3.14.0

Added

• Evals validity scenarios: Added the
  ull-hypothesis-rate-limiter\ and \ile-reconstruction-expense-tracker\ scenarios to validate the eval harness itself, and \�val scenario create\ templates now carry structural assertions through to the created scenario.

Changed

• Codex initialization: \jumbo init, repair, and \jumbo evolve --yes\ now install Codex-managed skills to the documented repository skill location .agents/skills\ while keeping .codex\ limited to Codex hooks/configuration.

Fixed

• Codex managed artifact preservation: Obsolete Jumbo-managed Codex skill copies under .codex/skills\ are removed only when byte-identical to the current managed templates, preserving customized skill content, user files, hook comments, unknown hook settings, and unrelated .codex\ configuration.
• Evals headless permissions: The Claude Code eval harness adapter now runs sessions with --dangerously-skip-permissions, since headless -p\ runs in isolated temp workdirs have no human to approve tool use and were leaving both A/B arms unable to write files.