I've previously used codex quite a bit (which also offers a bwrap based sandbox). I enabled it for dirge too expecting it to work in a similar way. However it seems like the sandbox is very rigid and it's not possible for the agent to ask for more permissions.
As an example I asked the model to get my WAN IP for which it used curl. As far as I understand, because of the strict --unshare-all there is no way for the agent to request additional permissions when operations fail.
I patched dirge to allow network within the bwrap sandbox and asked my agent to migrate a Python project from poetry to uv. It could use websearch to find information, but it couldn't run uv lock as the last step to create the lock file as that requires write access to ~/.cache/uv where uv caches package downloads.
I think it would be nice to allow the agent to request running outside the sandbox and/or additional file-system write permissions/network access for specific tool calls when running with bwrap sandboxing.
I've previously used codex quite a bit (which also offers a bwrap based sandbox). I enabled it for
dirgetoo expecting it to work in a similar way. However it seems like the sandbox is very rigid and it's not possible for the agent to ask for more permissions.As an example I asked the model to get my WAN IP for which it used curl. As far as I understand, because of the strict
--unshare-allthere is no way for the agent to request additional permissions when operations fail.I patched dirge to allow network within the bwrap sandbox and asked my agent to migrate a Python project from
poetrytouv. It could usewebsearchto find information, but it couldn't runuv lockas the last step to create the lock file as that requires write access to~/.cache/uvwhereuvcaches package downloads.I think it would be nice to allow the agent to request running outside the sandbox and/or additional file-system write permissions/network access for specific tool calls when running with bwrap sandboxing.