Give coding agents a tmux session with your real shell environment

Note: Using tmux and other more-or-less built for agent multiplexers is a slightly different approach to the one I describe here - I’m talking about using tmux from an agent instead of the agent running shell commands directly. Just saying that upfront to avoid confusion!

Coding agents spawn a fresh shell by default. That shell is usually set up slightly differently than how I set up ones I use from my terminal.

mise lives in my .zshrc, along with Homebrew on the PATH and the Postgres credentials I export there. A non-interactive agent shell often skips .zshrc entirely. zsh only sources .zshenv unless you force it. So the agent gets system Ruby, and node might not even be on the PATH. PGHOST isn’t set either. The agent can work away and resolve these issues itself, but it’s inefficient.

I don’t want to teach every agent how to bootstrap my machine. I already have a shell that’s set up the right way. tmux lets me hand that shell over.

Start a detached session from a terminal that’s already configured, in the project directory. You can also use tools like tmuxinator to create a session that’s already seeded with the windows and panes you want:

tmux new-session -d -s agent -c "$PWD"
# or
tmuxinator my-project

The session inherits that terminal’s environment. mise has already selected Ruby and Node, and PGHOST is set, etc etc.

Then tell the agent to run commands in that session rather than its own throwaway shell - for example, the agent can interact directly with tmux:

tmux send-keys -t agent 'bundle exec rspec' Enter
tmux capture-pane -pt agent

And I can attach while it’s working:

tmux attach -t agent

I don’t use this all the time. Mostly when I need a particular environment or shell setup. Being able to attach to the agent’s terminal and read the history is a nice side effect.