Runner Coverage 2026 10 26

Runner Coverage - 2026-10-26

Context

A coverage review (2026-10-26) estimated that Testosterone supported about 23% of test framework usage natively (pytest, Behave/BehaveX, Playwright) and could record about 94% through JUnit XML. A third of that 94% needed an extra reporter package first (jest-junit, a .NET JUnit logger, gotestsum, ...), and JUnit drops what the runner knew: describe blocks, retries, files that failed before any test ran, tags. The aim is for Testosterone to be the default results history for any test tool, so the common runners should work with the output they already write, with no extra install.

Decision

testo_core/reporting/junit_import.py stays the one entry for every result file (stage junit_xml globs, testo ingest, POST /api/v1/ingest). It now recognises a file by its content, not its name, and hands it to that runner's importer. Each importer writes the same Allure result dicts, so nothing after the import changes.

File	Written by	Importer
JUnit XML (<testsuite> / <testsuites>, and any other XML)	most runners	junit_import.py
Playwright JSON (suites list)	--reporter=json	playwright_import.py
Jest / Vitest JSON (testResults list)	jest --json, vitest --reporter=json	jest_import.py
TRX (<TestRun>)	dotnet test --logger trx (xUnit, NUnit, MSTest)	trx_import.py
go test -json lines (Action / Package events)	go test -json	gotest_import.py
Robot Framework output.xml (<robot>)	robot	robot_import.py
Allure *-result.json	any Allure adapter	allure_passthrough.py

• Identity. Each importer builds a stable fullName: file › describe › title for Jest/Vitest (paths relative to the repo, or to the report's common folder when made elsewhere), Namespace.Class.Method for TRX, package.Test/subtest for Go, Robot's long name. These are also the names the runners' own "run only this test" options take, which later rerun commands rely on.
• What JUnit lost. A Jest test that passed on retry is flaky; a test file that failed to load is one broken result; TRX data rows are one result each; a failed Go package with no failing test (build error) is one broken result; a Go parent test counts only when it failed on its own, so a failing subtest is not counted twice; Robot tags become Allure tags.
• Allure folders. testo ingest allure-results/ (a folder argument stands for its files) copies each result with the attachments it names, from the same folder only. Containers and the folder's settings files are skipped without a warning.
• Server uploads. POST /api/v1/ingest keeps a file's safe name and suffix (an Allure result names its attachments by file name) and reads every stored file by content.
• Native JavaScript stages. New equipment jest, vitest (frameworks/jest_adapter.py) and cypress (frameworks/cypress_adapter.py) run the runner through npx and add the report Testosterone reads, like playwright does: Jest's --json --outputFile, Vitest's json reporter (and vitest run, never watch mode), Cypress's built-in junit reporter. Jest and Vitest take --maxWorkers from workers; Cypress can't parallelise without its cloud, so workers warns there. A stage that sets its own reporter keeps it.
• Rerun commands for more runners. services/rerun_commands.py (the run page's "Rerun these tests") also uses the framework label each imported result carries, so an ingested file gets its own runner's command: Jest/Vitest (test files plus -t '^(describe path title|...)$'), go test -run '^(TestA|TestB)$' <packages> (top-level tests, which rerun their subtests), dotnet test --filter FullyQualifiedName=... (data rows rerun their method), robot --test <name> <root suite folder>, and for JUnit XML from a JVM build mvn test -Dtest=Class#m1+m2 or gradle test --tests Class.method when the run's target repo has a pom.xml or a Gradle build. The pytest guess for dotted JUnit names now needs a test_*.py / *_test.py file, the files pytest collects, so a Java class is no longer offered as a pytest module. Cypress gets none: its JUnit file does not name the spec file.
• Unknown files are skipped with a note in the stage log, as malformed JUnit always was. They never fail the stage on their own.

Consequences

• Jest, Vitest, .NET, Go and Robot suites can be recorded with the output their runner writes by default. JUnit XML keeps working for everything else.
• Detection reads each file once more than before (a sniff, then the parse). Result files are small next to the test run.
• historyId still hashes the stage's tool with fullName, so moving a suite from JUnit XML to its runner's own format starts a new history for its tests (different names).
• Allure containers (fixtures) are not shown; Testosterone has no place for them.

Addendum: RSpec, k6, XCTest and Android (2026-10-26)

A comparison with Qualflare (a hosted test-results service that lists XCTest, RSpec, k6 and Espresso among its 23+ frameworks) found these four still needed a JUnit reporter or weren't covered. Same pattern, same entry point:

File	Written by	Importer
RSpec JSON (examples list + summary)	rspec --format json	rspec_import.py
k6 summary (metrics + root_group)	k6 run --summary-export=..., or handleSummary writing JSON.stringify(data)	k6_import.py
xcresulttool JSON (testNodes tree)	xcrun xcresulttool get test-results tests --path X.xcresult (Xcode 16+)	xcresult_import.py
Android instrumentation JUnit (suite device property)	connectedAndroidTest (Espresso, UI Automator, Compose)	junit_import.py

• RSpec. full_name is spec file › full description. Pending examples are skipped; errors outside examples (a spec file that failed to load) become one broken result. Rerun: rspec <files> -e <description>, with bundle exec when the target repo has a Gemfile.
• k6. A load test has no test cases, so each check and each threshold becomes one result (checks › group › name, thresholds › metric › expression). A check with any failed iteration is failed ("3 of 120 failed"); a crossed threshold is failed with the metric's values. Both the --summary-export shape (keyed by name, threshold true = failed) and the handleSummary shape (lists, {"ok": false}) are read. Results have no duration (each would span the whole run). No rerun command: only the whole script reruns.
• XCTest. The .xcresult bundle is a database only Apple's tool reads, so the input is the JSON xcresulttool exports; reading the bundle directly would need macOS. full_name is Bundle/Suite/test(), the -only-testing: form. A test that passed on a repetition after failing is flaky. No rerun command: xcodebuild also needs a scheme and destination the results don't record.
• Android. The Android Gradle plugin's JUnit XML already worked; suites that name a device are now labelled android (and the device as host), so the run page offers ./gradlew connectedAndroidTest -Pandroid.testInstrumentationRunnerArguments.class=Class#method. In a project with several modules that have instrumentation tests, run it on the module's own task.

Builds on Command Adapter and JUnit Import, Playwright JSON Import and Ingest and Team Server.

---

Generated from docs/Specs & ADRs/Runner Coverage - 2026-10-26.md. Edit it there; changes made in the wiki are overwritten.