The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Ariadne listing page.
MCP Server for Affected Test Selection
ariadne is an MCP (Model Context Protocol) server that provides AI agents with the ability to identify affected tests. Powered by sazanami, it analyzes code changes and returns only the tests that need to be run.
[!IMPORTANT]
⚠️ MUST READ: Speed over Completeness
ariadne is built for the agent inner loop — edit, verify, commit — where fast feedback matters more than exhaustive selection. Static analysis cannot trace every execution path: reflection, DI frameworks, and data-flow indirection (e.g., Flux/MVI dispatch) can hide dependencies from any affected-test-selection tool, not just ariadne.
Always keep a final line of defense in CI. Run the full test suite (or a conservative selection) before merging. ariadne narrows what an agent runs while iterating; it is not a replacement for CI.
When ariadne detects changes it cannot analyze (build scripts, resources, unscanned source sets), it says so explicitly in the tool response instead of silently reporting "no affected tests".
git diff internallyThen register it with your MCP client — for Claude Code:
Or add to your MCP client configuration manually (e.g., Claude Desktop):
Mount the project you want analyzed and pass /workspace as project_path.
The image is also listed in the official MCP Registry as io.github.MikhailHal/ariadne.
Download ariadne-<version>-all.jar from Releases (requires JDK 21+) and configure your client with "command": "java", "args": ["-jar", "/path/to/ariadne-<version>-all.jar"].
Once configured, AI agents can use the get_affected_tests tool:
get_affected_testsParameters:
project_path (required) — Path to the Kotlin projectscope (optional, deep | shallow, default deep) — how far back to look:
deep — all changes since base_branch (committed and uncommitted). Safest; the whole branch is covered so nothing you already committed slips through unverified. May select more tests.shallow — uncommitted changes only (diff against HEAD). Fastest, for the tight edit loop. Verifying already-committed work is left to the caller. base_branch is ignored.base_branch (optional, deep scope only) — Branch to compare against. When omitted, ariadne uses the repository's default branch (origin/HEAD). If that is not set (e.g. a repo with no remote), it returns an error asking you to pass base_branch explicitly rather than guessing.Returns:
Agent request:
Response:
git diff --unified=0 against base branchMeasured against Now in Android (Google's reference Android app — 34 modules, ~268 Kotlin files):
| Metric | Result |
|---|---|
| Recall audit — 19 target functions across all layers | 18/18 valid targets detected (the 19th had no exercising unit test; correctly not selected) |
| End-to-end response time | ~4s (module discovery + call-graph build + BFS) |
| Module discovery | 34 modules via settings.gradle.kts, incl. nested modules and type-safe accessor dependencies |
| Source sets | main, debug, prod, benchmark, testDemo, … discovered per module (androidTest* excluded by design) |
Verified patterns include repositories behind project interfaces, a library-interface
override (androidx.datastore.Serializer), operator fun invoke use cases,
@Composable functions, extension mappers, ViewModel property-initializer chains,
and callable references. Two representative results:
core:common's asResult() selects 14 tests across three modules,
including ViewModel tests reachable only through val uiState = ...stateIn(...)PopulatedNewsResource.asExternalModel() selects 14 tests,
including 11 repository tests reachable only through .map(Type::mapper) chainsEvery unit-test class in Now in Android was measured by changing a function in the class it tests and checking whether that test class was selected:
| Test style | Selected |
|---|---|
| Plain unit tests (construct the object, call it) | 13 / 13 valid targets |
| Robolectric / Compose screenshot tests | 12 / 12 |
Framework-dispatched callbacks (lint Detector) | 0 / 2 — see below |
Robolectric turned out not to be a barrier: those tests call the composable
themselves (setContent { NiaTheme { ... } }), so the call exists in the source.
What decides coverage is not the test runner but whether the test's own code
contains the call.
The rule of thumb: if the framework calls your code instead of your test calling it, ariadne cannot connect them. These are limits of static analysis, not bugs — plan your CI safety net around them:
| Pattern | Status |
|---|---|
Framework-invoked callbacks — Fragment/Activity lifecycle (launchFragmentInContainer), lint Detector methods, Application.onCreate | Not traced: no call written in the test |
| Reflection / DI-container wiring | Not traced |
| UDF dispatch (Flux/MVI) | dispatch → collect is never an edge, but wiring in init (or a start() the test calls) is covered conservatively via constructor chains. Subscriptions started by DI/lifecycle are not covered (sazanami#38) |
stateIn / shareIn chains (map, onEach, flatMapLatest, combine) | Covered — verified with exact selection |
Instrumented tests (androidTest*) | Out of scope by design |
| Build scripts, resources, unscanned source sets | Not analyzed — reported explicitly in the tool response |
| Same-name top-level extensions in one package | Over-selected (receiver types are not part of top-level FQNs) — safe direction |
KMP source sets (commonMain, expect/actual) | Enumerated, but resolution quality unverified (#1) |
Full audit notes: sazanami#29, sazanami#38.
settings.gradle(.kts) includes,
enumerates src/<sourceSet>/{kotlin,java} layouts, and reads project(":x") /
type-safe accessor dependencies from build files. Dynamic includes,
projectDir remapping, custom srcDirs, and dependencies injected by convention
plugins are not detected — see #1See LICENSE for the full text.