Remote PR Review

Overview

Create a clean, throwaway worktree for a PR, then run two independent review passes (codex review and PAL precommit) and merge the feedback into one prioritized set of fixes.

Inputs

• pr: PR number or URL (preferred)
• repo: optional OWNER/REPO (used when pr is a number and you’re not already in the target repo)
• repo-dir: optional local repo path (required for non-openai/openai PRs if you aren’t already in that repo)
• extra: optional focus areas (perf, security, API, tests, etc.)

Quick start

• Prepare a pr-review-<pr-number> worktree and print JSON:
  • python "<path-to-skill>/scripts/prepare_pr_worktree.py" --pr "<number-or-url>" --json
  • If --pr is just a number and you’re not in the target repo, add --repo "<owner/repo>".
  • For non-openai/openai repos, add --repo-dir "<path-to-local-clone>" when needed.

The JSON includes at least: worktree_dir, repo, pr_number, base_ref, head_ref, pr_url.

Workflow

1) Create / reuse the PR worktree

• Run prepare_pr_worktree.py.
• If it fails due to a dirty worktree/repo, stop and either clean it up or delete the worktree and retry.

2) Run Codex review (CLI)

• cd <worktree_dir>
• Ensure base ref exists locally (pick the right remote for the repo; prefer upstream if present, else origin):
  • git fetch <remote> <base_ref>
• Run review and capture the full output as codex_review:
  • Default prompt (if the user didn’t provide one):
    • codex review --base "<remote>/<base_ref>" "Review for correctness, security, performance, tests, and maintainability. Prioritize issues (blockers vs suggestions vs nits) and reference files/paths when possible."
  • If the user provided extra focus areas, append them to the prompt.

3) Run PAL precommit review (tool)

• Call functions.mcp__pal__precommit against the PR base (not staged/unstaged):
  • path: <worktree_dir>
  • compare_to: "<remote>/<base_ref>"
  • precommit_type: "external"
  • severity_filter: "all"

Capture the output as pal_review.

4) Merge feedback and present a single review

• Deduplicate overlapping findings; reconcile disagreements (call them out explicitly).
• Prioritize into:
  • Blockers (must fix): correctness, security, data loss, breaking API/ABI, missing tests, CI failures.
  • High-signal improvements: maintainability, performance, edge cases, observability.
  • Nits: style/consistency (only if low-noise).
• Provide a short verification checklist (tests to run, manual steps, roll-out risk).

Cleanup (optional)

• openai/openai (workforest): wf rm pr-review-<pr-number> -y
• other repos (git worktree): wt rm "<worktree_dir>" -f (or git -C "<repo-dir>" worktree remove "<worktree_dir>")