codex-plan-improver
A plan written by one model has that model’s blind spots. codex-plan-improver gets a second opinion from OpenAI Codex before any code is written. Claude revises the plan from Codex’s feedback until Codex approves or five rounds pass.
Install
Section titled “Install”/plugin install codex-plan-improver@cc-toolkitIt also needs the OpenAI Codex CLI, installed and signed in, and jq and uuidgen on your PATH.
What it does
Section titled “What it does”| Command or skill | What it does |
|---|---|
| Plan-exit hook | Starts on its own when Claude calls ExitPlanMode. It denies that exit and tells Claude to run the review. The next exit goes through. |
/codex-plan-improver:codex-review [model] |
Sends the current plan to Codex and revises it round by round, up to five rounds. Codex runs in a read-only sandbox. |
Each round shows Codex’s feedback, ending in VERDICT: APPROVED or VERDICT: REVISE, and the revisions Claude made.
Claude skips any Codex suggestion that contradicts a requirement you stated, and tells you it did so.
- Work in plan mode as usual. When Claude leaves plan mode, the hook triggers the review, and you see the revised plan.
- Run
/codex-plan-improver:codex-reviewat any point when a plan exists in the conversation. Without a plan, Claude asks what to review. - Run
/codex-plan-improver:codex-review gpt-6-astrato use a different model. Any model name your Codex CLI accepts forcodex exec -mworks.
Options
Section titled “Options”| Setting | Default | Effect |
|---|---|---|
| Command argument | gpt-6.1-sol |
The model passed to codex exec -m. The model setting in ~/.codex/config.toml doesn’t apply. |
CC_TOOLKIT_CODEX_PLAN_REVIEW |
Unset (review on) | Any value other than 1 or true turns the plan-exit hook off. The manual command still works. |
The plugin README has every option.