The first step is to run each individual unittest function with coverage.

This is how Pants does test coverage atm. Each test file is a distinct process, and we use

--pytest-cov

to get coverage data for just that run. Then, we run

coverage combine

to merge all the distinct results into a single report

Then take the code coverage and run mutation on each line

Because each test runs in a chroot, I think this will be possible. That is, mutating the code will not impact your actual build root. It only mutates the chroot. It’s safe.

For each mutation, you will need to publish it to test runner such as buildfarm.

Is the remote execution only for performance, or some other motivation?

You will need a way to report false positive by converting the impacted code to AST then generalize in an AI model that are used in NLP.

That part we might not be as able to help with 😅