Piping output of a suggested command back to GPT for improved iterations

[adityakavalur]

Just want to make sure if the below is currently possible (and if not, if it is something that is in the planned feature list)

In chat/REPL mode, it seems like providing information is restricted to stdin and user prompts. There is no way to pipe the output of a command to the next user prompt, or have it be a prompt of its own. This becomes useful when the command/code snippet suggested by GPT is incorrect. Easily providing this stdout/stderr to GPT would speed up the development cycle.

(venv) [kavalurav@zfocus22 gpt]$ sgpt --repl temp5 --shell 
Entering shell REPL mode, type [e] to execute commands or [d] to describe the commands, press Ctrl+C to exit.
>>> can you compile hello.c and execute it?
gcc hello.c -o hello && ./hello
>>> e

hello.c:1:10: fatal error: mpi.h: No such file or directory
 #include <mpi.h>
          ^~~~~~~
compilation terminated.

>>> how do I fix the compiler error?
Without knowing the specific error, it's impossible to provide a specific solution. However, you can use the following command to compile your code and output the errors to a text file for further analysis:

gcc hello.c -o hello 2> errors.txt

In the above error if GPT was able to see the mpi.h error it would help out a lot. I get that it can become expensive to add tokens arbitrarily, an alternative would be to provide a scratch file that gets prepended every time as context? So the compile command can pipe the output to a file sgpt.scratch that would then get appended to the follow up question.

[adityakavalur]

Duplicate of #337 and #416