Race after `close`: immediate relaunch with a different launch config (`--headed`) fails ~50% with "Failed to connect: No such file or directory (os error 2)"

Summary

close returns while the session daemon is still alive (its <session>.sock / <session>.pid stay in ~/.agent-browser for ~1 s). If the very next command needs a different launch config — e.g. --headed after a headless run — it fails about half the time with:

✗ Could not configure browser: Failed to connect: No such file or directory (os error 2)

An immediate relaunch with the same config (headless → headless) never fails, so this looks like a race between the old daemon's shutdown and the CLI's "restart daemon for incompatible launch" path connecting to a socket that has just been unlinked.

Reproduction

agent-browser 0.37.1, macOS 15 arm64 (Homebrew), Chrome for Testing 153.0.8010.36 installed via agent-browser install. No profile, no plugins.

for i in 1 2 3 4 5 6; do
  agent-browser --session race open about:blank >/dev/null
  agent-browser --session race close >/dev/null
  agent-browser --session race --headed open about:blank 2>&1 | head -1
  agent-browser --session race close >/dev/null
done

Observed (6 runs): 3 × ✗ Could not configure browser: Failed to connect: No such file or directory (os error 2), 3 × ✓. Same ratio with --profile <dir>.

Control — same loop without --headed on the third command: 6/6 ✓.

State right after close returns:

agent-browser --session race open about:blank; agent-browser --session race close
ls ~/.agent-browser | grep '^race\.'      # race.pid race.sock still present
ps -p "$(cat ~/.agent-browser/race.pid)"  # daemon still alive
sleep 1; ls ~/.agent-browser | grep '^race\.'  # gone

Expected

Either close blocks until the daemon has actually exited and removed its socket, or the relaunch path treats ENOENT on the stale socket as "daemon gone" and proceeds to spawn a fresh one.

Workaround

sleep 1 (or a single retry) between close and the mode-switching command.