Grill runs six specialized agents (recon, architecture, error handling, security, testing, edge cases) in parallel against a codebase, then synthesizes their findings into one report with severity tags, file:line evidence, proposed changes, effort estimates, and explicit tradeoffs. Five review styles let you tune intensity from quick scan to Paranoid Mode. Eight optional add-on pressure tests target specific failure modes (concurrency, observability, dependency hygiene, etc.). Zero-findings is reported explicitly as `[GOOD]` — no padding with manufactured issues.