Ask to run code after improvement

[UmerHA]

Unlike after code creation, after code improvement(i.e., the -i flag) gtp-engineer doesn't ask if the code should be run.

Feature description

When using the -i flag and gpt-engineer has finished improving code, the use should be asked if they want to run the code.

Motivation/Application

In my view, a good workflow for gpt-engineer would be:

1. I enter a prompt, gpte asks clarifying questions, I answer, gpte creates a codebase
2. With my consent, gpte runs the code
3. I get asked if the code worked well. If not, I'm asked if I want to improve.
4. I enter an improvement prompt
5. gpte improves code
6. With my consent, gpte runs the code <- this issue
7. Repeat

[mahimairaja]

So you wish to add improvement on interactive mode? Am I right?

[UmerHA]

So you wish to add improvement on interactive mode? Am I right?

Yes, I think I'd be good if that existed. Not sure if I can contribute though.

[UmerHA]

fyi: requires #721 to be merged

[Ruijian-Zha]

I concur with @UmerHA's points. Additionally, automating the environment setup to handle dependencies would be beneficial, saving users from manual installations, especially when new dependencies are introduced post-improvement. This would streamline the process, making the -i flag for code improvement more user-friendly and efficient.

[ATheorell]

@Ruijian-Zha I'm not sure if I understand. If successful, the generated run.sh script should create a virtual environment and install dependencies. Hypothetically, by including the run.sh/requirement.txt/whatever in the improvement prompt, the agent should already be able to suggest addition of new dependencies. I may be missing something here.

[Ruijian-Zha]

@ATheorell The key idea is to run run.sh, requirements.txt, or similar scripts iteratively until the problem is resolved. The recently released Autogen can automatically install the necessary requirements in the current Conda sub-environment, aligning perfectly with our objectives. For more details, including my test code and results, you can visit my personal GitHub page: Autogen Direct Run & Iterative Debug.