用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/NVIDIA/cudf --skill review-cudf命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and multi-GPU DataFrame workloads.
Debug and fix pandas test suite failures under the cudf.pandas compatibility layer. Use when given pytest node IDs of failing pandas tests that need to be fixed for cudf.pandas compatibility.
Benchmark a cuDF branch, WIP changes, or a PR against the `main` branch
基于 SOC 职业分类
正在显示 SKILL.md
| name | review-cudf |
| description | Use this skill to review GitHub pull requests for cudf |
Use this skill when the user invokes /review-cudf with:
cudf GitHub repository is located at: https://github.com/NVIDIA/cudf
gh pr view <PR_NUMBER> --repo NVIDIA/cudf --json title,body,files,additions,deletions,baseRefName,headRefName
gh pr diff <PR_NUMBER> --repo NVIDIA/cudf
Hint: Check if GH_TOKEN (or GitHub CLI auth) is already configured in the environment (for example via your secret manager) so gh can authenticate and bypass rate limits; do not run gh auth token from within the agent. If gh auth is unavailable, fall back to GitHub's raw diff/patch URLs, git fetch of the PR ref, unauthenticated GitHub REST API with curl, or any other available methods.
Fetch review comments already posted for context on what's already been suggested and need not be repeated.
Read the Developer Guide (cpp/doxygen/developer_guide/DEVELOPER_GUIDE.md) — it is the authoritative reference for libcudf conventions. All rules in the guide apply during review. The checklist below calls out the most review-relevant rules and adds items not covered by the guide.
Analyze the changes against the checklist below, reading relevant source files as needed for context.
Produce a structured review using the output format at the bottom.
Dump the structured review to .agents/reviews/<PR_NUMBER>/review.md
For the detailed review checklist, read these files:
cpp/REVIEW_GUIDELINES.mdpython/REVIEW_GUIDELINES.mdcpp/doxygen/developer_guide/DEVELOPER_GUIDE.mdcpp/doxygen/developer_guide/TESTING.mdcpp/doxygen/developer_guide/BENCHMARKING.mdcpp/doxygen/developer_guide/DOCUMENTATION.mdUse all applicable checklist items from those files when reviewing the PR. The Developer Guide is the authoritative reference for libcudf conventions — all rules in the guide apply during review.
| Topic | Path |
|---|---|
| Developer guide | cpp/doxygen/developer_guide/DEVELOPER_GUIDE.md |
| Testing guide | cpp/doxygen/developer_guide/TESTING.md |
| Benchmarking guide | cpp/doxygen/developer_guide/BENCHMARKING.md |
| Documentation guide | cpp/doxygen/developer_guide/DOCUMENTATION.md |
| Profiling guide | cpp/doxygen/developer_guide/PROFILING.md |
Online libcudf API docs if needed: https://docs.rapids.ai/api/cudf/nightly/libcudf/
Structure your review as follows:
## PR Review: <PR title>
**PR:** <link>
**Summary:** <1-2 sentence summary of what the PR does>
### Findings
#### Critical
- **[file:line]** Description of issue that must be fixed before merge.
#### Suggestions
- **[file:line]** Description of improvement to consider.
#### Nits
- **[file:line]** Minor style or formatting issue. Keep these minimal, don't suggest adding comments around every line of code or obvious logic.
#### Highlights
- Highlight well-written code, good test coverage, or clever solutions.
### Verdict
One of: **Approve**, **Request Changes**, or **Comment**
With a brief justification.
If there are no findings in a category, omit that category.