[akash-aman]

Summary

• Cause: a -race binary sleeps for GORACE atexit_sleep_ms (1000 ms by default) when it exits with status 0, to catch races at exit. The fake agent-browser is the test binary run once per CLI call, so every fake call slept a second.
  • An empty func main(){} built with -race takes 1,087 ms against 6 ms without, using no CPU.
  • A stack sample shows the main thread in runtime.main → __tsan::Finalize → usleep.
  • CI was slower still (282 s for the tools package) because a 4-CPU runner runs 4 parallel tests at a time.
• Fix: the fake exits with syscall.Exit, which skips that shutdown (internal/testutil/fakecli/fakecli.go). The code under test is still race-checked in the test process; only the fake skips it.

go test -race (16-core Mac)	Before	After
internal/mcp/tools	78.9 s	9.6 s
internal/browser	6.0 s	1.8 s
internal/devtools	7.3 s	3.8 s
./... (wall)	about 80 s	11.9 s

Test plan

• go test -race -count=1 ./... five times in a row: all pass, tools 10.3–11.1 s
• go test -count=1 ./... and go vet ./..., gofmt -l clean
• CI: runs once Publish releases with GoReleaser on version tags #4 (which adds the CI workflow) is merged; the race-test step should drop from about 5 minutes to under one