How preloop works.
Technical notes from the team that builds preloop, on test selection, debugging and caching.
LOOP04beats
01Changesnapshot
02Verifyreplay
03Fixpause
04Mergepublish
3 min01
GitHub Actions now supports workflow lockfiles
A lockfile turns action tags into recorded commits. That removes a class of supply-chain risk, but only if CI checks it before a job runs.
preloop engineering4 min02How preloop decides which tests to skip
preloop skips a test only when its detected inputs are complete, unchanged, and safe to trust.
preloop engineering4 min03A fork PR can cancel your production deploy
A concurrency group keyed on a branch name lets any fork cancel a deploy. The same shape appears in workflow resolution, environment gates, OIDC and artifacts.
preloop engineering3 min04Pause a failed CI job instead of rerunning it
Start a CI job with --debug, and preloop keeps a failed step's machine and workspace available for debugging and retry.
preloop engineering4 min05What preloop caches and how it stays correct
preloop serves eighteen cache protocols from one content-addressed pool with immutable entries and scoped lookups.
preloop engineering3 min06Why we rewrote the GitHub Actions runner in Rust
The official runner carries a .NET runtime and two Node versions into every job. Ours is 36x smaller and idles at a seventh of the memory.
preloop engineering2 min07Typed CI with generated YAML
Generate provider-compatible workflow YAML from typed source, keep the YAML in the repository, and check for drift in CI.
preloop engineering3 min08Deciding which state to explore next in property testing
Property testing gives you an oracle. The hard part is choosing the next action, and a model-guided explorer found the right one about 3x more often than random.
preloop engineering3 min09The models I daily-drive for CI repair
On CI-bench, the models on the cost frontier are also the cheapest. For CI repair, reliability per dollar beats the largest model.
preloop engineering4 min10Do you really need remote CI?
Most of what remote CI does is the same work you could run where the change was made. The useful question is which checks still need to happen somewhere else.
preloop engineering5 min11CI-bench: how good are agents at diagnosing and fixing CI failures?
We ran 80 real CI failures through nine agent configurations, three times each. First-attempt scores hid real differences in reliability.
preloop engineeringWriting / 11 entries
All writing