AI Programming

DeepSeek Harness Turns LLMs Into an Execution Layer With User-Written Plugins

By 小虎AI生活 · Aug 21, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Harness shifts the LLM from a conversational tool into a programmable agent that can read files, write output, and chain tasks. The "everything is a plugin" architecture means any developer with basic TypeScript can extend the system's capabilities without waiting for upstream features.

Summary

DeepSeek Harness ("Black Whale") is not a new model but an execution layer built on the Cordis plugin framework. Every capability — models, tools, UI, session recording — is a plugin, and user plugins hold equal status with the 195 official packages. A minimal plugin requires only a named export, an injection declaration, and an apply function that registers a tool with a parameter schema and execute handler.

A practical daily-report plugin reads a workspace's todo.md, wraps it in a draft template marked "pending confirmation," and writes the output to a dated file — explicitly refusing to fabricate content the user never did. The plugin structure (single TypeScript file, YAML patch config) is simple enough that AI coding tools can generate complete plugins from a specification.

Version 0.1.0 is pre-release with unstable interfaces, and plugins run as local code with full system access, so only official or highly-starred community plugins are safe to install.

Takeaways

— Harness is an execution layer, not a model — it wraps LLMs and lets plugins define tools, interfaces, and session handling.

— The Cordis framework treats every capability as a plugin; the 195 official packages and all community plugins share identical status.

— A minimal plugin exports a name, an inject array, and an apply function that registers a tool with parameter/output schemas.

— Named exports are mandatory — export default silently discards injection declarations with no error.

— Object-type output schemas must include additionalProperties: false or registration fails.

— Plugin paths in cordis.yml must be absolute; relative paths are not supported.

— The daily-report plugin reads a todo.md, generates a draft explicitly marked "pending confirmation," and never fabricates content.

— Community plugins under the dsh-plugin tag exceeded 900 within the first week of release.

— Version 0.1.0 is pre-release with unstable interfaces; production deployment is not recommended.

— Plugins execute as local code — installing an untrusted plugin is equivalent to granting system access.

Conclusions

Harness's plugin architecture mirrors the extension model that made VS Code dominant: a thin core with all surface area exposed to third-party code.

The explicit "pending confirmation" design pattern in the daily-report plugin is a practical, low-tech guardrail against AI hallucination — it constrains the tool to only read and template existing data rather than generate claims.

Plugin count hitting 900 in one week suggests latent demand for turning LLMs from conversational interfaces into task-execution engines, but the pre-release instability and security model mean most of those plugins are experiments, not dependable tools.

Concepts & terms

Cordis

The plugin framework underlying DeepSeek Harness. It treats every system capability as a plugin — a TypeScript file exporting a fixed structure with name, inject, and apply — and gives user plugins equal status with built-in ones.

Harness (Black Whale)

DeepSeek's open-source execution layer that wraps large language models. It provides a plugin system where tools, interfaces, and session recording are all assembled from plugins, turning the model from a conversational interface into a programmable agent.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗