gitlab-issue-to-mr

Full workflow: fix a GitLab issue → create MR → monitor CI → auto-fix failures → repeat until green.

Extract the issue number from the user's message (e.g. "fix issue 42" or "#42" → 42). If not provided, ask before proceeding.

Use glab CLI and jq. Must run from a git directory with GitLab remote.

---

Phase 1: Fix the Issue

1. Read the issue:

   glab issue view <ISSUE_NUMBER> -F json
   glab issue view <ISSUE_NUMBER> -c   # read comments if description is unclear
   

2. Create a feature branch from the default branch:

   Derive a short title from the issue title: lowercase, max 4 words, spaces replaced with -, special characters removed. Example: "Fix login timeout error" → fix-login-timeout-error

   git checkout <default_branch> && git pull
   git checkout -b claude/<ISSUE_NUMBER>-<short-title>
   

   Example: claude/42-fix-login-timeout

3. Locate relevant code: Use Grep, Glob, Read to understand root cause.

4. Implement the fix: Minimal, focused changes. Add tests if applicable.

5. Commit and push (confirm with user first):

   git add <files>
   git commit -m "Fix #<ISSUE_NUMBER>: <concise summary>"
   git push -u origin claude/<ISSUE_NUMBER>-<short-title>
   

---

Phase 2: Create Merge Request

6. Create MR linked to the issue:

   glab mr create -i <ISSUE_NUMBER> --fill --remove-source-branch --push
   

   Capture and show the MR URL from the output.

---

Phase 3: Monitor CI and Auto-fix Loop

Repeat the following loop (max 3 iterations to avoid infinite loops):

3a. Wait for pipeline to start

# Poll every 15s until pipeline appears on the branch (max 2 min)
BRANCH="claude/<ISSUE_NUMBER>-<short-title>"
for i in $(seq 1 8); do
  PIPELINE=$(glab ci list -b "$BRANCH" -F json -P 5 | jq 'first(.[] | select(.status != "skipped"))')
  [ "$PIPELINE" != "null" ] && [ -n "$PIPELINE" ] && break
  echo "Waiting for pipeline... ($i/8)"; sleep 15
done
PIPELINE_IID=$(echo "$PIPELINE" | jq '.iid')
PIPELINE_ID=$(echo "$PIPELINE" | jq '.id')
echo "Pipeline IID: $PIPELINE_IID  ID: $PIPELINE_ID"

3b. Wait for pipeline to finish

# Poll every 60s until status is not running/pending/created (max 3 hours)
for i in $(seq 1 180); do
  STATUS=$(glab api "projects/:fullpath/pipelines/$PIPELINE_ID" | jq -r '.status')
  ELAPSED=$(( i * 60 ))
  echo "Status: $STATUS (${ELAPSED}s elapsed)"
  [[ "$STATUS" != "running" && "$STATUS" != "pending" && "$STATUS" != "created" ]] && break
  sleep 60
done
echo "Final status: $STATUS"

3c. If pipeline passed → done

if [ "$STATUS" = "success" ]; then
  echo "CI passed! MR is ready for review."
  # exit loop
fi

Report MR URL and summary. Stop here.

3d. If pipeline failed → analyze and fix

Get failed jobs and logs:

FAILED_JOBS=$(glab api "projects/:fullpath/pipelines/$PIPELINE_ID/jobs?per_page=100" \
  | jq '[.[] | select(.status == "failed") | {id, name, stage}]')
echo "$FAILED_JOBS"

# Fetch trace for each failed job (last 150 lines)
for JOB_ID in $(echo "$FAILED_JOBS" | jq '.[].id'); do
  echo "=== Job $JOB_ID ==="
  glab api "projects/:fullpath/jobs/$JOB_ID/trace" | tail -500
done

Analyze the error output and identify root cause.

Implement fix: Edit relevant files to resolve the CI failure.

Commit and push:

git add <changed_files>
git commit -m "Fix CI: <brief description of what was fixed>"
git push

Go back to 3a for the next pipeline.

---

Phase 4: Report

After the loop ends, report:

• MR URL
• Summary of all changes made (issue fix + any CI fixes)
• Final CI status
• If still failing after 3 iterations: show remaining errors and ask user how to proceed

---

IMPORTANT Rules

• Always confirm with the user before the first push and MR creation (Phase 1 step 5).
• After MR creation, proceed with CI monitoring without asking again unless CI fails 3 times.
• Do NOT use interactive/TUI commands (glab ci view, glab ci trace).
• Use :fullpath in glab api URLs — auto-resolves from git remote.
• Use jq for all JSON parsing.
• Max 3 CI fix iterations to prevent infinite loops; report to user if still failing.
• If CI fails with a flaky/infrastructure error (not a code issue), retry once before attempting a code fix.