Fetch, reproduce, and minimize GitHub issue reproductions for PyTorch bug triage.
npx skills add https://github.com/pytorch/pytorch --skill scrub-issue
Please help me install this Agent Skill. Skill: scrub-issue Source: https://github.com/pytorch/pytorch/tree/main/.claude/skills/scrub-issue Command: npx skills add https://github.com/pytorch/pytorch --skill scrub-issue
Triaging GitHub issues manually is slow and error-prone. This Skill automates fetching, verifying, and minimizing bug reproductions so maintainers can quickly determine whether an issue is actionable, still reproduces, or has already been fixed.
feature, tracker, needs reproduction, triaged).torch.compile crash. Use this Skill to fetch the issue, verify the crash still reproduces on the current PyTorch version, minimize the repro to under 20 lines, and post a concise summary comment on the issue.Use the scrub-issue skill to fetch, verify, and minimize the reproduction for PyTorch issue 12345.
High-intent search queries and answers about installing and using this skill.
Use the scrub-issue skill to fetch the issue, extract the repro, and apply ordered reduction strategies: remove unnecessary imports, shrink models, reduce tensor sizes, simplify backends, and remove control flow. The skill stops when the repro is under ~20 lines or two consecutive rounds fail to simplify further.
Run the original repro script with a 120-second timeout and compare the exception class and distinctive error substring against the reported error. Exit codes above 128 indicate signal kills (e.g., segfault = 139). For non-deterministic bugs, run 3-5 times with PYTHONHASHSEED=0 and torch.manual_seed.
Apply `feature` for feature requests, `better-engineering` for refactoring tasks, `tracker` for umbrella/meta issues, `needs reproduction` when repro code is missing, and `triaged` to update the last-modified timestamp. Use `gh issue edit --add-label` or `--remove-label` to manage labels.
Check for conda environments named `pytorch-<version>` (e.g., pytorch-2.6, pytorch-2.8) and binary-search across them. Run from /tmp with cleared PYTHONPATH using `conda run -n pytorch-<version> python /tmp/repro.py`. Test 2-3 evenly-spaced probe points in parallel each round.
Repro scripts from untrusted authors may contain network requests, file operations outside /tmp, shell execution, obfuscated code, or package installation. The skill checks for these patterns and verifies the author's collaborator status before executing any code.
NVIDIACreate GitHub issues from natural language prompts using the gh CLI.
OpenLaborReproduce bugs, locate root causes, and implement regression-tested fixes.
dannys-code-cornerMinimize compiler bug reproductions and file GitHub issue reports.
TorchedHatExplain torch.compile pipeline stages and intermediate representations for debugging.