Sort and label GitHub issues
For maintainers whose issue list has got out of hand. The assistant reads each new issue, labels it (bug, feature, question, needs steps to reproduce), points out likely duplicates, suggests who should look at it, and asks for reproduction steps when a bug report has none.
What you need
- Your GitHub repository and a GitHub token that may read issues, add labels and comment. Give the token nothing more: it is what really limits what the assistant can do.
How to build it
- Connect GitHub. Open APIs, click Connect an API, pick SDK / npm and add GitHub’s official package,
@octokit/rest. Then, under Tools, Create tool, SDK / npm package, make a tool for each thing it should do: list issues, add labels, comment. Keep the token in Credentials. - Create the assistant. Open Agents, click Create agent and switch to Autonomous. In Instructions, write your rules: “Label by type and area, link duplicates, and ask for steps to reproduce when a bug has none.” Pick the GitHub tools under Tools and APIs.
- Add firm rules. On its Constraints tab, add “Never close an issue” and “Never use the core-team-only labels”.
- Let it remember. Turn on Remember between conversations and set Whose memory to This agent’s own, so it recognises the same bug reported five different ways.
- Run it regularly. On its Overview tab, turn on Schedule and set how often it sweeps new issues, for example every 60 minutes.
Ideas
- Send suspected security reports to a private channel instead of labelling them in public.
- Turn a good triage run into a skill (Promote to skill on the run) so other repositories reuse the same rules.
For developers
To react the moment an issue opens instead of on a schedule, have a GitHub Action call the agent’s endpoint shown in the Integration card on its Overview tab. See the OpenAI-compatible API.