Implement @cleanup rollback, and stop --local displacing the app's after-startup microflow

[ako]

Follow-up to #109. Two changes to mxcli test, the second from field feedback.

1. @cleanup rollback is now real

The annotation has documented rollback as its default since the runner shipped, but TestCase.Cleanup was parsed and then used nowhere — every test committed.

That was not an oversight so much as a missing seam. The after-startup runner executes tests inside the startup action, so there is no context the runner owns to open a transaction on. The test endpoint creates that seam, because it builds the IContext each test runs on.

The handler now wraps the call in startTransaction() / rollbackTransaction() when the runner asks for it, in a finally — so a test that throws, the one most likely to leave half-written data, is rolled back too. @cleanup none commits, for when the writes are the point.

Verified against Postgres, not the endpoint's own claim. A suite with one rollback test and one @cleanup none test, run against an emptied table:

PASS the default is rollback — this Person does not survive the run [rolled back]
PASS stating rollback explicitly does the same thing               [rolled back]
PASS @cleanup none commits — the control                           [committed]

$ psql -c 'SELECT firstname FROM "app$person";'
 PersistedProbe        ← 1 row

Same microflow, same run, only the annotation differs. The none case is the control that rules out "nothing persisted for an unrelated reason".

Two failure modes this closes rather than opens:

• A typo is now a parse error. @cleanup rollbak would otherwise mean "don't roll back" — data left behind, run still green, nothing explaining why. Rejected at parse time so --list catches it and no runtime boots for a file that cannot run correctly. The .mdl and .md parsers are separate code paths and both are covered.
• A rollback that fails is reported, per test and in a summary. --verbose tags each result [rolled back] / [committed] / [ROLLBACK FAILED]. An endpoint too old to know the parameter is called out separately — plausible under --attach, where the host is whatever mxcli started it.

Behaviour change worth flagging: tests that previously committed now roll back by default. That makes behaviour match the documented contract, but a suite where one test depends on data seeded by an earlier one will start failing; @cleanup none is the escape hatch. Docker is unaffected — it always commits.

2. --local no longer displaces the app's after-startup microflow

Reported from a real project (mxcli-formula1 findings #19):`` a suite passed under --attach and failed under `--local`, with cached-service assertions seeing zero rows. The app loads its cache from an after-startup microflow, and the `--local` runner displaced that microflow with its own.

It was a deliberate choice — a test run wants a known baseline — but it was invisible: the run printed only After-startup set to MxTest.RegisterEndpoint, never that the user's startup logic had been displaced, so the failure looked like a bug in the code under test.

It was also inconsistent: the hosted --test-endpoint path already chained the project's own microflow, which is exactly why the two modes disagreed. --local now chains it too, so a suite behaves the same either way and tests see the app in the state it really boots into. --skip-app-startup opts out for a deterministic empty baseline, and the run always prints which of the two it did:

… (registers the endpoint; runs no tests, then runs your MyModule.ASU_Startup)
… (registers the endpoint; runs no tests; --skip-app-startup, so MyModule.ASU_Startup will NOT run)

Verified with the finding's own shape — a seeding after-startup microflow and a test asserting on its row: PASS chained, FAIL under --skip-app-startup, from an emptied table.

That control initially passed when it should have failed, which surfaced something worth documenting: the startup microflow's writes are not covered by @cleanup rollback — they happen at boot, outside any test's transaction. Now stated in the docs.

Also from the same report (#15)

mxcli test … --list bypassed resolveTestPaths, so a project-relative path resolved for execution but not for listing. Confirmed against the pre-fix binary:

$ mxcli-old test tests/ -p /…/SpikeApp.mpr --list
Error: stat tests/: no such file or directory     ← before
Found 1 test(s): …                                 ← after

Verification

• Every new test was checked to fail against a stubbed guard (removing startTransaction, defaulting to no-rollback, dropping the client parameter, disabling validation).
• Both behavioural claims verified against a live Mendix 11.13.0 runtime and its Postgres database, with a within-run control in each case rather than a code revert.
• Full suite green; merged with current main (Aggregates broke when SET became optional (upstream CI regression) #112) — the fix-issue.md symptom table merged via the union driver with both sides' rows intact and no duplicates.

Docs

CLI help for test, mxcli syntax test, the test-microflows skill, docs-site running-tests, CLAUDE.md, two symptom-table rows, and a worked example at mdl-examples/doctype-tests/cleanup-rollback.test.mdl.

Not addressed

The other two items in findings #15 are outside the test runner and untouched here: the MDL-ODATA01 hint omits Countable / SkipSupported / TopSupported, and .ai-context/skills/ goes stale after an mxcli upgrade with no staleness detection.

🤖 Generated with Claude Code

https://claude.ai/code/session_018hifgRSawfaRWXS44YKtSJ

---

Generated by Claude Code