tokencut folds recognized tool-output noise and leaves a reference to recover the original.

tokencut wraps a command (tokencut run -- pytest -v) and folds known-noisy output while keeping diagnostics, with omitted text retrievable by reference. Its built-in pytest fixture goes from 1,562 tokens to 174 including the recovery notice, and tokencut demo --json reproduces it with no model calls. The README is explicit that this is a local tokenizer estimate and not a billing claim. Default filtering preserves unfamiliar output and stronger truncation is opt-in, which is the right default for anything sitting between an agent and its test suite.