# PR Codex Review Loop

Use when a PR has been (or is about to be) pushed and needs the codex review loop, ...

- Repository: https://github.com/cubeplexai/cubeplex (574 stars)
- Category: Git and review
- License: see repository
- Last updated: 2026-11-01
- Runs scripts: yes, review before use

## Install

Claude Code: `npx skills add cubeplexai/cubeplex --skill pr-codex-review-loop -g -a claude-code -y`
Codex: `npx skills add cubeplexai/cubeplex --skill pr-codex-review-loop -g -a codex -y`
Cursor: `npx skills add cubeplexai/cubeplex --skill pr-codex-review-loop -g -a cursor -y`
Gemini CLI: `npx skills add cubeplexai/cubeplex --skill pr-codex-review-loop -g -a gemini-cli -y`

## Description

Use when a PR has been (or is about to be) pushed and needs the codex review loop, push → wait → fetch new comments via the bundled codex-poll.sh → fix actionable feedback → reply to every comment → re-tag @codex → repeat until clean. Triggers on phrases like "等 codex review", "看看 review", "@codex 直到 ok", "PR review 循环".

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# PR Codex Review Loop

A judgment-bearing wrapper around `scripts/codex-poll.sh` (bundled with
this skill). The poller is the deterministic part, single-shot,
cursor-based, dependable. This skill is the *strategy* layer: when to
poll, what counts as actionable, how to reply, when to stop.

## When To Use

- User says something like "提交 PR / push → 看 review / 等 codex → 改 → @codex 直到 ok".
- A PR was just pushed and the user expects the codex bot to weigh in.
- User explicitly invokes `/pr-codex-review-loop` or names the skill.

**Do not** auto-invoke after every push, only when the loop is actually
the intent.

## Prerequisites

- `gh` CLI is authenticated (`gh auth status`).
- `jq` is installed (the poller depends on it).
- The PR exists and the user has push rights (so follow-up commits land).
- You know **the PR number**. If unclear, ask once, then stop guessing.

## Scope Decision Belongs To The User

This skill does **not** decide whether to split a change into spec / plan /
code PRs. That decision is upstream:

- 1 PR for tightly-coupled changes (small fix, single concern).
- N PRs (spec → plan → code) when each artifact can be reviewed
  independently, or when the spec/plan needs validation before code.

If the user has not made the call, ask once with the trade-off (review
cost vs. coupling vs. rollback granularity). Once decided, run **one
loop per PR**.

## The Loop

For each PR you're driving:

The rest: https://github.com/cubeplexai/cubeplex/blob/HEAD/.agents/skills/pr-codex-review-loop/SKILL.md