AI Classification

How GitWire uses Claude to classify issues and pull requests.

The Pipeline

When a triage job is picked up by the Triage Worker:

1. Fetch issue data from the database (title, labels, existing metadata)
2. Build prompt with classification instructions and the issue context
3. Send to Claude via the Anthropic API
4. Parse response — extract type, priority, and summary
5. Upsert triage fields in the issues table
6. Apply labels on GitHub via the API

Classification Types

Claude assigns one of these types:

Type	Description	GitHub Label
bug	Something isn't working	bug
feature	New feature or request	enhancement
question	Further information requested	question
documentation	Improvements or additions to docs	documentation
enhancement	Improvement to existing feature	improvement
performance	Performance issue or optimization	performance
security	Security vulnerability or concern	security
other	Doesn't fit other categories	—

Priority Levels

Priority	Criteria	GitHub Label
critical	Security vulnerabilities, data loss, production outages	priority: critical
high	Breaking functionality, significant user impact	priority: high
medium	Non-critical bugs, feature requests with clear value	priority: medium
low	Minor issues, nice-to-have features	priority: low

Summary Generation

Claude generates a one-line summary of the issue, stored in triage_summary. This is used in:

• Dashboard issue lists (quick scan)
• /gitwire status command output
• Duplicate detection embedding text

Confidence

The triage worker does not currently assign a confidence score. All classifications are applied directly. Future versions may add a confidence threshold below which triage is skipped.

Re-triage

Issuing /gitwire triage on an issue triggers a re-classification. The previous triage fields are overwritten.

Worker Reference

See Triage Worker for implementation details.

→ Duplicate Detection