[Skills] .claude/skills: contradictions, stale references and structure issues needing maintainer decision #6156

An audit of .claude/skills found the problems below. They need a maintainer to decide which way to go, so the companion PR (fixes for the clear-cut bugs) does not change them.

Skills that contradict each other

1. Skip decorators. xpu-nightly-ci-fix allows a temporary @skipIfXpu with a tracking issue. issue-handler/issue-fix and xpu-issues-triaging forbid any skip.
2. Testing a torch-xpu-ops fix inside pytorch. issue-handler/references/environment-setup.md says to copy files, run ninja -C build and never touch third_party/xpu.txt. AGENTS.md, xpu-build-pytorch/reference.md and xpu-nightly-ci-fix say to set xpu.txt to HEAD and build only with pip install -e . -v --no-build-isolation.
3. setvars.sh. action/sycl/source-oneapi forbids it. issue-handler/references/environment-setup.md, the xpu-alignment setup and xpu-build-pytorch use it.
4. pr-review tone. The skill says "only report problems, no nits", but its detailed mode has a "Style nits" section and a mandatory "What looks good" section.
5. Triaging scope. xpu-issues-triaging says it is read-only with no execution, then allows un-skipping tests, adding prints and running git clone.
6. Stage order. issue-format sends interactive runs straight to triaging. The issue-handler orchestrator runs test-verification next.
7. Where the PyTorch checkout lives. Three skills use three locations: agent_space_xpu/pytorch, $PYTORCH_DIR and /tmp/pytorch.
8. Environment and commit rules vs AGENTS.md. xpu-alignment creates a .venv and runs pip installs. xpu-ops-pr-creation amends and pushes. AGENTS.md forbids installing tools and committing unless asked.

References to things that do not exist

9. Labels that intel/torch-xpu-ops does not have today:
   • agent:waiting-upstream (the repo has wait_upstream)
   • agent_test: ut (the repo has test: ut)
   • xpu-alignment, upstream-issue, upstream-pr, confirmed, related-failure, task
10. auto-label skill. CI passes this skill verbatim to the labelling model, so it was left unchanged.
    • It mentions test/xpu/core/, which does not exist.
    • The _windows_*.yml rule asks for windows_ci. auto_label.yml only accepts disable_* labels, so that label can never be emitted. The rule's disable_ut then blocks windows-ut in pull.yml.
    • Rule 6 ignores src/comm/.
11. pr-review "GitHub Actions Mode". It describes @copilot and @claude triggers that no workflow implements. The real trigger is @torchxpubot review in bot.yml. That trigger passes a PR number, so it uses the gh path.
12. Leftovers with no consumer in the repo:
    • the "Custom skills applied" markers
    • the pre-commit option in xpu-ops-pr-creation (the repo uses lintrunner)
    • the "machine-parsed" **Branch:** line in xpu-ops-pr-creation

Structure

13. A skill named setup. action/unitrace/setup has a name that is too generic for discovery. The upstream copy of this skill is named unitrace-setup.
14. Nested skills. action/** and issue-handler/* are nested, and issue-handler itself warns that nested skills may not be auto-discovered.
15. Licence. The four action/ skills declare license: MIT, but the repo is Apache-2.0. action/intel-gpu-device-selection/scripts/l0_igpu_check.py has no licence header.
16. Duplication. The xpu.txt development-override text is repeated in AGENTS.md, xpu-build-pytorch/SKILL.md, xpu-build-pytorch/reference.md, xpu-nightly-ci-fix/SKILL.md and issue-handler/*. These copies have already drifted apart (see item 2).

Each item needs a decision on which side is authoritative. After that, the change is a small doc edit.

Written with an AI assistant (OpenCode); the claims were checked against the repo and gh label list -R intel/torch-xpu-ops.