work-with-pr — Full PR Lifecycle

What it does

This skill treats "implement X" as a delivery problem, not just a coding problem. It decomposes the task into the smallest atomic PRs that each compile, pass, and stand alone for review, then runs this lifecycle per PR:

• Phase 0 — Setup: decide the PR split, create the branch off the base, and create a git worktree as a sibling of the repo. Your main working directory stays read-only so uncommitted work is never destroyed, and independent PRs are built concurrently via background subagents or a team.
• Phase 1 — Implement: delegates to the ulw-loop skill, which turns the brief into binary success criteria. The hard rule: green tests are not QA. Every criterion needs manual QA evidence written to disk (tmux / HTTP / browser / GUI) before you may commit or push, plus minimum-commit ratios (3 files → 2+ commits, 10 files → 5+ commits).
• Phase 2 — PR creation: generates a reviewer-friendly English body (Summary / Changes / QA & Evidence / Risks & Residuals / Related Issues), grouped by area rather than by file, with explicit warnings against pasting secrets, tokens or raw env dumps.
• Phase 3 — Verification loop: an uncapped loop over Gate A (gh pr checks --watch --fail-fast) and Gate B (Cubic bot must report "No issues found"). Any failing gate routes back to Phase 1 for a scoped fix plus fresh QA evidence. Cubic may be marked SKIPPED only when its quota is exhausted — never because it found issues.
• Phase 4 — Merge & cleanup: arms auto-merge by default, blocks until the PR is actually MERGED, syncs .omo state back to the main repo, then removes the worktree. On unrecoverable failure it deliberately keeps the worktree so you can resume.

It closes with a report of PR number, iteration count, gate outcomes and merge status, and includes a severity-ranked anti-pattern table.

Who it's for

• Developers who want an agent to go from prompt to merged PR without babysitting
• Tech leads pushing a small-PR review culture instead of 2000-line rubber stamps
• Repos with CI plus an AI review bot that need automatic fix-and-retry loops
• Anyone who wants agent runs isolated in git worktrees rather than in their live checkout

Examples

1. One issue, three PRs: ask it to "implement this issue and open a PR" and it fans out into schema, API handler, and UI PRs, running the two independent ones in parallel worktrees.
2. CI failure recovery: typecheck breaks after push, so it pulls gh run view --log-failed, fixes only that failure as an atomic commit, pushes, and re-enters from Gate A.
3. Handling bot review: Cubic flags three findings; it triages false positives, fixes the real ones, records new manual-QA evidence, then polls until a newer Cubic review lands.

Install the skill I found on Claude Skill Mart.
Copy the .opencode/skills/work-with-pr folder from the GitHub repo code-yeongyu/oh-my-openagent into my ~/.claude/skills/code-yeongyu-work-with-pr/.
When it's done, tell me in one line what this skill can do.