claude-code

name	claude-code
description	Drive Anthropic's Claude Code CLI (`claude -p`) as a non-interactive coding sub-agent from inside Codex. Use when you want to delegate a coding/analysis/refactor task to Claude, get a second opinion / adversarial review from another model, hand off long-context planning, or fan out parallel agents. Covers the exact `claude -p` flags, permission modes, output capture, JSON/stream-json modes, session resume, parallel fan-out, and the mandatory "delegate → capture → independently verify, never trust the self-report" supervision loop.
metadata	{"short-description":"Delegate tasks to the Claude Code CLI"}

Claude Code as a Sub-Agent

Claude Code is Anthropic's terminal coding agent. Its headless mode, claude -p (a.k.a. --print), makes it a disposable sub-agent: you write the metaprompt, Claude explores/edits/runs commands on its own, and returns a final message you capture and verify. You stay the planner; Claude is the implementer — it's a strong long-context planner, so this is useful for reviews, big-picture design, and cross-cutting refactors.

Golden rule: Claude's self-report is a claim, not the truth. Capture its output, then independently verify the artifacts yourself (read the files, re-run the tests) before treating the task as done.

Verified locally: claude 2.1.167. Confirm install with claude --version.

The core command

claude -p "your metaprompt here" \
  --add-dir <workdir> \
  --permission-mode <default|acceptEdits|plan|bypassPermissions> \
  --output-format json  >/tmp/claude_<rand>.json 2>>/tmp/claude_<rand>.log