OWNED SKILL · DELEGATE-TO-CODEX

Same work.
Someone else's meter.

Hands a scoped task to the OpenAI Codex CLI, running headless. It reads files, edits code, runs commands — and bills the whole thing to your ChatGPT subscription.

say "delegate to codex"

CLAUDE TOKENS

~0 spent

quota + context preserved

BILLED TO

ChatGPT sub

compute on OpenAI's dime

The problem

Token-heavy work — a long refactor, a research sweep, six independent tasks — burns Claude quota and context on exactly the jobs that deserve it least. The work isn't hard, it's just big. And the free lunch won't last forever.

The turn

Think of it as the Agent tool, but the compute is on OpenAI's dime. Codex runs as a headless sub-agent, does the real work, and hands back a result to relay. Because it authenticates with your ChatGPT subscription, none of it touches Claude tokens — that's the entire point.

Reach for it when

A task is token-heavy or long-running and quota matters.
You want an independent second opinion from a different model on a diff or a design.
Several independent tasks can fan out cheaply.

The receipt

One preflight proves the whole premise — if it doesn't say ChatGPT, the economics don't hold:

# want: "Logged in using ChatGPT" → billed to ChatGPT, not Claude
codex login status

The catch: logged in with an API key instead? Then it spends real OpenAI money. The saving still holds against Claude, but that's a different trade — so it gets flagged, never assumed.

Don't reach for it when

The user asked Claude to do it. Delegating isn't a loophole around the ask.
The job is trivial — the ~5–10s Codex startup costs more than it saves.
Codex isn't authenticated.

Part of Rascal's AI setup · one atom, justifying its own existence.

Same work.Someone else's meter.

The problem

The turn

Reach for it when

The receipt

Don't reach for it when

Same work.
Someone else's meter.