useful-dsh-plugins

中文

Community plugins for DeepSeek Harness. Independent project; not affiliated with DeepSeek. Requires Node.js 24 or newer; locally tested with DSH 0.1.5-rc.2 and Node.js 24.18 on Windows. DSH is a developer preview: a new host version needs compatibility checks.

Plugin features

Voice input: turn speech into an editable draft

• Hold the composer's microphone button or the configured shortcut to record. Release it to transcribe and append text to the draft, then edit and send it when ready.
• SenseVoiceSmall runs on the local CPU and supports Chinese, English, Japanese, Korean and Cantonese. After downloading the model and engine, recognition works offline without a cloud speech account.

Upload button: attach files to a conversation

• Click the upload button in the composer toolbar to choose one or more files and preview them as cards without changing the text being edited. Failed sends retain the cards for retrying. The configurable per-file limit defaults to 64 MiB.
• Remove attachments before sending, deduplicate identical uploads, and open sent attachments from their cards in chat history. This plugin transfers files; pair it with document companion when you want the agent to read document contents.

Vision reader: inspect images with an agent

• Pasting images into a conversation automatically saves local copies while retaining the original images in the message. PNG, JPEG, WebP and GIF are supported, with up to 10 images per operation.
• The separate vision tool takes an image path and a question to a configured vision model for tasks such as screenshot analysis, reading image text and explaining charts. This uses a model provider; the main conversation model must also accept image input.

Document companion: read and discuss documents side by side

• Click the book button in the composer toolbar to open a reading panel beside the chat. It supports PDF, DOCX, XLSX, CSV, TXT and Markdown, with continuous scrolling, single-page mode and adjustable panel width.
• Share the current reading position with the agent to ask about the visible page, read a specified page or block, search extracted text, jump to a position or switch documents. PDF pages are read through a vision model by default, requiring a configured, working model provider.

Plugin manager: manage third-party plugins and DSH updates

• Open Settings → Plugins → Manage to inspect third-party plugin versions, runtime state and data locations, enable or disable plugins, and restore disables made by the manager. Changes take effect live when the host and plugin support it.
• Check updates and reinstall supported npm-sourced plugins; update this collection through its shared installer. A separate DSH host update section supports standard global npm installations. Official plugins are excluded from ordinary plugin toggles; replacing package code or upgrading DSH still requires a restart.

Rigor 4 records requirements, plans, execution evidence, child tasks and independent reviews to help assess an agent's completion claims. It has its own repository and installer and is not included in the five-plugin collection above.

Install and remove

With DSH and pnpm already installed, one command installs the collection. On Windows, close DSH before replacing or removing packages so native dependency files are not held open; start dsh web afterwards.

npx --yes useful-dsh-plugins@0.5.1 setup

The same installer is also available directly from the public GitHub release:

npx --yes --package=https://github.com/lishize20040802-rgb/useful-dsh-plugins/releases/download/v0.5.1/useful-dsh-plugins-0.5.1.tgz useful-dsh-plugins setup

Use setup --preview to inspect the plan first. The installer verifies release checksums, stores archives and extracted packages inside <DSH_HOME>/third-party/, backs up the profile manifest, and calls the official DSH CLI to register the five packages. DSH keeps its native runtime resolution under profiles/<profile>/node_modules; the installer does not patch the host installation. Re-run the versioned command to repair or update this collection. Existing custom data paths are preserved.

<DSH_HOME>/
  third-party/
    archives/        downloaded and verified plugin packages
    packages/        extracted versioned packages
    sources/         optional development checkouts
    data/            voice runtime and plugin-manager recovery records
    backups/         local installation recovery records
  profiles/web/     native DSH plugin registrations and runtime dependencies

Remove the collection with npx --yes useful-dsh-plugins@0.5.1 uninstall. Personal data is retained. Optional --profile, --dsh-home, --dsh-package, and --store-dir support custom installations. The older aggregate package is removed during migration to avoid duplicate bundle rows; the new npm package is a zero-dependency installer.

For individual advanced installation, put the archive inside the same third-party archive directory and use the native CLI below.

Download the built .tgz for a plugin from Releases. Install it using the official CLI:

dsh plugin --profile web add ./dsh-plugin-voice-input-0.3.0.tgz
dsh web

To remove it:

dsh plugin --profile web remove dsh-plugin-voice-input

Repeat with the package name in the table. DSH registers and removes each package's dsh.bundle automatically. Restart after installing, updating, or removing package files. Disabling a loaded plugin is a separate operation; see the manager's documented live-reload conditions.

On Windows, this checkout also provides ./install.ps1 -Preview, ./install.ps1, and ./install.ps1 -Uninstall for all five plugins. Optional -Profile <name> and -StoreDir <existing-pnpm-store> use the same native CLI. Uninstalling code preserves documents, attachments, settings, and downloaded models.

The npm installer and GitHub assets use the same internal directory layout. The Windows wrapper runs the same installer.

Local speech model

The model release includes the SenseVoiceSmall INT8 ONNX weights (239,233,841 bytes), tokens, original model license, and SHA256 values. The voice plugin downloads them on first setup and recognizes audio locally thereafter. Its runtime is independent of the DSH installation, so updating DSH does not delete the model. See model provenance.

Automatic engine provisioning currently targets Windows x64. Other platforms require compatible sherpa-onnx binaries and explicit paths; they are not claimed to be tested.

New installations use <DSH_HOME>/third-party/data/voice-input; an existing legacy runtime and explicitly configured absolute paths remain supported. Relative plugin configuration paths resolve against DSH's data home, not the terminal's current directory. Package assets resolve relative to their own package.

Development and host upgrades

To update a standard npm-global DSH installation from the WebUI, open Settings → Plugins → Manage → DSH host update. The manager checks the official version, saves the previous host archive, and runs the update after the displayed confirmation. Restart DSH afterwards. Official modules remain excluded from ordinary third-party plugin operations. See the manager guide for supported installations and recovery limits.

Install root development dependencies, then each plugin's development dependencies with npm install --legacy-peer-deps. Run npm run build, npm test, npm run check:release, then npm run pack:plugins. Archives appear in artifacts/ and include built JavaScript. The SDK is provided by DSH at runtime, not copied into plugin bundles. Synthetic document fixtures are generated from source; personal files and recordings are excluded.

For local development, dsh plugin --profile web add link:./dsh-upload-button links a built package. Rebuild after source changes. Package updates, browser bundles, and host upgrades may require restarting DSH and refreshing the page even when profile patch reload is live.

Before a host upgrade, keep a local copy of the profile manifests and plugin archives, verify the new host in a separate profile, and run the plugin checks against that version. Update the pinned development SDK and testedDSHVersions only with test evidence. Plugins use public package exports and profile bundles; there are no patches to DSH's installed source. See compatibility and lifecycle.

Privacy and licenses

Plugin code is MIT. Model weights and downloaded inference binaries retain their upstream licenses. Speech recognition runs locally. Vision requests use the configured provider; document or image content sent to a model follows that provider's data handling. No personal settings, credentials, conversations, uploads, or recordings belong in this repository. The release checker reports suspicious paths and tokens without printing their contents; it complements code review and cannot prove that every possible secret is absent.