dsh-document-parser

A DeepSeek Harness Cordis bundle that registers the parse_document tool backed by LiteParse.

What it provides

The plugin lets an Agent parse a document in the current workspace into Markdown. It supports formats handled by LiteParse, including PDF, Office/OpenDocument files, and images.

The first version runs on the DSH Host side. It does not add a browser UI or a generic chat-attachment protocol.

Development

Requirements:

• Node.js 18 or newer
• pnpm
• A DSH installation compatible with @deepseek-ai/dsh-* version 0.1.0-rc.6

Install and build:

pnpm install
pnpm run check

Quality checks run automatically through Lefthook on commit and push:

• prettier --check validates staged formatting.
• oxlint validates TypeScript source.
• typecheck and build run before push.

The generated lib/ directory is ignored by Git and recreated by postinstall and build. It is included in the npm package through the files field, but should not be committed to source control.

Install into a local DSH profile

From the workspace where DSH should resolve document paths:

dsh plugin --profile web add file:/absolute/path/to/dsh-document-parser
dsh --profile web --dump-config
dsh web

The package declares a dsh.bundle entry, so DSH automatically adds its cordis.patch.yml layer after installation. For a headless profile, install it separately:

dsh plugin --profile headless add file:/absolute/path/to/dsh-document-parser

Agent usage

Ask the Agent to parse a file:

Parse docs/manual.pdf and summarize the main points.

Optional page selection is supported:

{
	"file_path": "docs/manual.pdf",
	"target_pages": "1-5"
}

Tool parameters

The parse_document tool accepts:

• target_pages: optional 1-based selection such as 1-5,10.
• ocr_enabled: whether to run OCR; defaults to true.
• ocr_language: Tesseract language code such as eng or chi_sim.
• output_format: markdown, text, or json; defaults to markdown.

DSH owns tool timeout, output retention, spill, and compaction policies. This plugin passes the requested parsing options to LiteParse without adding another plugin-level resource-limit configuration.

Office parsing may require LibreOffice on the host, as required by LiteParse.