LTP (Lean Task Protocol) compiles a natural-language goal into a deterministic
execution plan in a single LLM call, then runs it step by step in pure Python.
Unlike a ReAct loop, which re-invokes the model at every turn, LTP pays the model
cost once. This makes its token cost flat regardless of how deep the task chain is.
The kernelmcp Hybrid mode runs LTP first and falls back to a ReAct loop only when
the compiled plan produces an inadequate result, combining cost and reliability.
