@hackage / shikumi-optimize

The optimizer framework for shikumi: search procedures that automatically improve a Shikumi.Program.Program by rewriting its per-node optimizable parameters (instruction strings and few-shot demonstrations) and scoring each candidate with evaluate (EP-8) over a typed dataset and metric. Four strategies ship — labeled few-shot demo selection, bootstrap few-shot (recover demonstrations from a teacher's successful runs), instruction search (MIPRO/COPRO-style coordinate ascent under an explicit call budget), and ensemble search (combine complementary candidates by majority vote). Each optimizer returns a CompiledProgram (EP-9, MasterPlan integration point #6). Every acceptance assertion runs fully offline against a deterministic stub LM, so the whole package tests with no network and no API key.