dsh-tool-gbt9704

Model-facing gbt9704_convert tool: reformat a Markdown (.md) or Word (.docx) source into the GB/T 9704-2012 党政机关公文格式 (Party-and-government official document) body layout as a new .docx.

The converter is self-contained: it reads the source with Node's fs, parses it through a shared document model (mdast for Markdown, an OOXML extractor for existing .docx files), renders the model onto a hand-built OOXML document (a minimal .docx container via fflate), and writes the output file. It registers no service and keeps no state; a call is a pure path-in/path-out conversion.

Layout applied (GB/T 9704-2012 body rules)

All values are Config fields with the standard's defaults; a deployment may override fonts, sizes, margins, or line pitch from cordis.yml.

Writing conventions (约定位置解析)

The tool needs no metadata block; it reads the document by position:

• The first heading (or the first short paragraph when the source has no heading) is the 标题.
• Lines beginning with 一、 / （一） / 1. / （1） become the four 层次标题 levels, regardless of Markdown heading depth.
• The first short paragraph ending in a full-width colon is the 主送机关.
• The last two paragraphs are the 发文机关署名 and 成文日期 when the last one matches a date form (2024年1月5日 or 二〇二四年一月五日).
• GFM tables, ordered/unordered list items, code lines, blockquotes, and PNG/JPEG images convert to tables, body paragraphs, body paragraphs, body paragraphs, and centered images respectively.

Config

All fields are optional; the standard's defaults apply. z schema in src/index.ts.

Model Experience

Request context and condition

What the model sees

The tool's full description plus the standard's layout summary; the model supplies a source path and optional overrides. The description pins the conversion conventions verbatim so the model knows which Markdown forms map to which 公文 elements.

Token effect

One tool description, fixed; no data-dependent prompt growth. The model-facing result is a short summary (output path, detected title, block counts, warnings).

KV Cache effect

The tool description is a stable prefix; it does not invalidate cache reuse. Results vary per call but are result-turn content, not request context.

Known Limitations and Deferred Work

• Docx→docx fidelity — the reader extracts paragraph/table text and images only; complex Word features (headers/footers, comments, tracked changes, charts, text boxes, per-run fonts) are not carried into the rebuilt document. Use it to normalize plain official documents, not to round-trip arbitrary Word files.
• Images — only PNG and JPEG are embedded; inline (mid-paragraph) images and images inside table cells are dropped. Remote http(s) image URLs in Markdown are reported as warnings.
• Red-header and 版记 — the 版头 (发文机关标志, 发文字号, red separator line, 份号/密级) and 版记 (抄送/印发机关, 分隔线) are out of scope; the converter produces the 主体 body layout plus page numbers.
• Font availability — 仿宋_GB2312, 方正小标宋简体 and 楷体_GB2312 are common on Windows 公文 deployments but may be absent elsewhere; Word substitutes fonts at open time. Override via Config when a deployment standardizes on different names.
• No sandbox mediation — the tool performs its own fs reads and writes and does not go through the ctx.fs service, so its writes are not recorded as agent-observed file state and are not confined by the filesystem sandbox. Model-facing file mutation tools remain the sandboxed path; this tool is a dedicated converter.