DSH Cowork

READ + WRITE for office documents and Jupyter notebooks inside DeepSeek Harness — and anywhere else your agent runs.

DSH Cowork adds first-class document handling to coding agents:

Format	Read	Write (v1)
xlsx spreadsheets	✅ bounded, cell-addressed windows	✅ create + edit by cell ref
ipynb notebooks	✅ cells + inline outputs	✅ create + edit by cell index
pdf	✅ text windows (pages)	— (v2: form-fill)
docx	✅ paragraphs + word count	— (v2: generation)
pptx	✅ slides + shape ids	— (v2: generation)

中文版见 README.zh.md.

Why

DeepSeek Harness's built-in read is UTF-8 text only — it cannot open a spreadsheet, a PDF, a slide deck, or a notebook. Claude Code reads PDFs and notebooks natively; Codex has nothing. DSH Cowork closes that gap as an out-of-tree plugin (no fork, no PR needed — the ecosystem path CONTRIBUTING.md recommends).

Two ideas make "Cowork = READ + WRITE" coherent instead of two separate tools:

1. Stable addresses. doc_read returns cell refs for xlsx (A1, C12) and shape ids for pptx — doc_write consumes them. Line numbers cannot address binary formats; addresses can.
2. Bounded windows, explicit truncation. Every read is capped (pages / rows / slides / cells / bytes). Silent truncation is the cardinal sin: a > Truncated: notice always tells you the window was cut short.

Packages

Package	What it is
packages/core	Pure TS, zero DSH deps: sniff → extract → build, windowing, safety caps
packages/dsh	DSH bundle: doc_read / doc_write tools (install with dsh plugin add)
packages/mcp	MCP server over stdio — for Codex, Claude Code, any MCP client
packages/cli	doc-read / doc-write binaries + SKILL.md (the pi harness adapter)
packages/chatnode-wechat	DSH bundle: chat with / monitor / approve your DSH agents from WeChat (iLink gateway + conversation node)

Install into DeepSeek Harness

# GitHub delivery (not published to npm)
git clone https://github.com/Jesse-njx/dsh-cowork.git
cd dsh-cowork
pnpm install          # installs deps and builds all packages (prepare)
dsh plugin --profile <your-profile> add ./packages/dsh

Then doc_read / doc_write are available to the model. (A live agent session also verifies: ask the model to doc_read an .xlsx.)

Configure (all optional, sensible defaults):

# in your profile's cordis.patch.yml, override the cowork-docs row
- id: cowork-docs
  name: '@dsh-cowork/plugin'
  config:
    maxInputBytes: 67108864      # 64 MiB
    maxOutputBytes: 262144       # 256 KiB model-facing window
    maxDecompressedBytes: 536870912  # zip-bomb guard (512 MiB)
    maxZipEntries: 4096
    maxPages: 20
    maxSheetRows: 200
    maxSheets: 1
    maxSlides: 20
    maxCells: 200

Use from any agent (MCP)

# Codex / Claude Code MCP config
# { "mcpServers": { "cowork": { "command": "node", "args": ["<repo>/packages/mcp/lib/index.js"], "cwd": "<your working dir>" } } }

# or the plain CLI
doc-read report.xlsx --sheets Data --rows 50
doc-write edit report.xlsx --spec edit-spec.json

Safety model

• Zip bombs: entry-count + decompressed-size caps on every OOXML archive, checked before any codec expands bytes.
• Macros: .xlsm / .docm / .pptm (anything with vbaProject.bin) are rejected outright — Cowork never reads or writes macro formats.
• Read-only mode: in read-only sandbox mode, doc_write is hard-blocked while doc_read stays available.
• Edit guards: doc_write edit refuses unless the file was read this session (expected_version), and optionally fails on content change (expected_sha256) — a hash-check on every edit.
• No silent overwrite: creating over an existing file requires having read it (DSH) or force (CLI/MCP).
• Atomic writes: temp file + rename, never a partial target.
• Stale formulas: edited xlsx sheets get cached formula results cleared so Excel / LibreOffice recalculate on open (exceljs never recalculates).
• Untrusted input: formulas, hidden sheets, and speaker notes are treated as data, surfaced with the extraction, never executed.

Architecture

                   ┌──────────────────────────────────────────────┐
                   │                @dsh-cowork/core              │
                   │  sniff → readDocument / writeDocument → caps │
                   └───────┬──────────────┬──────────────┬────────┘
                           │              │              │
                ┌──────────▼───┐  ┌───────▼──────┐  ┌────▼───────┐
                │  packages/dsh│  │ packages/mcp │  │ packages/cli│
                │ DSH bundle   │  │ MCP stdio    │  │ + SKILL.md │
                └──────────────┘  └──────────────┘  └────────────┘

The DSH bundle routes reads through ctx.fs (bounded readBytes, sandbox-aware resolution, fs/observed events) and performs atomic byte writes itself because the fs service is text-only — re-observing the real post-write version so the built-in policy machinery stays coherent.

Development

pnpm install
pnpm -r build
pnpm -r test      # core 35 + plugin 15 + cli 6 + mcp 7 tests

• Fixtures: node packages/core/scripts/make-fixtures.mjs (needs macOS cupsfilter for the PDF; skip otherwise).
• Node 20+; tests run on Node's native TS type-stripping.

Roadmap

• v1 (this repo): read all five formats; write xlsx + ipynb.
• v2: ipynb write polish, docx/pptx generation, PDF form-fill (pdf-lib), PDF page-render-to-image for vision models, docx raw-OOXML edit.
• Ship: dsh-plugin GitHub topic, awesome-list entry, Discussions post.

Contributing

This is a dsh-plugin-topic project. Issues, PRs, and Discussions welcome — and per deepseek-harness CONTRIBUTING.md, associate your own plugins with the dsh-plugin topic too.