Allow AI to see output of the commands in repl mode

[moritz-t-w]

I think it would be very helpful to at least have the option to let the ai see the command output when executing, maybe instead of e you'd say o or something.

[endolith]

Yeah it's really infuriating because it seems to see some of the context of the conversation but not other parts.

So the issue is that it can see what I've asked before (purple) but not its own output (white)?

[moritz-t-w]

Yes that seems to be the case

[hcal]

I think it's a fantastic idea. I'd like to build upon this suggestion with a slight modification that could offer even more flexibility and control to the user.

Instead of having a different command like 'o' at the time of execution, my suggestion is to allow the user to prepend a command to the beginning of the next chat line. I suggest !! or something similar. This would indicate that we want to include the output of the previously run command in the chat text. Here’s how it might work:

Simple Use Case Example:

1. User asks "what are the details of the files this directory?"
2. Shell-GPT asks to execute a command: ls -lh
3. User types: e
4. System executes and shows output.
5. User then asks: !! What is the big file?
6. Shell-GPT includes the output of ls -lh in its context for responding.

Advantages:
Selective Inclusion: This approach avoids overwhelming the chat history with unwanted command outputs and keeps the interaction clean.
Ease of Use: It's intuitive and straightforward, requiring minimal changes to the current interaction model.
Enhanced User Experience: It provides a direct way to reference the output of commands, facilitating better assistance, especially in troubleshooting scenarios.

Is this the place to post a feature request or should it be posted in issues?

[moritz-t-w]

Just not !! because that already has a clearly defined functionality, which is to include the previous command, not its output.

[moritz-t-w]

This is the right place for feature requests, see pinned issue.

[moritz-t-w]

Shell-GPT includes the output of ls -lh in its context for responding.

@TheR1D I think it would still be nice to get the actual output of the command, so you can be sure chatgpt didn't make a mistake when reading it. But maybe that's a bit too coomplicated for a simple REPL interface. Maybe it could write the command outputs to a separate file that the user could tail on the side if they wanna check it.

[endolith]

@TheR1D Please implement something like this. I want a REPL-like thing that I can talk to, ask it questions in plain language and get responses in plain language, but it can also suggest commands and explain what they do, and then I can optionally execute them, and then it can see the command being executed and its output as part of the conversation context, to see if it worked correctly, suggest a new command that builds on the output of the previous, etc.

Currently it can suggest commands, and I can run them, but it can't see if I execute them or not, can't see the command being run, or the output of those commands, making it not very useful:

~/.c/shell_gpt> sgpt --repl temp --shell
Entering shell REPL mode, type [e] to execute commands or [d] to describe the commands, press Ctrl+C to exit.
>>> list files
ls
>>> e

functions  roles

─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
>>> what are the files?
ls -l
>>> no I'm asking you in english, what files did it output?
ls -l | awk '{print $9}'
>>> don't respond with a command, respond in english
The command output a list of files in the current directory.
>>> ok but which files?  do you see the output?
I apologize for the confusion. Without the specific output, I am unable to provide the exact names of the files listed.
>>> return the entire conversation so far
[… arguing with it …]
>>> don't respond as a command ,respond in plain text, showing the conversation starting from 'list files'
I apologize for the confusion. Here is the conversation starting from 'list files':

User: list files
Assistant: ls
User: what are the files?
Assistant: ls -l
User: no I'm asking you in english, what files did it output?
Assistant: ls -l | awk '{print $9}'
User: don't respond with a command, respond in english
Assistant: The command output a list of files in the current directory.
User: ok but which files? do you see the output?
Assistant: I apologize for the confusion. Without the specific output, I am unable to provide the exact names of the files listed.
User: return the entire conversation so far
>>>

[TheR1D]

ShellGPT can do that using function calling.

sgpt --install-functions

sgpt --repl temp "You are allowed to run shell commands only after my confirmation."
Entering REPL mode, press Ctrl+C to exit.

>>> Can you echo "hello world"?
Yes, I can do that. Please confirm if you would like me to proceed with executing the shell command to echo "hello world".

>>> run it
▌ @FunctionCall execute_shell_command(shell_command="echo "hello world"")
The command was executed successfully, and the output is:
 hello world

>>> what is in my home dir?
Please confirm if you would like me to execute the shell command to list the contents of your home directory.

>>> run it
▌ @FunctionCall execute_shell_command(shell_command="ls ~")
The contents of your home directory are:
 Applications
 Desktop
 Developer
 ...

See #416

[moritz-t-w]

ShellGPT can do that using function calling.

I mean that's pretty much all we're asking for. @endolith if you're happy with this imma close this discussion

[moritz-t-w]

@endolith:

>>> what are the files?
ls -l
>>> no I'm asking you in english, what files did it output?

Pro tip: I modified my prompt so that it translates my prompts to commands except if I add a question mark, in which case it answers the question. This is how I wrote it for me:

Provide only zsh commands for Linux/Arch Linux without any description except if the prompt ends in a question mark. If so, answer the question.
If there is a lack of details, provide most logical solution.
Ensure the output is a valid shell command.
If multiple steps required try to combine them together.

[endolith]

ShellGPT can do that using function calling.

From that example, it asks you permission to create a function and then calls it? So there's no way to vet the function before executing it? That seems backwards. It should just propose a function (without asking), explain what the function is meant to do, and then you should be able to confirm that you want it to actually execute or not.

Also can the AI see the output of the function?

[moritz-t-w]

@endolith:

From that example, it asks you permission to create a function and then calls it? So there's no way to vet the function before executing it?

this isn't entirely right, the AI calls an existing function, the function that it happens to be calling here then executes a command. But yes you can't vet that command. You could make your own function tho.
@endolith:

Also can the AI see the output of the function?

@TheR1D correct me if I'm wrong but I think the AI gets the output from the command and generates its response based on that.