dsh-book2skill: A Book-to-Skill DSH plugin for DeepSeek Harness

dsh-book2skill is a book-to-skill plugin (DSH plugin) for DeepSeek Harness, maintained by omdsh-dev, licensed under BSD-3-Clause, and last updated in 2026-08. It addresses the long task of turning a book into a reusable SKILL.md: from fetching the book, parsing chapters, deep reading, to generating the skill file and installing it, all in 5 stages with 3 human gates to keep every step controllable and traceable. Core capabilities include: deterministic steps (EPUB parsing, PDF extraction, OCR, install copy) handled by host tools, while understanding and generation (building knowledge maps, designing questions, deep-reading core chapters, writing SKILL.md with self-checks) are performed by the agent, with a browser panel providing a 5-stage timeline and gate approvals. Task state is stored in the host storage domain, resumable across sessions and restarts, and can be cancelled at any time.

What is the 5-stage workflow and 3 human gates of the dsh-book2skill DSH plugin?

The workflow has 5 stages: Fetch book (supports local path selection or z-lib search download, reusing book-downloader cookies, no hardcoded secrets); Parse chapters (EPUB to per-chapter md, merging <200 words, splitting >5000 words; PDF probed with PyPDF2, scanned ones go through OCR with per-page progress shown); Deep read (Gate 1: agent skims to build a map, posts up to 3 multiple-choice questions, user answers, then agent deep-reads 3-5 core chapters); Generate SKILL.md (Gate 2: editable preview + knowledge map + 3 self-check items, failures highlighted in red, one-click agent fix); Install (Gate 3: multi-select targets like ~/.claude/skills, ~/.codex/skills, etc., confirm then call book2skill_install, completion card with trigger example chips).

How does dsh-book2skill decouple OCR with a three-level detection mechanism?

The plugin does not inject paddleOcr; instead it probes level by level: first ctx.get('paddleOcr') host service (auto-activated when dsh-paddle-ocr publishes it); then /paddle-ocr loopback RPC (task channel task/start → status polling → commit, with per-page progress); finally HTTP direct call to stub (BOOK2SKILL_OCR_ENDPOINT, default http://127.0.0.1:8011/api/pdf/ocr). If all unavailable, it clearly errors and provides an agent fallback path: use paddle_ocr_layout tool to parse → book2skill_import_ocr to import.

What host tools and compatibility requirements does dsh-book2skill have?

There are 10 host tools: book2skill_start, book2skill_parse, book2skill_get_job, book2skill_stage_note (summary·knowledge-map·deep-read·draft·questions·selfcheck), book2skill_read_chapter, book2skill_zlib_search, book2skill_download, book2skill_import_ocr, book2skill_install, book2skill_cancel. It is compatible with DSH >=0.1.0-rc.3 <0.2.0 (using rc.3's ctx.shell capability contract), Node.js ^22.19.0 || >=24.0.0, and Profile Bundle automatically applies cordis.patch.yml.

How to install dsh-book2skill?

The installation command is as follows (ensure your DSH environment meets the version requirements first):

dsh plugin --profile web add github:omdsh-dev/dsh-book2skill

After installation, restart dsh web (host line requires process restart), refresh the browser, and the "Book to Skill" tab appears in the conversation area. Do not manually add the book2skill line to the profile; the plugin package already declares a Profile Bundle, and manual addition would create duplicate loader entries.

What red lines does dsh-book2skill have for security and traceability?

The plugin hardcodes no secrets: z-lib login state is read from ~/.claude/skills/book-downloader/auth/zlib-cookies.json, and OCR tokens are managed by dsh-paddle-ocr's credential vault. Generated content is traceable: SOP steps must come from the references section (self-check 1 enforces re-read validation), and both references index and trigger words have self-check items.

See more DSH plugins at dsh-plugin.org