Every rule the engine can emit. This page is generated from the code: a rule that exists but is undocumented, or documented but removed, fails the build and the test suite.
The reason this project exists. Each of these is a statement about what a change took away, which is why no tool that grades the end state can report them.
assertion_monotonicityrepairAn assertion on some value got weaker: == 42 became is not None, a specific exception became a bare Exception, an exact count became a lower bound. The subject is still asserted, so the test still passes and coverage is unchanged.
disabled_assertionrepairAn assertion still present in the source but no longer executed — commented out, moved behind a condition that is never true, or swallowed by a bare except.
empty_testrepairA test function whose body asserts nothing: a pass, a docstring, or setup with no check at the end. It runs, it is counted, it verifies nothing.
skip_markerrepairA @skip, @skipif, @xfail or equivalent added to a test that did not have one. The suite goes green by not asking.
vacuous_assertionrepairassert True, assert [1], assert x or True — an assertion whose truth does not depend on the code under test.
A refactor is only correct if everything that pointed at the old thing now points at the new one. These catch the half-finished move.
boundary_violationrepairAn import that the project's declared zones forbid. Configured in .yieldpoint.json; mature tools exist for this, and it ships here for convenience rather than as a reason to adopt.
dangling_referencerepairA name is gone and something still points at it. The classic half-finished rename, where the tests that happen not to exercise that path stay green.
export_removedrepairA name that was part of a module's public surface — in __all__, or imported elsewhere — is no longer there. Callers outside the change set cannot be seen by a diff, so this asks rather than assumes.
The cheapest way to make a suite pass is to stop running it. These watch the machinery rather than the code.
ci_check_disabledescalateA step is still there but can no longer fail: || true, continue-on-error, if: false. Reported as disabling the same step rather than as a removal plus an addition.
ci_check_removedescalateA step disappeared from a workflow file. Deleting the job that runs the tests is the most effective way to make the tests pass.
Maintainability, not correctness. These are reported but do not fail a run: a commit refused because a function is fifty-one lines teaches people to reach for --no-verify, and that habit mutes the rules that matter too. Set structure.gates true to make them binding.
change_too_largerepairAdded lines past the configured ceiling. Aimed at agents, which do not tire and will happily produce a forty-file diff nobody can read.
complexity_too_highrepairDefault 10.
duplicate_implementationrepairTwo functions with the same shape. Two copies drift apart, and a fix applied to one is a bug left in the other.
file_too_longrepairDefault 300 lines of code, excluding blanks and comments.
function_too_longrepairDefault 50 lines.
generated_file_editedrepairA file marked as generated was edited directly. The edit is lost on the next regeneration, which is a bug that appears later and somewhere else.
nesting_too_deeprepairDefault 4 levels.
too_many_parametersrepairDefault 5.
utility_modulerepairA utils-shaped file: unrelated functions with nothing in common but the fact that nobody knew where else to put them.
Every rule is sometimes right about the code and wrong about the intent. Say so in the source, with a reason, and the finding is suppressed for that line only — and counted, so suppression cannot rot quietly.
# yieldpoint: allow assertion_monotonicity - broadened to cover YAML clients
def test_every_client_resolves_a_path():
...One rule per comment, a reason required, and the comment must sit on the line the finding points at or within two lines above it.