| name | test-pruner |
| description | Audit and safely clean low-value, redundant, misleading, flaky, over-mocked, or AI-generated unit tests. Use when asked to clean useless tests, prune test suites, remove redundant tests, audit low-value tests, improve test signal, clean brittle snapshots, reduce test noise, review coverage theater, or decide whether tests should be kept, rewritten, consolidated, moved, quarantined, or deleted. |
Test Pruner
Delete false confidence, not useful protection. Tests are code: remove or rewrite them only when their maintenance cost exceeds their regression and documentation value.
Operating Mode
Default to audit-only. Do not delete or rewrite non-generated tests until the user approves the proposed actions.
Use safe-cleanup only after approval, and only for high-confidence items. Use deep audit when configured tooling exists or risk is high: mutation testing, necessist, assertion-density scripts, flake history, or coverage reports.
Workflow
- Scope target files, test framework, test command, changed files, and production code under test.
- Capture baseline:
git status, relevant test results, known failures, and configured coverage/mutation tools.
- Inventory tests: map each test to behavior, production unit, assertions, mocks, snapshots, skips, runtime, and fixtures.
- Classify each suspect test with
KEEP, REWRITE, CONSOLIDATE, MOVE LEVEL, DELETE, or QUARANTINE.
- Require evidence before deletion. Coverage stability alone is never evidence enough.
- Prefer preserving unique behavioral intent by merging edge-case data into stronger tests before removing duplicates.
- Validate affected tests before and after any approved cleanup. Run broader suite when shared fixtures/helpers changed.
- Report actions, evidence, validation, and remaining risk.
Evidence Gate
Before marking DELETE, show at least two evidence types:
- No unique behavior compared with a stronger test.
- No meaningful assertion or assertion cannot fail for wrong behavior.
- SUT is fully mocked or bypassed.
- Feature/contract is removed, dead, or obsolete.
- Snapshot is unreviewable and duplicates explicit assertions.
- Test is flaky or slow and lower-level or stronger coverage exists.
- Tool-backed signal: survived mutant, removable statement, assertion-density failure, or repeated flake history.
Safety Rules
- Never delete failing tests blindly; first decide whether they expose a real regression.
- Never delete public API, compatibility, migration, serialization, security, or bug-regression tests just because they look simple.
- Never delete a duplicate until unique data or edge cases are merged or proven irrelevant.
- Never blindly update snapshots; replace behavior-relevant snapshots with explicit assertions.
- Never quarantine without owner, issue or TODO, expiry, and fix/delete path.
- Treat AI-generated tests as suspicious, not worthless.
- If deletion risk is unclear, return an audit report and ask for approval.
References
- Read rubric.md before classifying tests for
DELETE, REWRITE, CONSOLIDATE, MOVE LEVEL, or QUARANTINE.
- Read patterns.md when a test shows weak assertions, mocked SUT, duplicate behavior, brittle snapshots, flaky behavior, obsolete behavior, or AI-generated test smells.
- Use report-template.md for audit reports with multiple findings or any proposed deletion.
- Do not load patterns.md for a single obvious generated artifact or naming-only question.
- Do not load report-template.md for quick yes/no triage unless deletion is proposed.
