release: v0.4.14 — Android element extraction, rebuilt

[justrach]

uitree reported what was in the uiautomator XML. What an agent needs is what can be acted on, named the way it reads on screen.

Prompted by a comparison against CursorTouch/Android-MCP. kuri already had roughly twice the command surface (27 vs 14 — they have no list-apps, uninstall, clear, openurl, keyevent, current-activity, logcat, screen-info, getprop, dumpsys, raw touch, multi-point gesture or batch), but their element handling was better. That gap is what this closes.

Bugs the comparison turned up

• android tap --label never worked. The tool table advertised the flag and android tools printed it, but the dispatcher only read positional coordinates — every documented invocation failed with missing argument: x y.
• No cli.zig test had ever run. main.zig aggregates tests through an explicit _ = @import(...) block, and both the Android and iOS dispatchers were missing from it. Their test blocks compiled but never executed; I confirmed it by deliberately breaking an assertion and watching the build still report success. That is how a flag could be advertised and unimplemented at once, and it is why the suite goes 44 → 59 here.
• Attribute values were never XML-decoded. A row labelled "Network & internet" was reported as Network & internet, so --label "Network & internet" could not match what is on screen. Found by running against a real emulator rather than by unit test — and my first fix was incomplete, because names synthesized from descendant text bypassed it.
• Unlabelled elements were dropped outright. The filter wanted clickable, or text, or an id, or a content-desc — which discards a scrollable RecyclerView or an unlabelled Switch. There was no way to discover the list could be scrolled.

What's new

android state	foreground app + screen size + every actionable element, one call instead of three plus manual filtering
uitree --interactive	interactivity detection across focusable/long-clickable/checkable/scrollable/selected/password + a narrow widget-class allowlist — 72 elements → 14 on a Settings screen
descendant naming	a clickable row whose label lives in child TextViews is addressable at all; the parser now tracks nesting to do it
selectors	--id / --class / --desc / --index on find and tap, AND-ed; resource-ids accepted short or fully qualified
type --clear	replaces a field instead of appending to it
notifications	reads package/title/text; --open pulls the shade down
wait <ms>	runs without a device, so it can't fail a script with "no device attached"

State is reported, not just capability: *checked/*unchecked, *password, *focused, *selected, *scrollable. Whether a switch is on is not something a label can tell you.

Verified live

Headless Android 16 emulator, not just unit tests:

• state on Settings returned rows named from their children — Network & internet Mobile, Wi‑Fi, hotspot — which previously listed as nameless
• tap --label Messages moved focus from NexusLauncherActivity to ConversationListActivity
• type --clear: "bluetooth" → "wifi", not "bluetoothwifi"
• notifications parsed Keyboard layout set to English (US). Tap to change. — nested parens intact

59/59 kuri-mobile tests (was 44, 8 of which never executed), 476/476 in the main project unchanged, clean ReleaseFast build.

Not included

The annotated / set-of-marks screenshot (their use_annotation) needs a PNG decode-draw-encode path in Zig; kuri has /screenshot/som browser-side but nothing reusable here. --json on state/find/uitree is probably the highest-value follow-up for agent consumers.

macOS assets

This repo still has no APPLE_* secrets, so release.yml will take its signed=false branch. The macOS tarballs will be signed and notarized locally via the codedb-notary profile and uploaded over CI's, as for v0.4.13.

🤖 Generated with Claude Code

https://claude.ai/code/session_017yzwAh6c95Ygb4LYttdVCs