Planning with Codex

For high-stakes features it is sometimes worth having two planning models argue with each other before you commit. ClosedLoop provides this via /code:plan-with-codex, which runs a bounded debate between Claude's plan-agent and Codex (default gpt-5.3-codex).

The command

/code:plan-with-codex [--max-rounds N] [--plan-file PATH]
                      [--codex-model MODEL] <prompt>

Debate rounds are capped by --max-rounds (default 15). The output is debate-plan.md.

How the debate runs

1. plan-agent (Opus) drafts an initial plan from the prompt.
2. Codex reviews the draft, raises objections, and proposes changes.
3. plan-agent either accepts, counters, or rewrites.
4. The debate continues until convergence, a clear disagreement is reached, or --max-rounds is hit.

Each round's state is persisted in three sidecar files so the debate can be resumed:

• {stem}.state
• {stem}.prompt
• {stem}.exclusions

Prerequisites

• codex must be on your PATH or configured under Settings → CLI Tools.
• The code:codex-review skill must be installed (it is, if you ran the one-line plugin installer).

When to use it

• Architecturally consequential features where you want an adversarial review before implementation.
• Plans that have failed judge evaluation repeatedly.
• Migrations and refactors where a second model's take can catch a whole-system risk.

When to skip it

• Small, well-scoped features. The debate adds real token cost; use it where it pays off.
• Any case where your PRD is still vague. Fix the PRD first.

Integration with the main workflow

The resulting debate-plan.md is a standalone artifact. To bring it into the main loop, promote it to plan.md (or run /code:code with --plan <debate-plan.md>). From there, the standard 18-phase flow takes over.

Related reading