DeepSeek Harness: Building and Extending AI Coding Agents Training Course

Overview of DeepSeek Harness

• Defining DeepSeek Harness and its role in the ecosystem
• The "everything is a plugin" design principle
• Web UI, CLI, and headless operational modes

Setup and Initial Execution

• Node.js prerequisites and package installation
• Starting the Web UI using npx
• Compiling and running from source code

Model Setup

• Configuring the DeepSeek API key
• Choosing and toggling between models
• Integrating OpenAI-compatible endpoints

Workspaces and Sessions

• Selecting and changing workspaces
• Using the session composer and managing conversation flow
• Reading, modifying, and generating files

Executing Coding Tasks

• Summarizing repositories and navigating codebases
• Executing commands and delegating tasks
• Maintaining a structured task plan

Permissions and Sandbox Environment

• Managing approval and permission policies
• Restricting access to files and commands
• Configuring local and remote sandbox backends

Native Tools and Integrations

• Filesystem, terminal, and shell utilities
• Support for the Model Context Protocol
• Integration with the Language Server Protocol

Subagents and Task Delegation

• Creating and forking subagents
• Utilizing product providers for external coding agents
• Delegating tasks across a session

The Cordis Plugin Framework

• Plugins defined as services, events, and effects
• Shared context and lifecycle management
• The absence of a privileged core system

Profiles, Bundles, and Configurations

• Managing profiles and bundle layering
• Modifying configurations via cordis.patch.yml
• Reviewing the initialized dependency tree

Core Subsystems

• The append-only session logging mechanism
• Assembly of system prompts and tool schemas
• Agent loop mechanics and turn processing

Events and Extension Points

• Handling session, agent, and capability events
• Implementing waterfall and interceptor events
• Mapping specific behaviors to extension points

Plugin Development

• Registering services using context keys
• Incorporating new tools and model adapters
• Publishing and discovering plugins

SDK, Headless Mode, and Automation

• Controlling the harness from external processes
• Using the TypeScript and Python SDKs
• Automating via CLI and headless operations