Refine a plan with Codex

Use the Claude-vs-Codex debate loop to pressure-test and refine an implementation plan.

/code:plan-with-codex runs a debate loop where Claude drafts an implementation plan and Codex reviews it, iterating until Codex approves or you hit the round budget.

Requirements

codex CLI installed and on PATH.
an existing plan or a prompt to generate one.

Command

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

Defaults:

--max-rounds 15
--codex-model gpt-5.4

How it works

debate-loop.sh alternates:

@code:plan-agent (opus) drafts or revises the plan.
@code:codex-review skill runs Codex against the plan.
Codex emits VERDICT:APPROVED or VERDICT:NEEDS_CHANGES with concrete feedback.
If NEEDS_CHANGES, the plan agent revises.
Loop until APPROVED or max rounds.

Session resume is supported via a {stem}.state sidecar file tracking round and Codex session ID.

Output

the refined plan file (or a new file if you did not specify one)
a record of each round's verdict and feedback

When to use it

when the plan touches unfamiliar territory and a second-model pressure test is valuable
when you want structural disagreement to surface blind spots
before launching a large loop, to avoid burning iterations on a flawed plan

Why two models disagree usefully

Claude and Codex have different priors. A plan that satisfies both tends to have fewer holes than a plan approved by one model alone. The debate is not consensus-seeking; it is adversarial, and that is what makes it useful.

After the plan is approved, hand it to /code:code --plan <file> to execute it through the standard loop.