Codex Delegate

Runs the codex CLI (user is already authenticated) as a single, sequential external subagent. One Codex conversation thread is opened per task and resumed for every follow-up step — never more than one Codex process running at a time, and never more than one thread per task.

When to use this

Use it for: a real implementation task, a multi-file refactor, a from-scratch investigation/audit, or anything where you'd otherwise spin up a heavyweight Agent subagent — i.e. medium-to-high importance work worth a second model's dedicated attention on the actual codebase.

Don't use it for: a one-line fix, a question you can answer by reading a file yourself, or anything where the overhead of a round-trip (tens of seconds, a fresh model spin-up) outweighs the benefit. Don't reach for this reflexively on every task — treat each delegation like you would a subagent dispatch: deliberate, not default.

Don't parallelize it. This skill is explicitly sequential: one Codex session, one turn at a time. If you have several genuinely independent tasks, either run them through this skill one after another, or use Claude's own Agent tool for concurrent work — don't run multiple codex exec processes against the same thread or the same working tree at once.

Prerequisites

The helper requires jq and a writable per-session state path. Existing state is never overwritten by start; resume it with send. Calls must be sequential.