The LocalShellBackent Script should use hard rules to load the correct path.

[Smile0524]

When the python script is executed in the localshell, the command in the execute of the localshellbackend often generates an incorrect path. Because the absolute path cannot be obtained during the deployment，so i can not specify an absolute path in the code.Relying solely on the llm to ensure the correctness of the script execution path is extremely unstable.

[Eugene Yurtsev (eyurtsev)]

Hi Smile0524,

You can introduce custom middleware to add information to the system prompt about the current working directory + any other relevant context. We do this in the deepagents-cli (you can check out the LocalContextMiddleware). it adds a bunch of useful context to make it easy for the agent to work on coding tasks.

Because the absolute path cannot be obtained during the deployment，so i can not specify an absolute path in the code.

1. You should be able to obtain the path using the strategy outlined above (i.e., add some logic in before_agent in custom middleware... or even just add information to the system prompt without any middleware).
2. In general, I'd be very careful deploying an agent that uses LocalShell. That shell would presumably allow a prompt injection attack to look up secrets or kill the agent process. (There are also other issues e.g., if you have multiple instances of the web server). Instead, we'd recommend using a Sandbox

Hope this helps!