// what it does
- 01Read the backlog. Fetch open issues in scope with their titles, descriptions, status, labels, owners, dates, links, and recent activity. Bound the run to the configured projects and thresholds.
- 02Identify candidates. Find likely duplicates, stale issues, missing owners, missing severity or product area, blocked items without a next action, and recently resolved work that still appears open.
- 03Verify each finding. Compare descriptions, comments, linked work, and recent activity. Treat similarity as a candidate signal, not proof, and exclude protected issues and states.
- 04Build an exact proposal. For every candidate, show the evidence, proposed field changes, duplicate relationship, or closure reason. Keep uncertain cases in a questions section with no mutation proposed.
- 05[Approval gate] Review the issue-by-issue cleanup proposal before changing labels, owners, states, comments, or duplicate relationships.
- 06Apply only approved edits. Make the exact approved changes in the configured tracker, preserve existing context, and add a concise comment when a closure or duplicate link needs an audit trail.
- 07Verify and summarize. Re-fetch changed issues, report successes and failures separately, and post a Slack summary linking the reviewed changes and unresolved questions.
// guardrails in the draft
Never close or merge an issue because its title merely resembles another issue.
Never change protected labels, protected states, or out-of-scope projects.
Do not invent severity, ownership, customer impact, or resolution status.
Make no tracker changes before the approval gate clears.
If an edit fails or the issue changed after review, stop on that issue and report the conflict instead of overwriting newer work.