@sharesummer

对文本内容在数据层进行脱敏（非仅视觉）：默认覆盖主体身份、合同主体（公司名称、地址栏、法定代表人/代理人等）、商业秘密与内部数据、技术凭证与系统权限、文件元数据、特定格式隐蔽内容；支持假名化占位、等长遮蔽或擦除；可选 contractCnV1（仅 pseudonymize）：甲/乙方等主体占位（单侧多法人时为【甲方名1】【甲方名2】类推）、合同编号略称、住址槽位与中段遮罩 18 位身份证等，且与内置 contractParty 规则互斥。可附加自定义正则与说明。不发起网络请求。

Government/law link discovery. chinacourt.cn/article/index/id/… list pages: only #articleList (to paginationControl) → /article/detail/YYYY/MM/id/*.shtml; excludes 要闻/right sidebar on same HTML. CAC /wxzw/zcfg/: article hrefs /YYYY-MM/DD/c_*.htm; followPagination adds POST /cms/JsonList for page 2+ (URL stays index_1.htm). PBOC tiaofasi: list→detail prefers detail HTML URL over attachments inside detail; CSRC/SAFE extractors; NDA /zwgk/zcfb/list/→detail only /zwgk/zcfb/ (excludes /zjjd/ /ytdd/). Generic crawl. Pagination, nested, collectPdf. maxTotalFetches caps GETs+JsonList. No filesystem.

基于知识库与 LLM 的合同条款级审查；输出 riskSummaryStructured 与 revisionSuggestions；扩展 extendedReviewReport（法规对照结论、材料完整性局限、基础形式审查；条款归类含必备与实务高频维；权利义务矩阵摘要；非常规条款标记）。高质量模式、变更优先审、分段与汇总去重、JSON 修复。

Local adaptive optimizer: Fitness, S/A/B/C grades, and declarative recommendations from aggregated de-identified metrics (no network, no FS in-WASM). Host companion applies policy. Full product spec: local-adaptive-self-evolution-scheme.md. 本地自适应：根据宿主聚合的脱敏指标计算 Fitness、S/A/B/C 分级与声明式优化建议 JSON；执行与快照由伴生完成。完整方案见同目录 local-adaptive-self-evolution-scheme.md。

genesis-l0-constraint gene (safety)

genesis-web-search gene (search)

Convert a single .docx or .pptx (OOXML) file from base64 to semantic HTML for downstream kb-article-normalize / ingest. No network; no filesystem. Layout and embedded objects are best-effort; complex charts may be omitted with warnings.

RAG answer: user question plus retrieved chunks (with ids) → markdown answer and citations. Host performs retrieval; keys via apiKeys only.

Parse WeChat-style or generic article HTML into structured fields: title, published_at, author, is_original heuristics, body_text, optional tags. No network; host fetches HTML then passes raw_html.

OpenAI-compatible text embeddings: batch strings in, vectors out. Host writes vectors to pgvector/Chroma. Keys via apiKeys only.

诉讼策略两阶段基因：第一阶段输出结构化策略 JSON；第二阶段在 previousStrategyOutput 约束下输出举证方案、质证提纲与代理意见草稿。知识库字段与合同审查基因区分（程序法、证据规则、可选案例）；脱敏由宿主侧完成。

PPT 工作流：refine 根据 existingPptxBase64 大纲与说明调用 LLM，生成 pptxgen 附录；skipAppendixBuild 时仅返回方案供 Node 侧用模板克隆。文书审查 Agent 可在 HTTP 层上传模板并合并。merge 见 ppt-tool-runner.mjs。

Analyze the writing style of reference WeChat articles and generate new articles matching that style on a given topic. Supports style extraction (tone, structure, rhetoric), outline generation, and full article writing via LLM.

Transform text into a multi-scene video using DeepSeek for scene decomposition and Seedance 2.0 for AI video generation. Each scene becomes a real 4-15s video clip with native audio. Only needs a DeepSeek key + Seedance key.