Benjy: looks like there might be a clear opportunity to optimize the output mode for python-eval? see e.g. https://travis-ci.org/pantsbuild/pants/jobs/215822589#L3820