Codex Delegate

Hand a task to the gpt-5.6 family through the Codex CLI. This is the delegation path per MODELS.md: the main loop (Root) invokes Codex directly through Bash - no Claude shim, and never a Claude worker whose only job is to launch, wait for, or relay a Codex call.

Delegation gate (from MODELS.md - check before delegating at all)

Delegate only when at least one shape applies: independence (fresh-eyes or adversarial value), context (repo/research mass stays out of the root context), contract (clear-spec implementation with a written contract, roughly >=2h), or parallel (independent taste-free lanes save wall-clock). Otherwise Root works inline. Cap concurrent model lanes at 2-3.

Never route design/UI/copy/naming/API-ergonomics or final architecture decisions to codex.

Model + effort routing (always pin BOTH -m and effort - never inherit defaults)