github-triage

This skill is a strictly read-only GitHub triage orchestrator for issues and pull requests. It solves the problem of large, unmanaged backlogs by fetching every open issue and PR, classifying each one, and spawning a background subagent per item that analyzes it and writes an evidence-backed report to a timestamped directory under /tmp. Its defining constraint is a zero-action policy: it never comments, closes, merges, labels, reviews, or otherwise mutates GitHub state.

The architecture maps one issue or PR to one background task and one subagent, running all items in parallel. Phase 0 sets up the repo name, report directory, and current commit SHA; Phase 1 paginates through all open items using gh list commands (fetching basic metadata first to avoid JSON parsing issues from control characters), explicitly processing the entire backlog rather than sampling. Phase 2 classifies items into question, bug, feature, other, or PR bugfix categories, and Phase 3 spawns subagents one by one. A strict allow-list permits only read operations: gh issue/pr view, gh api GET calls, Grep, Read, Glob, git log/show/blame, and Write limited to the /tmp report directory, while any POST, PUT, PATCH, or DELETE gh operation is forbidden as a critical violation. A mandatory evidence rule requires every factual claim to carry a GitHub permalink pinned to a specific commit SHA; unbacked claims are marked unverified and carry no weight.

Target users are maintainers and engineering teams who want a safe, thorough, auditable triage of their backlog without any risk of unwanted GitHub changes. The read-only design and explicit forbidden-command list make it a low-risk analysis tool.