DeepSeek Harness: A Plugin-Based Agent Framework

Architecture and Plugin Design

DeepSeek Harness (dsh) utilizes a modular architecture built on the Cordis plugin kernel. Unlike proprietary agent binaries that function as black boxes, this framework treats every component—including the agent loop, session logging, and UI rendering—as a swappable plugin. The system uses a boot profile stack where bundles contribute configuration rows, allowing users to override any behavior by modifying specific identifiers in a patch file. The framework is written in TypeScript and requires Node 22 or newer, enforcing 100% test coverage per file.

Migration and Compatibility

To lower the barrier for adoption, the harness includes bridges for existing Claude Code and Codex configurations. The Claude Code bridge reads an existing hooks.json file and executes those hooks at the corresponding lifecycle events, such as session start, prompt submission, and tool calls, without requiring code rewrites. The system also features an MCP client, a unified skills registry for local and remote services, and an experimental agent team module that provides sub-agents with a durable mailbox for inter-agent communication.

Security and Production Readiness

Despite the rapid accumulation of stars and commits, the project remains in a pre-release state. The repository documentation explicitly warns that the software has not undergone a security audit and should not be considered production-ready. The maintainers recommend running the harness within a disposable virtual machine to mitigate risks associated with sandboxing and tool execution. Furthermore, the project maintains no compatibility promises for on-disk formats, prioritizing architectural flexibility over long-term stability.