Create PR

Orchestrator skill that creates a PR and shepherds it through review. It invokes other skills as sub-agents and does its own work between them.

Input

Parse $ARGUMENTS:

If --mode is not specified, ask the user.

If --review-lens is invalid, stop and ask the user to choose a valid value. Do not silently reinterpret misspelled options.

If --pr is provided, resolve --mode, then continue directly to Step 5 (monitor) and Step 6 (iteration loop). Do not prompt for remote, merge, title, body, reviewers, labels, or self code review options; --review-lens has no effect because self code review only runs when creating a new PR.

Prerequisites

1. Verify gh is installed and authenticated (gh auth status). If not, link to https://github.com/cli/cli#installation and stop.
2. Check if PowerShell is available for dialog notifications. If unavailable, fall back to text-only alerts.
3. The merge step in Step 2 requires the babylon-skills:merge-and-resolve skill. Do not try to pre-check its availability — the agent's available skills list may be truncated, which causes false negatives. Always offer the merge option in Step 1a; if invocation in Step 2 fails because the skill isn't installed, warn the user and skip.

⚠️ Always pass --no-pager to git (e.g. git --no-pager log, git --no-pager diff, git --no-pager show). Without it, commands can launch an interactive pager (less) that blocks the shell. This applies to every git invocation in this skill, not just the examples shown below.

Step 1: Gather all inputs up front

If --pr was provided, only resolve --mode in this step. Skip the remaining Step 1 inputs and Steps 2–4, then jump to Step 5.

Every user prompt in this skill MUST use the ask_user tool — not plain chat text. Provide multiple-choice options where possible (e.g. yes/no, automatic/interactive, detected-remote/other). Ask one question at a time and wait for the answer before proceeding.

If the user replies with freeform feedback (e.g. "also add reviewer X", "change the title to X"), apply the change and re-ask the same question with the updated options. Only continue once the user picks a final option without further changes.

Collect everything before starting the workflow so it doesn't stop midway.

1a. Remotes, base branch, and merge

1. Detect git remotes (git remote -v).
2. Push remote (user's fork) — the remote whose URL contains the user's GitHub login (gh api user --jq ".login"). If --push-remote is specified, use it; otherwise present the detected fork and ask the user to confirm or change it.
3. Upstream remote (PR target) — the remote pointing at BabylonJS/Babylon.js (often named upstream or origin). If --upstream-remote is specified, use it; otherwise present the detected remote and ask the user to confirm or change it.
4. Base branch — if --base is specified, use it; otherwise default to the upstream's default branch (gh repo view BabylonJS/Babylon.js --json "defaultBranchRef" --jq ".defaultBranchRef.name", typically master) and prompt the user to confirm.
5. Once the upstream remote and base branch are confirmed, fetch so the base reference is current: git fetch <upstream-remote> <base-branch>.
6. If --merge not specified, ask: "Would you like to merge and resolve before creating the PR?" The babylon-skills:merge-and-resolve skill handles merging the upstream base into the feature branch.

Remember <push-remote>, <upstream-remote>, <base-branch>, and <self-login> (from gh api user --jq ".login") — they are reused in 1c, 1d, Step 2, Step 3, and Step 4.

1b. Mode

If --mode not specified, ask:

• Automatic — everything happens without pausing.
• Interactive — pauses at key points so you can review before changes are committed/pushed.

Resolve the self code review lens without prompting unless the user supplied an invalid value:

• <review-lens> from --review-lens, defaulting to both.

1c. PR title and body

1. Analyze only branch-specific changes using three-dot diff against <upstream-remote>/<base-branch> (resolved in 1a):

   # Branch commits with full messages (not just --oneline) so you can
   # read author intent, not just code deltas
   git --no-pager log <upstream-remote>/<base-branch>...HEAD
   
   # Files changed by the branch only (excludes merged-in upstream changes)
   git --no-pager diff <upstream-remote>/<base-branch>...HEAD --stat

   ⚠️ Do not use two-dot .. or compare against local master — either can include unrelated upstream commits or miss recent upstream work, inflating the file count.

2. Generate a proposed title and body using both the code diff and the commit messages — the messages often describe intent, motivation, and linked issues that the diff alone doesn't convey. The body should start with: > 🤖 *This PR was created by the create-pr skill.* Include a clear explanation of the changes, motivation, and any behavioral changes. Include links to related PRs or issues if referenced in commit messages.

3. Present to the user — they can accept, modify, or provide their own.

1d. Reviewers

Suggest the top 1–2 upstream-org members who authored or reviewed previous PRs touching the files/folders changed by this PR. Do the whole pipeline non-interactively. Reuse <upstream-remote>, <upstream-owner>, <upstream-repo>, <base-branch>, and <self-login> from Step 1a.

1. Collect recent commit SHAs on the base that touched the PR's changed files:

   git --no-pager diff --name-only <upstream-remote>/<base-branch>...HEAD
   git --no-pager log <upstream-remote>/<base-branch> --format="%H" -n 30 -- <file1> <file2> ...

2. For each SHA, map to its PR and collect author + review-submitters. Each appearance = +1 score for that login. Skip commits with no associated PR.

   gh api "/repos/<upstream-owner>/<upstream-repo>/commits/<sha>/pulls" --jq ".[0].number"
   gh api "/repos/<upstream-owner>/<upstream-repo>/pulls/<pr>" --jq ".user.login"
   gh api "/repos/<upstream-owner>/<upstream-repo>/pulls/<pr>/reviews" --jq "[.[].user.login] | unique | .[]"

3. Rank by score (highest first).

4. Walk the ranked list and filter. Drop and continue for:

   • <self-login>
   • Bots (logins ending in [bot])
   • Non-members of the upstream org: gh api "/orgs/<upstream-owner>/members/<login>" --silent (exit 0 = member). Fallback for user-owned repos: gh api "/repos/<upstream-owner>/<upstream-repo>/collaborators/<login>" --silent.

   Stop after the first 1–2 survivors.

5. Present the 1–2 candidates. Ask the user to confirm or change. If none survive, skip --reviewer on gh pr create.

Loop over commits explicitly — don't try to one-line the whole pipeline.

1e. Labels

Determine labels from the changed files using the rules in .github/instructions/pr-labels.md.

Present suggested labels and ask the user to confirm or change.

1f. Summary and plan

Present the summary and wait for confirmation:

Here's the plan:
- Push remote (fork):    <push-remote>
- Upstream remote:       <upstream-remote>
- Base branch:           <base-branch>
- Merge upstream first:  yes / skip
- Mode:                  automatic / interactive
- Self review lens:      instructions / agnostic / both
- Title:                 <title>
- Reviewers:             <reviewers>
- Labels:                <labels>

Steps:
1. Merge and resolve upstream changes (if selected)
2. Run self code review and quality checks locally (code-review skill)
3. Commit review fixes locally
4. Push once and create a standard PR
5. Monitor PR and apply fixes for review comments / CI failures

Ready to proceed?

Step 2: Merge and resolve (optional)

If the user opted to merge, invoke as a sub-agent, passing the resolved base branch and upstream remote from Step 1a:

/babylon-skills:merge-and-resolve <base-branch> <upstream-remote> --mode <automatic|interactive>

If invocation fails because the skill is not installed, warn the user "The babylon-skills:merge-and-resolve skill was not found — skipping merge step." and continue.

Step 3: Self code review

Complete the self review and validation locally before pushing the branch or creating a PR. This avoids triggering CI for code that the self review will immediately change.

1. Invoke the code-review skill, passing through the resolved base, mode, and review lens:

   /code-review --base <upstream-remote>/<base-branch> --mode <automatic|interactive> --lens <instructions|agnostic|both>

2. If interactive: pause after code-review completes and ask the user to review the changes before committing.

3. Run the Quality commands from .github/copilot-instructions.md, plus any tests relevant to the changed code. Apply any necessary fixes and iterate locally until they pass.

4. Commit all reviewed, uncommitted changes before pushing. If the worktree contained only fixes produced by code-review or validation, use:

   Code review fixes (automated by code-review skill)
   
   Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

   If the worktree also contained the original feature changes, use a concise commit message derived from the accepted PR title and body instead of describing the entire commit as review fixes.

5. Confirm the worktree is clean. Do not push in this step. The branch is pushed once, after local review and validation are complete, in Step 4.

Step 4: Create the PR

Use <push-remote>, <upstream-remote>, and <base-branch> from 1a.

1. Push the branch:

   git push -u <push-remote> HEAD

2. Get the current branch name and user login:

   git rev-parse --abbrev-ref HEAD
   gh api user --jq ".login"

3. Determine the upstream owner/repo from the upstream remote URL (e.g. git remote get-url <upstream-remote> → parse owner/repo). For this repo that is BabylonJS/Babylon.js.

4. Create a standard, review-ready PR. Write the title and body to temp files first so shell escaping doesn't mangle backticks, $, !, or backslashes in the markdown, then pass the single-line title with --title "$(cat ...)" and the multi-line body with --body-file:

   # Write files (exact markdown, no escaping needed)
   # ... create pr-title.txt and pr-body.md ...
   
   gh pr create \
     --repo "<upstream-owner>/<upstream-repo>" \
     --head "<user>:<branch>" \
     --base "<base-branch>" \
     --title "$(cat pr-title.txt)" \
     --body-file pr-body.md \
     --label "<label>" \
     --reviewer "<reviewer>"
   
   rm pr-title.txt pr-body.md

   gh pr create does not accept --title-file, so a short single-line title via $(cat ...) is fine; the multi-line body must use --body-file.

   --label and --reviewer can each be repeated (or take a comma-separated list) — pass one flag per label/reviewer chosen in 1d/1e.

5. gh pr create prints the new PR's full URL on success — surface it prominently in the chat (e.g. "PR created: ") so the user can click through to review it.

Step 5: Monitor the PR

Invoke the monitor-pr skill with the PR number (either from Step 4 or from the --pr argument):

/monitor-pr <pr-number>

monitor-pr does not accept a --mode argument — it just polls and prints status. The iteration loop below (Step 6) is what handles fixes in the mode chosen back in Step 1b.

Step 6: Iteration loop

Watch the monitor-pr output and react to actionable events.

Handling events

1. Unresolved review comments: Enumerate them with a single GraphQL query that captures everything you need to react, reply, and resolve (so you don't re-query later):

   gh api graphql -f query='
   query {
     repository(owner: "<owner>", name: "<repo>") {
       pullRequest(number: <pr-number>) {
         reviewThreads(first: 100) {
           nodes {
             id                 # thread id (PRRT_...) — used to resolve the thread
             isResolved
             isOutdated
             path
             line
             comments(first: 5) {
               nodes {
                 databaseId     # numeric id — used to reply to the comment
                 author { login }
                 body
                 url            # also encodes the databaseId as #discussion_r<databaseId>
               }
             }
           }
         }
       }
     }
   }'

   For each unresolved thread, read the comment, analyze if it needs a code change and/or response. Make fixes. Prepare a response prefixed with [Responded by Copilot on behalf of <user>] (user from gh api user --jq ".login"). Carry the thread id and parent comment databaseId forward — they're needed when posting the reply and resolving the thread.

2. Pipeline/CI failures (real, not flakes): Read CI logs, identify root cause, make the fix.

3. Test failures: Same as pipeline failures — fix the code, not the test (unless the test is wrong).

After making fixes

Always run the Quality commands from .github/copilot-instructions.md before pushing, plus any tests relevant to the changed code (e.g. integration/visualization tests for rendering changes). Iterate until they pass — do not push broken code.

Interactive mode — gate on user approval first:

1. Stage changes: git add -A

2. Do NOT commit, push, respond, or resolve yet.

3. Present separate tables for each category (only include tables that apply):

   Review comment fixes:

   #	Comment	Proposed Response	Code Changes
   1	<comment text>	<response>	<fix description + files>

   Keep the table compact — it's for the human to review. Persist the thread id and parent comment databaseId for each row in session-local storage (e.g. a review_threads SQL table keyed by row #) so the reply/resolve step can look them up without re-querying or cluttering the table.

   Pipeline failure fixes:

   #	Pipeline Failure	Code Changes
   1	<job — error>	<fix description + files>

   Test failure fixes:

   #	Test Failure	Code Changes
   1	<test — error>	<fix description + files>

4. Show a dialog (if PowerShell available):

   powershell -Command "Add-Type -AssemblyName System.Windows.Forms; [System.Windows.Forms.MessageBox]::Show('Fixes are staged and ready for review.', 'Fixes Ready', 'OK', 'Information')"

5. Wait for user approval, then proceed with the Automatic mode steps below.

Automatic mode: Commit (new commit, never amend), push, respond to comments, resolve threads (see "Posting responses to review comments" below for how to reply correctly). The monitor picks up new checks automatically.

Posting responses to review comments

When responding (in either mode), post the response as a reply to the review comment, not a top-level PR comment. Do not use gh pr comment — it creates an unlinked PR-level comment.

Use the parent databaseId and thread id persisted for this row in session-local storage (populated from the GraphQL query at the top of Step 6). Do not re-query — the IDs are already in hand. As a fallback, the parent databaseId is also encoded in the comment's URL as #discussion_r<databaseId>.

Post the reply via REST:

gh api -X POST \
  "/repos/<owner>/<repo>/pulls/<pr-number>/comments/<parent-id>/replies" \
  -F body=@reply.md
rm reply.md

Then resolve the thread via the resolveReviewThread GraphQL mutation using the captured thread id.

Exit conditions

• All checks pass, PR approved, all comments resolved → monitor-pr handles the "ready to merge" notification.
• User explicitly asks to stop.
• Same issue fails 3 times → stop and ask the user for help.

Guidelines

• Never force push. Never amend previous commits. Always create new commits.
• Always use the user's fork. Never push to upstream org repos.
• Never merge without explicit user approval.
• Prefix review responses with [Responded by Copilot on behalf of <user>].
• Keep PRs focused. Don't mix fixes with refactors.
• Update PR descriptions when the approach changes.
• gh pr merge --auto merges immediately if the repo has no branch protection — ask before using it.