What problem does it solve?
It reduces the risk that retikz specs pass while real AI-generated inputs still break the implementation, by systematically hunting for adversarial IR/JSON/zod-contract failures before documentation work begins.
Core Features & Use Cases
- Adversarial Bug Hunter sub-agent: dispatches an independent “opinion-shifted” agent to generate new input cases designed to make the implementation fail.
- Blocking vs non-blocking triage: classifies results into BLOCKING (must-fix) and WARNING/INFO (review/backlog) based on contract-critical failure modes like JSON round-trip, zod error clarity, schema invariants, and discriminator stability.
- Fix-to-test feedback loop: converts discovered BLOCKING issues into formal spec tests and re-runs the adversarial phase to confirm convergence.
Quick Start
Run the develop-test stage by selecting an ADR still in Proposed state, ensuring lint/tsc/vitest are currently all passing, then prompt the system to generate adversarial vitest cases and report any BLOCKING items.