GitHub PR Review Setup

Inputs

• •pr: PR number or URL (preferred)
• •repo (optional): OWNER/REPO (used when pr is a number and repo can’t be inferred)
• •repo-dir (optional): local clone path (required for non-openai/openai PRs if you aren’t already in that repo)
• •checkout-mode (optional): worktree (default) or inplace
• •worktree-name (optional): override worktree/branch name (default: pr-review-<number>)
• •out (optional): path to also write the JSON payload

Quick start

• •Prepare a PR review workspace + print JSON:
  • •python "<path-to-skill>/scripts/prepare_github_pr_review.py" --pr "<number-or-url>"
  • •Add --repo "<owner/repo>" when --pr is just a number and repo inference isn’t possible.
  • •Add --repo-dir "<path-to-local-clone>" for non-openai/openai PRs when needed.

Output (JSON)

The script prints one JSON object to stdout containing at least:

• •worktree_dir: local checkout path (worktree or in-place)
• •compare_to: base ref suitable for codex review --base and PAL compare_to (e.g. upstream/main)
• •checks: gh pr checks JSON
• •issue_comments, reviews, review_comments: all PR discussion content pulled via gh api --paginate

Workflow

1) Create / reuse a clean checkout (default: worktree)

• •Uses wf for openai/openai and wt for all other repos.
• •For --checkout-mode inplace, checks out the PR branch directly in the target clone (requires a clean repo).

2) Fetch the PR base ref (shared prerequisite)

• •Chooses remote: prefer upstream if present, else origin.
• •Fetches <base_ref> so compare_to exists locally.

3) Pull CI + discussion context (JSON only)

• •CI checks: gh pr checks
• •PR conversation: issue comments
• •Reviews + inline review comments

4) Run the actual review

Use the code-review skill with:

• •path = worktree_dir
• •compare_to = compare_to