has anyone experimented with giving agents something like a Store comonad that's generated from a jj workspace, and a few RPC commands for reading/transforming this, where write commands return a new Store?
seems like giving agents regular shell access is an antipattern, it's way too unstructured and prone to escape/globbing errors. obviously you can still expose arbitrary commands through a run() command that returns a tuple (exitcode,stdout,stderr)
i might build this for myself but if anyone has a working harness that goes in this direction, that would obviously be better