address

name	address
description	Run a PR's full review-fix cycle — wait for review and CI, fix failures, evaluate and reply to comments, then commit and push. Use whenever a PR was just opened or updated — the post-Ship phase of the Orchestrator loop.

Purpose

Run the full PR review-fix cycle in one invocation. This skill owns the review-and-reply logic and delegates the mechanics to their canonical owners so they never drift: committing → the commit skill; history cleanup + push → /rebase-clean.

Usage

/address [PR number | URL | file path]

• No argument: detect the current branch's PR via gh pr view --json number,url; if none, use review content from the previous conversation.
• PR number or github.com URL: fetch that PR's review via both endpoints in Phase 3.
• Otherwise: a file path containing review content.

Workflow

Run the cycle end to end. Exit early only if there is nothing to do — no review feedback (neither a summary body nor inline comments) and CI already green.

Phase 1 — Wait for review & CI

Wait for CI (status checks):

gh pr checks {pr} --watch --interval 30   # blocks until all finish; non-zero if any fail

Automated reviews (Copilot, Codex, any other review bot) are not status checks — wait for them separately. They fire on PR open / ready-for-review and post asynchronously, so wait until the set of bot reviews stops growing rather than for one specific bot; a slower reviewer is then never missed. If none arrives within the timeout, proceed:

# Review bots configured for the repo, matched case-insensitively against the reviewer login:
#   copilot → copilot-pull-request-reviewer[bot] / Copilot;  codex → chatgpt-codex-connector[bot].
# Extend this pattern as you wire up more reviewers.
BOTS='copilot|codex'
prev=-1
for i in $(seq 1 20); do
  n=$(gh api repos/{owner}/{repo}/pulls/{pr}/reviews --paginate \
        --jq "[.[] | select(.user.login | test(\"$BOTS\";\"i\")) | select(.body|length>0) | .user.login] | unique | length")
  [ "$n" -gt 0 ] && [ "$n" -eq "$prev" ] && break   # ≥1 bot review in, and none new since the last poll
  prev=$n
  sleep 15
done

Phase 2 — Triage CI

All green → continue. Failing → inspect (gh pr checks {pr}, then gh run view {run_id} --log-failed), reproduce and fix locally, and re-run the failed checks locally to confirm they pass. These fixes stay in the working tree and are committed with the review fixes in Phase 4.

Phase 3 — Address & reply

Fetch feedback from both endpoints. These queries are login-agnostic, so they already cover every reviewer — Copilot, Codex, and humans alike; skip items already addressed on a prior run:

# Summary bodies — empty-body reviews are inline-comment containers, skip them.
gh api repos/{owner}/{repo}/pulls/{pr_number}/reviews --paginate \
  --jq '.[] | select(.body|length>0) | {id, login: .user.login, body}'
# Inline diff comments — where bots put most of their concrete findings.
gh api repos/{owner}/{repo}/pulls/{pr_number}/comments --paginate

A bot's summary body is mostly boilerplate — Codex's "here are some suggestions" wrapper, Copilot's overview and per-file table, or a no-op notice (Copilot was unable to review … quota limit). Treat those as non-actionable; the real findings are the inline comments. Reply only to items that carry an actual request or question.

For each actionable summary body and inline comment:

1. Evaluate — technically correct? aligned with project conventions? a reasonable tradeoff? Don't blindly accept reviewer feedback — bot or human.
2. Implement only the justified fixes — trivial ones directly, substantial ones as an Implementer brief (/codex-work); verify each introduces no new issue.
3. Reply on GitHub to every comment and summary body — concise, in Japanese, with code snippets or references when they help:
   • Inline comment → gh api repos/{owner}/{repo}/pulls/{pr_number}/comments/{comment_id}/replies -f body="<reply>"; summary body (no reply thread) → gh pr comment {pr} --body "<reply>".
   • Addressed → state the fix (e.g. 対応しました。COMMENT ON COLUMN を追加しています。); declined → reasoning the reviewer can accept (e.g. こちらは意図的な設計です。理由: ...).

Phase 4 — Commit

Commit the CI and review fixes by following the commit skill, grouped into logical units.

Phase 5 — Clean up history & push

Run /rebase-clean: it regroups commits, rebases onto the latest main, and pushes with --force-with-lease. On this authorized open-PR follow-up push it runs unattended and self-checks the open PR and clean worktree — defer to it; don't re-implement its checks here.

Phase 6 — Summary

## Review Fix Summary

### CI
- <status before> → <status after fixes>

### Addressed
| Comment | Fix Applied | Reply |
|---------|-------------|-------|

### Not Addressed
| Comment | Reason | Reply |
|---------|--------|-------|

### Final Commits
- commit1: description

PR: <url>