-    raise AssertionError(f"Timed out: {description}")
+    raise AssertionError(f"Timed out after {seconds}s: {description}")
+
+
+def cpu_idle_share(interval=3):
+    def sample():
+        values = list(map(int, adb("shell", "head", "-1", "/proc/stat").split()[1:]))
+        return values[3] + values[4], sum(values)
+    idle1, total1 = sample()
+    time.sleep(interval)
+    idle2, total2 = sample()
+    return (idle2 - idle1) / max(1, total2 - total1)
+
+
+def settle():
+    # boot_completed fires long before package, launcher and dexopt work
+    # drains. Start the app on a quiet device instead of racing that work.
+    deadline = time.monotonic() + SETTLE_SECONDS
+    quiet = 0
+    idle = 0.0
+    while time.monotonic() < deadline and quiet < 2:
+        clear_foreign_dialogs()
+        idle = cpu_idle_share()
+        quiet = quiet + 1 if idle >= 0.6 else 0
+    RESULT["settle"] = f"{'quiet' if quiet >= 2 else 'still busy'} at {time.monotonic() - STARTED:.0f}s, {idle:.0%} idle"
+    step(f"settle: {RESULT['settle']}")
 
 
 def ui_dump(label):
@@ -47,12 +127,9 @@ def ui_dump(label):
     return ET.fromstring(text)
 
 
-def label_node(tree, label):
+def node_center(tree, matches):
     for node in tree.iter("node"):
-        # Android aggregates the decorative arrow into the button's accessible
+        if matches(node):
-        # text (observed in the API35 CI dump). Accept only that exact suffix.
-        names = (node.get("text", "").strip(), node.get("content-desc", "").strip())
-        if any(name in (label, f"{label} →") for name in names):
             bounds = re.fullmatch(r"\[(\d+),(\d+)\]\[(\d+),(\d+)\]", node.get("bounds", ""))
             if bounds:
                 x1, y1, x2, y2 = map(int, bounds.groups())
@@ -61,17 +138,53 @@ def label_node(tree, label):
     return None
 
 
-def tap_label(label, stage):
+def label_node(tree, label):
-    for attempt in range(5):
+    # Android aggregates the decorative arrow into the button's accessible
-        try:
+    # text (observed in the API35 CI dump). Accept only that exact suffix.
-            point = label_node(ui_dump(f"{stage}-{attempt}"), label)
+    return node_center(tree, lambda node: any(
-        except (subprocess.SubprocessError, ET.ParseError) as error:
+        name in (label, f"{label} →")
-            (EVIDENCE / f"{stage}-{attempt}.error.txt").write_text(str(error), encoding="utf-8")
+        for name in (node.get("text", "").strip(), node.get("content-desc", "").strip())
-            point = None
+    ))
-        if point is not None:
+
-            adb("shell", "input", "tap", str(point[0]), str(point[1]))
+
+def find_label(label, stage):
+    clear_foreign_dialogs()
+    try:
+        return label_node(ui_dump(stage), label)
+    except (subprocess.SubprocessError, ET.ParseError) as error:
+        (EVIDENCE / f"{stage}.error.txt").write_text(str(error), encoding="utf-8")
+        return None
+
+
+def press(label, stage):
+    """Tap a visible WebView action until the page shows it was clicked.
+
+    The page re-renders every second. At one frame per second the touch down
+    and up can straddle that re-render, and Chromium then drops the click: the
+    button stays exactly as it was and the handler never runs. Callers still
+    assert the service and listener effect strictly afterwards.
+    """
+    deadline = time.monotonic() + VISIBLE_SECONDS
+    taps = 0
+    last_tap = 0.0
+    attempt = 0
+    while time.monotonic() < deadline:
+        point = find_label(label, f"{stage}-{attempt}")
+        attempt += 1
+        if point is None and taps:
+            RESULT["taps"][stage] = taps
+            step(f"{label}: clicked after {taps} tap(s)")
             return
+        if point is not None and (not taps or time.monotonic() - last_tap >= RETAP_AFTER_SECONDS):
+            if taps == MAX_TAPS:
+                raise AssertionError(f"{label} stayed unchanged after {taps} taps")
+            adb("shell", "input", "tap", str(point[0]), str(point[1]))
+            taps += 1
+            last_tap = time.monotonic()
+            step(f"{label}: tap {taps} at {point}")
         time.sleep(1)
+    if taps:
+        raise AssertionError(f"{label} stayed unchanged {VISIBLE_SECONDS}s after {taps} tap(s)")
     raise AssertionError(f"Visible action not found: {label}")
 
 
@@ -85,7 +198,7 @@ def proxy_ready():
     # ADB forwards to emulator loopback. A real SOCKS5 greeting proves the
     # Rust listener is serving, beyond just a notification being displayed.
     try:
-        with socket.create_connection(("127.0.0.1", 11080), timeout=1) as peer:
+        with socket.create_connection(("127.0.0.1", 11080), timeout=3) as peer:
             peer.sendall(bytes([5, 1, 0]))
             return peer.recv(2) == bytes([5, 0])
     except (OSError, TimeoutError):
@@ -93,7 +206,7 @@ def proxy_ready():
 
 
 def launch():
-    adb("shell", "am", "start", "-W", "-n", COMPONENT)
+    adb("shell", "am", "start", "-W", "-n", COMPONENT, timeout=60)
     wait_for("Activity resumed", lambda: any(
         "ResumedActivity" in line and PACKAGE in line
         for line in adb("shell", "dumpsys", "activity", "activities").splitlines()
@@ -104,17 +217,20 @@ def launch():
 def main():
     apks = sorted(Path(sys.argv[1]).rglob("*.apk"))
     assert len(apks) == 1, f"Expected one x86_64 APK, got {len(apks)}"
-    adb("install", "-r", str(apks[0]), timeout=60)
+    adb("install", "-r", str(apks[0]), timeout=120)
     adb("shell", "pm", "grant", PACKAGE, "android.permission.POST_NOTIFICATIONS")
-    adb("logcat", "-c")
     adb("forward", "tcp:11080", "tcp:1080")
+    settle()
+    adb("logcat", "-c")
+    step("launch")
     launch()
     RESULT["launch"] = "passed"
-    deadline = time.monotonic() + 20
+    deadline = time.monotonic() + VISIBLE_SECONDS
     tree = None
     start_point = None
     attempt = 0
     while time.monotonic() < deadline:
+        clear_foreign_dialogs()
         try:
             tree = ui_dump(f"launched-{attempt}")
             start_point = label_node(tree, "Включить защиту")
@@ -132,23 +248,25 @@ def main():
             for node in tree.iter("node")
         ), "App Activity is resumed but its WebView is absent"
     if start_point is None:
-        RESULT["lifecycle"] = "skipped: WebView Start not accessible after 20s"
+        RESULT["lifecycle"] = f"skipped: WebView Start not accessible after {VISIBLE_SECONDS}s"
-        print("::warning::Activity launch passed; lifecycle skipped because UIAutomator did not expose Start after 20s")
+        print(f"::warning::Activity launch passed; lifecycle skipped because UIAutomator did not expose Start after {VISIBLE_SECONDS}s")
         return
     RESULT["lifecycle"] = "failed: lifecycle assertions incomplete"
-    tap_label("Включить защиту", "before-start")
+    press("Включить защиту", "before-start")
     wait_for("foreground service after Start", service_running)
     wait_for("Rust SOCKS listener after Start", proxy_ready)
     adb("shell", "input", "keyevent", "KEYCODE_HOME")
     time.sleep(5)
+    clear_foreign_dialogs()
     assert service_running() and proxy_ready(), "Proxy stopped after backgrounding"
+    step("ok: still serving after 5s in background")
     launch()
-    tap_label("Выключить", "before-stop")
+    press("Выключить", "before-stop")
     wait_for("foreground service after Stop", lambda: not service_running())
     wait_for("Rust listener after Stop", lambda: not proxy_ready())
     # Explicit restart and user force-stop, then relaunch. This tests the
     # user-stop contract, not Android's automatic low-memory process eviction.
-    tap_label("Включить защиту", "before-restart")
+    press("Включить защиту", "before-restart")
     wait_for("Rust listener after restart", proxy_ready)
     adb("shell", "am", "force-stop", PACKAGE)
     launch()
@@ -160,10 +278,21 @@ try:
     main()
 except Exception as error:
     RESULT["failure"] = str(error)
+    for name, args in (("failure-windows.txt", ("dumpsys", "window", "windows")),
+                       ("failure-activities.txt", ("dumpsys", "activity", "activities"))):
+        try:
+            (EVIDENCE / name).write_text(adb("shell", *args, check=False), encoding="utf-8")
+        except subprocess.SubprocessError:
+            pass
+    try:
+        (EVIDENCE / "failure.png").write_bytes(subprocess.run(
+            ["adb", "exec-out", "screencap", "-p"], capture_output=True, timeout=30).stdout)
+    except subprocess.SubprocessError:
+        pass
     raise
 finally:
     try:
-        logs = adb("logcat", "-d", "-v", "threadtime")
+        logs = adb("logcat", "-d", "-v", "threadtime", timeout=60)
         (EVIDENCE / "logcat.txt").write_text(logs, encoding="utf-8")
         crashes = adb("logcat", "-b", "crash", "-d")
         (EVIDENCE / "crash.txt").write_text(crashes, encoding="utf-8")
@@ -171,9 +300,9 @@ finally:
             RESULT["launch"] = "failed: application crash, ANR, or missing process"
             raise AssertionError("Application did not remain healthy")
     finally:
-        (EVIDENCE / "result.json").write_text(json.dumps(RESULT, indent=2), encoding="utf-8")
+        (EVIDENCE / "result.json").write_text(json.dumps(RESULT, indent=2, ensure_ascii=False), encoding="utf-8")
         summary = os.environ.get("GITHUB_STEP_SUMMARY")
         if summary:
             with open(summary, "a", encoding="utf-8") as report:
-                report.write("\nAndroid emulator smoke: " + json.dumps(RESULT) + "\n")
+                report.write("\nAndroid emulator smoke: " + json.dumps(RESULT, ensure_ascii=False) + "\n")
-        print(json.dumps(RESULT))
+        print(json.dumps(RESULT, ensure_ascii=False))