A DeepSeek Harness host plugin: the model-facing read_document tool. It reads document files that the plain read tool cannot handle — PDF, DOCX and XLSX — plus UTF-8 text files, all through the harness filesystem backend (ctx.fs), so it inherits the session workspace resolution, the sandbox policy and the fs-observation policy exactly like the built-in tools.

How it works

One tool, four formats, dispatched by file extension (or an explicit format argument):

Format	Extensions	Extractor
`text`	anything else (`.md` `.txt` `.csv` `.json` `.py` …)	UTF-8 decode, binary rejected (NUL sniff), BOM stripped
`pdf`	`.pdf`	`pdf-parse` (extracts the text layer only; see limitations)
`docx`	`.docx`	`mammoth.extractRawText`
`xlsx`	`.xlsx` `.xlsm`	`xlsx` → per-sheet TSV rows

Extracted text is windowed with the same semantics as the built-in read tool: 1-based line numbers, offset/limit paging, line-length and byte caps, and the OpenCode-style <path>/<type>/<content> envelope.

Config

Key	Default	Meaning
`readLimit`	`2000`	Default and maximum lines returned by one call.
`maxFileBytes`	`67108864`	Byte cap for one document read.
`sheetRowLimit`	`200`	Rows kept per worksheet (longer sheets get a `... (N more rows)` footer).

Installation

dsh plugin --profile web add dsh-plugin-doc-reader
# restart dsh web

Build & test

npm install --legacy-peer-deps   # pdf-parse / mammoth / xlsx
node --test                      # pure-parse + mock-ctx contract tests

Model Experience

System prompt

One guidance section (tool:read-document) points the model at read_document for PDF/DOCX/XLSX and notes offset/limit paging.

Tool schema

read_document(file_path, format?, offset?, limit?) with a structured { path, format, offset, lines, totalLines } output schema and the line-numbered envelope rendering.

Known Limitations

No image recognition (OCR) — this plugin extracts text only; it cannot read text out of images. Scanned/image-only PDFs (page photos with no text layer) come back empty, and standalone pictures (PNG/JPG screenshots) are refused as binary. OCR / 识图 is explicitly deferred (decision 2026-08: out of scope for now; if added later it would be an opt-in extractor, not a default, given its cost and formula-rendering quality).
Text mode is UTF-8 only — other encodings decode as garbage; binary content is refused rather than echoed.
Legacy .doc is not parsed — only OOXML .docx.
Large documents are windowed — the model reads one window per call; very long files need several paged calls.
Formula-heavy PDFs lose mathematical structure (plain text only; no LaTeX reconstruction).