GitTOU/dsh-tool-gbt9704

dsh-tool-gbt9704 is a Cordis-native plugin for the DeepSeek Harness (DSH) agent runtime, maintained by GitTOU, written in TypeScript, and released under the MIT license. It exposes a single model-facing tool named gbt9704_convert that reformats a Markdown (.md) or Word (.docx) source into a new .docx whose body conforms to the GB/T 9704-2012 党政机关公文格式 (Party-and-government official document) layout. Internally it is self-contained: the source is read with Node's fs, parsed through a shared document model (mdast for Markdown, an internal OOXML extractor for existing .docx files), then rendered onto a minimal .docx container built by hand using fflate; the plugin registers no service and keeps no state, so each invocation is a pure path-in / path-out conversion.

The typical workflow is straightforward: a model — or the operator driving DSH — supplies the source path and any optional Config overrides, and the tool re-lays the document out as official 主体. The converter infers 标题 from the first heading or short paragraph, recognises 一、 / （一） / 1. / （1） as the four 层次标题 levels regardless of Markdown heading depth, treats the first short paragraph ending in a full-width colon as 主送机关, and uses the last two paragraphs as 发文机关署名 and 成文日期 when the latter matches 2024年1月5日 or 二〇二四年一月五日. GFM tables, list items, code lines, blockquotes, and PNG/JPEG images map respectively to tables, body paragraphs, body paragraphs, body paragraphs, and centered images. The intended users are 文秘, drafters, and developers inside Chinese party-and-government units, agencies, or state-owned enterprises who use DSH as a writing assistant and want a one-step bridge from Markdown drafts to compliant 公文 layout that can be opened and printed in Word.

Dependencies and limits are spelled out plainly. Runtime needs Node plus the bundled mdast and fflate libraries; the layout values (A4 page, 天头 37 / 订口 28 / 地脚 35 / 翻口 26 mm, body 16 pt 仿宋_GB2312 justified, 22 lines × 28 characters, 579 twip fixed line pitch, title 22 pt 方正小标宋简体, 四号宋体 —1— page numbers) are Config fields defaulting to the standard and overridable from cordis.yml. Known constraints: docx→docx fidelity is intentionally limited to paragraph/table text and images, so headers/footers, comments, tracked changes, charts, text boxes, and per-run fonts are not preserved; only PNG/JPEG are embedded, inline and in-table images are dropped, and remote http(s) Markdown images become warnings; 版头 (发文机关标志, 发文字号, the red separator line, 份号/密级) and 版记 (抄送/印发机关, separators) are out of scope. Fonts 仿宋_GB2312, 方正小标宋简体, and 楷体_GB2312 are common on Windows 公文 deployments but may be missing elsewhere and will be substituted by Word on open. Crucially, the tool performs its own fs reads and writes and does not go through ctx.fs, so its writes bypass the agent's filesystem sandbox and are not recorded as observed file state — operators should still use sandboxed model-facing file mutation tools for ordinary edits and reserve this plugin as a dedicated GB/T 9704-2012 converter, with a first-run check that the host's font set matches the expected 公文 fonts.