dsh-pdf-reader: a PDF reading workspace for DeepSeek Harness

[vaxilicaihouxian]

dsh-pdf-reader: a PDF reading workspace for DeepSeek Harness

@deepseek-ai/dsh-pdf-reader adds PDF reading directly to a DeepSeek Harness session. It connects a local PDF to the model, keeps reading state and notes durable, and adds a browser reading view for a more focused workflow.

What it does

• Open a PDF from the mounted Harness filesystem and extract text by page range with bounded output.
• Ask the model to translate, summarize, explain, or compare extracted passages while preserving page numbers as source context.
• Save and restore the latest reading position through the session log.
• Save quoted notes with a source page, list them, and export them as ordinary Markdown.
• Load a PDF from the web sidebar before the first chat message. The browser view renders a selectable PDF.js text layer and keeps the selected file, page, and notes in browser storage.
• Send selected text to the model, save the answer as a note, edit or delete local notes, and export the notebook.
• Organize all browser notes into one Markdown summary with a dedicated subagent request, while retaining a browser-local history that can be restored.
• Add a final assistant answer to the current PDF notebook from the chat actions.

The host-side notes use a Markdown sidecar next to the source PDF, for example paper.pdf.dsh-notes.md. The browser reading path keeps the selected file local to that browser; it does not upload the original file or expose its absolute path to the host.

Install

Install the tagged release into the Harness web profile:

dsh plugin --profile web add "github:vaxilicaihouxian/dsh-pdf-reader#v0.1.1"

Restart the web profile after installation. GitHub installs build the plugin locally through prepare; with pnpm 10+, pnpm may ask you to authorize the package in the profile's pnpm-workspace.yaml before the build can finish. Only authorize source you trust.

Manual composition

The package can also be mounted from cordis.patch.yml:

- id: pdf-reader
  name: '@deepseek-ai/dsh-pdf-reader'
  config:
    maxPdfBytes: 134217728
    maxExtractCharacters: 80000
    notesDirectory: dsh-notes

The host composition must provide dsh-fs, dsh-session, and dsh-tools. The package is an out-of-tree plugin and uses the public Harness capability APIs; it does not patch agent-loop.

Current limitations

• Extraction requires an embedded text layer; scanned PDFs need OCR, which is not included yet.
• Host notes are sidecar Markdown rather than PDF annotations.
• Host document IDs are process-local and must be reopened after a process restart.
• There is no background whole-book translation or summarization job. Large documents are handled through bounded page extraction and normal model requests.

Feedback welcome

I would especially like feedback on OCR/provider integrations, note organization prompts, and workflows for long technical papers. Issues and discussion are welcome in the dsh-pdf-reader repository.

---

dsh-pdf-reader：DeepSeek Harness 的 PDF 阅读工作区

@deepseek-ai/dsh-pdf-reader 将 PDF 阅读直接带入 DeepSeek Harness 会话。它把本地 PDF 接入模型，持久化阅读位置和笔记，并提供更适合连续阅读的浏览器视图。

功能

• 从 Harness 挂载的文件系统打开 PDF，按页提取文本，并限制单次返回量。
• 让模型基于带页码的原文进行翻译、总结、解释和对比。
• 通过会话日志保存和恢复最近的阅读位置。
• 保存带引用和来源页码的 Markdown 笔记，查看笔记列表，并导出为普通 Markdown 文件。
• 在第一次聊天前从网页侧边栏选择 PDF。浏览器视图使用 PDF.js 渲染可选中的文本层，并把选中的文件、页码和笔记保存在浏览器本地。
• 将选中文本发送给模型，把回答保存为笔记，编辑或删除本地笔记，并导出整个笔记本。
• 使用独立 subagent 请求把浏览器中的全部笔记整理成一份 Markdown 总结，同时保留可恢复的整理前历史。
• 在聊天消息操作中把最终的 assistant 回答添加到当前 PDF 笔记本。

主机侧笔记以 Markdown sidecar 的形式保存在源 PDF 旁边，例如 paper.pdf.dsh-notes.md。浏览器阅读路径只在当前浏览器本地保存选中的文件，不会上传原始文件，也无法把文件绝对路径暴露给主机。

安装

将带标签的版本安装到 Harness 的 web profile：

dsh plugin --profile web add "github:vaxilicaihouxian/dsh-pdf-reader#v0.1.1"

安装后重启 web profile。通过 GitHub 安装时，插件会在本机通过 prepare 从源码构建；使用 pnpm 10 或更高版本时，pnpm 可能要求先在 profile 的 pnpm-workspace.yaml 中授权该依赖。只应授权你信任的源码。

手动组合

也可以通过 cordis.patch.yml 挂载：

- id: pdf-reader
  name: '@deepseek-ai/dsh-pdf-reader'
  config:
    maxPdfBytes: 134217728
    maxExtractCharacters: 80000
    notesDirectory: dsh-notes

宿主组合需要提供 dsh-fs、dsh-session 和 dsh-tools。这是一个 out-of-tree 插件，只使用 Harness 的公开 capability API，不修改 agent-loop。

当前限制

• PDF 必须包含文本层；扫描版 PDF 需要 OCR，而当前版本尚未提供 OCR。
• 主机侧笔记是 sidecar Markdown，不会写入 PDF 注释。
• 主机侧 document id 只在当前进程有效，进程重启后需要重新打开 PDF。
• 当前没有后台整本翻译或总结任务；大文档通过限制页数的文本提取和普通模型请求处理。

欢迎反馈

欢迎就 OCR/provider 集成、笔记整理 prompt，以及长篇技术论文的阅读流程提出建议。问题和讨论请提交到 dsh-pdf-reader 仓库。