Test Smell Detection

Test Smell Detection

Audit test code with the academic taxonomy, code evidence, calibrated framework idioms, and fixes native to the codebase.

Scope

• Audit only staged or named tests. Search the current workspace before asking for code; never claim a file is missing until that search finds no relevant test.
• Read production code only when it changes a verdict.
• For unfamiliar framework APIs, call test-analysis-extensions and read the matching language extension.
• Read the complete catalog when the caller requests all 19 smells, asks for citations, or the code may contain a smell outside the high-signal set below. Do not load it for a narrow question that this file answers.

Audit Workflow

1. Search for and read the staged tests; detect language, framework, boundaries, and integration markers. This is the first action even when no path is named.
2. Read the tests and only verdict-changing production context.
3. For each candidate, verify the executed check, choose the formal category, calibrate, then assign severity. Report proven non-catalog test-validity defects separately; do not relabel them as smells.
4. Rank confirmed findings by risk of false confidence or flakiness, then by maintenance cost.
5. Give a framework-correct replacement for each actionable finding. Never use .NET terminology or APIs in another ecosystem.

High-Signal Decisions

Calibration Rules

Apply these before assigning a finding:

• Mock-call verifications, snapshots, bare pytest assert, Pester Should -Invoke, and expected-exception constructs are assertions.
• A literal or snapshot assertion may expose a coverage gap, but is not Unknown Test or another smell without separate evidence.
• Count assertion statements. One assertion is never Assertion Roulette; missing messages alone are not a smell.
• Same-method tests are not Lazy Test when they cover distinct behaviors, boundaries, or state; require redundant equivalent paths.
• General Fixture requires shared lifecycle state. Repeated local construction is neither General Fixture nor Test Code Duplication by itself.
• Treat strings returned by the public API as observable contract unless production context or requirements make them display-only; interpolation alone is not Sensitive Equality.
• Magic Number Test requires an unexplained oracle value. Do not flag ordinary setup quantities whose role is locally obvious and irrelevant to the asserted behavior.
• Go table-driven subtests, pytest/JUnit/xUnit parameterization, Jest/Vitest .each, RSpec data tables, Pester -ForEach, and Catch2 SECTION/GENERATE are not Conditional Test Logic by themselves.
• Go's if err != nil { t.Fatal(...) } is idiomatic assertion flow, not Exception Handling.
• Integration markers legitimize declared external resources and multi-step flows, but not fixed sleeps or assertion-free execution.
• For integration tests, trace helper factories and repository return types before judging branch coverage. If a test branches on a subtype or state that the real helper can never produce, report the unreachable assertion path and explain the resulting false confidence. Do not stop at assigning the generic Conditional Test Logic label.
• A local temporary file still meets the formal Mystery Guest definition. Hermetic creation and cleanup reduce its severity; they do not change its taxonomy.
• A formatting name does not prove display text is the stable contract; confirm it from production behavior or requirements before clearing Sensitive Equality.
• Do not infer a smell from method names alone. Point to the statement or fixture relationship that proves it.
• Do not infer a high-severity non-catalog validity defect from a test name alone. Without production code or an explicit contract proving that the test is supposed to invoke another component, a name/body mismatch is at most an unranked observation, not evidence that the test silently passes broken production behavior.
• Catch2 SECTION and GENERATE are runner-controlled case expansion, and REQUIRE is a real assertion. When those are the only suspicious constructs, the academic-smell verdict is clean. Do not reverse that verdict because the test could have broader behavioral coverage.
• If no material smell remains after calibration, say that clearly. Never manufacture findings to fill a report.
• Never propose await for a void or otherwise non-awaitable API. If production work is synchronous, remove the sleep and assert immediately.