Skip to content

About

JUnit test history and failure triage for AI agents. Kotlin, SQLite, MCP, and cited evidence.

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

TestLedger

CI

Test-run history and failure evidence for AI agents, written in Kotlin.

TestLedger imports JUnit XML into a local SQLite database and exposes eight read-only MCP tools. An agent can compare builds, inspect grouped failures, discover tests with intermittent outcomes, and cite the exact observation behind a conclusion.

It also finds tests that got slower against a median of earlier runs. The same analysis is available as an MCP tool and a CI command with separate exit codes for regressions and insufficient history.

It handles a common triage mistake: a failing test that disappears from the next run is not marked as recovered. Skipped tests, new failures, regressions, and confirmed passes have separate categories.

What it does

  • Imports reports from Gradle, Maven, and other runners that produce JUnit XML.
  • Stores immutable runs with project, branch, commit, and timestamp metadata.
  • Groups failures by exception type and stack location, with evidence for each group.
  • Compares runs and tracks adjacent pass/fail transitions within one branch.
  • Finds intermittent outcomes and passes after retries across a branch's history, including tests absent from the selected head run.
  • Detects duration increases against a branch-specific median, with configurable thresholds, sample counts, and citations for each measurement.
  • Exposes MCP tools and a triage-tests prompt for an agent's investigation.
  • Produces a Markdown comparison report for a pull request or incident record.

There is no hosted model or API key inside TestLedger. Your MCP client supplies the agent and decides which tool results to send to its model provider. The included demo exercises the real protocol without making a model call.

Install

Requires Java 21 or later. Download and extract testledger-0.3.0.zip. The archive includes dependencies and launchers for Windows and Unix systems.

./testledger-0.3.0/bin/testledger --help

PowerShell:

.\testledger-0.3.0\bin\testledger.bat --help

Set JAVA_HOME to your JDK directory, or make java available on PATH. The following examples assume testledger is on your path; on Windows use the .bat launcher. No global Gradle installation is needed to build from source.

Import a run

testledger import \
  --reports ./build/test-results/test \
  --id build-142 \
  --project checkout \
  --branch main \
  --commit 9a13be2 \
  --at 2026-10-30T12:00:00Z

For Maven, point --reports at target/surefire-reports. For multiple modules, export their report directories under one root while preserving module paths. Use the same root layout on later imports so test identities remain stable. Reports are read recursively; build commands are never executed by TestLedger.

The database defaults to .testledger/ledger.db in the current directory. Every command accepts --db PATH. Keep the database outside your report directory. Existing database files must be recognized TestLedger ledgers. An unrelated SQLite database is rejected before its schema or journal mode changes. Earlier TestLedger ledgers remain readable and acquire an application identifier on their next import; stored run IDs and evidence stay unchanged. Importing the same run ID, reports, and metadata is idempotent. Reusing an ID with different data fails; earlier evidence cannot be silently overwritten.

testledger runs --project checkout --branch main
testledger summary --run build-142
testledger failures --run build-142
testledger compare --base build-141 --head build-142
testledger history --project checkout --branch main --test TEST_ID --window 20
testledger stability --run build-142 --window 20 --min-transitions 2
testledger test --run build-142 --test TEST_ID
testledger report --base build-141 --head build-142 > triage.md
testledger timings --run build-142
testledger timing-check --run build-142 --min-increase-ms 200 --min-increase-percent 50

Commands return JSON except report, which returns Markdown. Invalid input and missing data produce exit code 1. timing-check also returns 2 for detected timing regressions and 3 for insufficient history. runs, failures, compare, timings, and stability support --offset and --limit; JSON responses include nextOffset when another page exists.

See timing analysis for baseline selection, coverage, pagination snapshots, and CI usage. The stability guide explains discovery, ranking, and gaps in the evidence. Existing 0.1.0 and 0.2.0 databases work without migration.

Connect your agent

Import reports first, then add a stdio MCP server to your client's configuration. Use absolute paths because clients often start tools from a different directory.

{
  "mcpServers": {
    "testledger": {
      "command": "java",
      "args": [
        "-cp",
        "/absolute/path/testledger-0.3.0/lib/*",
        "io.github.almeida96.testledger.MainKt",
        "serve",
        "--db",
        "/absolute/path/checkout/.testledger/ledger.db"
      ]
    }
  }
}

On Windows, use forward slashes in JSON paths, such as C:/tools/testledger-0.3.0/lib/*. The Java launcher expands the classpath wildcard; no shell wrapper is needed. Client configuration locations vary.

Tool What the agent receives
list_runs Run metadata and final outcome counts for one project and branch
run_summary Passed, failed, skipped, and retry-hint counts
failure_groups Shared failure signatures with up to three example evidence IDs
compare_runs Paginated changes between two chronological runs
test_history Recent observations, missing runs, and pass/fail transitions
test_stability Ranked intermittent outcomes and passed retries, with cited timelines and coverage counts
test_evidence One testcase's failure text, source report, XML line, and retry hints
performance_regressions Slower passed tests, baseline medians, coverage, and cited duration samples

Select the triage-tests MCP prompt with a project and branch, or ask your agent:

Compare the latest two checkout runs on main. Inspect the new failures, check whether they appeared in earlier runs, and cite the evidence. Separate facts from hypotheses and list the next checks you would make.

To discover tests worth investigating without knowing their IDs:

Scan the 20 runs ending at build-142 for intermittent outcomes and tests that passed after retries. Include candidates missing from the head. Cite their timelines and distinguish observed changes from possible causes.

For a timing investigation:

Check build-142 for tests that grew by at least 200 ms and 50% against earlier runs on the same branch. Show the baseline, sample counts, and evidence. State how many tests lacked enough history before suggesting what to investigate.

Evidence IDs have the form testledger://runs/RUN_ID/tests/TEST_ID. They are stable references resolved through test_evidence using those two IDs, not HTTP links or advertised MCP resources. The tool also returns a relative XML report path and the testcase's starting line. The database retains the observation even if the original report is later removed.

The server opens SQLite read-only. It cannot import reports, run a shell, change code, or start a build. The owner refreshes the ledger with the CLI. This release supports stdio only; it does not expose a public HTTP endpoint.

Comparison rules

Category Base → head
regression Passed → failed/error
persistent_failure Failed/error → failed/error
new_failure Not present → failed/error
failed_after_skip Skipped → failed/error
recovered Failed/error → passed
now_skipped Failed/error → skipped
not_observed Failed/error → not present

Unchanged successful tests do not appear in the comparison. Runs must belong to the same project and branch, and the base timestamp must precede the head. These rules compare observations; they do not establish that a commit caused a failure. Keep separate project or branch labels for incompatible test environments.

test_history marks a sequence as intermittent after at least two adjacent pass/fail transitions in the selected window. Missing and skipped observations break the sequence, and equal timestamps cannot establish a transition. Different commits, dependencies, or environments may explain the changes, so this signal is not proof of a flaky test.

Reproduce the example

git clone https://github.com/Almeida96/testledger.git
cd testledger
./gradlew test installDist
npm install --global pnpm@11.25.0
pnpm install --frozen-lockfile
pnpm test:mcp
pnpm demo

Use gradlew.bat on Windows. JavaScript is used only for the independent MCP client and protocol tests; the application, parser, persistence, CLI, and server are Kotlin. These optional client scripts require Node 24 or later.

The demo imports four fictional shopping-service test runs and connects a TypeScript MCP client to the Kotlin server. It finds two regressions, four failure groups, and three outcome transitions for one test. It also keeps a missing test and a skipped test separate from the recovered test.

Results are saved to build/demo/report.md and build/demo/evidence.json. See the checked-in sample report, sample evidence, and demo walkthrough. These fixtures test known scenarios; they are not a benchmark of an LLM's diagnosis.

Input and privacy boundaries

The parser accepts UTF-8 JUnit-style XML with testsuite or testsuites roots. Nested suites, failures, errors, skipped cases, CDATA, and Surefire retry elements are supported. Suite-level aggregate counters are ignored; counts come from individual testcases. flakyFailure and flakyError entries record prior failed attempts without changing a successful final outcome. rerunFailure and rerunError preserve the final failure already present in the report.

Test identity combines the relative report path, suite path, class name, and test name. Renames and report moves create new identities. Duplicate identities in one import are rejected, so parameterized tests need distinct exported names.

DTD and external entities are rejected. Symlinks and hidden report paths are skipped. Imports are bounded to 1,000 XML files, 2 MiB per file, 25 MiB total, and 25,000 testcases. Failure details are capped at 12,000 characters and carry a truncation flag. Names are limited to 512 characters and suite paths to 1,024.

Suite properties, system-out, and system-err are not retained. Credential-like assignments, bearer tokens, URL credentials, and selected token formats are redacted before evidence is stored. This is pattern-based filtering, not a secret scanner guarantee. Review reports before sharing them with any agent or provider. The ledger contains test names and failure text; keep it out of source control.

Report contents are untrusted. The MCP prompt tells the agent to ignore embedded instructions, but client/model behavior remains outside the server's control.

Development

./gradlew test installDist distZip
pnpm test:mcp
pnpm demo

JUnit tests cover parser limits, external-entity rejection, credential redaction, retries, immutable imports, branch isolation, pagination, comparisons, and history. Timing tests cover median baselines, threshold boundaries, exclusions, coverage, backfilled runs, pagination snapshots, and finite extreme durations. Stability tests cover discovery, retries, missing tests, timestamp ties, scope isolation, and every combination of four final outcomes or missing observations. Protocol tests launch the installed JVM distribution and use the official TypeScript MCP client. They also check UTF-8 on Windows and nonzero CLI exits. CI runs Java 21 on Linux and Windows. JaCoCo reports are available under build/reports/jacoco/test.

Dependencies are pinned and Gradle dependency locks are committed. The Gradle wrapper distribution has a SHA-256 checksum. See architecture for storage, grouping, and consistency details.

Contribution guide. MIT license.

About

JUnit test history and failure triage for AI agents. Kotlin, SQLite, MCP, and cited evidence.

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages