PDF parsing toolkit for DeepSeek Harness: pdf_info, pdf_extract_text, pdf_render_page with hybrid pdfjs/built-in engines and system-font rendering for non-embedded CJK PDFs.
待人工精选——以下事实来自源码仓库。
PDF parsing toolkit for DeepSeek Harness: pdf_info, pdf_extract_text, pdf_render_page with hybrid pdfjs/built-in engines and system-font rendering for non-embedded CJK PDFs.
想用 DSH 获得这项能力的用户；装前建议先看源码和文档。
dsh plugin --profile web add github:zhtx2024/dsh-pdf 作者没有声明支持的平台。
pdf_info / pdf_extract_text / pdf_render_page) with hybrid engines, including system-font rendering for PDFs with non-embedded CJK fonts (e.g. EasyEDA / JLC EDA schematic exports). English | 中文 Tags: dsh-plugin · pdf · pdfjs · canvas · schematic · tools ## Install sh dsh plugin --profile web add github:zhtx2024/dsh-pdf Or from a local checkout during development: sh dsh plugin --profile web add link:<path-to-repo> Then restart DSH (or open a new session) — the three tools and an agent guidance block become available. ## Tools | Tool | What it does | |---|---| | pdf_info | Page count, page sizes (pt), document metadata, and the font list with per-font embedded flags. | | pdf_extract_text | Extracts text per page; optional page and withPosition (per-line x/y coordinates + font size). | | pdf_render_page | Renders one page to a PNG (local file) with configurable scale/outDir; engine auto-selection. | ## Why a custom renderer? pdfjs-dist cannot render text in PDFs whose fonts are not embedded and lack a ToUnicode map — the classic case is JLC EDA / EasyEDA schematic exports (SimSun/SimHei Type0 fonts with UniGB-UCS2-H encoding): pdfjs emits translateFont failed and drops the glyphs. dsh-pdf solves this with a two-engine design: - pdfjs engine — used when all fonts are embedded (standard PDFs). - sysfont engine — a built-in content-stream renderer that decodes the PDF text operators (BT/ET, Tf, Tm, Td, Tj, TJ) itself, maps UniGB-UCS2-H/UTF-16BE strings to Unicode, and draws them with system fonts (SimSun, SimHei, Microsoft YaHei, …) via @napi-rs/canvas. Text extraction uses the same auto-selection: embedded fonts → pdfjs text layer; non-embedded fonts → the built-in decoder (which correctly recovers CJK text pdfjs loses). ## Example text pdf_info("D:/Downloads/SCH_SA-V11A.pdf") → 3 pages, A3 landscape, jsPDF producer, 6 fonts (all non-embedded) pdf_extract_text("D:/Downloads/SCH_SA-V11A.pdf", { page: 1 }) → net names, pin numbers, and Chinese labels ("主控板", "嘉立创", …) pdf_render_page("D:/Downloads/SCH_SA-V11A.pdf", { page: 1, scale: 2 }) → D:/Downloads/SCH_SA-V11A-pages/page-1.png (2396×1698, engine: sysfont) ## Architecture - lib/index.js — cordis plugin entry (name, inject, apply); registers the three tools as native tool objects (plain JSON-Schema parameters and output — no @deepseek-ai/dsh-tools dependency needed) and injects the agent guidance block (systemPrompt.section, order 160). - lib/pdf-core.js — self-contained engine: PDF object parser, font scanner, pdfjs loading with fs-based CMapReaderFactory/StandardFontDataFactory (Node 24's global fetch doesn't speak file://), the built-in text extractor, and the sysfont renderer. - cordis.patch.yml — the dsh.bundle.patch insert row. ## Limitations - Scanned/image-only PDFs have no text layer — extraction returns nothing (rendering still works, as a plain image). - The sysfont renderer assumes Type0 CID codes equal Unicode code points (true for UniGB-UCS2-H exports); exotic CID fonts fall back to pdfjs. - Large PDFs are parsed fully into memory. ## License MIT不看 Star，人工一条条挑的。
把 DSH 的 settings.yaml 做成可视化看板：所有已注册的 settings namespace——包括官方界面从未覆盖的第三方插件配置——都渲染成可编辑表单。项目还早，但方向很对：告别手改 YAML。
给 DSH 的 DIY 轨迹可视化：把 agent 的计划与执行画成图，不用翻原始日志也能看清它在干嘛。项目还早，但补上了「一眼看懂 agent 在做什么」这个真实缺口。
从 Claude Code / Codex 把活派给 DSH：在宿主内拉起带分级预设的 DSH agent 会话，看原生子代理进度，还能借它的多模态桥给纯文本的 DSH 补上视觉和生图。编排型插件，把 DSH 变成其他编码 agent 的后端。