DeepSeek Harness v0.1.5 version is now available. The new version is deeply integrated with the DeepSeek V4.1 Flash model training, and the model has been specifically trained and optimized under different configurations of DeepSeek Harness.

image.png

File Upload and Sidebar Preview Launch, Plugin Expansion Entry Standardized

In terms of adaptation for V4.1 Flash, this model has been specifically trained and optimized for Harness, covering standard mode, programmatic tool call (PTC) mode, and minimal mode; when using this model, the new version supports updating system prompts while retaining existing KV Cache.

File processing and preview: The Web interface supports uploading images, PDFs, and other file types, and the model can read files as needed through file tools. The right Sidebar supports browsing the workspace file tree and directly previewing Markdown, HTML, PDF, common code, and image formats. Files generated by the model can be viewed in the Sidebar, or opened with the default system application or located in the file manager.

Plugin expansion: The Web interface has established standardized plugin expansion entries on both sides. Plugins can register global entries between "New Session" and "Workspace" in the left Sidebar and interact with the central content panel; the right Sidebar supports plugins registering new Tab types and content pages, and displaying preview content or tool interfaces through multiple tabs, split screens, and full-screen methods, laying the foundation for future plugin management features.

Long Session Performance and Sub-Agent Collaboration Optimization

In terms of long session performance and experience, the new version optimizes the loading, recovery, and continuous conversation performance of long sessions, reduces memory usage, and optimizes the storage format; session content folding, complete history navigation, and interactions such as main text width, font size, and operation statistics have also been improved.

In terms of sub-agent communication and task control, the parent agent and controllable sub-agents support bidirectional communication, allowing additional information to be added and task direction adjusted during execution; users can queue messages, edit or delete pending messages, interject or stop tasks, and the main agent can choose models and inference intensity for sub-agents.

The experimental feature Agent Teams allows the main agent to create multiple team members, split, assign, and track work through shared task lists, team members can send messages to each other, and the main agent can view progress, wait for results, interrupt tasks, and consolidate them uniformly. This feature is provided as an experimental plugin, disabled by default, and needs to be added to the Profile via command line for use, and it will consume additional Tokens.

In terms of installation, systems that have installed the Node.js toolchain can quickly start the Web UI using the npx command, or obtain the complete source code from the GitHub repository. According to the introduction, future versions plan to add a built-in plugin management panel and provide more expansion entries for plugin developers.