DeepSeek Harness 0.1.5 Released with DeepSeek V4.1 Flash Support and UI Enhancements
DeepSeek has released DeepSeek Harness v0.1.5, introducing specialized optimizations for the DeepSeek V4.1 Flash model alongside new file upload and sidebar file preview capabilities. The update also adds standardized UI extension points for plugins, improved long-session performance, and an experimental Agent Teams feature. This release strengthens DeepSeek's open-source agent infrastructure by allowing LLM agents to execute complex, multi-sub-agent workflows more efficiently. By supporting KV Cache retention during system prompt updates and Programmatic Tool Calling (PTC), developers gain lower latency and lower token consumption for complex agentic tasks. The update allows updating system prompts while keeping existing KV Cache for DeepSeek V4.1 Flash across standard, Programmatic Tool Calling (PTC), and minimal modes. Additionally, developers can quick-start the Web UI using `npx @deepseek-ai/dsh web` or enable experimental multi-agent collaboration features via CLI profiles.
## BACKGROUND
DeepSeek Harness (dsh) is an open-source agent framework developed by DeepSeek AI on top of the Cordis plugin architecture, designed to orchestrate LLMs, code execution, and developer tools. Key concepts in LLM inference include KV Cache (which stores generated key-value pairs to accelerate token generation) and Programmatic Tool Calling (PTC, where LLMs generate executable code to combine multiple tool calls locally rather than making individual API roundtrips).