test-anti-patterns

Test Anti-Pattern Detection

Quick, pragmatic analysis of test code in any supported language for anti-patterns and quality issues that undermine test reliability, maintainability, and diagnostic value.

Language-specific guidance: Call the test-analysis-extensions skill to discover available extension files, then read the file matching the target codebase (e.g., extensions/dotnet.md, extensions/python.md, extensions/typescript.md, extensions/go.md). The extension file tells you which sleep / time / random / skip / setup-teardown / mystery-guest APIs to look for in that language.

When to Use

• User asks to review test quality or find test smells
• User wants to know why tests are flaky or unreliable
• User asks "are my tests good?" or "what's wrong with my tests?"
• User requests a test audit or test code review
• User wants diagnostic findings before deciding what to improve

When Not to Use

• User wants to write new tests from scratch (use code-testing-agent)
• User wants direct implementation fixes rather than a diagnostic review (use the relevant write/edit skill)
• User asks to fix swapped Assert.AreEqual argument order in MSTest (use writing-mstest-tests)
• User asks to convert MSTest DynamicData from IEnumerable<object[]> to ValueTuple (use writing-mstest-tests)
• User wants to run or execute tests (use run-tests for .NET)
• User wants to migrate between test frameworks or versions (use migration skills)
• User wants raw .NET coverage collection (use run-tests), non-.NET coverage collection or analysis (use native tooling), project-wide .NET coverage/CRAP metrics (use coverage-analysis), or named-target .NET CRAP (use crap-score)
• User asks whether tests would catch a bug or wants behavioral/pseudo-mutation gaps (use test-gap-analysis)
• User wants test-mix or happy-vs-error-path classification, standardized tagging, or trait/category distributions (use test-tagging)
• User wants a deep formal test smell audit with academic taxonomy and extended catalog (use test-smell-detection)

Inputs

Workflow

Step 1: Detect language and load extension

Identify the target codebase's language and test framework. Call the test-analysis-extensions skill and read the matching extension file. The extension file documents framework-specific anti-pattern markers — what counts as a sleep/wait, a test marker, a skip, a setup/teardown, a shared-state hot spot, and an integration boundary — so this skill stays language-neutral.

Step 2: Gather the test code

Read the test files the user wants reviewed. If the user points to a directory or project, scan for all test files using the discovery markers in the loaded language extension file (e.g., [TestClass]/[Fact]/[Test] for .NET, test_*.py / def test_* for pytest, *.test.ts / it() for Jest, *Test.java / @Test for JUnit, *_test.go / func TestXxx for Go, *_spec.rb for RSpec, #[test] for Rust, *.Tests.ps1 / Describe for Pester, TEST(...) for GoogleTest, TEST_CASE(...) for Catch2/doctest).

If production code is available, read it too -- this is critical for detecting tests that are coupled to implementation details rather than behavior.

Step 3: Scan for anti-patterns

Check each test file against the anti-pattern catalog below. Report findings grouped by severity. The examples are .NET-centric but the patterns generalize — use the loaded language extension file to map each pattern to the framework you are auditing.

Critical -- Tests that give false confidence

High -- Tests likely to cause pain

Medium -- Maintainability and clarity issues

Low -- Style and hygiene

Step 4: Calibrate severity honestly

Before reporting, re-check each finding against these severity rules:

• Critical/High: Only for issues that cause tests to give false confidence or be unreliable. A test that always passes regardless of correctness is Critical. Flaky shared state is High. Missing-await on async assertions is Critical (silent pass).
• Medium: Only for issues that actively harm maintainability -- 5+ nearly-identical tests, truly meaningless names like Test1 / test / it1.
• Low: Cosmetic naming mismatches, minor style preferences, assertion messages that could be better. When in doubt, rate Low.
• Use the caller's severity vocabulary consistently. If the caller asks for Critical / Warning / Info, map reliability risks to Warning and maintenance/cosmetic concerns to Info rather than silently collapsing every item into Critical. Severity describes the demonstrated failure mode, not how much prose a finding receives.
• Separate a systemic finding from its instances. Coverage touching across a facade is one Critical systemic finding whose evidence lists every affected test. All assertion-free instances, including the last facade method, retain the same false-confidence severity. Report 1 finding / 6 affected tests, not six findings plus a seventh summary finding, and do not downgrade one instance merely to manufacture multiple tiers.
• Do not severity-rank ordinary missing cases as anti-patterns. Adjacent untested branches, exception paths, and boundaries may be useful coverage opportunities, but list them separately from the anti-pattern counts unless a weak existing test specifically creates the gap. They are not Critical merely because the suite has a systemic Critical issue.
• Not an issue (per-language nuance):
  • Go and Rust table-driven loops with sub-tests (t.Run / for case in cases { ... }) are idiomatic, not "Conditional Test Logic". Do NOT flag.
  • pytest bare assert is the canonical assertion form, not a missing assertion library. Do NOT flag.
  • Go tests use if got != want { t.Errorf(...) } as canonical equality. Do NOT flag as ad-hoc.
  • Separate tests for distinct boundary conditions (zero vs. negative vs. null). Do NOT flag as duplicates.
  • Explicit per-test setup instead of [TestInitialize] / beforeEach (this improves isolation).
  • Tests that are short and clear but could theoretically be consolidated.
  • Round-trip or serialization equality with non-trivial input. It is valid metamorphic evidence; suggest an independent representation assertion when two implementations could share the same bug.
  • Clone value equality. Keep it, and add distinct-reference or mutation- independence evidence when the contract promises a deep copy.
  • A validator or accessor returning the original value when pass-through is the production contract. Missing invalid-input cases are a coverage gap, not proof that the existing assertion is tautological.

IMPORTANT: If the tests are well-written, say so clearly up front. Do not inflate severity to justify the review. A review that finds zero Critical/High issues and only minor Low suggestions is a valid and valuable outcome. Lead with what the tests do well.

Step 5: Report findings

Depth bar — a tidy report that is shallower than an unassisted review is a failure. Before writing, satisfy all five:

1. Account for every test in scope. Build the complete method/field inventory before summarizing. For a systematic pattern such as coverage touching, enumerate every affected test at least once rather than giving representative examples. A finding table that silently skips tests (or fixtures like an unused static HttpClient field) is incomplete. State the number reviewed.
2. Verify the production contract before judging the oracle. Inspect the actual transformation, DTO fields, and promised identity/clone semantics. Never invent fields or require lossless round-tripping when production is intentionally lossy.
3. Make every Critical/High fix complete and specific. Give the replacement assertion with the exact expected value (the computed discount, the exact CSV line, the full expected object), not a // assert something here placeholder.
4. Name the adjacent gaps the tests should also cover — untested error paths, boundary values, and round-trip/culture-sensitivity risks in the same class. These are part of "what's wrong with my tests", and omitting them is the most common way this review loses to an unassisted one.
5. Keep the report internally consistent. Summary counts must equal the enumerated findings. Publish a settled conclusion: do all reconsidering before you write, and never leave "wait, that's wrong" / "this should fail but doesn't" reasoning in the output.

Present findings in this structure:

1. Summary -- Total issues found, broken down by severity (Critical / High / Medium / Low). If tests are well-written, lead with that assessment.
2. Critical and High findings -- List each with:
   • The anti-pattern name
   • The specific location (file, method name, line)
   • A brief explanation of why it's a problem
   • A concrete fix (show before/after code when helpful)
3. Medium and Low findings -- Summarize in a table unless the user wants full detail
4. Positive observations -- Call out things the tests do well (sealed class, specific exception types, data-driven tests, clear AAA structure, proper use of fakes, good naming). Don't only report negatives.

Before publishing, assign each finding a stable identity. A grouped row counts as one finding regardless of how many methods it lists; separate rows count separately. Recompute the summary from those rows. Keep affected tests as a different number so a bundled finding cannot create a hidden count mismatch.

Step 6: Prioritize recommendations

If there are many findings, recommend which to fix first:

1. Critical -- Fix immediately, these tests may be giving false confidence
2. High -- Fix soon, these cause flakiness or maintenance burden
3. Medium/Low -- Fix opportunistically during related edits

Validation

• Every test method in scope is accounted for (reviewed count stated; none silently skipped)
• Identity and round-trip findings match the production contract and use only real fields
• Every finding includes a specific location (not just a general warning)
• Every Critical/High finding includes a concrete fix with exact expected values
• Adjacent untested error paths and boundary values are called out
• Summary counts match the enumerated findings
• Grouped findings distinguish finding count from affected-test count
• Adjacent coverage opportunities are not inflated into Critical anti-pattern findings
• Report covers all categories (assertions, isolation, naming, structure)
• Positive observations are included alongside problems
• Recommendations are prioritized by severity

Common Pitfalls