Navigate App Adb

Skill: Navigate ApkAnalyzer via adb

Drives the running app by dumping the on-screen UI tree, locating an element, and tapping its center — the same loop repeated for each hop. Use this after the app is installed and running (see the run-app skill for that part).

The loop: dump → locate → tap → verify

1. Dump the current screen

adb shell uiautomator dump /sdcard/window_dump.xml
adb pull /sdcard/window_dump.xml .

(uiautomator dump captures every window, including Compose bottom sheets and dialogs — you don't need a separate step for those.)

2. Find your target node

Everything on screen is locatable through two standard AccessibilityNodeInfo attributes that show up directly in the dump — no special app wiring required for either one:

• text= — almost everything: chip labels, button text, list rows (an app row shows both its app name and package name), toolbar titles.
• content-desc= — the handful of icon-only controls that render no visible text at all (every IconButton and the recently-viewed Switch — see the reference below).

grep -o '<node[^>]*text="Filter"[^>]*>' window_dump.xml
grep -o '<node[^>]*content-desc="Sort"[^>]*>' window_dump.xml

Tapping the bounds of a text= node still triggers its parent's click handler — the tap only needs to land somewhere inside the clickable ancestor. Pull the bounds="[x1,y1][x2,y2]" attribute off whichever node you matched.

3. Tap it

adb shell input tap $(( (x1 + x2) / 2 )) $(( (y1 + y2) / 2 ))

For text fields, follow the tap with:

adb shell input text "search%squery"   # %s = literal space; adb input text can't take real spaces

To go back instead of tapping a back button node:

adb shell input keyevent 4

4. Verify

Either dump again (step 1) and confirm the expected text/description is present, or grab a screenshot for a visual check:

adb exec-out screencap -p > screenshot.png

Then use the Read tool on screenshot.png to look at it directly.

Gotchas

• On Windows, drive adb from PowerShell, not the Bash tool. Git Bash rewrites /sdcard/... into a Windows path, so uiautomator dump /sdcard/window_dump.xml writes to C:/Program Files/Git/sdcard/... and the following adb pull fails. Either use the PowerShell tool or prefix the command with MSYS_NO_PATHCONV=1.
• Never redirect adb exec-out screencap -p into a file from PowerShell. PowerShell re-encodes the stream and corrupts the PNG. Capture on the device and pull it: adb shell screencap -p /sdcard/screen.png then adb pull /sdcard/screen.png.
• Floating overlays steal taps. A Messenger chat head or similar bubble sits above the app and swallows a tap aimed at what's underneath — which can open someone's private conversation. Tap toward the left of a row rather than its centre, and re-dump after every tap to confirm you are where you expect before tapping again.
• A locked device answers adb but shows nothing. dumpsys window | grep mDreamingLockscreen reports true when the keyguard is up; a secure lock can't be dismissed from adb, so ask the user to unlock rather than retrying taps into a black screen.
• Off-screen elements don't dump. A list item below the fold won't appear until you scroll (adb shell input swipe <x1> <y1> <x2> <y2> <durationMs>) or it's within the initial viewport.
• Matches aren't always unique. Two rows can share a label (e.g. two apps with the same display name) — if a match gives more than one hit, narrow using a more specific string on the same row (an app's package name is usually unique) or the row's bounds to disambiguate. Localized strings also change with the device's language.
• A genuinely icon-only control needs a contentDescription, not a workaround. If you add a button with no visible label, give it one via IconButton's contentDescription parameter (a stringResource in the module that owns the screen — never a hardcoded literal) or, for a component that doesn't expose the parameter (e.g. Switch, or a Chip with an empty label), apply Modifier.semantics { contentDescription = "..." } directly at the call site. Don't reach for position-based taps — they break the moment layout shifts. Only do this for elements that render no text anywhere in their subtree; everything else is already findable by the text it already shows.