scripting-adb-for-ci

name	scripting-adb-for-ci
description	Use this skill to wire `adb` reliably into CI — bash idioms, exit codes, parallel device fan-out with `xargs -P`, port forwarding (`adb forward` LOCAL REMOTE vs `adb reverse` REMOTE LOCAL — opposite argument order, the most common scripting bug), test-runner status codes (`-1` error, `-2` failure, `-3` ignored, `-4` assumption-failure), `am instrument -w -r` with `--num-shards` / `--shard-index`, the `timeout` wrapper (since `adb -t` is transport-id, NOT timeout), retry-on-transient-error with `adb kill-server`, idempotent setup (`pm clear` + animations to 0), `trap` cleanup, capture-on-failure (`screencap` + `logcat -d`), and Test Orchestrator wiring through `androidTestUtil("androidx.test:orchestrator:1.6.1")` (NOT `androidTestImplementation`). If the user mentions "adb forward vs reverse argument order", "am instrument exit code 0 even on failure", "adb timeout flag", "retry adb kill-server", "trap cleanup adb", or "androidTestUtil orchestrator", use this skill.
license	Apache-2.0. See LICENSE for complete terms.
metadata	{"author":"Jaewoong Eum (skydoves)","keywords":["android-testing","adb","ci","bash","port-forwarding","test-orchestrator","parallel-fanout","exit-codes","retries","timeout"]}

Scripting adb For CI — Bash Idioms, Exit Codes, Fan-out

This skill is the connective tissue that turns the other ADB skills into a CI pipeline: exit-code propagation, retries, timeouts, parallel device fan-out, port forwarding, idempotent setup, capture-on-failure, and Test Orchestrator wiring. Each section cross-links to the skill that owns the underlying primitive.

When to use this skill

• A CI job invokes adb shell am instrument and "passes" even when tests fail — it gates on $?, but am instrument always exits 0; the real status lives in INSTRUMENTATION_STATUS_CODE lines on stdout (and -w is needed just to get that stdout at all).
• The team has multiple physical devices on a single agent and wants to run sharded tests in parallel.
• adb forward and adb reverse are confused — most teams have at least one bug from the opposite argument order.
• Transient error: device not found / error: closed failures are not retried; the job fails on the first hiccup.
• A script lacks cleanup so leaked port forwards or modified settings put global ... survive into the next run.
• The team wires Test Orchestrator with androidTestImplementation — the build silently fails to install the orchestrator APK.

When NOT to use this skill

• The single primitive (logcat, screencap, pull, push, settings, input, am instrument) is not yet understood. Read the relevant sibling skill first.
• The CI itself (Gradle plugin, Firebase Test Lab cloud config). Use FTL's own docs for that layer.
• Compose / instrumentation test mechanics. Use ../../../compose/synchronization/synchronizing-with-idle/SKILL.md and friends.

Prerequisites

• Bash (zsh works for most of this; macOS users should brew install coreutils for gtimeout if relying on timeout).
• Platform Tools 24+ on the host. On older toolchains adb shell exit codes are unreliable.
• Devices on API 24+ for reliable shell exit-code propagation. Pre-Nougat use a sentinel __ADB_RC=$? in stdout and grep on the host.

Exit codes — what adb propagates

Operation	Behavior
adb push / adb pull	0 on success, non-zero on transport or filesystem failure. Always check.
adb install / adb uninstall	0 on success. Modern adb returns non-zero on INSTALL_FAILED_*, but old versions exit 0 even on package-manager failure — verify by piping stderr.
adb wait-for-device (and wait-for-usb-device, wait-for-local-recovery, etc.)	Block until the state is reached, then exit 0. Useful boot gate: `adb wait-for-device shell getprop sys.boot_completed
adb shell <cmd>	On platform-tools ≥ 24 and device adbd ≥ Android 7.0 (Nougat / API 24), the device-side command's exit code is propagated. Older devices always returned 0 regardless.
adb -t <transport-id>	Transport ID, NOT a timeout. Most common scripting confusion.
am instrument	Always exits 0, pass or fail — AOSP Instrument.java ends run() with an unconditional System.exit(0) and AndroidJUnitRunner calls finish(Activity.RESULT_OK, …) regardless of test outcome. -w makes the shell wait for the runner so its stdout is complete; without -w the shell returns immediately with nothing usable. Either way, gate CI on parsed INSTRUMENTATION_STATUS_CODE lines (below), never on $?.

adb shell exit-code propagation can be disabled with -x:

shell [-e ESCAPE] [-n] [-Tt] [-x] [COMMAND...]
   -x: disable remote exit codes and stdout/stderr separation

You almost never want -x in CI. The default is what you want.

AndroidJUnitRunner status codes

am instrument -w -r emits per-test INSTRUMENTATION_STATUS_CODE: <int> followed by a final INSTRUMENTATION_CODE: <int>:

Code	Symbolic name	Meaning
1	REPORT_VALUE_RESULT_START	Test started
0	REPORT_VALUE_RESULT_OK	Test passed
-1	REPORT_VALUE_RESULT_ERROR	Unexpected error (process death, runner failure)
-2	REPORT_VALUE_RESULT_FAILURE	Assertion failed
-3	REPORT_VALUE_RESULT_IGNORED (AndroidJUnitRunner extension)	@Ignored / filtered out
-4	REPORT_VALUE_RESULT_ASSUMPTION_FAILURE (AndroidJUnitRunner extension)	JUnit assumeXxx returned false

Don't treat -3 or -4 as failures — they are skips; only -1 and -2 are failures. The trailing INSTRUMENTATION_CODE line is unrelated: it's an Activity.RESULT_* value (-1 = RESULT_OK = no runner-level crash; 0 = RESULT_CANCELED = runner died) and, like the shell $? (always 0), it does not reflect test pass/fail. Grep stdout for per-test INSTRUMENTATION_STATUS_CODE: -1 / -2 (see the WRONG/RIGHT pattern below).

Port forwarding — adb forward vs adb reverse

The argument order is opposite. This is the most common scripting bug.

adb forward [--no-rebind] LOCAL  REMOTE        # host → device.  LOCAL  REMOTE
adb reverse [--no-rebind] REMOTE LOCAL         # device → host.  REMOTE LOCAL

Direction	Use case	Recipe
adb forward	Host needs to drive a device-side service.	adb forward tcp:9222 localabstract:chrome_devtools_remote
adb reverse	Device needs to reach a local-only host service (CI runner, dev laptop).	adb reverse tcp:8080 tcp:8080 (Metro/local backend)

Token grammar (from adb help):

Token	Where
tcp:<port>	Either side. tcp:0 on LOCAL = pick any free port; chosen port is printed to stdout.
localabstract:<name>	Linux abstract Unix domain socket. Used by Chrome DevTools / Stetho / Flipper.
localreserved:<name>	Reserved Unix socket namespace.
localfilesystem:<name>	Filesystem-backed Unix socket.
jdwp:<pid>	REMOTE only. Java Debug Wire Protocol for the given device-side pid.
vsock:<CID>:<port>	REMOTE only. virtio-vsock (emulator).
dev:<char-device>	Character device passthrough.
acceptfd:<fd>	LOCAL only — listen on a pre-opened fd.

Cleanup is the same surface: --list, --remove, --remove-all. Always remove what was added in a trap so re-runs on the same agent don't fail with "cannot bind: address already in use".

Timeouts — wrap with timeout, not adb -t

adb -t <transport-id> is unrelated to timeout. The host has no per-command adb timeout flag. Wrap with the GNU/BSD timeout utility:

timeout 60s adb -s emulator-5554 shell am instrument -w -r ...
echo $?
# 124 = timeout reached (process killed); 137 = had to escalate to SIGKILL.
# 0 just means am instrument exited (it always does) — see the timeout pattern below for pass/fail parsing.

On macOS timeout is gtimeout from coreutils (brew install coreutils).

For background captures, set a max duration so a stuck job doesn't fill the disk:

( timeout 600s adb logcat -v threadtime > artifacts/logcat.txt ) &
LOGCAT_PID=$!
trap 'kill $LOGCAT_PID 2>/dev/null' EXIT

Retries on transient errors

The usual offenders: error: device 'XYZ' not found, error: closed, error: protocol fault, error: device offline. A kill-server + reconnect resolves most.

retry() {
  local n=0 max=3 delay=2
  until "$@"; do
    n=$((n+1))
    [ "$n" -ge "$max" ] && return 1
    echo "[retry] $* failed (attempt $n); restarting adb"
    adb kill-server; adb start-server
    sleep "$delay"
  done
}

retry adb install -r app.apk
retry adb shell am instrument -w -r ...

For am instrument retries specifically, also pm clear the test pkg first — otherwise transient state leaks across attempts. Note retry only re-runs on transport errors (the adb host binary exits non-zero); a genuine test failure leaves am instrument exiting 0, so it won't be retried — detect those by parsing the output, not by exit code.

Parallel device fan-out

adb devices \
  | tail -n +2 \
  | awk '$2=="device"{print $1}' \
  | xargs -I {} -P 4 adb -s {} <command>

Key fragments:

• tail -n +2 skips the List of devices attached header.
• awk '$2=="device"' filters out unauthorized, offline, recovery, etc.
• -P 4 runs 4 in parallel; tune to physical USB-hub count.
• adb -s {} targets a serial. $ANDROID_SERIAL is the env-var equivalent inside the spawned shells.

For tests with shared device-state (logcat clear, app install), serialise the prep then parallelise the run:

SERIALS=$(adb devices | awk '$2=="device"{print $1}')
echo "$SERIALS" | xargs -I {} -P 0 adb -s {} install -r app-debug.apk
echo "$SERIALS" | xargs -I {} -P 0 adb -s {} install -r app-debug-androidTest.apk

# Distribute shards across devices (one shard per device)
i=0
for S in $SERIALS; do
  adb -s "$S" logcat -c
  adb -s "$S" shell am instrument -w -r \
       -e numShards $(echo "$SERIALS" | wc -l) -e shardIndex "$i" \
       com.example.test/androidx.test.runner.AndroidJUnitRunner \
       > "results-$S.txt" &
  i=$((i+1))
done
wait

# am instrument always exits 0 — derive pass/fail from the captured output:
if grep -qE '^INSTRUMENTATION_STATUS_CODE: -[12]$' results-*.txt; then
  echo "test failures detected"; exit 1
fi

(xargs -P 0 runs as many concurrently as there are inputs.)

Sharding via numShards / shardIndex

"If you need to parallelize the execution of your tests ... use the -e numShards option to specify the number of separate shards to create and the -e shardIndex option to specify which shard to run." — AndroidJUnitRunner docs.

Distribution is hash-bucketed by test name — deterministic across runs given the same test set. Adding/removing tests reshuffles buckets.

Per-device artefact paths

Always namespace by serial:

artifact_dir() { echo "artifacts/${1//[:_]/-}"; }
for S in $(adb devices | awk '$2=="device"{print $1}'); do
  D=$(artifact_dir "$S"); mkdir -p "$D"
  adb -s "$S" logcat -d -v threadtime > "$D/logcat.txt"
  adb -s "$S" shell screencap -p /sdcard/last.png
  adb -s "$S" pull /sdcard/last.png "$D/last.png"
done

Idempotent setup + cleanup with trap

Never assume clean state from a previous run. Bake cleanup into setup, and use trap to enforce teardown.

Minimal setup preamble

adb wait-for-device
adb shell input keyevent 82                         # wake; no-op if already on
adb shell pm clear com.example.app                  # wipe app data + cache
adb shell pm clear com.example.app.test             # wipe test process state
adb shell am force-stop com.example.app

# Hermetic animations
adb shell settings put global window_animation_scale 0
adb shell settings put global transition_animation_scale 0
adb shell settings put global animator_duration_scale 0

(See ../../control/injecting-input-and-state/SKILL.md for the full pm clear vs am force-stop reasoning.)

trap cleanup

cleanup() {
  adb shell am force-stop com.example.app                       || true
  adb shell pm clear     com.example.app                        || true
  adb uninstall          com.example.app                        || true
  adb forward --remove tcp:6100                                 || true
  adb reverse --remove-all                                      || true
  adb shell settings put global window_animation_scale 1        || true
  adb shell settings put global transition_animation_scale 1    || true
  adb shell settings put global animator_duration_scale 1       || true
}
trap cleanup EXIT

Each step is || true — partial-state cleanup must never mask the original failure exit code.

Capture-on-failure

set +e
./gradlew connectedDebugAndroidTest
RC=$?
set -e

if [ "$RC" -ne 0 ]; then
  mkdir -p artifacts
  adb shell screencap -p /sdcard/fail.png
  adb pull /sdcard/fail.png artifacts/fail.png
  adb logcat -d -v threadtime              > artifacts/logcat.txt
  adb logcat -d -b crash    -v threadtime  > artifacts/crash.txt
  adb logcat -d -b events   -v descriptive > artifacts/events.txt
  adb shell dumpsys activity               > artifacts/dumpsys-activity.txt
fi
exit $RC

(See ../../capture/capturing-screenshots-and-screenrecord/SKILL.md and ../../observability/extracting-logs-with-logcat/SKILL.md for the underlying primitives.)

Test Orchestrator wiring

"Android Test Orchestrator collects JUnit tests at the beginning of your test suite run, but it then executes each test separately, in its own instance of Instrumentation." — developer.android.com.

Each @Test runs in its own Instrumentation invocation, so process state, statics, and Application singletons are reset between tests.

Gradle wiring — androidTestUtil, NOT androidTestImplementation

android {
  defaultConfig {
    testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
    testInstrumentationRunnerArguments clearPackageData: 'true'
  }
  testOptions {
    execution 'ANDROIDX_TEST_ORCHESTRATOR'
  }
}

dependencies {
  androidTestImplementation 'androidx.test:runner:1.7.0'
  androidTestUtil           'androidx.test:orchestrator:1.6.1'   // NOT androidTestImplementation
}

clearPackageData: 'true' makes the Orchestrator run pm clear <pkg> between tests — the strongest isolation knob.

Direct ADB invocation (non-Gradle)

adb install --force-queryable -r path/to/orchestrator-1.6.1.apk
adb install --force-queryable -r path/to/test-services-1.6.0.apk

adb shell 'CLASSPATH=$(pm path androidx.test.services) app_process / \
 androidx.test.services.shellexecutor.ShellMain am instrument -w -e \
 targetInstrumentation com.example.test/androidx.test.runner.AndroidJUnitRunner \
 androidx.test.orchestrator/.AndroidTestOrchestrator'

--force-queryable is needed on API 30+ so the Orchestrator can bindService into the target despite package-visibility rules.

Tradeoffs

• Pros: each test starts from a known-good Application; one crash doesn't kill the rest of the run; pm clear between tests is real, not "I cleaned up in @After".
• Cons: per-test process spin-up adds ~1–2 s. A 500-test suite goes from ~5 minutes to ~15+ minutes. Reserve for integration suites where state leakage actually bites.

Patterns

Pattern: WRONG — gating CI on am instrument's exit code

# WRONG
adb shell am instrument -w -r com.example.test/androidx.test.runner.AndroidJUnitRunner
RC=$?
[ "$RC" -eq 0 ] && echo "tests passed" || echo "tests failed (rc=$RC)"
# WRONG because: am instrument calls System.exit(0) (AOSP Instrument.java) and AndroidJUnitRunner's
# finish() reports RESULT_OK regardless. $? is 0 even when tests fail — with OR without -w.

# RIGHT — keep -w (so stdout is complete) and -r (raw, parseable), then parse the stream
adb shell am instrument -w -r com.example.test/androidx.test.runner.AndroidJUnitRunner \
  | tee instrument.log
# -1 = error, -2 = failure; -3 (ignored) and -4 (assumption failure) are skips, not failures.
if grep -qE '^INSTRUMENTATION_STATUS_CODE: -[12]$' instrument.log; then
  echo "tests failed"; exit 1
fi
echo "tests passed"

Pattern: WRONG — confusing adb forward and adb reverse argument order

# WRONG — copy/pasted the forward syntax for reverse
adb reverse tcp:8080 tcp:8080      # ambiguous (works because both are tcp:8080)
adb reverse tcp:9000 tcp:8080      # this means: device port 9000 → host port 8080
# Developer expected: host service on 8080 reachable as 9000 on device.
# WRONG because: reverse argument order is REMOTE LOCAL. The device tries to dial 8080 on
# the host, but the dev meant to expose the local 8080 to the device on port 9000.

# RIGHT — adb reverse REMOTE LOCAL
# Expose host service on 8080 to the device as 9000:
adb reverse tcp:9000 tcp:8080
# In the app code: open a connection to localhost:9000 — it terminates at the host's :8080.

# Or, simpler, identical port:
adb reverse tcp:8080 tcp:8080      # device 8080 ⇆ host 8080

Pattern: WRONG — androidTestImplementation for the orchestrator APK

// WRONG
dependencies {
  androidTestImplementation 'androidx.test:orchestrator:1.6.1'
}
// WRONG because: the orchestrator is a SEPARATE APK, not a library. androidTestImplementation
// merges it into the test APK's classpath, which AGP rejects (or silently ignores). The
// correct configuration is androidTestUtil, which AGP installs as an additional APK.

// RIGHT
dependencies {
  androidTestImplementation 'androidx.test:runner:1.7.0'
  androidTestUtil           'androidx.test:orchestrator:1.6.1'
}

Pattern: timeout an instrumentation invocation

timeout 600s adb shell am instrument -w -r \
  com.example.test/androidx.test.runner.AndroidJUnitRunner | tee instrument.log
RC=${PIPESTATUS[0]}   # exit of `timeout`, not `tee`
case "$RC" in
  124) echo "timed out at 600s"; exit 124 ;;
  137) echo "timed out and had to be SIGKILLed"; exit 137 ;;
esac
# RC is 0 whether tests passed or failed (am instrument always exits 0); pass/fail comes from the stream:
if grep -qE '^INSTRUMENTATION_STATUS_CODE: -[12]$' instrument.log; then
  echo "failed"; exit 1
fi
echo "passed"

Pattern: retry-with-server-bounce

retry adb -s "$SERIAL" install -r app-debug.apk
retry adb -s "$SERIAL" shell am instrument -w -r \
  -e clearPackageData true \
  com.example.test/androidx.test.runner.AndroidJUnitRunner

The retry helper bounces the adb server on transient failure (see "Retries on transient errors" above).

Pattern: parallel sharded fan-out with per-device artefacts

SERIALS=( $(adb devices | awk '$2=="device"{print $1}') )
N=${#SERIALS[@]}
mkdir -p artifacts

pids=()
for i in "${!SERIALS[@]}"; do
  S="${SERIALS[$i]}"
  D="artifacts/${S//[:_]/-}"
  mkdir -p "$D"
  (
    adb -s "$S" logcat -c
    adb -s "$S" shell pm clear com.example.app
    adb -s "$S" shell am instrument -w -r \
         -e numShards "$N" -e shardIndex "$i" \
         com.example.test/androidx.test.runner.AndroidJUnitRunner \
         > "$D/results.txt"
    # am instrument exits 0 regardless — pass/fail is in the captured stream.
    if grep -qE '^INSTRUMENTATION_STATUS_CODE: -[12]$' "$D/results.txt"; then
      adb -s "$S" shell screencap -p /sdcard/fail.png
      adb -s "$S" pull /sdcard/fail.png "$D/fail.png"
      adb -s "$S" logcat -d -v threadtime > "$D/logcat.txt"
      exit 1
    fi
  ) &
  pids+=($!)
done

rc=0
for p in "";    || rc=1;    
 

Mandatory rules

• MUST pass -w -r to am instrument so the shell waits for the runner and emits raw, parseable output. Without -w the shell returns before any output is produced.
• MUST NOT gate CI on $? from am instrument — it calls System.exit(0) regardless of test outcome (with or without -w). Detect failures by grepping stdout for INSTRUMENTATION_STATUS_CODE: -1 (error) and -2 (failure).
• MUST distinguish -3 IGNORED and -4 ASSUMPTION_FAILURE from real failures when parsing INSTRUMENTATION_STATUS_CODE. -1 and -2 are the failure codes.
• MUST remember that adb forward is LOCAL REMOTE (host first) and adb reverse is REMOTE LOCAL (device first). The argument orders are opposite.
• MUST wrap long-running adb commands with timeout <N>s (or gtimeout on macOS). adb has no built-in timeout flag — adb -t is transport ID, not timeout.
• MUST use androidTestUtil("androidx.test:orchestrator:1.6.1"), not androidTestImplementation. The orchestrator is a separate APK, not a library.
• MUST retry transient adb errors (device not found, closed, protocol fault) with a bounded retry helper that calls adb kill-server; adb start-server between attempts.
• MUST install a trap cleanup EXIT that removes port forwards (adb forward --remove-all, adb reverse --remove-all), restores animation scales, and pm clears the app under test. Each step || true so cleanup never masks the test exit code.
• MUST pm clear <pkg> (not just am force-stop) before each test run. See ../../control/injecting-input-and-state/SKILL.md.

Verification

• Every am instrument invocation in scripts/Gradle has -w -r, and no script branches on its $? as a pass/fail signal — failures are detected by parsing INSTRUMENTATION_STATUS_CODE: -1/-2 from stdout.
• adb forward and adb reverse appear with arguments in the documented order; CI passes a smoke test that the host can reach the device service and vice versa.
• androidTestUtil('androidx.test:orchestrator:...') (not androidTestImplementation) — ./gradlew :app:dependencies --configuration androidTestUtil shows the orchestrator.
• trap cleanup EXIT is installed in the top-level CI script.
• All long-running adb commands are wrapped with timeout.
• CI run completes 50 iterations without a "leaked port forward" or "previous run's animation scale" flake.
• Status-code parsing treats -3 and -4 as skips, not failures.
• Per-device artefact directories are namespaced by serial; CI artefact archive shows them in the expected layout.
• Retry helper bounces adb kill-server; adb start-server between attempts on transient errors.

References

• developer.android.com/tools/adb — forward / reverse syntax, transport, server.
• developer.android.com/studio/test/command-line — am instrument -w -r, -e numShards / -e shardIndex, AndroidJUnitRunner arguments, output dirs.
• developer.android.com/training/testing/instrumented-tests — Test Orchestrator overview.
• developer.android.com/reference/android/app/Instrumentation — REPORT_VALUE_RESULT_* constants (OK = 0, START = 1, ERROR = -1, FAILURE = -2).
• firebase.google.com/docs/test-lab/android/command-line — Firebase Test Lab CI exit codes (0 pass, 10 test failure, 15 indeterminate, 18 incompatible, 20 infra error).
• Research note tasks/research/A3-adb-observability-automation.md — full bash idioms, exit-code propagation matrix, port-forward token grammar, parallel fan-out, retries, capture-on-failure, Test Orchestrator wiring.
• Research note tasks/research/A2-adb-shell-commands.md — am instrument arguments, status codes, AndroidJUnitRunner -3/-4 extensions.
• Sibling skill: ../../architecture/understanding-adb-architecture/SKILL.md — server / daemon / adb kill-server.
• Sibling skill: ../../devices/connecting-to-devices/SKILL.md — device states, wait-for-device.
• Sibling skill: ../../devices/connecting-over-wifi/SKILL.md — adb pair / adb connect for headless CI.
• Sibling skill: ../../apps/installing-and-managing-apps/SKILL.md — pm install / pm uninstall / pm clear.
• Sibling skill: ../../tests/running-instrumented-tests-via-adb/SKILL.md — am instrument -w -r deep dive.
• Sibling skill: ../../control/injecting-input-and-state/SKILL.md — hermetic preamble (animations to 0, pm clear, am force-stop).
• Sibling skill: ../../capture/capturing-screenshots-and-screenrecord/SKILL.md — / capture-on-failure.