dsh-office
编辑精选维护状态: 活跃omdsh-dev/dsh-office
办公三件套!Office document tools for DeepSeek Harness (dsh): generate, read, and edit spreadsheets (.xlsx), PDFs, and presentations (.pptx).
15
星标
2
Fork
TypeScript
语言
Apache-2.0
许可证
2026-08-14
创建于
2026-08-16
最近推送
README
@huiliyi37/dsh-office
English | 中文
Office document tools for DeepSeek Harness (dsh): generate, read, and edit spreadsheets (.xlsx), PDFs, presentations (.pptx), and Word documents (.docx).
Ported from the office plugins of the Tianshu terminal coding agent (Apache-2.0 licensed upstream), adapted to the dsh cordis tool model.
Tools
| Tool | What it does |
|---|---|
xlsx_read |
List sheets of a .xlsx, or read one sheet as a markdown table (range-limited for large files; formula text preserved) |
xlsx_write |
Write a 2D array to a new .xlsx (formula cells, header bold, column widths, number formats) |
xlsx_edit |
Edit an existing .xlsx: add sheets, update cells (value or formula), append rows |
xlsx_recalc |
Recalculate every formula with a lightweight pure-TS engine and report error values (#REF!, #DIV/0!, #VALUE!, #N/A, #NAME?, #NUM!) with locations |
xlsx_audit |
Statically audit formula structure: array-formula traps, aggregation ranges missing rows, formulas overwritten by hardcoded values, inconsistent formulas in a column, self-references, division by zero |
pdf_create |
Generate a real PDF with headings, paragraphs, tables, lists, code blocks and footer page numbers; CJK text renders via an auto-detected system font |
pdf_read |
Extract text from a PDF for reading into context |
pdf_merge |
Merge multiple PDFs into one, in order |
pdf_split |
Split a PDF into single-page files, or extract pages by spec ("1,3,5-7") |
pptx_create |
Generate a .pptx deck from slide definitions (title / section / content / two-column / image / table / chart), with optional theme and speaker notes |
pptx_read |
Extract slide text as markdown, optionally including speaker notes; include adds structure (summary/layouts/images/tables: shape names & cm positions, image targets, table dims) |
pptx_edit |
Find/replace text inside an existing .pptx (<a:t> surgery), preserving all layout and styling |
docx_create |
Generate a real .docx Word document from content blocks (heading / paragraph / table / code / list), with optional striped tables, page background color, and diagonal text watermark |
docx_read |
Extract text from a .docx for reading into context |
Install & load
Compatibility: requires
dsh≥0.1.0-rc.5(bundles@deepseek-ai/dsh-tools≥0.1.0-rc.5). Installing on an older core whosedsh-toolsis a0.0.1release installs a seconddsh-toolscopy and crashes every tool call withCannot read properties of undefined (reading 'prepare').
Full load (all 14 tools)
dsh plugin --profile <name> add @huiliyi37/dsh-office
dsh --profile <name>
The first dsh plugin call initializes the profile (@deepseek-ai/dsh-base
is its first bundle) and appends this package to the profile's bundles list.
Launching the profile then registers all tools automatically.
Load only the families you need
Pass a config row in your profile's cordis.patch.yml to enable/disable
per family. Omitted families stay enabled; set a family to false to
exclude it (useful to keep the tool surface small):
# cordis.patch.yml
- insert:
- id: dsh-office
name: '@huiliyi37/dsh-office'
config:
enable:
xlsx: false # skip spreadsheet tools
pdf: true # keep PDF tools
ppt: false # skip presentations
docx: true # keep Word tools
Families: xlsx (read/write/edit/recalc/audit), pdf (create/read/merge/split),
ppt (create/read/edit), docx (create/read).
Uninstall
dsh plugin --profile <name> remove @huiliyi37/dsh-office
Manual alternative
Install the package anywhere Node resolution can find it and reference it
from your own cordis.patch.yml (same config.enable switches apply):
npm install @huiliyi37/dsh-office
# cordis.patch.yml
- insert:
- id: dsh-office
name: '@huiliyi37/dsh-office'
Skill
The package ships a usage skill (skills/SKILL.md, anthropics-compatible
format) that teaches the model large-file pagination and generation
discipline. Install it into a skill discovery root:
mkdir -p ~/.dsh/skills && cp -r node_modules/@huiliyi37/dsh-office/skills/dsh-office ~/.dsh/skills/
Usage examples
// xlsx_write — create a workbook
{ "file_path": "report.xlsx", "data": [["Name", "Score"], ["Alice", 92]], "header_bold": true }
// pdf_create — document with a heading, table and list
{
"destination_path": "doc.pdf",
"title": "Quarterly Report",
"content": [
{ "type": "heading", "text": "Summary" },
{ "type": "table", "headers": ["Region", "Revenue"], "rows": [["APAC", "120"]] },
{ "type": "list", "items": ["Alpha", "Beta"] }
],
"page_numbers": true
}
// pptx_create — a deck with a title slide and a bullet slide
{
"destination_path": "deck.pptx",
"slides": [
{ "type": "title", "title": "Roadmap 2026" },
{ "type": "content", "title": "Highlights", "items": ["Plugin runtime", "Office tools"] }
]
}
Development
npm install
npm run build # tsc → lib/
npm test # vitest: round-trip tests through the tool execute path
License
Apache License 2.0. Tool logic ported from the Tianshu office plugins (also Apache-2.0 licensed, copyright Tianshu contributors); see file headers for per-module provenance.
更多「知识与研究」插件
mirage
作者 strukto-ai
全球首个面向 AI 智能体的统一虚拟文件系统。
graph-memory
作者 adoresever
Deepseek Harness、Openclaw知识图谱记忆插件。2026年4月受邀发布在清华大学讨论会。Knowledge Graph + Memory;Knowledge Graph Context Engine for OpenClaw — extracts structured triples from conversations, compresses context 75%, enables cross-session experience reuse
memtrace-public
作者 syncable-dev
面向编码智能体的结构化记忆,双时态图、MCP 原生、零 LLM 调用,支持多款主流编码工具。
anolisa
作者 alibaba
阿里巴巴开源的 Agentic OS,提供运行时、安全、可观测性,并通过无 Token 响应压缩降低用量与成本。
