DeepJIT

Every day in dsh, you probably catch yourself re-running nearly identical multi-tool sequences: call tool A, wire up B's parameters, trigger C's workflow. The steps are almost the same every time, yet you still execute them manually. DeepJIT was built for exactly this kind of repetitive friction: it quietly watches agent execution traces in the background, identifies recurring high-frequency workflows, and uses an LLM to compile them into reusable Markdown skills or step-by-step flow templates—then hot-reloads them straight into the live session with no restart.

The compilation pipeline borrows ideas from JIT compilers. An AOT pass first validates flows against the live tool registry and constant-folds literal arguments. A tiering layer then promotes frequent, reliable skills into flow templates and deoptimizes underperforming ones. A GC pass prunes stale artifacts that have gone unused beyond a configurable retention window. All artifacts live locally, and JIT deliberately never compiles its own tools, ruling out self-referential loops.

If you spend a lot of time in dsh repeating similar multi-step workflows and want to turn that muscle memory into versionable, reusable assets, DeepJIT handles the accumulation for you—so you can stay focused on the next task instead of re-running the last one.