Skip to content

What Mjölnir checks ​

Mjölnir looks for evidence that your verification cannot be trusted. Not style, not formatting — the specific places where a test suite or a pipeline reports success it hasn't earned.

The complete, always-current list lives in the rule catalog — generated from the live registry, filterable by severity, tier and language. This page explains the shape of what it looks for and how to read it.

The ten families ​

Each family name links into the rule catalog, pre-filtered.

FamilyPrefixWhat it catches
CI integrityQA-CI-continue-on-error, || true, swallowed exit codes — the tricks that turn a red pipeline green
Test hygieneQA-TEST-Committed .only, unjustified skips, empty bodies, tests with no assertions
Test qualityQA-TQUAL-Tautological assertions, mock-only verification, unawaited promises
PlaywrightQA-PW-Brittle selectors, networkidle waits, hardcoded environment URLs, debug artifacts
Python / pytestQA-PY-time.sleep(), non-strict xfail, random/time dependence without a freeze
Java / JUnit · TestNGQA-JV-@Disabled, Thread.sleep(), blanket route mocks
C# / .NETQA-CS-[Ignore], Task.Delay, WaitForTimeoutAsync()
CypressQA-CYP-Cypress-specific false greens: fixed waits, focused and forced interactions
SeleniumQA-SE-Selenium WebDriver anti-patterns: hard sleeps and brittle locators
EnvironmentQA-ENV-Environment coupling that makes a test pass only on one machine

The false-positive firewall ​

Every rule ships with must-fire and must-not-fire fixtures. A rule that fires on its own negative fixture cannot ship — that is the firewall, and it is enforced by mjolnir doctor, not by convention.

Detection also runs on a comment- and string-free view of the source (TypeScript rules use the compiler AST). A pattern inside a prose comment or a doc-example string is documentation, not a finding.

Tiers — and what they actually mean ​

The tiers are internal. What you see on a finding is GATE (it counts against your gate and can fail CI) or WARN (reported, never gating), assigned from the measured false-positive rate, not from how confident the author feels:

TierMeaningDefault scan--include-warn
core≤ 10 % measured FP✅✅
extended≤ 30 % measured FP✅✅
quarantineabove 30 %, or not yet measured (n < 10)❌✅

--include-warn was --strict until 6.0. The old name is deprecated and warns; it widens what the scan sees, never what fails.

How much of this is measured ​

73 of the 79 active rules carry a false-positive rate measured against real OSS code; the remaining 6 are explicitly unmeasured and non-core:

  • The rule catalog marks every rule either with its measured rate or as on assumption — filter with Measured only to see the audited set.
  • Every scan footer reports how many of the rules that fired are measured.
  • mjolnir explain --list --unmeasured lists the ones that aren't.
  • Every rule's mjolnir explain page states its status.

Rates are published even when they're unflattering — a rule that audits badly gets quarantined for it. Method and the full table: false-positive audit.

Language maturity ​

LanguageAdapterCoverage today
TypeScript / JScompiler AST (ts-morph)broadest, most measured — mostly core/extended
Python / pytestcomment/string-masked regexbroad, corpus-audited — mostly core/extended
Javacomment/string-masked regexnewer — mostly extended/quarantine
C# / .NETcomment/string-masked regexnewer — mostly extended/quarantine

TypeScript and Python have the broadest measured coverage. Java and C# ship, are documented, and stay out of the headline number until a real consumer suite — not a binding library's own tests — has been audited.

Getting the catalog locally ​

bash
mjolnir explain --list --md          # full catalog as Markdown
mjolnir explain --list --unmeasured  # only the rules running on assumption
mjolnir explain QA-CI-001   # what / why / fix + FP status for one rule

Released under the MIT License.