dsh-evidence

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-evidence
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.

Future npm beta — not available yet

Once the repository release gates and npm trusted publishing are complete, installation becomes:

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

This is documented as a future command. It does not work today.

The profile/plugin contract follows the official Harness plugin reference and bundle publishing guide.

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

	search_documents	read_document
Use it to	find where the answer is	read what is at a place
Key inputs	file_paths, optional short query	file_path + coordinate/version, or offset/limit, or sheet/cell_range/list_sheets
No query	indexes changed content and returns a compact inventory — good for "read these first"	—
Returns	ranked evidence blocks, each with path, format, text, coordinate and version	a paged window of the document, or one coordinate expanded exactly

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

Input	Local projection	Stable coordinate	Current boundary
Text	UTF-8, UTF-16 BOM, high-confidence BOM-less UTF-16, GB18030	line:S-E, optional chars:S-E	Other encodings and binary files are rejected
PDF	Text-layer extraction, page-preserving	page:N, optional local line/character range	No OCR for scanned or image-only pages
DOCX	Body, paragraphs, tables, headers, footers, footnotes, endnotes	line:S-E, optional chars:S-E	No image OCR; not a pixel-faithful Word renderer
XLSX	Sheet inventory, cell values, ranges, row/column coordinates	quoted Sheet!A1:F40	No formula calculation, chart/shape interpretation, or macro execution
PPTX	Slide-order DrawingML text and speaker notes	slide:N, optional local line/character range	No slide-image OCR, chart data, SmartArt, animation, or embedded objects
JPEG/PNG/WebP/GIF	Native Harness image attachment	Harness attachment identity	Needs a vision-capable model; not parsed by read_document

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.

Capability details — composer, retrieval, reads, vision handoff

Composer and upload

• Paperclip multi-select, whole-folder selection, and page-level drag and drop.
• Finder multi-selection merges DataTransfer.items and DataTransfer.files, so a mixed batch is not silently reduced to the first recognized file.
• Documents are captured before Harness' image-only drop handler; pure JPEG/PNG/WebP/GIF drops stay on the native image path.
• Bounded parallel uploads (default 4); one failure does not cancel the batch.
• Byte-sniffed file cards with uploading / AI-readable / failed states.
• @ candidates cover uploaded and workspace files, projected as workspace-relative paths rather than host absolute paths.
• Per-session quota, SHA-256 dedup, TTL cleanup, safe file-name normalization, recursive folder cleanup.

Local retrieval

• CJK runs are indexed as overlapping bigrams and queried as phrases, so 流程绩效 does not match a block containing only 绩效流程. Single characters use a bounded substring fallback; ASCII tokens such as Q3 stay whole.
• SQLite FTS5 is used only after a startup capability probe succeeds; otherwise the JS memory backend is selected explicitly and the tool result says so.
• Query persistence is off by default.

Exact document reads

• XLSX supports workbook inventory, one-based sheet selection, coordinate-preserving A1 ranges, merged headers, hidden/sparse sheets, and explicit detected-value counts.
• PPTX follows presentation relationship order, not ZIP filename order.
• The system prompt tells the model to use these tools first and fall back to Python/shell only after an explicit error or unsupported-feature notice.

Native vision handoff

• JPEG/PNG/WebP/GIF use Harness' native composer attachment rail and the provider-neutral base64 image_url path.
• The selected model must still declare image input support; this plugin does not make a text-only model visual.

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

Setting	Default	Meaning
maxFileBytes	24 MiB	Maximum bytes for one document read
uploadMaxBytes	24 MiB	Maximum bytes for one upload body
maxUploadBytesPerSession	512 MiB	Session upload quota; 0 explicitly disables it
readLimit	2,000 in the bundled patch; schema fallback 800	Maximum lines returned by one read_document call
maxOutputChars	24,000	Base character window; narrative formats get smaller format-specific windows
maxConcurrentUploads	4	Admitted upload bodies
uploadTtlMs	7 days	Uploaded-file retention
retrievalEnabled	true	Enables search_documents; read_document remains available when false
retrievalMaxFiles / retrievalMaxResults	12 / 12	Files per search call / evidence blocks returned
retrievalQueryLogEnabled	false	Persist normalized model queries for local tuning
trustedHosts	[]	Additional reverse-proxy authorities; empty means loopback only

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.

Where data lives, and what leaves the machine

Store	Default	Contains	Lifecycle
Uploads	<session-workspace>/.dsh-filess/<storageKey>/	Uploaded file bytes	Per-session quota, SHA-256 dedup, default 7-day TTL sweep
Retrieval index	$DSH_HOME/dsh-files/index	Search projection, coordinates, versions; queries only when explicitly enabled	Private permissions, document/query TTL, JS memory fallback when persistence is unavailable

• Parsing and indexing are local to the Harness host; the plugin makes no external parsing request of its own.
• The evidence a tool returns does enter the conversation and may be sent to your configured model provider. Native image attachments are also sent to the selected vision provider.
• Do not sync the retrieval directory to cloud storage, and do not commit it to Git.

Known limits — what this deliberately does not do

• Scanned PDFs and images embedded in office files are not OCR'd.
• Office layout is projected into text and coordinates, not rendered pixel-for-pixel.
• XLSX formulas are not calculated and macros never execute.
• PPTX charts, SmartArt, animations and embedded objects are not interpreted.
• Upload quota locking is process-local, not a cross-process transaction.
• Portable Node exposes no complete dirfd/openat chain, so a hostile same-UID process racing ancestor replacement remains an OS-isolation boundary.
• Compatibility is pinned to the tested Harness prerelease train until a newer runtime passes the same acceptance workflow.

Project status

Project	Public source beta, independently maintained by Cooberped
Harness baseline	Tested against npm @deepseek-ai/dsh@0.1.1-rc.2 with the web profile
Runtime acceptance target	OpenCode Go — DeepSeek V4 Flash
npm	Not published. Package metadata targets @cooberped/dsh-evidence@0.6.0-beta.1; scope ownership, trusted publishing and the first-release license gate remain open
Compatibility	Newer Harness source trains are not claimed compatible until separately tested

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.