Android JVM unit-test coverage

Produce the coverage report Android Studio shows for a JVM unit-test run — module total, per-package, and per-class Class / Method / Line / Branch percentages — straight from the command line, without modifying any committed build files.

How it works

./gradlew testDebugUnitTest produces no coverage on its own; Gradle has no built-in coverage engine. This skill applies JaCoCo at runtime through a Gradle init script. JaCoCo instruments the compiled bytecode in-process during the normal JVM test run (no device, no androidTest) — the same engine AGP's enableUnitTestCoverage uses internally.

Two bundled files do the work:

• assets/jvm-coverage.init.gradle.kts — applied via --init-script. It is non-invasive: it touches no build.gradle.kts and produces no git diff. It attaches JaCoCo to each module's unit-test task and registers a jvmCoverage task that emits a JaCoCo XML report. It auto-selects the right test task (testDebugUnitTest for Android modules, test for pure Kotlin/JVM modules).
• scripts/render_coverage.py — parses that XML into the Android-Studio-style table.

Always run this in a sub-agent

A coverage run compiles the module, executes its tests, and emits a large amount of Gradle output before any number appears. None of that belongs in the main conversation. So dispatch the run to a single sub-agent and let it return only the finished coverage table(s).

Dispatch one general-purpose sub-agent and give it:

• the module Gradle path(s) and matching directory(ies) (see the table below),
• the absolute path to this skill's assets/jvm-coverage.init.gradle.kts and scripts/render_coverage.py,
• the instruction: run steps 2–3 of the workflow, then return only the rendered table(s) — and, if the Gradle run fails, the BUILD FAILED block verbatim instead.

If you are already running inside a sub-agent, perform the steps directly: sub-agents cannot dispatch further sub-agents in Claude Code.

Presenting the result

The renderer's output is the report. When you relay the sub-agent's result to the user, reproduce the rendered table verbatim inside a fenced code block so its monospace column alignment (the Class … Method … Line … Branch … layout and n/a cells) is preserved exactly as shown under "Output shape".

Do not:

• reformat the table into a Markdown table or any other layout,
• rename, reorder, or merge columns,
• add a "key findings", "gaps", or "what stands out" analysis section unless the user explicitly asks for interpretation.

Report the numbers as produced; if the user wants analysis, they will ask.

Workflow

Let SKILL_DIR be this skill's directory. Run all Gradle commands from the project root (where gradlew lives).

1. Identify the module

Map what the user named to a Gradle module path and its directory:

(The init script handles the Android-vs-Kotlin difference automatically; you only need the path and directory.)

2. Run the coverage task

./gradlew :core:crypto:jvmCoverage \
  --init-script "$SKILL_DIR/assets/jvm-coverage.init.gradle.kts"

The XML lands at <module-dir>/build/reports/jvm-coverage/coverage.xml.

To cover every module in one pass, run the task unqualified — Gradle invokes it on each subproject that registered it:

./gradlew jvmCoverage \
  --init-script "$SKILL_DIR/assets/jvm-coverage.init.gradle.kts"

3. Render the report

python3 "$SKILL_DIR/scripts/render_coverage.py" \
  core/crypto/build/reports/jvm-coverage/coverage.xml \
  --module :core:crypto

For a whole-project run, call the script once per module's XML (loop over the coverage.xml files under each module's build/reports/jvm-coverage/).

To focus the per-class detail on one package, add --package (repeatable); the module total still reflects the whole module:

python3 "$SKILL_DIR/scripts/render_coverage.py" \
  core/crypto/build/reports/jvm-coverage/coverage.xml \
  --module :core:crypto --package com.example.core.crypto.internal

Output shape

MODULE :core:crypto  —  Class  90%  Method  82%  Line  78%  Branch  64%

  com.example.core.crypto
    Class  88%  Method  80%  Line  76%  Branch  61%
      AesGcm                            Class 100%  Method  90%  Line  85%  Branch  70%
      Sha256                            Class 100%  Method 100%  Line 100%  Branch n/a

  com.example.core.crypto.internal
    Class 100%  Method 100%  Line  95%  Branch  80%
      CharArrayUtf8                     Class 100%  Method 100%  Line  95%  Branch  80%

n/a means JaCoCo emitted no counter of that kind for the node (e.g. a class with no branches has no Branch %), exactly as Android Studio leaves such cells blank.

Notes

• Generated code is excluded to match what Android Studio treats as "your code": R/BuildConfig/Manifest, DI factories, di/ packages, Compose ComposableSingletons, and kotlinx.serialization $$serializer classes. Adjust the coverageExcludes list in the init script if a project needs different rules.
• No coverage rows? Confirm the module actually has a src/test suite. Modules with only src/androidTest have no JVM unit tests and produce an empty report — this skill does not cover on-device instrumented tests.
• Robolectric modules (e.g. those with unitTests.isIncludeAndroidResources = true) work unchanged — their unit tests still run on the JVM.
• The init script and XML live under build/, which is git-ignored, so running the skill never dirties the working tree.