Paper Repro Triage

中文论文复现执行工作流。用于用户上传或提供深度学习、机器学习、LLM、CV、NLP、多模态、数据集、benchmark、prompt 工程或 agent 论文的 PDF、arXiv 链接、论文主页、项目页、标题摘要或源码线索，并要求判断可复现性、搜索官方代码、检查本地源码、追踪数据集论文源码、定位数据处理代码、自...

This skill is internally consistent with its stated purpose, but note the following before installing or invoking it: - It will create and write files under paper-repro-workspace/<paper-slug>/ and will auto-clone repositories from GitHub/GitLab into that workspace (bootstrap_repo.py restricts clones to GitHub/GitLab URLs). - It actively searches and reads local code directories (including the current agent workspace and any paths set in the environment variable PAPER_REPRO_LOCAL_CODE_ROOTS). If you have sensitive files in those locations, the skill may enumerate/read them as part of its local-code discovery — review or unset PAPER_REPRO_LOCAL_CODE_ROOTS if needed. - The skill does not request credentials, but cloning SSH-style URLs ([email protected]) may attempt to use available SSH keys; it does not exfiltrate secrets or send data to arbitrary endpoints. The scripts print snippets (e.g., README head) into reports, so inspect generated repro-report.md and generated scaffold before sharing. - The skill is explicit about NOT installing dependencies, downloading datasets, or running training unless the user explicitly requests run/execute actions; it is safe in that it avoids automatic execution of heavy or potentially disruptive operations by default. Recommended actions: review the workspace path(s) the agent will scan, ensure no sensitive directories are exposed via PAPER_REPRO_LOCAL_CODE_ROOTS, and inspect any cloned repositories and the generated repro-report.md before performing further actions (like installing dependencies or running training).

Type: OpenClaw Skill Name: paper-repro-triage Version: 1.0.0 The skill bundle provides a structured workflow for triaging and preparing academic paper reproductions. It includes Python scripts (e.g., bootstrap_repo.py, scaffold_repro_project.py) for cloning repositories, searching local directories, and generating PyTorch project templates. While the skill utilizes potentially risky capabilities like shell execution (git clone) and file system writes, these are necessary for its stated purpose and are accompanied by explicit safety instructions in SKILL.md that prohibit the agent from automatically installing dependencies, downloading large datasets, or executing training scripts without user consent. No evidence of malicious intent, data exfiltration, or obfuscation was found.