tddlisted

# Cross-model test-driven implementation Tests-first is the single biggest quality lever in agentic coding, and tests designed by a model that is NOT the implementer are measurably more objective (the implementer's tests inherit its own misunderstandings). You implement; Codex designs tests. ## Procedure 1. **Extract the behavioral contract.** From the approved plan/spec, write the contract: inputs, outputs, invariants, edge cases, failure modes. This is WHAT, never HOW - do not include your intended implementation approach. If `${RITUAL_INVARIANTS_FILE:-.ritual/invariants.md}` exists and has bullets, fold in every invariant the change touches - each must yield at least one test. 2. **Codex designs the test list.** Call the `codex` MCP tool with ONLY the behavioral contract (not your implementation plan - the test designer must not inherit your bias): > Design a test list for this specification. Cover: happy paths, boundary/edge cases, failure modes and error handling, concurrency/ordering issues if applicable, and candidates for property-based tests. For each test: a name, given/when/then in one line each, and what real bug it could catch. Flag anything in the spec that is untestable as written. If the `codex` tool fails because the MODEL is unavailable (model-not-found / unsupported model - NOT an auth error), retry the same call ONCE with `model: "gpt-5.5"` (verified fallback) and note the downgrade when reporting the merged list. Codex's default is deliberatel