feat(equity-research): Add AI-powered company research agent plugin - feat(equity-research): 新增 AI 驱动的公司研究报告 Agent 插件

[xyangryr-cyber]

Summary

An OpenClaw-based AI agent that generates institutional-grade equity research reports from a single
Feishu message. Turns hours of analyst work into minutes.

Pain Points Solved

• Manual research is slow: Analysts spend 4-8 hours gathering data from filings, earnings
  calls, and market sources for a single company update. This agent automates the entire data
  collection → analysis → writing pipeline.
• Report quality is inconsistent: Without strict structural contracts, reports vary wildly in
  depth, citation quality, and risk analysis. This plugin enforces 9-chapter structure, mandatory
  inline citations, risk severity layering, and financial data tables.
• Hallucination in AI-generated reports: Most LLMs fabricate financial numbers when asked to
  write reports. This agent has hard rules: no tool call = no report. Every data point must come from
  real-time web search, not parametric knowledge.
• Delivery friction: Reports stuck as local files never reach stakeholders. This agent
  automatically creates Feishu Docs and delivers the URL — zero manual copy-paste.

Highlights

• Multi-model orchestration: qwen3-max for data collection, Claude Sonnet 4 for analysis &
  writing — optimizing cost and quality at each phase
• Evidence contract system: Tiered source hierarchy (official filings > market data > news),
  with automatic degraded mode when evidence is insufficient
• 4-class request taxonomy: Class A (initiating coverage) through Class D (quick lookup), each
  with calibrated depth, structure, and citation requirements
• Battle-tested guardrails: Fixes for 3 real production bugs baked into the agent rules —
  process message leaking, empty Feishu docs, async subagent delivery failure
• Sector-aware KPI checklists: Reference files for consumer, internet, semiconductor, and more
  — ensuring domain-relevant metrics in every report

Plugin Structure

• equity-research/docs/ — workflow, evidence/output contracts, compliance & quality checklists,
  degraded mode specs
• equity-research/skills/initiating-coverage/ — report templates, sector KPI references, example
  excerpts
• equity-research/openclaw-agent/ — OpenClaw runtime config (AGENTS.md, BOOTSTRAP.md, models.json
  with credential placeholders)

Test Plan

• TSMC Class B report — 9 chapters, 15 inline citations, financial tables, risk layering ✅
• Zhipu AI Class A report — 10 specific data sources, scenario analysis, comparable companies
  ✅
• No credential leaks — all secrets use ${PLACEHOLDER} format ✅

🤖 Generated with Claude Code

---

概述

基于 OpenClaw 的 AI Agent，一条飞书消息即可生成机构级公司研究报告。将分析师数小时的工作压缩到几分钟。

解决的痛点

• 人工研究太慢：分析师写一份公司更新需要 4-8 小时收集财报、业绩会、市场数据。本 Agent
  自动化了从数据采集 → 分析 → 写作的全流程。
• 报告质量参差不齐：没有结构化约束，报告在深度、引用质量、风险分析上差异巨大。本插件强制执行 9
  章节结构、行内引用、风险分层（高/中/低）、财务数据表格。
• AI 生成报告的幻觉问题：大多数 LLM 直接编造财务数字。本 Agent 硬规则：没有工具调用 =
  不允许写报告。每个数据点必须来自实时搜索，禁止使用参数知识。
• 交付断裂：报告卡在本地文件，永远到不了使用者手里。本 Agent 自动创建飞书文档并交付
  URL，零手动操作。

项目亮点

• 多模型协作：qwen3-max 负责数据采集，Claude Sonnet 4 负责分析与写作——在每个阶段优化成本和质量
• 证据合约体系：分层信源（官方披露 > 可复核市场数据 > 新闻），证据不足时自动切换降级模式
• 4 级请求分类：Class A（首次覆盖）到 Class D（快速查询），每级有对应的深度、结构和引用要求
• 实战验证的防护规则：3 个真实生产环境 Bug 的修复已固化到 Agent
  规则中——过程消息泄露、飞书空文档、异步子 Agent 交付失败
• 行业感知的 KPI 清单：覆盖消费、互联网、半导体等行业的参考文件，确保每份报告包含领域相关指标

插件结构

• equity-research/docs/ — 工作流、证据/输出合约、合规与质量清单、降级模式规范
• equity-research/skills/initiating-coverage/ — 报告模板、行业 KPI 参考、示例摘录
• equity-research/openclaw-agent/ — OpenClaw
  运行时配置（AGENTS.md、BOOTSTRAP.md、models.json，凭证已替换为占位符）