arxiv-daily

每日 arXiv 论文获取加整理技能。当用户想要获取每日 arXiv 论文相关领域论文、设置论文订阅、管理论文推送时间、添加新的论文领域时触发。也用于用户询问"arxiv"、"每日论文"、"论文推送"等相关话题时。

Security Scan

The skill's code and instructions match its stated purpose (fetching and preparing daily arXiv papers); there are no unexpected credentials, remote endpoints, or obfuscated/hidden behavior, but several operational responsibilities (scheduling, pushing messages, summarization) are left to the agent and are not implemented in the shipped scripts.

Purpose & Capability

Name/description (daily arXiv fetch and summarization) align with the provided scripts. fetch.py scrapes arXiv list/abstract pages and writes per-day text files; run.sh runs the script and installs Python deps if missing. No credentials or unrelated subsystems are requested.

Instruction Scope

SKILL.md instructs the agent to read references/ and data/ files, create configuration like 'like-<category>.txt', schedule timers, generate translations/summaries, and push results via user channels (e.g., WeCom). The included scripts only perform fetching and saving text; summarization, translation, scheduling, and push delivery are expected to be done by the agent/platform (not implemented in code). This is coherent but important: the agent will need permission to read/write the referenced directories and to send messages through whichever channel the user chooses.

Install Mechanism

No install spec in the registry; this is instruction + small scripts. run.sh will optionally pip-install requests and beautifulsoup4 if missing (standard packages). Nothing is downloaded from arbitrary URLs or executed from unknown hosts.

Credentials

The skill declares no required environment variables or credentials. SKILL.md references channels/account/chat_id (e.g., WeCom) but does not request tokens — it assumes the agent/platform provides messaging capability. That is proportionate but the user should confirm the agent has the necessary channel permissions before enabling push behavior.

Persistence & Privilege

always:false (no forced always-on presence). The skill writes files under data/ and references/ as part of its functionality; it does not modify other skills or global agent config. There is no autonomously persistent installer component included.

Assessment

What this skill will do and what to check before installing: - It scrapes arXiv (https://arxiv.org) and saves plain text files under the specified data/ directories. The provided fetch.py and run.sh are straightforward and readable — there are no hidden network endpoints or credential-exfiltration patterns. - The skill does NOT implement scheduling, summarization/translation, or delivery to chat channels itself. SKILL.md expects the agent/platform to: (a) create timers/cron jobs, (b) read data files and produce translated summaries (likely using the LLM), and (c) send messages through your chat channel (WeCom or others). Confirm your agent environment supports those capabilities and that you are comfortable granting the agent write access to the data/references directories and permission to send messages on your behalf. - run.sh can auto-install Python packages via pip if requests/beautifulsoup4 are missing. Installing packages from PyPI is normal but be aware it requires network access and pip privileges on the host. - Minor inconsistencies to note: SKILL.md mentions automatic data retention/cleanup (7 days) and generating .md summary files, but the shipped scripts do not implement retention cleanup or summarization/translation; those must be handled by the agent or additional scripts. Also the push workflow references channel/account identifiers but the skill does not request or store credentials — delivery depends on your platform's existing integrations. - If you accept these operational constraints (agent will perform scheduling, summarization, and pushing), this skill appears coherent and proportionate. If you prefer an all-in-one solution, request or require implementations for scheduling, summary generation, and secure channel delivery before enabling automated pushes.

每日 arXiv 论文整理技能。

功能概述

这个技能帮助用户在指定时间自动获取整理 arXiv 论文，并在指定时间推送总结给用户。

工作流程：

1. 提取（如 09:38）- 调用 fetch.py 获取当天指定领域的论文
2. 推送（如 14:00）- Agent 读取论文信息，生成总结并推送给用户

目录结构

arxiv-daily/
├── SKILL.md              # 本文件
├── scripts/
│   └── fetch.py          # 获取论文的脚本
├── references/           # 用户配置文件目录
│   └── like-<category>.txt   # 用户喜好配置
└── data/                 # 论文数据存储目录
    └── <category>/
        ├── YYYY-MM-DD-<category>.txt   # 每日论文数据
        └── YYYY-MM-DD-<category>.md   # 每日论文的翻译总结

首次使用 / 初始化配置

当用户首次安装、询问或调用此技能时，通过对话引导用户完成初始化：

需要询问用户的信息：

1. 感兴趣的领域

   • 引导用户访问 https://arxiv.org/ 查看分类
   • 确认领域关键字，如 cs.AI、cs.CV、cond-mat.mtrl-sci 等

2. 提取时间

   • 建议在人少的凌晨，如 9:38
   • 记录为 cron 表达式格式
   • 提醒用户注意时区，arxiv在工作日的00:00 UTC更新。

3. 推送时间

   • 如 14:00
   • ⚠️ 重要：如果提取时间和推送时间相近，提醒用户拉长间隔
   • 建议：要么更早提取（如 9:00），要么更晚推送（如 20:00）

4. 推送配置

   • 如： channel: WeCom, account: papa_wecom, chat_id: wecom-agent:LinHaoWei
   • 建议：询问用户是否要通过现在正在使用的聊天渠道推送，并向用户展示当前对话的 channel，account，chat_id

5. 语言偏好

   • 如：中文

6. 翻译内容

   • 选项：仅摘要 / 标题+摘要

初始化流程：

1. 询问上述6个问题
2. 生成配置文件 like-<category>.txt 保存到 references/ 目录
3. 引导用户创建定时器（提取任务 + 推送任务）
4. 确认初始化完成

配置文件格式

references/like-<category>.txt 示例：

# arXiv 论文推送配置
category: cs.CV
fetch_time: 38 9 * * *      # 每天早上 9:38
push_time: 0 14 * * *       # 每天下午 14:00
channel: WeCom, account: papa_wecom, chat_id: wecom-agent:LinHaoWei
language: zh-CN
translate: title+abstract  # 可选: abstract-only, title+abstract

每日提取流程

1. 读取配置

   • 扫描 references/like-<category>.txt 获取对应领域用户配置

2. 调用脚本

   • 根据读取的配置信息，调用fetch.py脚本

3. 读取论文数据

   • 根据 category 读取最新的 data/<category>/YYYY-MM-DD-<category>.txt

4. 生成总结

   • 对txt文档的每篇论文生成 2-3 句话的摘要总结
   • 包含：标题、链接、摘要总结、Comment
   • 保存到 data/<category>/YYYY-MM-DD-<category>.md
   • 注意要处理txt文档中所有论文，不要只处理一部分。
   • 总结文档的格式为：

今日 arXiv 论文速递 (cs.CV - 2024-01-15)

共 61 篇新论文

1. **论文标题**
   🔗 https://arxiv.org/abs/xxxx.xxxxx

   💡 摘要：这篇论文提出了...（2-3句话总结）

   💬 Comment：CVPR 2026 Main Track

2. **论文标题**
   🔗 https://arxiv.org/abs/xxxx.xxxxx

   💡 摘要：...

   💬 Comment：Project page:this https URL

...

61. **论文标题**
   🔗 https://arxiv.org/abs/xxxx.xxxxx

   💡 摘要：...

   💬 Comment：IEEE

📊 查看全部：https://arxiv.org/list/cs.CV/recent

内容要求：

• 标题：原文标题（或翻译后的标题）
• 链接：arXiv 论文链接
• 摘要：2-3句话总结论文核心内容
• Comment：文章中的comment内容

每日推送流程

当到达推送时间时：

1. 读取配置

   • 扫描 references/like-<category>.txt 获取对应领域用户配置

2. 推送：

   • 根据读取的配置，按照约定的渠道将处理好的当天md文件发送给用户。
   • 发送文案为：

🌅 下面将今日 arXiv 论文速递 (cs.CV - 2024-01-15)发送给您。

添加新领域

当用户想要添加新的论文领域时：

1. 参考「初始化配置」流程，询问同样的5个问题
2. 生成新的配置文件 like-<new-category>.txt
3. 引导用户创建新的定时器
4. 确认添加完成

脚本使用

run.sh

./rscripts/run.sh cs.CV data/cs.CV/ 2026-03-19

fetch.py

python scripts/fetch.py cs.CV data/cs.CV/ 2026-03-19

参数：

• arXiv 分类，如 cs.CV, cs.AI
• 保存路径
• 日期（留空为今日）

注意事项

1. 时间间隔：提取时间和推送时间至少间隔 3 小时以上，确保数据已准备好
2. 多领域支持：用户可以有多个 like-*.txt 配置文件
3. 数据保留：论文数据保留 7 天，自动清理过期文件
4. 错误处理：如果某天的数据获取失败，推送时告知用户
5. 周六周日：arxiv在周六和周日不更新，设置定时任务时可以跳过周六周日。