HoangNguyen0403 Quality Engineering: Test Healing
Priority: P0 (CRITICAL)
Failure Classes
SELECTOR_DRIFT (id/locator changed) · TIMING_SYNC (race/wait, no state change) · DATA_ENV (fixture/seed/env stale) · INFRA (network/runner/flake) · REAL_REGRESSION (product behavior actually changed).
Classify from evidence: trace/screenshot/DOM diff supports a drift/timing/data explanation, or the product diff shows an intentional behavior change (REAL_REGRESSION).
Allowed Repairs
Move the locator up the selector ladder (e.g., getByRole if the ID changed); replace a sleep with an explicit state wait; fix a stale fixture/seed; retry only for INFRA.
Forbidden Repairs
Never weaken an assertion. Never add test.skip/fixme; a test that must stop gating goes to quarantine per quality-engineering-flaky-triage with a ticket + expiry, and keeps running. Never widen a matcher. Never inflate a timeout by more than 2x. Never blind --update-snapshots. Never catch-and-continue. Never touch production code — that is REAL_REGRESSION, not a heal.
Verdicts
HEALED (repair verified by 3 consecutive green reruns, ASSERTION_DELTA: none) · REAL_BUG_DO_NOT_HEAL (route to dev-fix) · QUARANTINE_CANDIDATE (flaky, route to flaky-triage) · BLOCKED (no evidence artifact, or no stable locator target: route to specialist-testid-inserter).
Red Flags
"the assertion was too strict anyway" · "the product changed so update the expected value" · "just add retries so it goes green" — do not heal with these rationalizations. All three are REAL_BUG_DO_NOT_HEAL or QUARANTINE_CANDIDATE in disguise, never HEALED.
References
- Failure Taxonomy Signals
- Repair Catalog
- Forbidden Repairs
- Verdict Contract
- Flaky Triage — quarantine contract for
QUARANTINE_CANDIDATE - Visual Baseline — screenshot failures before
REAL_REGRESSION