risk-sentiment-scanner

企业舆情信用风险扫描。当用户提供企业名称列表时，自动抓取各企业最新公开舆情（新闻、公告、监管信息），结合风险评分模型输出结构化信用风险报告。触发场景：（1）用户说"扫描风险"、"舆情分析"、"信用风险评级"；（2）用户提供了一份企业名单需要批量评估；（3）用户上传含企业名称的CSV/TXT文件；（4）定期复检提醒...

Security Scan

The skill's description promises automated web scraping and LLM-driven risk scoring, but the included script does not perform the web fetching itself and the runtime instructions reference external tools and a Feishu push without declaring required credentials — this mismatch warrants caution.

Purpose & Capability

The skill claims end-to-end scraping + rating. SKILL.md instructs the agent to call batch_web_search and extract_content_from_websites to gather articles. The included scripts/scan.js only performs local text analysis and file I/O (reads a watchlist, writes reports) and explicitly expects the agent to supply fetched news texts. This is a coherence mismatch (placeholder analysis code vs claimed full pipeline) but can be explained by design (agent performs network steps).

Instruction Scope

Instructions direct the agent to read memory/risk-watchlist.md, accept uploaded CSV/TXT, use external search/extraction tools, and optionally push reports to Feishu. Reading the memory file and saving reports are expected, but the skill also tells the agent to transmit data to third-party endpoints (Feishu) without declaring required credentials. The agent will therefore have discretion to call external endpoints and access user memory — review what tool permissions the agent will get before enabling.

Install Mechanism

No install spec; skill is instruction-only with a small local script. No downloads or archive extraction are specified, so there is low installation risk. scripts/scan.js is plain JS and performs only local processing and filesystem writes.

Credentials

SKILL.md references pushing to Feishu (Feishu Skill interface) and using 'batch_web_search'/'extract_content_from_websites' tools, but the skill declares no required environment variables or credentials. If you intend to enable Feishu pushes or other third-party integrations, credentials will be needed — their absence in requires.env is an inconsistency to clarify.

Persistence & Privilege

The skill does not request always:true and is user-invocable only. It reads/writes files under the agent's memory and reports directories (memory/risk-watchlist.md, memory/risk-reports, reports/...), which is expected for this use case; it does not modify other skills or system-wide settings.

What to consider before installing

This skill appears to be a judgment/analysis layer that expects the agent to perform web searches and extract article text — the bundled script only scores text and writes reports. Before installing: (1) Confirm what tools/permissions your agent will get (batch_web_search, extract_content_from_websites) so you understand what external calls will be made; (2) If you plan to push to Feishu, prepare and review the required Feishu credentials and only grant them if you trust the skill; (3) Test the skill on non-sensitive / dummy company names to verify it actually fetches content and doesn't produce misleading 'low risk' defaults (the script returns low-risk when no news is provided); (4) Be aware the skill will read memory/risk-watchlist.md and write reports to disk — avoid storing sensitive secrets in those files. If the owner/source is unknown and you need guaranteed accountability for data handling, treat this as untrusted until provenance and credential usage are clarified.

对目标企业进行公开舆情信息抓取 + LLM 驱动信用风险评级，输出结构化 JSON 报告。

工作流程

Step 1 — 接收企业列表

支持三种输入方式：

方式 A（对话直接提供）：

蚂蚁集团
贵州茅台
碧桂园

方式 B（上传文件）： 读取用户上传的 .txt 或 .csv 文件，每行一个企业名称，自动去重。

方式 C（Cron 定时触发）： 从 memory/risk-watchlist.md 读取企业名单（可由用户提前维护）。

Step 2 — 舆情信息抓取

对每个企业执行以下操作：

# 搜索近6个月相关舆情（多关键词组合）
企业名称 + "风险"
企业名称 + "违规 / 处罚 / 调查"
企业名称 + "债务 / 重组 / 违约"
企业名称 + "监管 / 合规"
企业名称 + "经营 / 财报 / 转型"

使用 batch_web_search 工具，每个企业最多抓取 8 条搜索结果。

重点来源优先级：

1. 证券时报 (stcn.com)
2. 21财经 (21jingji.com)
3. 新浪财经 (finance.sina.com.cn)
4. 腾讯新闻 (news.qq.com)
5. 东方财富网 (eastmoney.com)

Step 3 — 正文提取

对每家企业，取搜索结果中 最新且相关的 3 条 URL，使用 extract_content_from_websites 提取正文内容。

提取策略：

• 每篇正文最多读 2000 字（截断后半部分）
• 保留：标题、时间、风险相关段落
• 丢弃：广告、评论区、导航内容

Step 4 — LLM 风险评级

将整理后的舆情文本发给 LLM，按以下结构输出（每家企业独立评分）：

{
  "company": "企业名称",
  "risk_level": "R1-低风险 | R2-中低风险 | R3-中高风险 | R4-高风险",
  "risk_score": 0-100,
  "risk_trend": "上升 | 稳定 | 下降",
  "key_positive_factors": ["利好因素1", "利好因素2"],
  "key_negative_factors": ["风险因素1", "风险因素2"],
  "red_flags": ["需重点关注信号1", "信号2"],
  "news_summary": "近6月舆情摘要（100字内）",
  "recommended_action": "业务合作建议",
  "review_frequency": "月度 | 季度 | 半年度",
  "data_sources": ["来源1 URL", "来源2 URL"],
  "last_updated": "YYYY-MM-DD"
}

Step 5 — 风险评分模型（参考）

维度	权重	评分逻辑
监管/处罚	30%	有处罚记录+R25，有重大违法+R40
财务压力	25%	债务重组进行中+R30，违约记录+R50
公司治理	20%	高管被查/反腐+R25，持续动荡+R40
舆情情绪	15%	主流负面报道占比>60%+R20
经营状况	10%	营收/利润持续下滑+R15

最终等级：

• R1（0–25分）：低风险，舆情正面
• R2（26–50分）：中低风险，有可管理风险敞口
• R3（51–75分）：中高风险，偿债能力承压
• R4（76–100分）：高风险，业务可持续性存疑

Step 6 — 输出与推送

输出内容（JSON 格式）：

{
  "report_date": "2026-03-18",
  "total_companies": 3,
  "summary": {
    "R1_count": 0,
    "R2_count": 1,
    "R3_count": 1,
    "R4_count": 1,
    "high_risk_companies": ["碧桂园", "贵州茅台"]
  },
  "companies": [ /* 每家企业的完整评级 JSON */ ]
}

输出位置（按用户偏好）：

• 默认：直接输出到当前对话
• 可选：保存至 memory/risk-reports/YYYY-MM-DD.md
• 可选：推送至飞书文档（通过 Feishu Skill 接口）

Step 7 — 异常告警

若扫描结果中出现 R4 级企业，自动在报告顶部输出红色告警：

🚨 【高风险预警】碧桂园 — R4(78分)
   触发原因：177亿美元债务重组仅支付2%本金，净资产极度薄弱
   建议动作：建议回避新增敞口，存量业务降级处理

---

维护企业监控名单

用户可通过以下方式管理要跟踪的企业列表：

1. 对话更新：直接告诉我"把 XXX 公司加入监控列表"
2. 文件维护：编辑 memory/risk-watchlist.md，每行一个企业名称+备注

格式示例：

# 风险监控名单
蚂蚁集团        # 科技/金融，头部
贵州茅台        # 白酒，R3观察中
碧桂园          # 房地产，R4，存量处置
宁德时代        # 新能源，季度复检

---

注意事项

• 本 Skill 仅基于公开信息，不构成正式信用评级
• 风险评分为 LLM 判断结果，供参考，不当作为唯一决策依据
• 涉及投资、信贷等重大决策前，建议咨询专业金融顾问