Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

Verified Research

v1.2.1

执行多源核实的深入研究,拆解复杂问题,优先采集高可信度来源,核实信息时效性并生成结构化报告,支持三天缓存。

0· 57·0 current·0 all-time
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Suspicious
medium confidence
Purpose & Capability
The skill's name/description (deep multi-source research with 3-day cache) aligns with the provided scripts: research.sh, claim-card.sh, manifest.sh, finalize.sh, cleanup.sh implement caching, manifesting, report generation, and auto-archive. However, the skill touches /root/.openclaw/workspace/MEMORY.md (a global workspace memory), which is not declared under required config paths. Also the scripts require Python 3 and common CLI tools (md5sum, stat, sed, xxd, etc.) but the registry lists no required binaries — a mismatch between claimed requirements and actual runtime needs.
!
Instruction Scope
SKILL.md and the scripts instruct the agent to create per-session cache dirs under /tmp/deep-research-cache, produce reports, and schedule a 3-day cleanup. cleanup.sh will append a summary into /root/.openclaw/workspace/MEMORY.md and then delete the cached directories. This means (a) user research summaries are automatically persisted into the agent's global MEMORY.md without a separate explicit consent step at cleanup time, and (b) cached data is irreversibly deleted after the retention window. SKILL.md stated the full report is not copied to the workspace unless asked, but cleanup.sh still writes a summary to MEMORY.md — a documented contradiction that users should notice.
Install Mechanism
There is no install spec (instruction-only with shipped scripts), which minimizes supply-chain risk. The only risk is runtime: the scripts rely on Python 3 and common POSIX utilities but the skill metadata does not declare these as required binaries. No remote downloads or archive extraction are used.
!
Credentials
The skill declares no required environment variables or config paths, yet the scripts read/write /root/.openclaw/workspace/MEMORY.md and create /tmp/deep-research-cache/* directories. Writing into a global MEMORY.md (workspace memory) is effectively granting persistent storage access and could surface user data across sessions. No credentials are requested, which is good, but the implicit access to the agent's workspace is not disclosed in metadata.
Persistence & Privilege
always:false (normal). The skill does persist data to disk (temporary cache in /tmp) and appends summaries to the global MEMORY.md on cleanup. Cached directories are removed after the retention period. The automatic write to MEMORY.md and automatic deletion are significant behavioral privileges but are consistent with the skill's stated 3-day cache design — the concern is that this behavior is not declared in the metadata and may be unexpected to users.
What to consider before installing
This skill is broadly coherent with its stated purpose (multi-source research), but it has a few important discrepancies you should consider before installing: - The scripts write summaries into /root/.openclaw/workspace/MEMORY.md automatically during cleanup. That means research summaries will be appended to the agent's global memory without an extra consent step; if you do not want your research summaries persisted globally, do not install or edit the scripts to disable that behavior. - The metadata declares no required binaries or config paths, but the scripts require Python 3 and common CLI tools (md5sum, stat, sed, xxd, tail, etc.) and expect the workspace path /root/.openclaw/workspace to exist. Consider adding these as declared requirements or run in an environment where these are available. - cleanup.sh deletes cached research directories after archiving the summary. If you want to keep full reports, explicitly copy them to your workspace before the 3-day window expires. - If you plan to use this skill in a multi-tenant or sensitive environment, review/modify the scripts to: (a) prompt before writing to global MEMORY.md, (b) avoid hard-coded /root paths, and (c) explicitly declare runtime dependencies in the skill metadata. - Suggested mitigations: run the skill in a sandboxed agent instance first, inspect/modify the code to change the MEMORY.md write behavior or target a user-controlled path, and confirm Python 3 is available. If these points are acceptable or are intentionally intended behavior, the skill appears functionally coherent; otherwise treat it as suspicious and proceed with caution.

Like a lobster shell, security has layers — review code before you run it.

latestvk97e6jnk8k1e4zy94hc8pjt2kx84bg56methodologyvk97e6jnk8k1e4zy94hc8pjt2kx84bg56multi-sourcevk97e6jnk8k1e4zy94hc8pjt2kx84bg56researchvk97e6jnk8k1e4zy94hc8pjt2kx84bg56verificationvk97e6jnk8k1e4zy94hc8pjt2kx84bg56
57downloads
0stars
1versions
Updated 2w ago
v1.2.1
MIT-0

Deep Research Skill

Multi-source research methodology — executes when user asks "研究/调查/搜索" or similar research tasks. Supports 3-day research cache + auto-archive to MEMORY.md.


Research Cache System(研究缓存系统)

工作原理

每次研究启动时,在 /tmp/deep-research-cache/{slugified_topic}/{unix_timestamp}/ 下创建独立工作目录:

/tmp/deep-research-cache/
  {topic-slug}/
    {timestamp}/
      claims/          ← 每条证据卡(claim_NNN.md)
      rounds/          ← 每轮研究日志(round_NNN.md)
      manifest.json    ← 所有证据的索引 + 可信度统计
      report_final.md  ← 最终报告
      .cleanup_scheduled ← 3天清理倒计时标记

3天清理逻辑

  1. finalize.sh 生成报告后,写入 .cleanup_scheduled 标记(记录"报告生成时间")
  2. cleanup.sh 每运行一次,检查所有研究目录:
    • 有活动(用户在3天内继续研究,新增加了 claim/round)→ 删除 .cleanup_scheduled,取消清理计划
    • 无活动且距报告生成已 > 3天 → 将摘要写入 MEMORY.md,删除整个目录
  3. MEMORY.md 写入内容:主题 + 日期 + 核心结论(1-3句)+ T1/T2/T3/T4 来源统计 + 完整报告所在路径(待删除前还有效)

如何继续研究

用户只需说"继续研究{原话题}",AI agent 会:

  1. 找到 /tmp/deep-research-cache/ 下该 topic 的最新目录
  2. 从 manifest.json 读取已有证据
  3. 在最新目录继续追加 claim + round
  4. 更新 manifest + 重新生成报告

注意:完整报告在清理前仅存在于 /tmp/deep-research-cache/,不复制到 workspace。如需保留报告,请告知用户或主动复制到 workspace。


触发条件

用户说以下类型的话时触发:

  • "帮我研究一下..."
  • "你去查一查..."
  • "搜索..." + 涉及多个来源的问题
  • "帮我调查..."
  • "核实..."
  • "确认一下..."
  • "继续研究..."

注意:简单的事实查询("今天天气如何")不需要触发此技能。


核心原则

  1. 不搜到哪算哪 — 研究开始前先制定搜索计划
  2. 多源交叉验证 — 每个关键结论至少 3 个独立来源
  3. 来源分级可信度 — 优先使用高可信度来源
  4. 反复核实 — 同一事实用不同关键词核实至少 2 次
  5. 明确存疑 — 无法核实的内容必须标注,不得臆造

来源可信度分级(4级)

等级类型权重说明
T1官方文档/学术论文/一手数据/官方API最高原始出处,未经转述
T2权威媒体报道/行业白皮书/官方公告经核实的主流权威来源
T3技术博客/社区讨论/垂直论坛/GitHub Issues需要额外核实,可作为线索
T4社交媒体/论坛/无法溯源的引用极度存疑,仅作辅助线索

优先级规则:T1 > T2 > T3 > T4。优先使用 T1/T2;T3/T4 必须与 T1/T2 交叉验证。


研究流程(8步)

Step 0:问题类型识别

拿到问题后,先判断属于哪类:

类型特征示例
事实型有明确答案,可直接查证"XX工具的最新版本号"
比较型需要多维度对比"A工具 vs B工具,哪个更好"
探索型开放性,需要归纳"XX行业的市场趋势是什么"
操作型需要步骤指引"如何在XX上配置XX"

Step 0.5:时效性评估(AI/科技类必做)

  • AI/科技/开源项目话题:先核实信息时效性
  • 搜索时优先看最新发布的内容(近6个月内)
  • 明确标注每条信息的发布时间
  • 如果来源时间不明,该来源可信度降级

Step 1:问题拆解 & 边界定义

将问题拆解为若干子问题,明确:

  • 研究的范围(什么要查,什么不查)
  • 预期的输出形式
  • 需要查几个维度

Step 2:来源分级 & 权威锁定

  • 针对每个子问题,列出预期可用的来源类型
  • 优先查找 T1/T2 来源
  • 每个关键事实至少找到 2 个不同级别的来源

Step 3:事实提取 & 证据卡

每个关键结论提取为"证据卡":

[证据卡]
- 结论:...
- 来源:[来源名称]([等级],[时间])
- 原文摘要:...
- 可信度:✅确认 / ⚠️存疑 / ❌矛盾

使用缓存系统时,每张证据卡通过 claim-card.sh 写入缓存目录。

Step 4:构建对比框架

比较型问题必须建立对比矩阵:

[对比维度]
| 维度 | 来源A | 来源B | 来源C | 结论 |
|------|-------|-------|-------|------|

Step 5:参考对齐 & 冲突处理

  • 同一事实多个来源结论一致 → ✅ 确认
  • 多个来源结论矛盾 → ⚠️ 标注冲突,明确说明各方说法
  • 仅单一来源 → ⚠️ 标注"单来源,待验证"

Step 6:事实→结论推导链

明确推导过程:

结论X ← 依据事实A(来源)+ 依据事实B(来源)
         推导逻辑:[为什么A+B得出X]

Step 6.5:独立核实(BLOCKING 步骤)

在输出结论前,必须完成此步骤:

  • 不同关键词重新搜索关键结论,验证一致性
  • 每个关键结论至少核实 2 次(不同来源或不同角度)
  • 如果核实结果与原结论矛盾 → 返回 Step 5 重新评估
  • 完成后才可进入 Step 7

Step 7:使用场景验证(合理性检查)

  • 结论是否符合常识?
  • 是否存在明显的反例?
  • 对用户的实际场景是否适用?

Step 8:输出格式化

按以下格式交付(详见下方输出模板)。报告由 finalize.sh 生成。


输出模板

# 研究报告:[问题标题]

## 研究概要
- 问题类型:
- 研究时间:
- 来源数量:X 个( T1:X / T2:X / T3:X / T4:X )
- 核心结论(1-3句话):

## 关键结论

### 结论1:[标题]
- 置信度:🟢 高 / 🟡 中 / 🔴 低
- 证据:
  - [来源A](T1,2024年)— ...
  - [来源B](T2,2023年)— ...
  - [来源C](T3,2024年)— ...
- 多源核实:[✅ 一致 / ⚠️ 存疑 / ❌ 矛盾]
- 对用户场景的意义:...

### 结论2:(同上格式)

## 存疑项 & 待验证
- ❓ [未核实清楚的点] — 来源:[来源],风险:[...]
- ❓ [需要用户确认的内容]

## 研究局限性
- [本次研究的局限]
- [未覆盖的范围]

## 主要参考来源
1. [来源名] — [URL/平台] — [T1/T2/T3/T4]
2. ...

---
*本报告基于多源交叉核实,如有疑问请标注具体结论反馈核实。*

核实次数最低要求

结论来源等级最少核实次数
T1 单一来源2 次不同关键词/角度核实
T2 单一来源2 次不同来源核实
T1 + T2 多源一致1 次额外核实
T3/T4 结论必须升级到 T1/T2 来源,否则标注"低可信度"

脚本使用指南

启动研究

cd /root/.openclaw/workspace/skills/deep-research/scripts
TOPIC="你的研究主题" SESSION_ID="可选ID" bash research.sh

记录证据卡

CLAIM_ID="claim_001" \
TOPIC="研究主题" \
CONTENT="具体结论内容..." \
SOURCE="https://..." \
SOURCE_TIER="T2" \
VERIFICATION_STATUS="pending" \
ROUND="1" \
CACHE_DIR="/tmp/deep-research-cache/xxx/xxx" \
bash claim-card.sh

更新索引

CACHE_DIR="/tmp/deep-research-cache/xxx/xxx" \
TOPIC="研究主题" \
SESSION_ID="xxx" \
bash manifest.sh

生成报告

CACHE_DIR="/tmp/deep-research-cache/xxx/xxx" \
TOPIC="研究主题" \
SESSION_ID="xxx" \
bash finalize.sh

清理检查(可定时运行)

bash cleanup.sh

触发此技能时的用户沟通规则

研究开始时(简短告知):

"收到,我来研究一下。计划:[拆解成X个维度,查找X个来源],完成后给你完整报告。"

研究过程中(如果发现方向偏差):

"我发现原问题的方向可能需要调整——[说明原因],建议改成[新方向],可以吗?"

完成时:按上方模板输出完整报告,并告知:

"报告已生成,缓存保留3天,如继续研究请在3天内继续对话。"


质量门控

以下情况必须不上报结论,先继续研究:

  • 关键结论不足 3 个来源
  • 存在未解决的来源冲突
  • 时效性无法确认(AI/科技类)

以下情况必须标注"存疑"

  • 单一来源
  • 来源为 T3/T4
  • 发布时间超过 1 年(科技类话题)
  • 不同来源结论存在分歧

Self-Improving 集成

每次研究完成后,在 .learnings/research-log.md 追加一条:

[日期] | 问题:[主题] | 来源数:X | 质量:🟢/🟡/🔴 | 教训:[如果有]

如果研究过程中发现方法论问题 → 更新本 SKILL.md。


本技能参考 wshuyi/deep-research 方法论 + 学术来源可信度分级体系构建。

Comments

Loading comments...