github-pr-triage_skill

Overview

This skill triages GitHub pull requests using a streaming, agent-per-PR architecture. It fetches every open PR with exhaustive pagination, launches one independent background task per PR, and streams each analysis as it completes. The skill applies conservative auto-close logic and produces a final comprehensive report.

How this skill works

First, it collects all open PRs via exhaustive pagination. For each PR it launches exactly one background task that performs a focused analysis (comments, reviews, files changed, branch checks) and returns a strict, machine-parsable report. Results are streamed in real time as each task finishes; auto-close actions are conservative and require confirmation unless explicitly allowed. At the end it aggregates all analyses into a final comprehensive report.

When to use it

• You need fast, parallel triage across many open PRs with live progress updates.
• Maintainers want per-PR isolation so failures or slow analyses don’t block others.
• You want conservative automated cleanup of clearly obsolete PRs with reviewable actions.
• You need an auditable final report summarizing merge-ready, review-needed, stale, and closeable PRs.

Best practices

• Always run exhaustive pagination to ensure no open PRs are missed.
• Enforce 1 PR = 1 background task; never batch multiple PRs into one task.
• Stream results immediately as each background task completes to surface urgent items early.
• Use conservative auto-close criteria and require explicit confirmation before closing unless auto-close was pre-approved.
• Rate-limit polling for background outputs to avoid excessive API / orchestration calls.

Example use cases

• Nightly triage run that analyzes hundreds of open PRs in parallel and streams results to a dashboard.
• On-demand cleanup: find and conservatively close PRs already implemented elsewhere or clearly abandoned.
• Pre-release sweep to identify merge-ready PRs and surface blockers in real time.
• Assign reviewers by streaming items that need human review while other analyses continue running.