Skills
skills/stablyai/orca/auto-pr-merge

auto-pr-merge

Installation
SKILL.md

auto-pr-merge

End-to-end autonomous PR workflow: create PR, wait for CI checks to pass, fix any issues (typecheck, code review), update PR, and merge with admin override. Runs fully autonomously without user confirmation.

IMPORTANT: Execute this entire process autonomously without asking the user for confirmation at any step. Just do it.

Process Overview

  CREATE PR  →  WAIT FOR CHECKS  →  FIX LOOP (max 3)  →  MERGE --admin
     │               │                    │                     │
  /create-pr    1min + 1min polls    diagnose CI           gh pr merge
                 (up to ~6min)       fix issues             --admin
                                     pn typecheck           --squash
                                     /review-code           --delete-branch
                                     commit & push
                                     wait for checks

Step 1: Create PR

Run the /create-pr skill to create or update the pull request.

Use the Skill tool: skill: "create-pr"

After /create-pr completes, extract the PR number and URL:

gh pr view --json number,url --jq '.number, .url'

Store the PR number for later use.

CRITICAL: DO NOT STOP after /create-pr completes. Continue to Step 2.

Step 2: Wait for PR Checks

Use the Check Polling Procedure (defined below) to wait for CI checks. Based on the result:

  • PASSED → Skip to Step 4 (MERGE)
  • FAILED → Proceed to Step 3 (FIX LOOP)
  • NO_CHECKS → Skip to Step 4 (MERGE) — no CI configured, nothing to wait for

Step 3: Fix Loop (Max 3 Iterations)

Run up to 3 iterations of the diagnose-fix-review-push cycle.

3a. Diagnose Failures

Identify which checks failed and get their logs:

# Get the most recent failed run ID for this branch
FAILED_RUN=$(gh run list --branch $(git branch --show-current) --limit 5 --json databaseId,conclusion,name --jq '[.[] | select(.conclusion == "failure")] | .[0].databaseId')
echo "Failed run: $FAILED_RUN"

# View failed step logs (truncated to last 200 lines to avoid context bloat)
if [ -n "$FAILED_RUN" ] && [ "$FAILED_RUN" != "null" ]; then
    gh run view "$FAILED_RUN" --log-failed 2>&1 | tail -200
fi

If gh run view --log-failed returns too much output or doesn't work, fall back to:

gh run view "$FAILED_RUN" --json jobs --jq '.jobs[] | select(.conclusion == "failure") | {name, steps: [.steps[] | select(.conclusion == "failure") | {name, conclusion}]}'

3b. Fix Issues

Based on the failure diagnosis:

  1. Read the error output from the failed checks
  2. Identify the root cause (build error, lint error, test failure, type error, etc.)
  3. Apply fixes using the Edit tool
  4. Delegate to specialized skills when appropriate:
    • Lint errors: Run /fix-lint skill
    • Build errors: Run /fix-build skill
    • Test failures: Read failing test, fix the code or test manually

3c. Run Typecheck

After applying fixes, verify types are clean:

pn typecheck 2>&1
  • If typecheck passes: continue to 3d
  • If typecheck fails: fix type errors and re-run (up to 3 attempts within this step)
    • Read the error output
    • Fix the type errors
    • Re-run pn typecheck
  • If still failing after 3 attempts: move on to 3d anyway (CI will catch remaining issues)

3d. Run Code Review (Quick Round)

Run only /review-code to catch any issues introduced by the fixes. Do NOT run /review-correctness, /review-via-codex, /review-algorithm-architecture, or any other review skill — only /review-code:

Use the Skill tool: skill: "review-code"

After review completes:

  • If Critical or High issues found: fix them using the Edit tool, then re-run pn typecheck to verify the review fixes don't introduce type errors
  • If only Medium or Low issues: acceptable, continue
  • If no issues: continue

3e. Update PR

Commit and push only if there are actual changes:

# Check if there are changes to commit
if [ -n "$(git status --porcelain)" ]; then
    git add -A && git commit -m "fix: address CI failures and review feedback"
    git push --force-with-lease
else
    echo "NO_CHANGES"
fi
  • If NO_CHANGES: No fixes were needed/possible. Exit the fix loop and proceed to Step 4.
  • If changes were pushed: continue to 3f.

3f. Wait for PR Checks Again

Use the Check Polling Procedure (defined below) to wait for CI checks. Based on the result:

  • PASSED → Exit fix loop, proceed to Step 4
  • FAILED → Next iteration of fix loop (back to 3a)
  • NO_CHECKS → Exit fix loop, proceed to Step 4

If max 3 fix iterations reached, proceed to Step 4 anyway.

Step 4: Merge with --admin

Merge the PR using admin override with squash merge.

IMPORTANT: Worktree detection. When running from a git worktree, --delete-branch will fail because gh tries to checkout the default branch locally, but it's already checked out in the main worktree. Detect this and handle accordingly:

# Check if we're in a worktree (not the main working tree)
IS_WORKTREE=false
if [ "$(git rev-parse --git-dir)" != "$(git rev-parse --git-common-dir)" ]; then
    IS_WORKTREE=true
fi

BRANCH=$(git branch --show-current)

if [ "$IS_WORKTREE" = "true" ]; then
    # In a worktree: merge WITHOUT --delete-branch, then delete remote branch separately
    gh pr merge --admin --squash
    # Delete the remote branch manually (local cleanup happens when worktree is removed)
    git push origin --delete "$BRANCH" 2>/dev/null || true
else
    # Normal repo: use --delete-branch as usual
    gh pr merge --admin --squash --delete-branch
fi
  • --admin bypasses branch protection rules (required reviews, status checks)
  • --squash squashes all commits into one clean commit
  • --delete-branch cleans up the feature branch after merge (skipped in worktrees to avoid checkout conflict)

If merge fails:

  1. Check the error message
  2. If the error contains 'master' is already used by worktree or similar: retry without --delete-branch and delete the remote branch manually with git push origin --delete <branch>
  3. If merge conflicts: use Skill tool with skill: "resolve-conflicts", push, then retry merge once
  4. If other error: report the full error to the user

Check Polling Procedure

This is the shared polling logic used by Step 2 and Step 3f. Do NOT run this as a single long bash command — the Bash tool will timeout. Instead, run each poll as a separate Bash call.

IMPORTANT: Minimize token waste. Do NOT add verbose commentary between polls. Just run the wait, run the check, and act on the result. No cheerful status messages, no "patience is key", no filler text.

Initial wait

After a push, CI takes time to run. Wait 1 minute before the first poll:

echo "Waiting 1 minute for CI checks to run..." && sleep 60

IMPORTANT: You MUST actually execute the sleep 60 command and wait for it to complete. Do NOT skip the sleep or claim you waited without running the command. The sleep ensures CI has time to finish.

CRITICAL: When calling the Bash tool for sleep 60, you MUST set timeout: 120000 (2 minutes) on the Bash tool call. Similarly, for sleep 60 poll interval calls, set timeout: 120000 (2 minutes).

Poll loop (run each poll as a separate Bash call)

For each poll attempt (1 through 5):

CHECKS=$(gh pr checks --json name,state,bucket 2>&1)
echo "$CHECKS"

# Parse results using 'bucket' field (pass, fail, pending, skipping, cancel)
if echo "$CHECKS" | jq -e 'length == 0' >/dev/null 2>&1; then
    echo "RESULT:NO_CHECKS"
elif echo "$CHECKS" | jq -e '[.[].bucket] | all(. == "pass" or . == "skipping")' >/dev/null 2>&1; then
    echo "RESULT:PASSED"
elif echo "$CHECKS" | jq -e '[.[].bucket] | any(. == "fail" or . == "cancel")' >/dev/null 2>&1; then
    echo "RESULT:FAILED"
else
    echo "RESULT:PENDING"
fi

Decision logic after each poll

  • RESULT:PASSED → Return PASSED. Stop polling.
  • RESULT:FAILED → Return FAILED. Stop polling.
  • RESULT:PENDING → Wait 1 minute, then poll again (use timeout: 120000 on the Bash tool call): 
    sleep 60
  • RESULT:NO_CHECKS → Return NO_CHECKS (no CI configured).

After 5 polls (timeout)

If checks are still pending after 5 polls (~6 min total), return FAILED to enter the fix loop for investigation.

Handling stale checks after re-push (Step 3f only)

After pushing new commits in Step 3e, old check results may linger briefly. To avoid reading stale results:

  1. Record the latest commit SHA before pushing: OLD_SHA=$(git rev-parse HEAD)
  2. After pushing, the first 1-2 polls should verify the check suite is for the new commit. If gh pr checks still shows the old results (check names match but they completed instantly), wait an extra 30s.
  3. Alternatively, look for checks with state: "IN_PROGRESS" or "QUEUED" as a signal that fresh checks have started.

Exit Conditions

  • Success: PR merged successfully
  • Max fix iterations: After 3 fix loop iterations, attempt merge with --admin regardless
  • No changes to fix: Fix loop produced no changes, merge with --admin
  • No CI checks: Repo has no checks configured, merge immediately
  • Unrecoverable error: PR creation fails, merge fails after retry, or gh CLI issues

Progress Tracking

Display progress after each major step:

╔════════════════════════════════════════════════════════════╗
║  AUTO PR MERGE                                             ║
╠════════════════════════════════════════════════════════════╣
║  Step 1 - Create PR:       ✅ PR #123 created              ║
║  Step 2 - Wait for checks: ❌ 2/5 checks failed           ║
║  Step 3 - Fix Loop:                                        ║
║    Iteration 1/3:                                          ║
║      Diagnose:   ✅ Type errors in 2 files                 ║
║      Fix:        ✅ Fixed type errors                      ║
║      Typecheck:  ✅ Passed                                 ║
║      Review:     ✅ No critical issues                     ║
║      Push:       ✅ Updated PR                             ║
║      Checks:     ✅ All passed                             ║
║  Step 4 - Merge:            ✅ Merged with --admin --squash ║
╠════════════════════════════════════════════════════════════╣
║  PR URL: https://github.com/org/repo/pull/123              ║
║  Result: Successfully merged!                              ║
╚════════════════════════════════════════════════════════════╝

Critical Instructions

  1. DO NOT ask for user confirmation - Execute the entire workflow autonomously
  2. DO NOT stop after /create-pr - The PR creation is just step 1 of 4
  3. DO NOT stop after checks pass - Must complete the merge step
  4. DO NOT run polling as a single long bash loop - Each poll must be a separate Bash call to avoid the 2-minute Bash timeout. Use sleep in its own Bash call between polls.
  5. DO wait the full polling period - Don't skip check waiting; CI needs time
  6. DO run typecheck before review - Catch type errors early
  7. DO run /review-code in each fix iteration - Ensure code quality
  8. DO check for actual changes before committing - Skip commit/push if git status --porcelain is empty
  9. DO use --admin for merge - This is intentional to bypass protection rules
  10. DO use --squash for merge - Keep git history clean
  11. DO delete branch after merge - Clean up with --delete-branch
  12. Maximum 3 fix iterations - Don't loop forever; after 3, merge with --admin
  13. DO NOT run /review-correctness or /review-via-codex - This skill only uses /review-code for quick checks in the fix loop. Codex-based reviews are slow and belong in /auto-review-fix, not here.
  14. DO actually execute sleep commands - When the skill says sleep 60, you MUST run the bash command and wait for it to complete. Do not skip or fabricate the wait.
  15. DO set timeout on Bash tool for sleep commands - sleep 60 requires timeout: 120000. The default 2-minute Bash timeout is sufficient for 1-minute sleeps but set it explicitly for safety.

Error Handling

  • If /create-pr fails: Report error and exit
  • If gh CLI is not installed or authenticated: Report and exit
  • If no CI checks are configured: Skip waiting and merge directly
  • If check polling times out: Enter fix loop to investigate
  • If typecheck loops more than 3 times within a single fix iteration: Move on to review
  • If fix loop produces no changes: Exit loop and merge with --admin
  • If merge fails due to conflicts: Try /resolve-conflicts, push, then retry merge once
  • If merge fails for other reasons: Report the error to the user with the full error output
Weekly Installs
3
Repository
stablyai/orca
GitHub Stars
1.5K
First Seen
1 day ago
Security Audits
Gen Agent Trust HubFail
SocketWarn
SnykPass