A rerun is a clue

If the same revision passes and fails without intentional changes, capture what differs: timing, data, ordering, dependencies, or runner resources.

Preserve the evidence

Keep traces, screenshots, and relevant logs from the original failure. A passing retry should not erase the information you need to investigate.

Look for shared state

Tests that work alone but fail in a suite often share mutable resources. Use unique data and clean boundaries instead of relying on execution order.

Restore trust deliberately

Track flaky scenarios, give each an owner, and record fixes. Retries can keep delivery moving temporarily, but the goal is to make the first result meaningful again.