---

name: pr description: DraftForge PR lifecycle workflow. Use when the user runs /pr, opens or pushes a pull request, or asks to monitor CI checks and Copilot review on a branch. Creates the PR (or reuses an existing one), polls GitHub Actions and Copilot up to ~15 min, then analyses CI failures and Copilot comments and reports findings grouped by file and skill.

/pr — PR lifecycle workflow

Purpose

Drive a PR from "branch pushed" to "actionable review summary" without the user babysitting it. The skill:

1. Resolves PR state (open existing or create new).
2. Requests Copilot as a reviewer if missing.
3. Polls GitHub Actions checks and the Copilot review (up to 15 min).
4. Pulls CI failure logs and Copilot inline comments.
5. Categorises findings by file → matched .github/instructions/*.instructions.md → canonical skill, then surfaces them to the user with concrete next steps.

When to trigger

• User types /pr.
• User says "open a PR" / "create a PR" / "spin up a PR" on the current branch.
• User says "watch CI" / "monitor checks" / "wait for Copilot" on a branch with an open PR.
• A push has just happened and the user is asking what's next.

Do not trigger automatically just because a branch has uncommitted work — confirm first.

Workflow

Phase 1 — Resolve state

branch=$(git branch --show-current)
pr_json=$(gh pr view --json number,state,headRefName,baseRefName,url,reviewRequests,reviews 2>/dev/null || echo "")

• If pr_json is empty → no PR yet → go to Phase 2 (create).
• If PR exists and is OPEN → go to Phase 3 (request Copilot if absent, then monitor).
• If PR is MERGED / CLOSED → report and stop; ask whether to open a new one.

Before creating, verify the branch is pushed (git rev-parse --abbrev-ref --symbolic-full-name @{upstream} succeeds) and is ahead of main.

Phase 2 — Create the PR (if missing)

Pre-fill title/body from commits when the branch has 1–3 commits; otherwise prompt the user once for a title. Always use a HEREDOC body with ## Summary and ## Test plan sections, per repo convention. Full patterns and the exact gh pr create invocation: references/creating-the-pr.md.

Phase 3 — Request Copilot reviewer (idempotent)

PR=$(gh pr view --json number -q .number)
# Skip if Copilot is already in reviewRequests OR has already reviewed
need_copilot=$(gh pr view $PR --json reviewRequests,reviews \
  --jq '(.reviewRequests + .reviews) | any(.login // .author.login | tostring | test("copilot"; "i")) | not')
if [ "$need_copilot" = "true" ]; then
  gh pr edit $PR --add-reviewer @copilot
fi

The exact bot login varies (copilot-pull-request-reviewer[bot] / Copilot). The case-insensitive test("copilot"; "i") match is intentional — don't hardcode an exact login.

Phase 4 — Monitor checks and Copilot review

Poll every 30 s, max 30 iterations = 15 min. Break early when both are done. Full polling pattern, jq queries, and "is Copilot done?" heuristics: references/monitoring-ci-and-copilot.md.

While polling, narrate progress in single-line updates ("iter 4/30: 3 checks running, copilot pending"). Do NOT print large JSON blobs to the user — those go to logs.

If the 15 min budget exhausts:

• If checks are still running: report which jobs and proceed to analysis with what completed.
• If Copilot is still pending: report and analyse CI-only; tell the user they can re-run /pr later to pick up the Copilot review.

Phase 5 — Analyse and report

Two analyses run in parallel; one merged report:

1. CI failures — gh run view <id> --log-failed for each failed run, extract the error lines (look for FAIL, Error, Traceback, assert, lines ending with :[0-9]+: for compiler errors). Cap at 10 lines per failed step.
2. Copilot inline comments — gh api repos/{owner}/{repo}/pulls/{N}/comments → for each comment, match the path against every .github/instructions/*.instructions.md applyTo glob and tag it with the matched skill(s).

Output format and severity heuristics: references/analyzing-results.md.

What this skill does NOT do

• Does not push for the user (assumes branch is already pushed; tells the user to push if not).
• Does not fix the issues Copilot raises — only surfaces them with the canonical-skill pointer so the user / next skill invocation can act.
• Does not enable automatic Copilot review at the repo level — that's a one-time UI step (Settings → Rules → Rulesets).
• Does not stream raw gh run view --log-failed output into the chat — extracts just the error lines.

References

• creating-the-pr.md — gh pr create patterns, title/body autofill, push-check
• monitoring-ci-and-copilot.md — polling loop, jq queries, "done" heuristics
• analyzing-results.md — CI log extraction, Copilot-comment → skill matching, output format