Explore

Build a real eval suite for an agent or prompt: cases, graders, a baseline, and a regression gate

Find why a test is flaky rather than retrying it: classify the cause, prove it, fix it, and quarantine honestly

Structured PR review with risk classification, a correctness/security/test/perf checklist, and severity-tagged output

Red/green loop that picks an untested module, writes a failing test, implements it, and iterates with coverage deltas