codex-dispatch

codex-dispatch

Route a prompt to OpenAI Codex (gpt-5.6 sol / luna / terra) from Claude Code and get a synthesized answer back — like delegating to a subagent. The Claude main agent dispatches a self-contained task, Codex runs it (with real file access), and Claude digests the result into a concise synthesis with its own judgment layered on top.

Install

/plugin marketplace add 0-to-1-Labs/claude-marketplace
/plugin install codex-dispatch@0to1-labs

Requires the codex CLI installed and authenticated:

codex login

Usage

/codex <prompt>                      # defaults to gpt-5.6-sol, workspace-write
/codex luna <prompt>                 # pick a variant
/codex terra <prompt>                # deepest reasoning
/codex ro <prompt>                   # read-only / advisory — Codex can inspect but not edit files
/codex ro terra <prompt>             # modifiers combine, in any order

Variants

Variant	Model id	Use for
sol	gpt-5.6-sol	Fast default driver
luna	gpt-5.6-luna	Balanced
terra	gpt-5.6-terra	Deepest reasoning — hard bugs, architecture

Modes

• Default (workspace-write): Codex runs in your current directory and can read and modify files — a real worker, not just an oracle.
• ro / read-only: Codex inspects files but only returns advice. Nothing is modified.

How it works

1. /codex (or the codex-dispatch skill, invoked autonomously) parses the variant and mode.
2. skills/codex-dispatch/dispatch.sh runs codex exec non-interactively (approval_policy=never, so it never hangs), streaming the trace and writing two artifacts to the session scratchpad:
   • the clean final message (-o)
   • the full event log (reasoning + tool calls + output)
3. The Claude main agent reads the final message, synthesizes it in its own words, flags anything questionable, lists any files Codex changed, and hands you the raw-log path.

Configuration

The helper honors these env vars:

• CODEX_DISPATCH_DIR — where artifacts are written (default: $TMPDIR/codex-dispatch)
• CODEX_SANDBOX — workspace-write (default), read-only, or danger-full-access

If OpenAI renames a variant, edit the case map in skills/codex-dispatch/dispatch.sh.