Preface¶

The common approach to having models read long articles, PDFs, or entire books in DeepSeek Harness (DSH) is to directly input the full text into the conversation or let the model freely summarize. This type of output often lacks traceable sources: how viewpoints correspond to the original text, which inferences are based on evidence, and which content is not supported by the original text, making it difficult to verify afterward.

DeepRead (xiehuan123/dsh-deepread) changes the reading process to be “evidence-first”: extracting traceable claims, evidence, confidence levels, knowledge maps, and review questions from articles, books, PDFs, files, or pasted text. Viewpoints not supported by the original text are clearly marked, and are not silently elevated to themes or fabricated evidence.

What Is This¶

DeepRead is published by maintainer xiehuan123 and classified as “Model Reasoning” in the SkillHub directory. The project is licensed under MIT, with the current npm version being 1.0.0, requiring Node ^22.19 || >=24.

The plugin is provided in two compatible forms:

• Portable Agent Skill: For Agent Skills-compatible tools like Codex and Claude Code, with zero runtime dependencies. Agents invoke file and web tools themselves according to the evidence-first workflow.
• Host plugin package: For DeepSeek Harness Web/headless and dsh-TUI, providing the deepread tool, PDF extraction, optional persistence/background tasks/Web routing, batch comparison, cost preview, and HTML/XMind-compatible exports; the browser client serves as an optional web entry point.

The GitHub repository currently has approximately 37 stars and 3 forks.

Core Features¶

Five Reading Modes¶

Input and Export¶

Supports three types of input: WeChat Official Account article URLs (stable links from mp.weixin.qq.com), local files (.txt/.md/.html/.pdf, PDFs use a built-in pure JS extractor with Chinese ToUnicode mapping and page markers), and pasted text.

Results are displayed within the session by default; export optionally supports md / mm (FreeMind, importable into XMind) / html (web report with light/dark themes) / all, written to the workspace deepread-output/.

Host Plugin Additional Capabilities¶

• Batch Comparison: batch takes 2–10 documents (url/path/text), outputting per-document summaries and a cross-document comparison matrix, conflicts, complementarities, and a synthesized report.
• Citation Traceability: Reports carry page/paragraph sources, and the citation table can be traced back to the original text’s 【Page N】 markers.
• Cost Preview: estimate: true previews token consumption, model call counts, and estimated time without invoking the model.
• Progress Transparency: Long articles, large PDFs, and batch tasks are converted to official background jobs, with progress streams pushed segment by segment; PDF parsing displays progress per page in a streaming manner.
• Web Panel: deepread tool result cards (four-level confidence legend, collapsible blocks) and a 📖 Reading Panel next to the input area; the panel supports budget pre-checks (POST /api/deepread/budget) and local recent reading history (localStorage).

DeepRead does not silently elevate themes to claims when the original text lacks support, nor does it fill gaps with fabricated evidence.

Installation and Enabling¶

Portable Agent Skill¶

For environments like Codex and Claude Code:

npx skills@latest add xiehuan123/dsh-deepread

DeepSeek Harness Full Plugin¶

Install the Host plugin package in the DSH Web environment:

dsh plugin --profile web add dsh-deepread

After installation, restart dsh web, and you can access via the 📖 Reading Panel or call the deepread tool in conversation.

Typical Usage¶

Agent Skill Scenario¶

After installing the Skill, you can directly make requests to the agent:

Deep-read docs/architecture.pdf in knowledge-map mode.
For every important claim, show the supporting evidence and source location.

DSH Host Scenario¶

After restarting the Web service, call the deepread tool in chat, or use the Reading Panel to select a link/path/text, mode, and export format, then start with one click.

The repository provides reproducible example outputs (not placeholder screenshots), such as:

• deep mode: examples/claude-code-token-optimization.md
• map mode: examples/ad-fact-check-knowledge-map.md
• deep mode: examples/vivo-tauri-architecture.md

Use Cases and Notes¶

Who It’s For

• Users who need to systematically read long articles, PDFs, and books in DSH or Agent Skills environments and produce verifiable, structured reports.
• When writing research notes, fact-checking before citation, or deeply learning a topic using the Feynman method, the map / feynman / book modes correspond to different depths and output formats.
• When needing to compare viewpoint differences and complementarities across multiple documents, you can use batch for cross-document comparison.

Pre-use Notes

• The plugin runs with the permissions of the current dsh process. Before installation, you should read the GitHub source code and the MIT license to confirm that its behavior aligns with your security policies.
• The Host plugin requires Node ^22.19 || >=24; full functionality (PDF extraction, background jobs, Web panel, budget API) depends on the DSH Web/headless or dsh-TUI environment, while the Portable Skill form has no Host runtime dependencies.
• SkillHub (https://www.skillhub.cn) is a community-independent directory with no official affiliation to DeepSeek or High-Flyer.

Conclusion¶

DeepRead transforms “finished reading” into “read clearly”: each important viewpoint is paired with evidence and confidence where possible, and clearly states when lacking original text support. If you are already using DSH for document-related tasks, you can install and try it out using the commands above.

• Directory Page: https://www.skillhub.cn/plugins/xiehuan123/dsh-deepread
• GitHub: https://github.com/xiehuan123/dsh-deepread
• Chinese Documentation: https://xiehuan123.github.io/dsh-deepread/