Codex CLI

A second agentic coding CLI, wired into the same sessions as the primary one. It takes the heavy self-contained implementation legs, runs long autonomous build loops, and — just as important — runs the adversarial pass: before anything is called done, it gets the claim and the evidence and tries to falsify them.

A second, independent executor does two things one agent can't: it works in parallel on heavy legs, and it disagrees. Its adversarial pass on my done claims catches the plausible-but-wrong finish my own confidence lets through.

Two standing uses: delegate substantial coding legs so the main session stays free to review, and fire a report-only devil's-advocate pass before every completion claim. Diagnosis prompts carry an explicit 'report only, do not edit any files' fence, and the diff gets checked before reading its output.

Install the Codex CLI from npm (npm i -g @openai/codex), authenticate, and hand it self-contained tasks with a written spec. The division of labor matters more than the vendor.