codex-plan-review

Invoke scripts via ${CLAUDE_PLUGIN_ROOT}/bin/<name>. That env var is set by Claude Code for every plugin context and gives the absolute path to the marketplace clone — no PATH or cwd assumptions. Bare-command invocation (e.g. plan-review.sh) is NOT reliable: the plugin never puts its scripts on PATH, so bare calls exit 127.

Respond in the user's preferred language (detect from their recent messages, or fall back to the language setting in CLAUDE.md).

Send a design document to Codex CLI for review, then revise the document based on the findings. Automatically loops until no critical issues remain or 5 rounds are completed.

Arguments: $ARGUMENTS

• A file path to a design document or plan
• No arguments: default to the most recently modified .md file under Docs/

Execution Flow

1. Determine the Target File

Try in order of priority:

1. If the user specified a file path, use it
2. Otherwise check the current conversation for a Claude-generated plan (typically under Docs/ or similar), use the most recent one
3. If no plan file exists either, review the current conversation context — find the most recently discussed design proposal, refactoring suggestion, or review conclusion, write it as a temporary spec file (Docs/qq/<branch-name>/tmp-review-spec_<YYYYMMDD-HHmm>.md, timestamped). Get branch name with: git branch --show-current | tr '/' '_'
4. Last resort: use ls -t Docs/**/*.md | grep -v '/qq/' | head -1 to find the most recently modified design document (excluding qq-generated artifacts)

2-6. Automatic Review Loop

Loops automatically without prompting the user each round. Loop termination conditions (stop when any is met):

• No [Critical] issues in the Codex review result
• 5 rounds have been completed

Each round:

2a. Send to Codex for Review

Run the following command using the Bash tool with run_in_background: true:

${CLAUDE_PLUGIN_ROOT}/bin/plan-review.sh <file_path>

The script calls codex exec --sandbox read-only, outputting results to stdout and <filename>_review.md. The script automatically reads the project root's CLAUDE.md and includes the coding standards in the Codex prompt. Codex review typically takes 5-10 minutes. With background execution, the system automatically notifies you when done — no sleep or polling needed. Inform the user that the background task has been submitted and will continue automatically upon completion. You may continue other conversation with the user while waiting.