overthinkos/adb

ADB — Android Debug Bridge

Overview

ov eval adb <method> is the host-side ADB client. The host ov binary connects to the running container's host-published ADB server port (container :5037 → host's HOST_PORT:5037, e.g. 35002 on the eval-android-emulator-pod deploy) using github.com/zach-klippenstein/goadb and dispatches operations against the emulator backing that server.

Same architecture pattern as ov eval mcp: host-side protocol client, no container-side helper, works against any deploy that publishes the adb-server port — pod / vm / host / nested all work transparently because the connection is plain TCP to 127.0.0.1:<host-port> and the portforward / passt / etc. layer handles the rest.

Also as a declarative verb

Every ov eval adb <method> is authorable as an adb: verb inside a eval: block. The method name becomes the verb's YAML value; method- specific args are sibling fields (apk:, property:, args:, artifact:). Shared matchers (stdout:, stderr:, exit_status:, artifact_min_bytes:) work like other verbs. Deploy-scope only — needs a running container with a host-mapped ADB server port; the validator rejects build-scope use at ov image validate time.

Example:

- id: emulator-is-up
  scope: deploy
  adb: devices
  stdout:
    - contains: "emulator-5554"
- id: boot-done
  scope: deploy
  adb: getprop
  property: sys.boot_completed
  stdout:
    - contains: "1"
  timeout: 300s

See /ov-eval:eval for the full per-verb schema notes; this skill is the adb-specific method reference.

Quick Reference

| Subcommand | CLI form | Required modifier | Description | |---|---|---|---| | devices | ov eval adb devices <image> | — | List devices/emulators with state | | shell | ov eval adb shell <image> -- <cmd...> | args: (list) | Run a shell command on the emulator | | install | ov eval adb install <image> --apk <path> | apk: | Install an APK from the host filesystem | | uninstall | ov eval adb uninstall <image> <package> | args: [pkgid] | Remove a package by id | | getprop | ov eval adb getprop <image> <property> | property: | Read a system property | | screencap | ov eval adb screencap <image> --artifact <png> | artifact: | Capture a PNG screenshot to a host file | | logcat-tail | ov eval adb logcat-tail <image> [--lines N] [--filter F] | — | Dump recent logcat lines (uses logcat -d) | | wait-for-device | ov eval adb wait-for-device <image> [--timeout 60s] | — | Block until sys.boot_completed=1 | | wait-ui-settled | ov eval adb wait-ui-settled <image> [--timeout 600s] | — | Block until the foreground is not an ANR dialog (dismissing via HOME) | | current-focus | ov eval adb current-focus <image> | — | Print the foreground window line (mCurrentFocus) | | keyevent | ov eval adb keyevent <image> <key> | key: (arg) | Send a key event (input keyevent KEYCODE_…) |

--serial <serial> selects a specific device (default emulator-5554). -i <instance> addresses a <base>/<instance> pod deploy.

Method allowlist + required modifiers

| Method | Required modifiers | Notes | |---|---|---| | devices | — | Output is one line per device: <serial>\t<state> | | shell | Args | Args[0] is the program; Args[1:] are positional args. Prefix -- is auto-added so flags don't leak into the outer Kong invocation. | | install | Apk | Pushes the APK via the sync protocol to /data/local/tmp/, then pm install -r. Best-effort cleanup of the staged file. Asserts Success in pm output. | | uninstall | Args | Args[0] = package id (e.g. com.example.android.apis). Same Args:[0] convention as shell to keep the modifier surface flat. | | getprop | Property | Property key like sys.boot_completed, ro.build.version.release. | | screencap | Artifact | Runs screencap -p \| base64 on the device, decodes host-side, writes to Artifact. The base64 round-trip is mandatory — raw binary stdout gets mangled by goadb's shell stream on some emulator builds. Pairs with artifact_min_bytes: for size assertion. | | logcat-tail | — | Uses logcat -d (dump-and-exit). Amount (--lines) trims to the last N lines host-side; Query (--filter) is the literal filter spec (e.g. MyApp:I *:S). | | wait-for-device | — | Polls getprop sys.boot_completed every 2s up to Timeout (default 60s). Lighter than the wire-protocol wait-for-device because that returns the moment the device ATTACHES (well before sys.boot_completed). | | wait-ui-settled | — | Polls mCurrentFocus (dumpsys window); while a system "Application Not Responding" dialog holds focus it dismisses it with KEYCODE_HOME and keeps polling, up to Timeout. Prints settled on success. Set timeout: generously (the android-emulator gate uses 600s) — the shared 30s default is too short for a churning emulator. Pure goadb — no in-container shell. | | current-focus | — | Prints the mCurrentFocus window line. Assert the foreground app (stdout: contains: io.appium.android.apis) or detect a stuck ANR dialog. | | keyevent | KeyName | input keyevent <key> via key: KEYCODE_HOME / KEYCODE_BACK / a numeric code. Generic input building block; wait-ui-settled uses the same call internally to dismiss dialogs. |

Authoring gotchas

apk: must be a host-side path

The host ov binary reads the APK from its own filesystem and pushes via the sync protocol — apk: is a path on the host running ov, NOT a container-internal path. (Compare with appium: install-app where apk: is the in-container Appium server's view of the path because Appium reads the file itself.)

args: for shell commands; prefix -- is automatic

Always write shell args as a YAML list:

adb: shell
args: [pm, list, packages, io.appium.android.apis]

The -- prefix is automatic — you don't have to escape -l / -p / -fsS etc. flags. Don't reach for command: here; command: is a different verb.

args: [pkgid] for uninstall

Mirrors shell's convention to avoid adding a dedicated package: field. The validator enforces Args non-empty at config time.

Host-side screencap needs screencap -p | base64

A raw screencap -p produces binary PNG bytes that goadb's shell stream may mangle. The implementation always pipes through base64 and decodes host-side. This is invisible to authors — adb: screencap "just works" — but worth knowing when debugging.

wait-ui-settled — UI readiness past sys.boot_completed

A freshly-booted google_apis_playstore emulator runs minutes of GMS post-boot churn (Play Store auto-update, Chimera config, Heterodyne sync) that starves the GMS-coupled system UI (Pixel Launcher via AiAi, systemui). It ANRs, and the "Application Not Responding" dialog occludes whatever app is foreground — so a UI probe (appium find, monkey) fired right after sys.boot_completed silently fails. sys.boot_completed is necessary-but-not-sufficient readiness; wait-ui-settled is the sufficient half: it polls the focused window and dismisses any ANR dialog with KEYCODE_HOME until the foreground is clean. It is load-independent (the ANR is GMS churn, not host load) and adapts to a load-dilated churn purely via timeout:. Pure Go over goadb — no in-container shell, so it is immune to the stdin-heredoc hazard that breaks shell-based settle loops (see /ov-eval:eval "in-container command: stdin").

# Order matters — wait-for-device (boot) BEFORE wait-ui-settled (UI),
# wait-ui-settled BEFORE any UI-interacting check (appium / monkey).
- id: emulator-ui-settled
  scope: deploy
  adb: wait-ui-settled
  timeout: 600s
  stdout:
    - contains: settled

current-focus and keyevent are the building blocks wait-ui-settled composes — usable on their own to assert the foreground app or inject a key (adb: keyevent + key: KEYCODE_BACK).

--filter is a free-form logcat spec

--filter "MyApp:I *:S" passes MyApp:I *:S to logcat verbatim (each word becomes one positional arg). Don't quote individual words; the shell-fields split handles separation.

Long-running wait

wait-for-device blocks up to --timeout (default 60s) — set timeout: on the eval check to a value at least as large as --timeout, plus a small margin. Common pattern:

adb: wait-for-device
timeout: 120s

Implementation

Source: ov/adb.go (CLI subcommands + Run() methods using github.com/zach-klippenstein/goadb); allowlist + dispatcher: ov/evalrun_ov_verbs.go (adbMethods map + runAdb); validator: ov/validate_eval.go (case "adb" in validateOvVerb).

The host port is read from podman's NetworkSettings.Ports via InspectContainer — same source-of-truth used by the eval test runner's ${HOST_PORT:N} substitution. Host-networked containers expose the container port AS the host port.

Related skills

• /ov-eval:appium — sibling verb for higher-level UI automation against the same emulator (W3C WebDriver).
• /ov-eval:android — the kind: android device + apk: package format + target: android deploy. ov eval adb install / install-app are thin wrappers over the SAME shared installer (ov/android_install.go) the apk format drives — so the verb, the format, and the deploy target never drift.
• /ov-eval:eval — the unified eval system and the Check struct that holds every verb discriminator + modifier.
• /ov-tools:android-emulator (when authored) — the image these verbs target.

When to Use This Skill

MUST be invoked for any task involving ov eval adb commands or adb: declarative checks. Invoke this skill BEFORE reading the verb's Go source or reaching for command: adb ... workarounds.