功能描述

深度研究技能，用于进行领域调研、文献调研、survey研究。当用户说"做一个survey"、"深度研究一下XX"、"文献调研"、"研究一下XX领域的最新进展"、"帮我调研XX"、"学术调研"时自动触发。支持arXiv、PubMed、PMC、Google Scholar等多个数据源，自动下载PDF并解析全文，生成三...

使用说明 (SKILL.md)

Skill: Adaptive Depth Research v6.0 Universal

Name: Deep Research v7
Author: xueylee-dotcom

版本：6.0.0 描述：领域无关 | 配置驱动 | 数据源自适应 | 三层输出

🎯 核心设计原则

数据源自适应：arXiv/PMC 自动下 PDF，付费期刊自动切摘要模式
领域无关：提取逻辑不依赖特定术语，靠配置驱动
输出分层：执行摘要 + 验证清单 + 完整报告，各取所需
一次配置，全领域复用

📦 架构图

【配置层】
config/research-config.yaml  # 定义领域、关键指标、数据源优先级

      ↓

【检索层】(自动分流)
├─ arXiv → 下载 PDF → 全文解析 → 高可信提取
├─ PMC → 下载 PDF → 全文解析 → 高可信提取
├─ PubMed → 仅摘要 → 方向性提取 + 标注"需验证"
└─ Web → 网页抓取 → 关键结论提取

      ↓

【提取层】(通用 Prompt)
prompts/extract-universal.txt  # 不依赖领域术语

      ↓

【输出层】(三层报告)
├─ reports/executive-summary.md  # 决策者用，≤1 页
├─ reports/validation-checklist.md  # 执行者用，可操作
└─ reports/full-report.md  # 审计用，完整溯源

🚀 触发命令

# 完整研究流程
bash scripts/run-research.sh "\x3C主题>" --domain "\x3C领域>"

# 示例
bash scripts/run-research.sh "transformer efficiency" --domain "machine learning"
bash scripts/run-research.sh "telemedicine cost savings" --domain "healthcare"

📁 文件结构

skills/deep-research/
├── SKILL.md
├── config/
│   └── research-config.yaml      # 领域配置
├── prompts/
│   ├── extract-universal.txt     # 通用提取Prompt
│   ├── cluster-cards.txt         # 卡片聚类
│   └── write-brief.txt           # 主题简报
├── templates/
│   ├── executive-summary.md      # 执行摘要模板
│   ├── validation-checklist.md   # 验证清单模板
│   └── full-report.md            # 完整报告模板
└── scripts/
    ├── run-research.sh           # 完整研究流程
    ├── fetch-and-extract.sh      # 自动分流提取
    ├── extract-from-pdf.py       # PDF解析
    └── check-sourcing.sh         # 溯源验证

🔧 配置说明

修改研究领域

编辑 config/research-config.yaml:

research_domain: "your_domain_here"

key_metrics:
  performance:
    - accuracy
    - AUC
    - F1
  # 添加你的指标...

📊 三层输出说明

1. 执行摘要 (executive-summary.md)

受众：决策者
长度：≤1页
内容：
- 核心结论（已验证 vs 待验证）
- 可直接行动
- 需验证后行动

2. 验证清单 (validation-checklist.md)

受众：执行者
格式：操作导向
内容：
- 缺失指标汇总表
- 具体验证路径
- 通用验证方法

3. 完整报告 (full-report.md)

受众：审计/存档
内容：
- 方法论说明
- 已验证结论 + 证据
- 待验证线索
- 战略建议（短/中/长期）
- 完整卡片索引

🎯 执行流程

Step 1: 检索

# 自动检索 arXiv + PubMed
bash scripts/run-research.sh "\x3C主题>"

Step 2: 提取

# 自动分流提取
bash scripts/fetch-and-extract.sh \x3Csource_url>

# 或手动提取
python3 scripts/extract-from-pdf.py card-001 "\x3Cpdf_url>"

Step 3: 验证

# 溯源检查
bash scripts/check-sourcing.sh reports/full-report.md sources/

Step 4: 生成报告

三层报告自动生成在 reports/ 目录

✅ 质量门禁

数据完整度标注：high/medium/low
缺失指标清单：每个卡片明确列出
验证路径具体：可操作，非模糊建议
溯源验证：所有数据可溯源到卡片

📈 v6.0 vs 之前版本

维度	v5.x	v6.0 Universal
领域	医疗特化	领域无关，配置驱动
数据源	固定PubMed	自适应分流
提取	依赖医疗术语	通用Prompt
输出	单一报告	三层输出
配置	硬编码	YAML配置

💡 用户只需

修改配置：改 research_domain 和 key_metrics
运行命令：bash scripts/run-research.sh "\x3C主题>"
阅读摘要：5分钟了解核心结论
按清单验证：30分钟补全关键数据
推进决策：基于验证结果

🔑 核心哲学

让工具适应人，而不是人适应工具。
你专注领域知识和商业决策，工具负责广度扫描、结构化整理、诚实标注。

Skill版本：6.0.0 Universal | 最后更新：2026-03-19

v7.0 融合版 (2026-03-21)

新增能力

Research Claw 桥接：

搜索：arXiv, Google Scholar, PubMed, Semantic Scholar
PDF解析：使用pymupdf
网页读取：Web Reader Tool

使用方法

from scripts.research_claw_bridge import ResearchTools

tools = ResearchTools()

# 搜索论文
papers = tools.search_papers("LLM agent", max_results=5)

# 完整流程（搜索+下载+提取）
results = tools.full_research_flow("商保控费")

文件

scripts/research_claw_bridge.py - 桥接模块

安全使用建议

What to consider before installing/running this skill: - Review the code before running: the skill contains many executable Python scripts that will perform arbitrary web fetches, PDF downloads, parsing, and call a ResearchTools bridge. Inspect scripts/research_claw_bridge.py and cookie-management code to see what remote endpoints and credentials they use and whether any data is transmitted to unexpected hosts. - Credentials and API keys: the bundle depends on 'openai' and mentions LLM usage but does not declare required environment variables (e.g., OPENAI_API_KEY) or other API credentials. If you run it, supply only minimal, scoped keys in an isolated environment; do not expose high-privilege keys on your main machine. - Network & paywalled content: INTEGRATION.md and scripts mention using cookies to bypass logins and scraping paywalled content. Decide whether you are comfortable with that behavior and confirm it complies with terms of service and local laws. - Isolation: run the skill in a sandbox or disposable environment (container/VM) with no access to sensitive host files. Pay attention to paths the scripts write to (they use /root/.openclaw/workspace/... by default). Change output directories to a safe location before running. - Cross-skill imports: scripts insert absolute workspace paths and import code outside the skill; ensure those paths are expected and review the referenced external code. That import pattern can allow the skill to read other skills' data. - If you need to use it: 1) open and read scripts/research_claw_bridge.py and cookie_manager.py first; 2) search for any 'requests.post' or custom endpoints within the code to see where data is sent; 3) run with limited permissions and network monitoring; 4) prefer providing API keys via temporary/scoped credentials. Given these mismatches (undeclared credential needs, absolute workspace access, and scraping-with-cookie hints), treat the skill as suspicious until you verify the code and runtime behavior in an isolated environment.

功能分析

Type: OpenClaw Skill Name: deep-research-v7 Version: 7.0.0 The skill bundle contains a hardcoded Baidu Qianfan API key ('bce-v3/ALTAKSP...') within 'scripts/deep_research_v8.py' and 'scripts/research_claw_bridge.py', which constitutes a significant credential leakage vulnerability. While the tool's logic for academic research, PDF parsing, and report generation appears legitimate and highly sophisticated, it performs extensive high-risk operations including downloading remote files, scraping web content via proxies like 'r.jina.ai', and executing subprocesses. The hardcoded credentials and the broad network/filesystem access required for its functions make it suspicious, although no clear evidence of intentional malice or data exfiltration was found.

能力评估

⚠ Purpose & Capability

The declared purpose (literature/survey research, PDF download, parsing, report generation) matches the included scripts, but the package also lists and appears to expect LLM calls (openai) and external-search integrations while declaring no required environment variables or credentials. Several files import or read from absolute paths under /root/.openclaw/workspace/research-claw/... which implies the skill expects/attempts to access other workspace resources outside its own directory. That cross-path access is not explained by the SKILL.md and seems disproportionate to the stated simple 'survey' trigger.

⚠ Instruction Scope

SKILL.md instructs running many scripts that fetch web pages, download PDFs, and parse full text; integration docs and scripts explicitly reference cookie usage and 'bypass login' techniques and batch web scraping. The runtime instructions reference files like v9_papers_filtered.json and write into /root/.openclaw/workspace paths. The instructions therefore direct access to arbitrary URLs, potential paywalled content (with cookie hints), and local workspace files outside the skill - broader scope than the manifest declares.

ℹ Install Mechanism

There is no install spec (instruction-only), which lowers supply-chain risk. However, the bundle includes many executable Python scripts that will be run directly. No external archives or remote installers are fetched at install time, but runtime actions can perform network IO and file writes.

⚠ Credentials

clawhub.yaml lists dependencies including 'openai' and the SKILL.md / scripts call tools that likely invoke LLMs, but the skill declares no required env vars (e.g., OPENAI_API_KEY) or credentials (Semantic Scholar, APIs, or cookie tokens). Scripts and docs reference using cookies to access protected sites, yet no credential inputs are declared. Asking for web scraping and LLM access while not declaring needed secrets is an incoherence and operational risk (credentials might be requested or used ad-hoc).

⚠ Persistence & Privilege

The skill is not set always:true (good), but many scripts perform file reads/writes to absolute agent workspace locations (/root/.openclaw/workspace/...), and they insert that workspace path onto sys.path to import ResearchTools from a different project. That indicates the skill will access other skills' or global workspace files and could read or modify files outside its own scope. This cross-skill/workspace access is a privilege concern.

版本历史

v7.0.0

v7.0: Research Claw桥接版，支持多数据源搜索(arXiv/Semantic Scholar)和Survey报告生成，自动触发技能

元数据

Slug deep-research-v7

版本 7.0.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 1

常见问题

Deep Research v7 是什么？

深度研究技能，用于进行领域调研、文献调研、survey研究。当用户说"做一个survey"、"深度研究一下XX"、"文献调研"、"研究一下XX领域的最新进展"、"帮我调研XX"、"学术调研"时自动触发。支持arXiv、PubMed、PMC、Google Scholar等多个数据源，自动下载PDF并解析全文，生成三... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 103 次。

如何安装 Deep Research v7？

在 OpenClaw 或 Claude Code 对话框中运行命令「/install deep-research-v7」即可一键安装，无需额外配置。

Deep Research v7 是免费的吗？

是的，Deep Research v7 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。

Deep Research v7 支持哪些平台？

Deep Research v7 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。

谁开发了 Deep Research v7？

由 xueylee-dotcom（@xueylee-dotcom）开发并维护，当前版本 v7.0.0。

Deep Research v7