TokenReducer: An Experiment in Context Placement

Field notes

The workflow keeps the parent coordinator and its selected model while moving a large read into a fresh scripted worker process. A gate checks the real line window, with delegation guidance triggered above 350 lines; the worker then returns a bounded answer or a file confirmation. The page reports 545 passing tests and records a 4,096-byte compact-answer limit, while explicitly distinguishing reduced parent input from reduced total processing.