carterdea/gh-address-pr-comments

Address PR Comments

Resolve actionable review comments for a PR, one-by-one.

Steps

0. Resolve PR_NUMBER

• If user supplied a PR number/URL, use it.
• Otherwise look up the open PR for the current branch: gh pr view --json number,state,headRefName -q 'select(.state=="OPEN") | .number'
• If that returns nothing, try: gh pr list --head "$(git branch --show-current)" --state open --json number -q '.[0].number'
• Only ask user if both lookups fail or return multiple PRs.

1. Checkout PR gh pr checkout {PR_NUMBER}

2. Fetch PR data + comments (include author + position info) gh pr view {PR_NUMBER} --json title,body,state,author,headRefName,baseRefName,url,reviews gh api repos/{OWNER}/{REPO}/pulls/{PR_NUMBER}/comments --jq '.[] | {id, path, line, position, body, user: .user.login, user_type: .user.type}' gh api repos/{OWNER}/{REPO}/issues/{PR_NUMBER}/comments --jq '.[] | {id, body, user: .user.login, user_type: .user.type}'

3. Filter out noise before classifying. Skip entirely (do not list, do not act on):

• Outdated review comments: review comments where position is null (line no longer exists in diff — GitHub UI labels these "Outdated"). Issue comments don't have position; only applies to pulls/comments endpoint.
• Vercel bot comments (login matches vercel[bot], vercel-bot, or body mentions Vercel preview/deployment status)
• Bare agent mentions: body is just @claude, @codex, or short variants like @codex review, @claude please review, @claude take a look (no actionable content beyond the tag). Heuristic: strip mentions + whitespace; if <= ~3 words remain and none describe a change, skip.
• Already addressed this session: comments whose requested change you already made earlier in this same session. Safe to skip — briefly note them as "already addressed" in the final summary rather than re-applying or re-verifying.

4. Classify each comment by author:

• Human (user_type == "User"): trust default. Assume correct unless obviously wrong. Verify scope + intent, then apply.
• Bot (user_type == "Bot" OR login matches cursor[bot], chatgpt-codex-connector, claude[bot], coderabbitai[bot], github-actions[bot], *-bot, *[bot]): skeptical default. Bots hallucinate, flag non-issues, miss context. For each bot comment:
  • Read cited code + surrounding context before acting
  • Ask: is claim factually correct? Does fix improve code or just silence bot?
  • Reject if: false positive, stylistic noise, conflicts with project patterns, suggests broken refactor
  • If rejected, note reason in summary. Do not "address" via no-op reply commit.

5. Present numbered list grouped by author type (Human first, Bot second). Flag bot items with [BOT: skeptical]. Ask user which to handle.

6. For each selected item:

• Show relevant code context
• For bot items: state verdict (valid / false positive / partial) before coding
• Make smallest correct change
• Add/update tests when needed
• Skip silently-rejected bot items — list in final summary

7. Summary git status --short git diff --stat