uroboroslisted

# uroboros ## Governing law This is skill law for the planner seat on every invocation, every wave, and every fix. It is not per-run content: never restate it in `TASK.md`, and no plan can waive it. A plan that appears to authorize skipping or combining a step is itself the defect. 1. **Build** — Codex writes in isolation. The planner never implements. 2. **Evidence verification** — Confirm the true exit code of every declared command. The harness runs each once and records full output in `__uro_evidence/`; exit codes are evidence for the seats, never a verdict. Stdout text is not status. Never accept a piped exit code: a pipe reports the last exit code, not the command's. 3. **Adversarial review** — Cursor writes ONE holistic report covering the full hunt list: correctness, regressions, security, edge cases, test adequacy, and violations of intent, scope, or invariants — an unmet requirement is a finding like any other. Claude then validates each blocking finding read-only. An invalid finding is recorded as overruled; an unavailable or unreadable arbiter preserves the objection. 4. **Correction loop** — The planner authors *finding → fix design → mutation pin*; Codex implements; repeat. A mutation pin proves that a new test could have failed: inject the defect, observe the specific assertion fail, restore the implementation, observe green, and record both the failing and green counts. 5. **Scoped re-verify** — Cursor re-checks the correction