A DeepSeek Harness PDF production skill, reimplemented from the capability set of the ZCode built-in pdf plugin. It covers four routes — reports, creative visuals, academic LaTeX, and existing-PDF processing — and converts Office documents through DSH's own LibreOffice engine.

The upstream skill is proprietary, so this package is a clean-room reimplementation: the workflow and capability boundary were used as a specification, and no upstream text, script, template, or asset was copied. See NOTICE.

Routes

Route	Tool	Dependency
Report (report, resume, cover)	`scripts/report.py`	`reportlab`
Creative (poster, visual)	`scripts/html2pdf.mjs`	Playwright + Chromium
Academic (LaTeX)	Tectonic	Tectonic, optional
Process (merge, split, extract, fill)	`scripts/pdf.py`	`pypdf`; `pdfplumber` / `PyMuPDF` for tables and page images
Conversion (Office or HTML to PDF)	`scripts/office2pdf.mjs`	`@deepseek-ai/libreoffice-kit`

Install

dsh plugin --profile <profile> add dsh-zcode-pdf
dsh plugin --profile <profile> add ./packages/pdf

The skill registers as pdf. Its core Python packages install on demand into the harness interpreter:

<python> <skill-directory>/scripts/pdf.py env.check
<python> <skill-directory>/scripts/pdf.py env.fix

Office conversion additionally needs the LibreOffice engine, installed on demand as a plain profile dependency (about 150 MB):

dsh plugin --profile <profile> add @deepseek-ai/libreoffice-kit

A deployment that already ships the kit is found automatically, or through DSH_LIBREOFFICE_KIT pointing at the kit package directory. Chromium (an installed Google Chrome or Microsoft Edge is reused automatically) and Tectonic stay optional and are installed only when a task needs them.

Design notes

office2pdf.mjs uses @deepseek-ai/libreoffice-kit, so Office conversion needs no system LibreOffice.
html2pdf.mjs prints through Chromium, which keeps real CSS layout and honours @page.
report.py accepts a Markdown subset (headings, lists, pipe tables, quotes, fenced code, page breaks, inline emphasis) and switches to a CJK CID font automatically.
pdf.py always writes a new file and re-inspects the result; it never strips encryption and never overwrites the input.
The core Python packages are pure Python (reportlab, pypdf) so they install into the harness interpreter. Table extraction and page rendering need native extensions, which a hardened interpreter can refuse to load; pdf.py env.venv builds a dedicated environment and prints the interpreter to use.

Known limitations

@deepseek-ai/libreoffice-kit is an optional peer installed on demand; without it the conversion route reports the exact install command.
Tectonic is not bundled and is a large first-run download.
The skill reimplements the upstream routes, so upstream-specific templates and prompts are not available.