dd-report

name	dd-report
description	Assemble a full company due-diligence report — entity resolution, registry profile, ownership and related parties, business and industry-chain position, risk records, adverse media, and (for issuers) financial red flags. Triggers on "尽调", "尽职调查", "due diligence", "DD报告", "排查这家公司", "供应商准入", "交易对手调查".

Company Due-Diligence Report

A structured, evidence-first DD report. Every section states what was checked, what was found, and what could not be covered.

Workflow

Step 1: Entity resolution

Resolve the exact legal entity via registry lookup (天眼查 search_companies). Confirm 统一社会信用代码/注册号, full legal name, and whether the target is the group, the listed vehicle, or a subsidiary. If multiple plausible entities exist, stop and confirm with the user. Keep the resolved company_id — the capability list in later steps requires it.

Step 2: Registry profile (登记信息 section)

Registration status, capital (认缴 vs 实缴 — label which), establishment date, legal representative, registered address, business scope, and recent registry changes (法人变更、注册资本变动、经营范围变更 — frequent changes are themselves a flag). 天眼查 get_company_basic_profile returns the aggregated picture directly (registry, contacts, tags, scale, former names) and needs no capability list; go through the capability list for get_company_registration_info and get_change_records when you need the itemised change history.

登记信息 is a section heading, not a provenance tag: every field here is [披露] with its [n]. Any ratio you derive from these fields (实缴/认缴 占比, 变更频次) is [测算], and "frequent changes suggest instability" is [推断].

Step 3: Ownership and related parties

Run the related-party-map skill's core steps: shareholders (with percentages and dates), outbound investments, brother companies, and — via 天眼查 (get_suppliers_and_customers, get_relation_graph) — supply-chain and funding-chain relationships. Summarize the map here; attach the full map as an appendix or separate deliverable if large.

Step 4: Business and industry position

• What the company actually sells, to whom; key customers/suppliers from 天眼查 supply-chain links (label data vintage).
• Industry-chain position inferred from the business description and supply-chain links ([推断] — no structured chain-mapping source), with concentration risk where the chain data exposes it.
• For listed or bond-issuing entities: recent announcements (万得 wind-docs.get_company_announcements) and issuer financial profile (hexin-bond.bond_basic_info, bond_financial_data) — leverage, profitability, contingent liabilities.
• Agency credit rating, where the entity has issued bonds — hexin-bond.bond_special_data. Returns 债项评级, 主体评级(主评机构), 主体评级展望, 主体最新评级变动方向 (上调/下调/维持), 评级机构, 评级类型, 最新评级日期. A rating downgrade or a 展望 moved to 负面 is first-order evidence for "can we do business with them" and outranks most media findings. Quote it attributed to the named agency with its 评级日期; this skill still issues no rating of its own.
  • Read 主体评级类型 back every time. On a guaranteed bond it can return 债券担保人信用评级 — the guarantor's rating, typically from a different agency than the 债项评级; on a non-guaranteed bond the same column returns 主体长期信用评级, the entity's own. Carrying a guarantor's AAA as the counterparty's own rating makes a credit-dependent entity look standalone — the exact error a DD report exists to prevent. Where the type is a guarantor rating, label it as the guarantor's, name the guarantor, and record the entity's own rating as 检索范围内未发现.
  • Non-issuers have no rating here; that is 不适用, not 检索范围内未发现.

Step 5: Risk records (风险记录 section)

Run the risk-scan skill for the target AND its controlling shareholder / actual controller / key related entities identified in Step 3. Each of those is a separate subject: pull a fresh capability list per entity, because an absent dimension on the target says nothing about its controller.

For natural persons (实控人、法定代表人、董监高), 天眼查 get_company_people lists them and get_person_risk_profile returns that person's 风险总览、失信、被执行、限制消费、终本案件、司法协助 — direct entry points, no capability list needed. Company-oriented checks miss these entirely. Related-party risk transmits: a clean target with a defaulted controller is not clean, and the report says so rather than leaving the reader to join two sections.

Join the two directions on the name. risk-scan's 限制高消费 check returns a restrictedPerson per record — the natural person the court restricted, typically the 法定代表人 or an 实际控制人. Match that name against the Step 3 roster and against Step 2's 法定代表人: a 限高 naming the current legal representative is a live finding about the company's own management, while one naming a predecessor is dated and says so. Where the name matches nobody in either list, say that rather than dropping the record.

风险记录 is likewise a section heading, not a tag — a recorded 失信/涉诉/处罚/质押 item is [披露]. Each check carries one of 有记录 / 检索范围内未发现 / 源不可用 into Step 7's coverage table.

Step 6: Adverse media (舆情 section)

万得 wind-docs.get_financial_news over the trailing 24 months: 违约、被查、爆雷、纠纷、监管 keywords plus the company name. Media findings are [媒体] and stay [媒体] until a disclosed record corroborates them, at which point they become [披露] and cite that record.

Step 7: Assemble

Short-form by default: Markdown in-session, per the house formatting policy. A full DD report goes to PDF via the report-render skill — never hand-rolled with weasyprint, wkhtmltopdf, pandoc, or a bare reportlab script, because those do not emit [n] as PDF link annotations and the citations arrive unclickable. A batch of subjects screened side by side goes to .xlsx via xlsx-author. Word is the answer only when the user asks for it or the reader will edit the file — report-render's DocxReport builds it with the same calls; unspecified long-form stays PDF (the house formatting policy). 要出 Word 就先载入 report-render 技能，再动手建。 DocxReport 与 Report 同一套调用；手搓 python-docx / docx-js 会丢掉标签配色、[n] 跳转与封面页，机制与实测记录在那个技能里，不在这里重述。

# [公司全称] 尽职调查报告
主体: [全称] / 统一社会信用代码 [码] · 检索于: [timestamp] · 委托口径: [用途,如供应商准入/授信/投资]
标签口径: [披露] 登记在册或主体披露 · [测算] 本文推导 · [预期] 第三方具名预期 · [推断] 分析师推论 · [媒体] 媒体报道未经记录佐证

## 结论摘要
- 总体印象一段(不下"通过/不通过"结论,列出决定性事实)
- 分级列表,前置至多 3 条 🔴: 🔴 高(决策前须澄清) / 🟡 中(记录并跟踪) / ⚪ 低·信息
  分级给到单条发现,不给到主体本身——本技能不出评级

## 一、主体信息(登记信息)
## 二、股权与关联方
## 三、业务与产业链地位
## 四、风险记录
## 五、舆情
## 六、(如适用)财务概览

## 覆盖范围与局限
检索于: [timestamp] · 口径/委托用途: [用途]

| 检查项 | 结论 | 源 | 检索于 |
|---|---|---|---|
| [逐项一行] | 有记录(N 项) [n] / 检索范围内未发现 / 源不可用 | [系统名] | [date] |

本次未能覆盖: [失败或未授权的源,以及它们本应覆盖的检查项]
数据滞后性: [所用各源的已知滞后,如登记变更公示、判决上网、处罚传输]

## 来源
[n] 〔一手|二手〕发布主体 · 文档或系统名 · 日期(发布日; 检索于) · URL

Sources entries follow the citation policy: 〔一手|二手〕 is mandatory, a 二手 entry names what it relays (二手 · 财新 · 转引 [原始报告] · [date](发布); 检索于 [date]), and a live database query with no publication date carries 检索于 [date] only. The count of distinct [n] markers equals the number of entries; 资料来源：工商数据、新闻检索 is not a citation.

Guardrails

• "检索范围内未发现" ≠ "无风险"/"无此事"/"通过" — always the former phrasing. A source that could not be queried is 源不可用, never folded into 未发现.
• Percentages, amounts, dates: only from retrieved records, each with [n]. Anything you computed from them is [测算].
• Deliver as Markdown by default; PDF/Word on request (embed CJK fonts for PDF).