中文 · English

md2chat share

给 DeepSeek Harness（Web 与 Desktop）每条完成的回复加一个 「分享到聊天」 按钮。回复里的 Markdown 会走 md2chat 自己的渲染管线——表格、代码、公式、Mermaid 变图片，正文保持文字——再按目标平台需要的样子写进剪贴板。

目标	剪贴板里得到什么
微信	按内容自动决定：没有要转图的块就复制纯文本；正文夹少量图文块时复制一份 HTML 富文本（文字 + 图片）；结构复杂或篇幅很短时生成一张长图
企业微信	同样的策略，但文字和图片按顺序分段交付——企微输入框不能一次粘贴图文混排。第一段在点击时复制，其余在对话框里逐段复制
图片	始终是一张长图 image/png；超过 md2chat 的像素上限时按行拆成多张
邮件	可编辑表格的 HTML，外加纯文本备选
Word	直接下载 .docx：表格、列表、代码块都是 Word 原生可编辑结构，不经过浏览器渲染

转换全部在本机完成：只有 Markdown 里的远程 ![](https://…) 图片会联网，正文不上传。

安装

dsh plugin add md2chat-dsh-share

也可以直接从仓库装：

dsh plugin add github:yingqi-studio/md2chat-dsh-share

需要 DeepSeek Harness Web 或 Desktop。

工作原理

• index.js —— Host（Node）半边。在 Harness 的连接鉴权围栏内注册三个路由：
  • GET /api/md2chat-share/vendor/*，按需提供 md2chat 的 MathJax、Mermaid、highlight.js；
  • POST /api/md2chat-share/word，把回复的 Markdown 解析成块，用内置的 OOXML 生成器写出 .docx 字节流。
• src/docx/ —— Word 导出，纯 Node、零依赖、无子进程：markdown.mjs 解析块与行内样式，ooxml.mjs 拼 document.xml / styles.xml / numbering.xml，zip.mjs 手写 ZIP，index.mjs 抓取远程图片并装配全部 part。不依赖 Chrome、不需要 macOS，Windows 和 Linux 上一样能用。
• src/client/ —— 浏览器半边。
  • index.ts 往 conversation.chat.assistant-actions 槽位注册入口。
  • reply.ts 从 Chat 快照里读出回复的文本块。
  • convert.ts 直接调用 md2chat 的核心文件而不改动它们（chat-payload.js、chat-markdown.js、renderers.js、table-images.js、chat-targets.js、clipboard-image.js），并负责投递策略。
  • ShareAction.tsx 是按钮和对话框。
• scripts/build.mjs 用 esbuild 产出 Harness 客户端 bundle 格式的 lib/client.js（window.__ModuleLoader__.load({ id, factory })）；React 与 @deepseek-ai/* 走壳层的模块表。按需加载的 vendor 渲染器复制到 lib/vendor/，由 Host 半边提供。

转换复用了 md2chat 桌面应用的核心文件，所以输出与它一致。对话框打开时开始转换，点击目标即刻写剪贴板；ClipboardItem 的 promise 让这次点击的用户激活在较慢的 Mermaid / MathJax 渲染期间保持有效。

开发

npm install
npm run build        # lib/client.js + lib/vendor + lib/docx + lib/docx
npm test             # 构建 + 单元测试
npm run typecheck

限制

• 只转换文本块，与 Harness 内置的「复制」一致。Markdown 里的远程图片会去抓取；本地路径和图片附件暂不支持。
• 不提供钉钉、飞书的原生剪贴板格式。
• Word 导出只认 PNG / JPEG / GIF 图片，WebP 会被丢弃；公式保持原始 LaTeX 文本，不转成 Word 公式对象。

---

English

Adds a Share to chat button under each finished reply in DeepSeek Harness (Web and Desktop). The reply's Markdown is converted with md2chat's own pipeline — tables, code, formulas and Mermaid diagrams become images, prose stays text — and the result is written to the clipboard in the shape each destination needs.

Destination	What lands on the clipboard
WeChat	Depends on the reply: plain text when there is nothing to rasterise; one HTML rich-text paste (text + images) for prose with a few blocks; a long image when the reply is structured or short
WeCom	The same policy, but text and images are delivered as ordered segments — the WeCom input cannot mix them in one paste. The first segment is copied on click, the rest one by one from the dialog
Image	Always one long image/png, split at line breaks into several images when it exceeds md2chat's pixel limit
Email	Editable-table HTML plus the plain-text alternative
Word	A .docx download: tables, lists and code blocks stay native, editable Word structure — nothing is rasterised

Conversion happens on your machine: only remote ![](https://…) images in the Markdown are fetched, the text itself never leaves it.

Install

dsh plugin add md2chat-dsh-share

Or straight from the repository:

dsh plugin add github:yingqi-studio/md2chat-dsh-share

How it works

• index.js — the Host (Node) half. Inside the Harness connection authentication fence it registers:
  • GET /api/md2chat-share/vendor/*, serving md2chat's lazily loaded MathJax, Mermaid and highlight.js bundles;
  • POST /api/md2chat-share/word, which parses the reply's Markdown into blocks and writes .docx bytes with a built-in OOXML writer.
• src/docx/ — Word export, plain Node, no dependencies and no subprocesses: markdown.mjs parses blocks and inline styles, ooxml.mjs assembles document.xml / styles.xml / numbering.xml, zip.mjs writes the ZIP, and index.mjs fetches remote images and lays out the package. No Chrome and no macOS: it works the same on Windows and Linux.
• src/client/ — the browser half.
  • index.ts registers an entry in the conversation.chat.assistant-actions slot.
  • reply.ts reads the reply's text blocks from the Chat snapshot.
  • convert.ts calls md2chat's core files without modifying them (chat-payload.js, chat-markdown.js, renderers.js, table-images.js, chat-targets.js, clipboard-image.js) and owns the delivery policy.
  • ShareAction.tsx is the button and the dialog.
• scripts/build.mjs uses esbuild to produce lib/client.js in the Harness client-bundle format window.__ModuleLoader__.load({ id, factory }); React and @deepseek-ai/* come from the shell's module table. The lazily loaded vendor renderers are copied to lib/vendor/ for the Host half to serve.

The conversion reuses the same core files as the md2chat desktop app, so output matches it. Conversion starts when the dialog opens; clicking a destination writes the clipboard immediately, and a ClipboardItem promise keeps the click's user activation alive across slow Mermaid/MathJax renders.

Develop

npm install
npm run build        # lib/client.js + lib/vendor + lib/docx + lib/docx
npm test             # build, then unit tests
npm run typecheck

Limitations

• Only text blocks are converted, the same as the Harness's built-in Copy. Remote ![](https://…) images in the Markdown are fetched; local paths and image attachments are not supported yet.
• DingTalk and Feishu native clipboard formats are not part of this plugin.
• Word export embeds PNG / JPEG / GIF images only; WebP is dropped. Formulas stay as LaTeX text instead of becoming native Word equations.