User Guide

Master DeepSeek Harness: CLI & profiles, agent presets, skills & goals, subagents & workflows, plugins, and the Python SDK

Welcome to the user guide. If you already got DeepSeek Harness running through installation and walked the Web GUI quickstart, this section helps you shape it to your needs: launch flags, per-session agent composition, skills, goals, multi-agent orchestration, and programmatic access.

Chapter overview

CLI & Profiles

Launcher flags, profile layers, headless one-shot tasks, plugin management, and dump-config

Agent Presets

Compose one agent per session: persona, tools, prompt sections, and skills from a single agent.cordis.yml

Skills & Goals

Load reusable instruction packs on demand; keep a session driving toward one long-running objective

Subagents & Workflows

Delegate work to child agents, or orchestrate large multi-agent collaborations with workflow scripts

Third-party Plugins

Install and use dsh-plugin packages inside a profile

Python SDK

Skip the Web UI and drive the harness programmatically from Python through the same API

Migrate from Codex / Claude Code

Three gradual paths: model-level compatibility, hooks.json bridges, and native Cordis plugins

Config Reference

A curated lookup of settings, credentials, environment variables, model and permission config

Tools Reference

A curated lookup of the tools available to the agent: files, commands, delegation, goals, plugins

Reading order

To understand every dsh launcher flag and how profiles layer, start with CLI & Profiles.
To give different sessions different personas and tools, see Agent Presets.
To attach reusable instructions or drive one long-running objective, see Skills & Goals.
To delegate to child agents in parallel or orchestrate multi-agent flows, see Subagents & Workflows.
Migrating from Codex or Claude Code? See the migration guide.
Looking up config keys or the tool list? See the config reference and tools reference.
To extend the harness itself, start with the developer guide.

Developer preview

DeepSeek Harness is in developer preview and iterating rapidly. Breaking changes are expected. This documentation describes current behavior.