moulwyse

Lattice sits on top of Codex and controls what the model sees at each step. Same model, same behavior - 76% less text sent per task, 62% faster, 4/4 tasks accepted. Long refactors finish instead of dying at 60% of the window.