DeepSeek Harness Plugins (DSH)

A curated collection of production-grade plugins and extensions for DeepSeek Harness (DSH).

These plugins turn DeepSeek Harness into a full-featured AI development environment: run agents on remote servers over SSH, browse files and run commands in an integrated Web terminal, enforce security guardrails on tool execution, inspect MCP servers and skills, and dynamically inject environment variables into system prompts.

---

Plugin Directory

Package	Category	Description	Web GUI	Tests
dsh-guard	Security & Guardrails	Monotonic tool interception, destructive command prevention, and approval gating	—	13 passed
dsh-workbench	Developer Tools	File tree explorer, live editor with syntax highlighting & gutter, visual diffs, and terminal emulator	Yes	17 passed
dsh-chat-ux	UX & Controls	Claude Code-style conversation layout, tool call folding, and fork/rollback actions	Yes	13 passed
dsh-git	Git & Worktrees	Worktree manager, branch switcher, dirty status indicator, and GitHub PR integration	Yes	13 passed
dsh-ssh	Remote Execution	Remote agent execution over SSH with ACP bridge and browser approval relay	Yes	80 passed
dsh-console	Observability	Skills & MCP servers inspector panel with /skills and /mcp slash commands	Yes	3 passed
dsh-env	System Prompt	Dynamic environment variables (today, platform, harness_version, etc.)	—	5 passed
dsh-updater	Lifecycle & Updates	Automated update detection, one-click Web GUI upgrades, and auto-restart	Yes	11 passed

---

Featured Capabilities

🛡️ dsh-guard — Tool Guardrails & Security Policies

Protect your workstation and infrastructure against accidental destructive operations before tool execution:

• Instant Denials (ctx.tools.guard): Synchronously rejects catastrophic commands (rm -rf /, raw disk writes dd of=/dev/sd*, mkfs, fork bombs) without consuming token rounds.
• Human Approval Gating: Intercepts dangerous operations (git force-pushes, branch deletions, server restarts, package publishing) and prompts for interactive user approval via Cordis tools/pre-execute waterfall hooks.
• Filesystem Confinement: Restricts write and edit tools strictly to permitted workspace boundaries, preventing traversal into sensitive directories (/etc, ~/.ssh).

💻 dsh-workbench — File Explorer & Web Terminal

An integrated developer workbench embedded right into the DeepSeek Harness Web GUI:

• Directory Tree: Browse files, inspect contents, view modified files, and preview git diffs directly in the sidebar.
• Terminal Emulator (xterm.js / PTY): A real, responsive terminal session running alongside your agent chat.
• Runnable Code Blocks: Adds an interactive "Run in Terminal" action button to markdown code blocks returned by the model.
• Safe Traversal Checks: Uses hardened path confinement logic to prevent directory traversal and symlink escapes.

🌐 dsh-ssh — Remote Agent Execution via ACP

Run DeepSeek Harness agents directly on remote staging, cloud VMs, or production hosts over SSH:

• Agent Client Protocol (ACP): Connects to remote machines over SSH stdio (ssh host dsh --profile acp) with bidirectional JSON-RPC streaming.
• Local Workspace Shadows: Indexes and caches workspace mirrors under ~/.dsh/remote/<host>/, allowing local session history, search, and context persistence while operations run remotely.
• Approval Relay: Transparently routes remote tool approval requests back to your workstation's local browser UI.
• Web GUI Host Switcher: Switch target servers on the fly from the chat header, browse remote folders, and probe ping latency.

🔍 dsh-console — Skills & MCP Server Inspector

Complete visibility into your agent's active tooling and runtime configuration:

• Settings Panel: Adds a dedicated "Skills & MCP" panel to the Web GUI settings.
• Slash Commands:
  • /skills: Lists all discovered workspace and global skills, instructions, and invocation arguments.
  • /mcp: Displays active Model Context Protocol (MCP) servers, client transports, and exposed tools.
• Privacy-Safe: Redacts authorization tokens, bearer headers, and environment secrets from inspection output.

⚙️ dsh-env — Dynamic Prompt Environment Variables

Inject real-time runtime facts into agent system prompt templates:

• Built-in Variables: Automatically maintains today (date boundary), platform (OS name/architecture), os_version, and harness_version.
• Custom Variables: Define key-value pairs in configuration to supply deployment-specific context.

🔄 dsh-updater — Update Manager & One-Click Upgrades

Keep DeepSeek Harness and all community plugins continuously updated right from the Web GUI:

• Automatic Version Detection: Queries the npm registry for new @deepseek-ai/dsh releases and monitors git remotes for plugin updates.
• Web GUI Settings & Header Badge: Adds an "Updates" settings page and an alert badge in the conversation header when a new version is published.
• Automated Symlink Repair: Automatically fixes virtual store symlinks when updating @deepseek-ai/dsh in pnpm workspaces.
• Seamless Auto-Restart: Restarts the server gracefully in the background while the browser UI automatically detects when it is back online and refreshes.
• Slash Commands: Use /update to inspect release versions or /update now to upgrade directly from chat.

---

Installation & Usage

1. Clone and Link Plugins

Clone this repository and run the automated installer to link all plugins into your DeepSeek Harness profile (web by default):

git clone https://github.com/artemiroshnichenko/dsh-plugins.git
cd dsh-plugins
pnpm install
pnpm build
pnpm run link-plugins

2. Enable in Configuration

Add the desired plugins to your local ~/.dsh/cordis.patch.yml or ~/.dsh/profiles/web/cordis.patch.yml:

- insert:
    - id: workbench
      name: dsh-workbench

    - id: chat-ux
      name: dsh-chat-ux
      config:
        dedupSystemPrompts: true
        pageByTurn: true

    - id: git
      name: dsh-git
      config:
        ghEnabled: true

    - id: guard
      name: dsh-guard
      config:
        blockSshMutations: true

    - id: console
      name: dsh-console

    - id: env
      name: dsh-env

    - id: updater
      name: dsh-updater

# To route sessions over SSH (replaces default local agent loop):
- delete:
    - id: agent-loop
- insert:
    - id: ssh
      name: dsh-ssh
      config:
        hosts:
          - name: staging
            ssh: ubuntu@staging.example.com
            cwd: /var/www/app
            port: 22

Restart your DeepSeek Harness server or reload your profile:

dsh web

---

Architecture & Standards

All plugins in this repository follow DeepSeek Harness architectural standards:

• Cordis DI Architecture: Pure service, plugin, and effect registrations via @deepseek-ai/cordis.
• Strict TypeScript: Compiled with TypeScript 5.5+ in strict mode (noImplicitAny, strict null checks).
• Runtime Schemas: Validated at configuration boundaries using @deepseek-ai/schemastery.
• Typert RPC Gateway: Clean host-to-client remote procedure calls using @deepseek-ai/dsh-typert-protocol.
• Comprehensive Unit Testing: 112 unit tests powered by Vitest, running against real fixtures without mocking network boundaries.

---

Monorepo Development

# Clone the repository
git clone https://github.com/artemiroshnichenko/dsh-plugins.git
cd dsh-plugins

# Install workspace dependencies
pnpm install

# Build all packages and Web client bundles
pnpm build

# Run unit tests across all packages
pnpm test

# Typecheck the entire codebase
pnpm typecheck

---

Frequently Asked Questions (FAQ)

Can I use dsh-ssh with password authentication?

We strongly recommend configuring SSH key authentication or an SSH agent (`ssh-add`) so that DSH can spawn non-interactive background SSH sessions without blocking on terminal password prompts.

Does dsh-guard slow down tool execution?

No. Static security rules and pattern inspections run synchronously in sub-millisecond time before commands are dispatched to the shell.

How does dsh-workbench isolate file operations?

All file access is validated against real canonical paths (`fs.realpath`) to ensure operations cannot escape authorized project roots via symlinks or .. paths.

---

Contributing

Contributions, feature requests, and issue reports are welcome! Please check out CONTRIBUTING.md to get started.

License

MIT © 2026 Artem Miroshnichenko