Predictive Test Selection for Normal-Sized CI Pipelines Without Meta's Infrastructure
Meta published Predictive Test Selection in 2018: train a model on historical test outcomes, select the ~30% of tests relevant to a given diff, catch 99.9% of regressions. Seven years later, no off-the-shelf tool brings this to teams outside FAANG. TestImpact.io shut down, Launchable pivoted, Buildkite Test Engine exists but is narrow and expensive, Gradle Enterprise is JVM-only. AI-assisted development is pushing CI bills up 3–5x (more PRs, more agents, more commits) and a December 2025 Ask HN thread explicitly asks for 'an LLM tool that can sit on a CI pipeline to propose what tests should be blocking.'
Forget the LLM framing — the original Meta approach is a gradient-boosted decision tree, which is fine. What's new is 'GitHub Actions reusable workflow you add in 3 lines, we slurp your coverage data + PR history, we send back a set of test IDs to run.' Monetize per-CI-minute saved; that pricing sells itself to the CFO.
landscape (5 existing solutions)
The technique is seven years old and openly published. Nobody has turned it into a product a 15-engineer team on GitHub Actions can drop in with an action reference. The CI-bill-shock from AI-generated PR volume is forcing this conversation right now — every team with a 40-minute test suite is quietly bleeding.