triaging-issues

An issue-triage skill for the PyTorch GitHub repository that routes issues to oncall teams, applies labels, and leaves first-line responses. It is used when processing new PyTorch issues or when asked to triage a specific issue, and it operates through GitHub MCP tools: issue_read for details, comments, and labels; issue_write to apply labels or close issues; add_issue_comment to redirect questions; and search_issues for context.

Triage follows an ordered sequence. Issues that already carry any oncall label are skipped entirely because they belong to a sub-oncall team. Questions are closed with a forum-redirect template, while unclear reports get a request for more information. Issues that need reproduction, for example those linking to external files such as .zip, .pt, .pth, .pkl, .safetensors, .onnx, or model-hub downloads, have those links redacted from the body, are labeled needs reproduction, and receive a template asking for a self-contained repro; the same applies to hardware-specific, version-upgrade, or complex-setup reports lacking a minimal script. Issues belonging to another repo are transferred. PT2 and torch.compile issues get special handling: oncall pt2 is not a redirect, so they continue through full triage and must receive at least one module label, falling back to module compile ux. Other secondary-oncall redirects apply exactly one oncall label and stop.

Guardrails are strict. Only labels present in labels.json may be applied, never invented ones, and categories like ciflow, test-config, release notes, sev, merge blocking, and actionable are prohibited. A PreToolUse hook validates labels against the allowlist and a PostToolUse hook adds a bot-triaged marker; when a label is blocked, the skill adds only triage review and stops for a human. It must not override labels a human has already applied, and it emphasizes labeling by root cause rather than by keywords. Supporting files include labels.json, pt2-triage-rubric.md, and templates.json.