cld — Claude for Codex

Full-surface Claude Code integration for Codex: delegate, review, fan out, resume, and act on structured Claude results. Codex drives the claude CLI directly through a context-safe bundled runner — no middleman runtime.

The inverse of cdx (which lets Claude Code drive Codex).

What you get

Skill	Purpose
driving-claude	The brain: invocation contract, flag heuristics, sessions, fleet, verification. Auto-triggers on any "ask/have/let Claude…" request
claude-task	Delegate one task with full flag control
claude-review	Schema-backed code review with verified findings
claude-fleet	2–4 parallel Claude workers (decomposition, angles, A/B)
claude-session	List, resume, fork Claude sessions
claude-setup	Health-check install/auth/runner
prompting-claude	How to prompt Claude models well
claude-structured-output	--schema patterns + bundled schemas

Invoke explicitly with $ in Codex (e.g. $claude-review) or just describe the task ("get a second opinion from Claude on this diff").

Requirements

• Claude Code CLI installed and authenticated (claude on PATH; run it once interactively to log in). Verified against Claude Code 2.1.207.
• Node.js (for the bundled runner).
• Codex ≥ 0.144 / ChatGPT desktop app with plugins.

Install

ChatGPT desktop app

Add the Nuko Nova marketplace and install:

codex plugin marketplace add nuko-nova-dynamics/marketplace
codex plugin add cld@nuko-nova-tools

Then restart the ChatGPT desktop app; cld appears under Plugins. Bundled skills load in new Codex chats and CLI sessions.

How it works

Every delegation goes through scripts/claude-run.mjs, which wraps claude -p --output-format json and prints a compact summary (session id, status, tokens, cost, permission denials, final message) plus artifact paths for the full result. Permission tiers:

Tier	Claude permissions
--sandbox ro	read files + read-only shell (git diff/log/…, ls, cat, rg) + web; nothing else
--sandbox write	auto-accepted file edits + shell
--sandbox full	--dangerously-skip-permissions (explicit user intent only)

--sandbox full gives the delegated Claude process unrestricted command and filesystem access and bypasses its normal permission prompts. Use it only after an explicit request, in a trusted workspace, and with the exact task and target reviewed. Prefer ro for inspection and write for ordinary implementation.

Structured output: --schema <file> forces the result to validate against a JSON Schema (bundled: review-findings, verdict, task-report, patch-plan). Any claude flag the runner doesn't wrap passes through verbatim with --raw <arg>. Full verified reference: skills/driving-claude/references/flag-map.md.

Smoke test

node tests/smoke.mjs   # runs 3 cheap live probes against claude (costs a few cents)

License

Apache-2.0