DeepJIT: Compiling Repeated Agent Workflows into Reusable Skills and Processes for DSH

DeepJIT is a DSH (deepseek-harness) plugin maintained by fly3366, designed to transform repetitive workflows in agent execution into reusable assets. It achieves this by monitoring agent execution traces to mine high-frequency "hot workflows" and utilizes LLMs to compile them into markdown skills or flow step templates. These artifacts support hot-loading into running DSH instances without requiring a restart. The plugin provides comprehensive compiled-based lifecycle management, including AOT validation, tiering, deoptimization, and garbage collection (GC), while simultaneously recording metrics and emitting GenAI semantic-convention spans for monitoring.