17.2 Installation & 4 Run Modes

🐋 "One command starts an Agent; four profiles decide its capability boundary."

Four Profiles at a Glance

DeepSeek Harness uses profile-based startup — dsh --profile <name> loads a preset plugin set:

Mode A: standard — Daily Work

npx @deepseek-ai/dsh web          # Web UI at 127.0.0.1:3080
dsh chat                          # CLI
dsh gateway                       # Telegram/Discord etc.

Config (~/.dsh/config.json):

{
  "profile": "standard",
  "llm": {
    "provider": "deepseek",
    "model": "deepseek-v4-pro",
    "apiKey": "${DEEPSEEK_API_KEY}",
    "fallback": ["anthropic-claude-4-7", "openai-gpt-4.1"]
  },
  "workspace": "/workspace",
  "mcpServers": { "github": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"]} },
  "permissions": {
    "shell": ["ls","cat","grep","find","pwd"],
    "shellBlocklist": ["rm -rf /","sudo","chmod 777","curl | sh"]
  }
}

Mode B: minimal — Benchmarking

dsh --profile minimal eval --benchmark swe-bench

Only shell + file edit — controls for variables so model capability itself is measured, not tool misuse. Any paper using DeepSeek Harness should state the profile explicitly.

Mode C: ptc — Programmatic Tool Calling

PTC lets the LLM generate a program (Python/TS) that the harness executes to call tools — N tool calls collapse into one LLM output:

Normal loop:  LLM → tool_call(read) → execute → result → LLM → tool_call(grep) → ...
PTC:          LLM outputs Python (read files, find bugs, edit) → harness executes all

Pros: token savings, readability, composability (conditionals/loops/try-except). Cons: needs stricter sandbox (PTC sandbox blocks outbound network, non-workspace writes, subprocess, os.system by default).

Mode D: create — Plugin Prototyping

dsh --profile create dev

Watches ./plugins/**/*.ts; auto-reloads in ~200ms on save — sub-second feedback loop for plugin development.

Install: 3 Ways

# 1) npx
npx @deepseek-ai/dsh web

# 2) source
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness && corepack enable && pnpm install && pnpm dsh web

# 3) Docker
docker run -d --name dsh -v ~/.dsh:/root/.dsh -v ~/workspace:/workspace \
    -e DEEPSEEK_API_KEY=sk-... -p 3080:3080 deepseek/harness:latest web

Repo layout (per main): packages/ (core/llm/shell/terminal/fs/edit/lsp/web/skill/subagent/workflow/mcp/scheduler/...), plugins/, skills/, profile/ (4 JSON), docs/.

Section Summary

Next section: 17.3 Architecture: Cordis Microkernel & Plugin Topology