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.

Severity

Severity follows demonstrated risk, not a fixed label copied from the catalog:

• High: can silently pass while behavior is broken, creates nondeterministic failures, or hides unexecuted assertion paths.
• Medium: makes failures ambiguous or couples tests to unstable details.
• Low: primarily maintenance debt, such as a reasoned skip or over-broad cheap fixture.

State the reason for the assigned severity. Downgrade or omit a finding when the surrounding test type makes the pattern intentional.

Output Contract

Scale the response to the input:

• For one to three files, give a verdict and one compact table: severity, formal smell, evidence, risk, and fix.
• For larger suites, add counts and a short priority order. Do not repeat findings across dashboards, prose, and plans.
• Show code only when it clarifies a fix; omit unchanged setup.
• Add brief Not findings only for plausibly suspicious idioms.
• Put proven non-catalog validity defects after the academic-smell verdict. On a clean input with no production contract, do not assign severity to optional coverage observations or let them overturn the clean verdict.
• Do not narrate discovery or catalog loading; return the audit directly.

Every reported smell must have a formal taxonomy name, precise location, evidence from the code, practical risk, and a concrete framework-correct fix.

Validation

• Every finding is supported by code, not a keyword or method name.
• Unknown Test and Empty Test remain distinct.
• Every disabled test remains Ignored Test, and every local file dependency remains Mystery Guest; rationale and hermetic cleanup change severity only.
• Framework idioms and integration boundaries were calibrated before reporting.
• Clean tests and suspicious-but-valid idioms are not turned into filler.
• Clean framework idioms are not converted into high-severity non-catalog findings from naming or absent production context.
• Fixes use the target framework's APIs and preserve the behavior under test.
• Claims about files reviewed, builds, or test runs match actions actually performed.