AI coding has made CI a bottleneck, so we reworked ours to keep up

AI-driven development sped up how fast code lands, but CI became the bottleneck at Linear, so the engineering team reworked CI to cut pull-request wait time and runner costs even as the test suite almost quadrupled. They focused on four areas: upgrade infrastructure and toolchain, optimize jobs that gate other work, eliminate repeated setup, and make test execution more efficient. Key wins: moving off GitHub Actions to third-party runners made jobs 34% faster on average (some tsc runs 52% faster); switching to tsgo cut median tsc time by 73%; rewriting TypeScript-dependent lint rules to use AST-only checks reduced API lint time 68% and full-repo lint 55%, and Oxlint further lowered runner-minutes. Overall PR wait dropped from over six minutes to just over five, and runner time per test roughly halved.

They reduced critical-path delays by fetching only what each job needs (change-detection median 26→8s, p90 31→12s, slowest 138→37s), replacing checkout with a retrying composite action to avoid hangs, and moving cache writes off the merge path (saving ~42s per API PR). Repeated setup was cut by shipping a CI base image with preinstalled deps, narrowing pnpm installs to the API package (44-73s → 16-18s), and abandoning slow caches; per-shard setup fell ~44% (110-140s → 67-73s). They batch small checks (saving ~87,000 runner-minutes/month, ~11.8% of CI use), split large test files and increased shards (3→8) to reduce the slowest shard 5.25→4.33 minutes, and used opt-in vitest projects to safely share module state within workers.