Use this skill to run instrumented Android tests directly through `adb shell am instrument -w -r` without going through Gradle. Covers the required `-w` (wait — REQUIRED for meaningful exit codes) and `-r` (raw output) flags, the `-e` argument table (`class`, `class#method`, `package`, `size`, `numShards`/`shardIndex`, `debug`, `annotation` / `notAnnotation`, `listener`, `clearPackageData`, `targetInstrumentation`), the canonical runners `AndroidJUnitRunner` and `AndroidTestOrchestrator`, the orchestrator wrapping pattern (target = orchestrator, `-e targetInstrumentation <pkg>/<runner>`), and the output framing (`INSTRUMENTATION_STATUS_CODE` 1=start, 0=ok, -1=error, -2=failure, -3=ignored, -4=assumption-failure; `INSTRUMENTATION_RESULT`; `INSTRUMENTATION_CODE`). Use when the user mentions `am instrument`, `AndroidJUnitRunner`, "run tests from CI without Gradle", "Orchestrator", `clearPackageData`, `targetInstrumentation`, exit codes from `am instrument`, or `INSTRUMENTATION_STATUS_CODE`.
Installation
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Use this skill to run instrumented Android tests directly through `adb shell am instrument -w -r` without going through Gradle. Covers the required `-w` (wait — REQUIRED for meaningful exit codes) and `-r` (raw output) flags, the `-e` argument table (`class`, `class#method`, `package`, `size`, `numShards`/`shardIndex`, `debug`, `annotation` / `notAnnotation`, `listener`, `clearPackageData`, `targetInstrumentation`), the canonical runners `AndroidJUnitRunner` and `AndroidTestOrchestrator`, the orchestrator wrapping pattern (target = orchestrator, `-e targetInstrumentation <pkg>/<runner>`), and the output framing (`INSTRUMENTATION_STATUS_CODE` 1=start, 0=ok, -1=error, -2=failure, -3=ignored, -4=assumption-failure; `INSTRUMENTATION_RESULT`; `INSTRUMENTATION_CODE`). Use when the user mentions `am instrument`, `AndroidJUnitRunner`, "run tests from CI without Gradle", "Orchestrator", `clearPackageData`, `targetInstrumentation`, exit codes from `am instrument`, or `INSTRUMENTATION_STATUS_CODE`.
license
Apache-2.0. See LICENSE for complete terms.
metadata
{"author":"Jaewoong Eum (skydoves)","keywords":["am-instrument","AndroidJUnitRunner","AndroidTestOrchestrator","clearPackageData","targetInstrumentation","INSTRUMENTATION_STATUS_CODE","test-sharding","numShards","shardIndex","run-listener"]}
Running Instrumented Tests via ADB — am instrument without Gradle
adb shell am instrument is the underlying command Gradle invokes; running it directly is the right tool for CI scripts that already manage their own APKs, for sharding fan-out across many devices, and for tight-loop debugging of a single test method. The pitfall most CI scripts fall into: omitting -w, which makes the exit code meaningless. The second-most-common pitfall: flipping the orchestrator-target relationship.
When to use this skill
The user wants a single test method to run from a script: adb shell am instrument -w -r -e class com.example.MyTest#myMethod ....
The user wants to shard a test suite across N devices using numShards / shardIndex.
The user is wiring AndroidX Test Orchestrator with clearPackageData true and gets the target/targetInstrumentation order confused.
The user's CI script reports green when tests actually failed because $? is 0 even though INSTRUMENTATION_STATUS_CODE: -2 shows a failure.
The user wants the on-device runner to wait for a debugger attach before running tests.
When NOT to use this skill
The user wants to install or reset the app under test — use ../../apps/installing-and-managing-apps/SKILL.md.
The user wants to choose a device, wait for boot, or set up Wi-Fi debugging — use ../../devices/connecting-to-devices/SKILL.md and ../../devices/connecting-over-wifi/SKILL.md.
The user is writing the JUnit4 / AndroidJUnit4 test class itself — use ../../../instrumentation/runner/running-instrumented-tests-with-androidjunit4/SKILL.md.
The user wants Gradle to do the run for them (./gradlew connectedDebugAndroidTest) — that path also reads instrumentation runner args from testInstrumentationRunnerArguments.
Prerequisites
App APK installed (com.example.app) and test APK installed (com.example.app.test) — see ../../apps/installing-and-managing-apps/SKILL.md.
The test APK was built with -t allowed (android:testOnly="true").
For Test Orchestrator: androidx.test.orchestrator APK installed (typically via androidTestUtil("androidx.test:orchestrator:1.6.1") and adb install -r androidx.test.orchestrator.apk).
Device in device state and animations disabled for stable runs (see docs/CORPUS.md §I.6 hermetic test setup).
Workflow
1. Compose the canonical command. Verbatim from developer.android.com/studio/test/command-line and am help:
-w "Wait for instrumentation to finish before returning. Required for test runners." Without it, the shell returns immediately and $? is meaningless.
-r "Print raw results (otherwise decode report_key_streamresult)." Pair with -w for CI-friendly output.
Common runners:
androidx.test.runner.AndroidJUnitRunner — the default for AGP androidTest.
androidx.test.orchestrator.AndroidTestOrchestrator — the wrapper that runs each test in its own instrumentation invocation.
Bare invocation against the AndroidJUnitRunner:
adb shell am instrument -w -r \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
2. Use the -e argument table for selection, sharding, and runner behavior. The full table understood by AndroidJUnitRunner / AndroidX Test:
Key
Value
Meaning
package
<java_package>
"Fully qualified Java package name. Any test class using this package executes." Takes precedence over class. Comma-separate for multiple.
class
<fqcn>
"Only this test case class executes."
class
<fqcn>#<method>
"Only this method executes." Hash separator.
class
<fqcn1>,<fqcn2>#m,...
Comma-separate to combine multiple selectors.
size
small / medium / large
"Runs test methods annotated with @SmallTest, @MediumTest, or @LargeTest."
annotation
<fqcn>
Filter to tests carrying this annotation.
notAnnotation
<fqcn>
Exclude tests carrying this annotation.
numShards
<N>
Total shard count for parallel runs.
shardIndex
<i>
This shard's 0-based index (0 ≤ i < N).
debug
true
"Runs tests in debug mode." Waits for debugger attach before each test.
log
true
"Loads and logs all specified tests but doesn't run them." Use to verify filter combinations.
listener
<fqcn>
Register a RunListener. Comma-separate for multiples.
filter
<fqcn>
Register a custom JUnit Filter.
runnerBuilder
<fqcn>
Custom RunnerBuilder.
disableAnalytics
true
Disable AndroidJUnitRunner usage stats.
clearPackageData
true
AndroidX Orchestrator only: pm clear between tests.
targetInstrumentation
<pkg>/<runner>
AndroidX Orchestrator only: the wrapped instrumentation.
coverage
true
Collect JaCoCo coverage.
coverageFile
<path>
Override device coverage file location.
Recipes:
# One class.
adb shell am instrument -w -r \
-e class com.example.app.LoginTests \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# One method.
adb shell am instrument -w -r \
-e class com.example.app.LoginTests#login_succeeds \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# @LargeTest only.
adb shell am instrument -w -r -e size large \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# Shard 0 of 4 (run on 4 devices in parallel for 4× speedup).
adb shell am instrument -w -r -e numShards 4 -e shardIndex 0 \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# Wait for debugger before each test.
adb shell am instrument -w -e debug true \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# Custom RunListener.
adb shell am instrument -w -r \
-e listener com.example.app.MyRunListener \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
3. Wire AndroidX Test Orchestrator correctly. The orchestrator runs each test in its own instrumentation invocation, isolating crashes and (with clearPackageData) wiping app state between tests. To use it:
The target (last positional arg) is the orchestrator: androidx.test.orchestrator/.AndroidTestOrchestrator.
The wrapped runner is passed via -e targetInstrumentation <pkg>/<runner>.
Common copy-paste mistake is to flip the two — putting the AndroidJUnitRunner as the positional argument and the orchestrator inside targetInstrumentation. That fails with confusing errors because the orchestrator is the runner and the AndroidJUnitRunner is the target.
4. Read the output framing. Verbatim format:
INSTRUMENTATION_STATUS: <key>=<value> # repeated per Bundle entry per status frame
INSTRUMENTATION_STATUS_CODE: <int> # one per status frame
...
INSTRUMENTATION_RESULT: <key>=<value> # repeated per Bundle entry of the final result
INSTRUMENTATION_CODE: <int> # final, exactly once
Status code values (per Instrumentation.REPORT_VALUE_RESULT_* plus AndroidJUnitRunner additions):
Code
Meaning
1
Test started (REPORT_VALUE_RESULT_START).
0
Test passed (REPORT_VALUE_RESULT_OK).
-1
Process error / unexpected throw (REPORT_VALUE_RESULT_ERROR).
-2
Assertion failure (REPORT_VALUE_RESULT_FAILURE).
-3
Ignored (@Ignore) — added by AndroidJUnitRunner.
-4
Assumption failure (org.junit.AssumeViolatedException) — added by AndroidJUnitRunner.
AndroidJUnitRunner adds -3 and -4 on top of the framework's four canonical values; pure framework code only emits 1, 0, -1, -2 for per-testINSTRUMENTATION_STATUS_CODE. The final INSTRUMENTATION_CODE line is separate and uses Android's Activity.RESULT_* constants — -1 (RESULT_OK) when the run completes without a runner-level error and 0 (RESULT_CANCELED) when the runner itself errored. Neither this line nor the shell $? reliably reports test pass/fail: AndroidJUnitRunner calls finish(Activity.RESULT_OK, results) regardless, and AOSP frameworks/base/cmds/am/.../Instrument.java ends run() with unconditional System.exit(0) — so $? is 0 even when tests fail. CI scripts MUST parse INSTRUMENTATION_STATUS_CODE: -2 (failure) and INSTRUMENTATION_STATUS_CODE: -1 (error) lines from stdout to detect failures. See step 6 for the canonical grep idiom.
Standard Bundle keys per status frame (from the Instrumentation reference):
REPORT_KEY_NAME_CLASS — current test FQCN.
REPORT_KEY_NAME_TEST — current test method.
REPORT_KEY_NUM_CURRENT — 1-based index of the current test.
REPORT_KEY_NUM_TOTAL — total test count.
REPORT_KEY_STACK — failure stack trace.
REPORT_KEY_STREAMRESULT — pretty stream output (decoded by default; pass -r to keep raw).
5. Do NOT trust the shell exit code as a pass/fail signal. With -w, $? reports 0 on overall pass AND on test failures — am instrument ends with System.exit(0) and AndroidJUnitRunner's finish() reports RESULT_OK regardless of test outcome. Without -w, the shell returns immediately and $? is even less meaningful. MUST parse stdout for INSTRUMENTATION_STATUS_CODE: -2 (failure) and -1 (error) lines — see step 6. -w -r is still required (the wait + raw flags AGP/Gradle and Android Studio invoke with), just not for its exit code.
6. Parse the output for CI. Two practical options:
Let Gradle write the XML at app/build/outputs/androidTest-results/connected/<variant>/TEST-*.xml — required when integrating with JUnit XML test reporters (most CI dashboards). Use ./gradlew connectedDebugAndroidTest for that path. The two approaches are not mutually exclusive; many pipelines run am instrument directly for sharded execution, then post-process raw status streams into JUnit XML themselves.
7. Disable animations and reset state for hermetic runs. Per docs/CORPUS.md §I.6:
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
adb shell pm clear com.example.app # before each run; orchestrator does this per-test
Or via Gradle: testOptions.animationsDisabled = true and Test Orchestrator's -e clearPackageData true.
8. Be aware of shell exit-code propagation limits.adb shell exit-code propagation is reliable only since API 24 / Platform Tools 24. On older combinations, scripts that depend on $? from adb shell <cmd> need to capture the value on-device (echo $? > /sdcard/exit) and pull it back. For modern devices and Platform Tools 24+, adb shell <cmd>; echo $? propagates the device-side exit code as expected — but note that for am instrument that code is always 0 regardless of test results (step 5), so propagation is irrelevant to pass/fail detection.
Patterns
Pattern: WRONG vs RIGHT — exit code without -w
# WRONG
adb shell am instrument -r \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
echo $?
# 0# WRONG because: -w is omitted, so the adb shell returns immediately. $?# reflects the launch, not the test outcome. Failures in the test run are# silently lost.
# RIGHT — gate on stdout, NOT on $?
output=$(adb shell am instrument -w -r \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner)
ifecho"$output" | grep -qE 'INSTRUMENTATION_STATUS_CODE: -[12]$'; thenecho"FAILED"; exit 1
fi# am instrument exits 0 even when tests fail (System.exit(0) in AOSP Instrument.java),# so the only reliable signal is parsing INSTRUMENTATION_STATUS_CODE: -2 (failure) or -1 (error).
Pattern: WRONG vs RIGHT — orchestrator target/targetInstrumentation flip
# WRONG
adb shell am instrument -w -r \
-e clearPackageData true \
-e targetInstrumentation androidx.test.orchestrator/.AndroidTestOrchestrator \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# WRONG because: the orchestrator is the runner, not the target. AndroidJUnitRunner# does not know how to invoke an "orchestrator target", and clearPackageData is# not a recognised arg of AndroidJUnitRunner — only of the orchestrator.
# RIGHT
adb shell am instrument -w -r \
-e clearPackageData true \
-e targetInstrumentation com.example.app.test/androidx.test.runner.AndroidJUnitRunner \
androidx.test.orchestrator/.AndroidTestOrchestrator
# Orchestrator is the runner (last positional arg).# AndroidJUnitRunner is the wrapped target (passed via -e targetInstrumentation).
Pattern: WRONG vs RIGHT — sharding across two devices
# WRONG
adb -s emulator-5554 shell am instrument -w -r -e numShards 2 -e shardIndex 0 \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner &
adb -s emulator-5554 shell am instrument -w -r -e numShards 2 -e shardIndex 1 \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner &
wait# WRONG because: both shards run on the same device (-s emulator-5554). They# serialize against each other through am, defeating the parallelism goal of# sharding. Worse, they may collide on the same app under test.
# RIGHT
adb -s emulator-5554 shell am instrument -w -r -e numShards 2 -e shardIndex 0 \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner &
adb -s emulator-5556 shell am instrument -w -r -e numShards 2 -e shardIndex 1 \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner &
wait# One shard per device. Each device runs roughly half the suite.
Pattern: WRONG vs RIGHT — running a single method
# WRONG
adb shell am instrument -w -r -e class com.example.app.LoginTests:login_succeeds \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
# WRONG because: the separator between class and method is `#`, not `:`.# AndroidJUnitRunner treats the whole string as a class name and finds nothing.
# RIGHT
adb shell am instrument -w -r -e class com.example.app.LoginTests#login_succeeds \
com.example.app.test/androidx.test.runner.AndroidJUnitRunner
Mandatory rules
MUST pass -w to am instrument for any CI/script invocation so the shell waits for the runner to complete (necessary for stdout parsing). MUST NOT gate on $? — am instrument calls System.exit(0) regardless of test outcome.
MUST pair -w with -r for CI consumption — raw output is parseable; the decoded report_key_streamresult form is not.
MUST use <class>#<method> (hash separator) for single-method selection. Colons or dots fail silently.
MUST put the orchestrator as the runner (positional last arg) and the AndroidJUnitRunner as the target via -e targetInstrumentation — not the other way around.
MUST NOT rely on am instrument exit codes when -w is omitted.
MUST NOT assume clearPackageData true works without Orchestrator — it is an Orchestrator-only argument.
PREFERRED: use Test Orchestrator + clearPackageData true for hermetic test isolation in CI; this maps to Gradle's testOptions.execution = "ANDROIDX_TEST_ORCHESTRATOR".
PREFERRED: disable animations (settings put global *_animation_scale 0) before the run, or rely on Gradle's testOptions.animationsDisabled = true.
Verification
No script gates on $? from am instrument — it is always 0 (System.exit(0) in AOSP Instrument.java; finish(RESULT_OK, …) in AndroidJUnitRunner). Pass/fail is detected by grepping stdout for INSTRUMENTATION_STATUS_CODE: -1 (error) / -2 (failure).
Output stream contains exactly one INSTRUMENTATION_CODE: <int> line at the end.
Per-test frames carry matching INSTRUMENTATION_STATUS_CODE: 1 (start) followed by 0 (ok) or -1/-2/-3/-4.
When using Orchestrator, the runner positional arg is androidx.test.orchestrator/.AndroidTestOrchestrator and -e targetInstrumentation carries the AndroidJUnitRunner FQN.
When sharding, -e numShards N -e shardIndex i runs on distinct devices (-s differs per shard).
-e class <fqcn>#<method> runs exactly one test method (cross-check with INSTRUMENTATION_STATUS: numtests=1).
Animation scales are 0.0 on the device (adb shell settings get global window_animation_scale).
