A bare wait also waited for tee

I set up a Claude Code routine that picks up an approved ticket, makes the change in a client’s Rails app, and opens a PR. It runs in a cloud container that has no MySQL, no Chrome, and not the project’s Ruby 3.4, so a setup script installs them first. Claude starts only after that script exits. Mine took two fixes before it did, one for Ruby and one for a bare wait. The first run never got to Claude:

Setup script failed with exit code -1. Script output: ... Setting up google-chrome-stable ...

It stopped about 8 minutes into the session. The script compiled Ruby with mise, and that alone took minutes. So I switched to a prebuilt Ruby, the tarballs from ruby/ruby-builder that setup-ruby uses on GitHub’s runners. Unpacked into mise’s install folder, it didn’t run:

/root/.local/share/mise/installs/ruby/3.4.10/bin/ruby: error while loading shared libraries:
libruby.so.3.4: cannot open shared object file: No such file or directory

Those builds expect to live at their build path, /opt/hostedtoolcache/Ruby/<version>/x64. Unpacking it there and pointing mise’s version folder at it with a symlink worked:

mkdir -p "/opt/hostedtoolcache/Ruby/${RUBY_VERSION}" "$HOME/.local/share/mise/installs/ruby"
tar -xzf /tmp/ruby.tgz -C "/opt/hostedtoolcache/Ruby/${RUBY_VERSION}"
ln -sfn "/opt/hostedtoolcache/Ruby/${RUBY_VERSION}/x64" \
        "$HOME/.local/share/mise/installs/ruby/${RUBY_VERSION}"

In an amd64 Ubuntu container, ruby -v printed ruby 3.4.10 ... [x86_64-linux], the OpenSSL extension loaded, and Bundler installed.

Then the script never ended. It logged everything through tee, ran two install functions in the background, and called wait:

exec > >(tee -a "$LOG") 2>&1
packages &
ruby_and_node &
wait

In a local test container it had been running for an hour and 27 minutes, with two processes left:

PID  PPID  ELAPSED  COMMAND
2483    1  01:26:35 bash /work/routine_setup.sh
2484 2483  01:26:35 tee -a /var/log/setup.log

Both install jobs were done. wait with no arguments was also waiting for the tee from the process substitution. tee exits when the script’s output closes, and that only happens when the script exits. Waiting on the job PIDs fixed it:

packages &
PACKAGES_PID=$!
ruby_and_node &
RUBY_PID=$!
wait "$PACKAGES_PID" "$RUBY_PID"

A small test script with the same exec > >(tee ...) and the PID wait exited in 3.5 seconds. The full script installed Chrome, MySQL 8.0, and Ruby 3.4.10 and exited. That was bash 5.2 on Ubuntu 24.04.

One more thing from that day. “Setup script failed” wasn’t the end of the run. The cloud started the same session again 4 minutes later with the setup cached, and that run opened the PR. Claude had already marked the ticket as failed, so a setup failure is worth a second look before anything gets closed.