Odoo Test Runner

Odoo Test Runner

A green test on the dev DB you've run all week is a lie — it passed against records, half-applied columns, and hand-fixed data the customer's DB won't have. A throwaway DB, installed clean and dropped after, is the only signal you can believe that the development does what the ticket asked. These are stances to carry, not a checklist to run: the exact odoo-bin invocations and test-class patterns live in references/. When a task skips planning (odoo-grill), these reflexes carry verification alone.

Running — make the run credible

Fresh DB, or don't trust the green

Pollution looks like correctness: a leftover record from yesterday makes today's search return 1 instead of 0, and the check passes for the wrong reason. A pass on your long-lived dev DB therefore proves nothing about a clean install. The only run you can believe is a throwaway DB carrying just your module's dependencies, built fresh and dropped after — even when the run failed, because the DB you keep "for inspection" is the next run's pollution. After a fix, re-run on a fresh DB: the fix itself may lean on dev-DB state the customer won't have. One binary owns the whole loop (odoo-bin db init/drop) because dropping the DB takes its filestore with it — raw dropdb orphans the filestore and re-pollutes. (Invocation, flags, ~/.odoorc → references/odoo-bin-commands.md.)

Prove the ticket, not the checkbox

--test-enable going green proves the code runs without crashing — not that it does what was asked. Before verifying, restate the acceptance criteria in one sentence; if you can't, you don't yet know what you're proving. "Verified" means every claim in the ticket has a concrete artifact behind it — a passing test, a shell output, or a screenshot — and the suite is green on a fresh DB. Anything less is "it didn't crash."

The new behavior earns a new test

A suite that still passes after your change proves you didn't break the old behavior — not that you delivered the new one. The bug you fixed earns a test that fails before the fix and passes after; the feature you built earns a test asserting its new contract. Leaning on an existing test that happens to still pass is how a feature ships untested. A surprise found while poking in the shell or UI earns a unit test before you fix it — that's the move that turns a one-off into a regression that can never return.

Match the layer to the risk

Unit (--test-enable) for logic that breaks unnoticed — computes, constraints, security, migrations, anything with branches. Shell (odoo-bin shell) for data-shape questions — did the field compute, does this domain return those records — faster than clicking. UI (:8069) for what the customer actually clicks — views, wizards, reports. Not exclusive: a new computed field on a form wants all three; a pure migration wants only the unit test.

Isolate the module under test

Install only your module and its dependencies, never -i all — a failure then traces to your code, not a neighbor's. Name the DB tmp_test_<module>_* so orphan cleanups stay trivial, and tag your suite so --test-tags runs just it instead of letting a hundred unrelated tests drown your signal.

When it fails, fix the code — not the test

Failures cascade, so read the first trace; the rest are noise. Never relax the assertion to match the buggy output — a test bent to accept the bug is worse than the bug, because it now hides every future one too. (For the debugging discipline itself, lean on systematic-debugging.)

Writing — tests that run and mean something

Pick the lightest class — and let it roll back

TransactionCase is the default: one transaction per test, rolled back, so every test starts from the same clean state — that roll-back is the fresh-DB discipline enforced inside the suite, and a stray commit() breaks it. Step up to HttpCase only for the HTTP layer — controllers, sessions, tours. SavepointCase is gone (merged into TransactionCase, v16+) — don't import it.

A test that never runs is a green lie

Several silent no-ops pass while testing nothing: @tagged("-at_install") with no post_install never executes; a new test file not imported in tests/__init__.py leaves the runner discovering nothing and "passing" empty; and a tour logging in as admin/demo/portal silently fails to authenticate, because a demo user's login is not their password and a fresh DB leaves it empty — set the passwords first. Each looks green; none proves anything. (Class table, @tagged, the tests/__init__.py rule, the full Python+JS+manifest tour wiring, and the password SQL → references/writing-tests.md.)

Assert what won't drift, and the specific failure

Assert stable record values (assertRecordValues for many fields at once), never UI labels — labels are translatable and move under you. Assert the specific exception (assertRaises(UserError)), never Exception — the broad catch passes on the next, unrelated bug too.

References (consult, don't memorize)

For the JS half of a tour (registering the tour, step trigger/run shapes), see odoo-js. For wrapping up the whole ticket — review + test + customer-readiness — see odoo-task-completion; for the review pass itself, odoo-code-review.