I'm building DiffCI, an open-source project exploring a simple question:
How much CI/CD work can we safely avoid when a code change affects only a small part of a repository?
We tested the approach against historical changes from projects including Turborepo, Nx, Biome, Cal.com, and DeepSeek Harness.
A few interesting things happened:
On Nx, testing exposed a bug where a directly modified test wasn't selecting itself. We fixed it.
On Biome, our analyzer initially failed because it assumed a root TypeScript config existed. Real repositories aren't that simple.
On Cal.com, DiffCI identified just 2 of 424 tests as affected — but Vitest workspace behavior still executed essentially the whole suite.
That last result was particularly important.
Knowing what doesn't need to run and actually preventing CI tooling from running it are two different problems.
We're now expanding the work beyond test selection toward builds, linting, type checking, packaging, and other CI/CD stages.
The goal isn't to report impressive theoretical percentages. It's to measure actual compute, time, electricity, and eventually emissions avoided while falling back safely whenever DiffCI isn't certain.
DiffCI is open source. We're looking for more real-world repositories to test.
GitHub: https://github.com/DiffCI/core
Top comments (0)