dsh-files

Attach files in DeepSeek Harness and let the model actually read them.

Upload files or a whole folder from the Web composer. Parsing and indexing stay on your machine. The model searches for compact evidence and expands only the exact page, slide, line range or spreadsheet range it needs — instead of pasting whole documents into the prompt or shelling out to Python. Raster images stay on Harness' native vision path.

[!IMPORTANT] Source beta — not published to npm yet. @cooberped/dsh-evidence@beta does not exist on npm at this time. Use the source install below.

How it works

1. Upload once into the active session workspace.
2. Index locally — no full document enters the model context.
3. Retrieve compact evidence for a concrete question.
4. Expand a versioned coordinate only when more context is needed.
5. Answer from evidence, or say the evidence was not found.

Files become addressable evidence rather than prompt baggage. That is the whole design.

Install from source

You need the DeepSeek Harness CLI with the web profile (validated baseline: npm @deepseek-ai/dsh@0.1.1-rc.2), Node.js >=22.13.0, and pnpm on PATH.

git clone https://github.com/Cooberped/dsh-evidence.git
cd dsh-files
pnpm install --frozen-lockfile
pnpm build

dsh plugin --profile web add .      # link this checkout into the web profile
dsh --profile web --dump-config     # confirm the bundle layer is present
dsh web                             # restart

The install is a link to this checkout: after pulling updates, re-run pnpm install --frozen-lockfile && pnpm build and restart. Remove it with dsh plugin --profile web remove @cooberped/dsh-evidence.

dsh plugin --profile web add @cooberped/dsh-evidence@beta
# restart dsh web

Use it

1. Open the Harness Web composer.
2. Add files: paperclip for multi-select, the folder button for a directory, or drop anything on the page.
3. Check that every intended file has its own card showing AI-readable.
4. Ask a concrete question — the model calls the tools on its own.

Prompts that work well:

Index these three files first. Do not summarize them yet.

Across these files, find the definition and formula for the Q3 retention metric.
Give the source file and exact page, slide, line range, or Sheet!Range for every claim.

Compare the meeting decision with the workbook target.
If the documents do not contain enough evidence, say what is missing instead of guessing.

The two tools

Two rules hold the loop together:

• Coordinates carry a version (source hash + parser/block schema). Supplying a coordinate makes version mandatory, and it is checked before any content is returned. Edit the file and the old coordinate fails loudly instead of quietly reading the wrong lines.
• Zero recall is not permission to guess. The tool tells the model to retry with different terms, read sequentially when justified, or report that the files contain no evidence.

Formats and honest boundaries

The format is decided from the bytes, never the extension — an executable renamed to .pdf is rejected. Long output is paged and character-bounded, and every truncation is visible.

Configuration

The bundle ships conservative defaults in cordis.patch.yml. Harness applies user profile overlays after bundle layers, so inspect the composed result before boot:

dsh --profile web --dump-config

Less common settings and their authoritative defaults live in src/index.ts. An explicit retrievalIndexDir must be an absolute private path; ~ is not expanded.

Runtime backend. The package requires Node.js >=22.13.0 — the floor pdfjs-dist sets, and Node 20 reached end of life on 2026-04-30. A persistent retrieval index additionally needs the runtime to provide node:sqlite with FTS5 compiled in; when the startup probe finds either missing, the complete but process-local JS backend takes over. Tool output reports the selected backend — the fallback is a supported mode, not a silent partial success.

Security and privacy

• Byte-level truth. Extensions are hints only. PDF headers and OOXML ZIP members decide the parser; known foreign binaries and spoofed files are rejected.
• Bounded OOXML. ZIP member count and name length, declared XML sizes, aggregate XML expansion, workbook rows/cells and sparse-sheet dimensions are all capped before parser allocation.
• Workspace containment. Reads go through ctx.fs; when a session cwd exists, the target must stay inside the active session workspace.
• Safe upload storage. Pre-existing symlinks and special files are rejected; creation uses exclusive/no-follow flags where supported; quotas and deletion fail closed.
• Loopback by default. Upload and workspace endpoints require a loopback host and same-origin checks. trustedHosts exists only for a deployment-controlled reverse proxy.

trustedHosts is not authentication. The current Harness WebServer plugin contract cannot prove that a supplied session ID belongs to the caller. Do not expose this plugin as an unauthenticated public multi-tenant upload service. Use loopback-only self-hosting, or an authenticated proxy that also constrains session access.

Project status

This is not an official DeepSeek plugin and is not affiliated with or endorsed by DeepSeek. It is also not a clean-room rewrite: the MIT-licensed history of taxueseek/dsh-files is retained, while Cooberped independently develops the retrieval, coordinate, security, performance and release layers described here.

Development

pnpm install --frozen-lockfile
pnpm typecheck
pnpm test
pnpm benchmark:retrieval
pnpm license:check
pnpm package:check

pnpm release:check   # everything above, for a final candidate

release:check covers type checking, both bundles, focused regression tests, dual-backend retrieval correctness, license policy and the npm tarball contract. The benchmark uses deterministic synthetic PDF/DOCX/XLSX/PPTX fixtures — real business documents, answer sets and model outputs stay outside the repository (see benchmark/README.md).

Contributing

Issues and pull requests are welcome. Contributions are reviewed and merged by maintainers after required checks; GitHub does not merge community code automatically.

Before opening a PR: read CONTRIBUTING.md and sign off commits for DCO; add focused tests for behavior changes; run the smallest relevant checks locally; keep real documents, credentials, private paths and model outputs out of Git; and record every new visual asset in assets/README.md.

Security reports follow SECURITY.md. Release ownership and provenance gates are in RELEASING.md.

License, lineage and marks

Project code and the original SVG documentation graphics are licensed under the MIT License. Upstream history and copyright notices are retained. Dependency and bundled-data notices are in THIRD_PARTY_NOTICES.md.

DeepSeek, DeepSeek Harness, OpenCode Go and other third-party names or marks belong to their respective owners and are used only to identify compatibility or a test target. Asset provenance is recorded in assets/README.md.