test results are now ga in depot ci and github actions. ingest junit xml from your ci to get pass/fail/error/skip counts on every job, plus org-wide analytics for flaky, slow, and recurring failures.
kyle wrote up something we've been chewing on: the branch-pr-review-merge workflow was built for humans at human speed. agents blew past that, and github's collaboration model is now the bottleneck.
best test for a process you inherited: ask what breaks if you delete it. "nothing, but people would be annoyed" means it's a ritual with no principle left. watts on keeping the why alive as you scale.
we cut our slowest ci job from 6m 40s to 2m 36s. maxed vitest parallelism to use spare cpu (3m 54s), then sharded across 2 jobs with depot test splitting, now in beta.
fingerprint a failure too strictly and every run looks like a brand-new bug. too broadly and unrelated failures get lumped together. getting that balance right was the whole challenge, and we wrote up how we approached it.
