herdr

herdr runs an agent interactively in a terminal pane. The pane inherits herdr's environment, so the committed project layer applies.

Set up

boxer install claude-code
git add .claude .mcp.json && git commit -m "Run agent commands in boxer"

herdr starts the agent itself, so you cannot pass --permission-mode. Grant permissions in the project settings, and include Skill. Using boxer's skill needs its own permission, and a pane waiting on Use skill "boxer"? looks like a hang.

Run the agent itself in the sandbox

Put the directory from boxer shim install --harness claude,codex first on herdr's PATH. Then herdr agent start --kind claude runs boxer shell claude.

The repository has a herdr plugin in adapters/herdr that adds a pane action for boxer shell claude and warms the sandbox when herdr creates a worktree:

herdr plugin link adapters/herdr

A sandboxed shell in every pane

boxer shim install --shell writes boxer-bash, a shell whose commands all run in the sandbox. Set herdr's terminal.default_shell to it for panes you type in. herdr cannot attach its own agent to such a pane, because it identifies a pane by its foreground process, which is then boxer.

Tested in the matrix.