I have a couple ideas for improvements here now that we're past the MVP:

1. Detect if we're only running a single file, and disable parallel mode automatically. Running a single file in parallel mode will always (AFAIK) be slower than running in serial. Add a command-line option to disable this behavior (a general-purpose, contextual --force might be helpful)
2. Automatic optimization via duration caching
   • Cache per file durations (cache could live in node_modules/.cache/mocha, which is an unofficial convention).
     • This would be timing the run() call in lib/nodejs/worker.js, from beginning to end.
     • Maybe calculate the mean over n most recent runs?
   • On subsequent runs, execute slowest test files first. This will help avoid the case at the end of the run where there's only a single worker process, munching on a meaty test file, and the other workers are idle. TypeScript uses a strategy like this in their custom tooling around Mocha (look at their implementation for ideas).