AutoGPT
AutoGPT copied to clipboard
Significant-Gravitas
Executing python code directly from AI output
I've encountered the same problem as #101. GPT Thinks that evaluate_code will execute python:
CODEREVIEW-GPT THOUGHTS: Let's find some code to review and suggest improvements.
REASONING: We need to search for files with common code extensions that are not too large and more recent than a certain date.
PLAN:
- Use the 'evaluate_code' command to run a Python script that searches for files with common code extensions in the current folder and subfolders.
- Use the 'memory_add' command to save the file paths for later reference and reuse.
- Review the saved files using 'read_file' command to ensure they are indeed code files.
- Choose a file to review and suggest improvements.
CRITICISM: We may need to adjust the search criteria if we don't find relevant code or if too many files match the criteria.
NEXT ACTION: COMMAND = evaluate_code ARGUMENTS = {'code': 'import pathlib\nfrom typing import List\n\ndef find_code_files(root: str, extensions: List[str], max_size: int, min_date_modified: float) -> List[str]:\n """\n Returns a list of file paths for code files with the specified extensions, maximum size, and minimum date modified\n rooted in the given directory and its subdirectories.\n """\n files = [str(path) for path in pathlib.Path(root).rglob("*") if path.suffix in extensions\n and path.stat().st_size <= max_size\n and path.stat().st_mtime >= min_date_modified]\n return files\n\ndef print_files(files: List[str])
-> None:\n """\n Prints the list of file paths to the console, one file per line.\n If no relevant files are found, prints a message indicating so.\n """\n for file in files:\n print(file)\n else:\n print("No
relevant files found.")\n\n\ndef analyze_code(code: str) -> None:\n """\n Analyzes the given code and provides suggestions for improvements.\n """\n pass\n\nif __name__ == \'__main__\':\n root: str = \'.\'\n extensions: List[str] = [\'.py\', \'.js\', \'.java\']\n max_size: int = 10000 # bytes\n min_date_modified: float = 1609459200 # seconds since epoch (01/01/2021)\n\n files: List[str] = find_code_files(root, extensions, max_size, min_date_modified)\n print_files(files)\n'}
On one hand, that's a bug, which is addressed on #101.
On the other hand... that's a very interesting idea from GPT. Perhaps giving it the ability to execute python code could allow it to execute a lot of tasks in a way more dynamic way.
Is the file argument a path or actual code?
The feature I'm proposing is the direct execution of code. "file" makes it seem like it's a path to a file already in the system.
Same problem here too :
NEXT ACTION: COMMAND = execute_python_file ARGUMENTS = {'file': '<path_to_python_script>'}
Executing file '<path_to_python_script>' in workspace 'auto_gpt_workspace'
SYSTEM: Command execute_python_file returned: Error: Invalid file type. Only .py files are allowed.
I asked it to create a python script and it just try to execute "path_to_python_script". Any solution ?
Not a duplicate. This is a feature request for direct python code execution.
As execute_python_file? I'm not sure if that's the same thing as I stated on my original message.
As I said:
Is the file argument a path or actual code?
The feature I'm proposing is the direct execution of code. "file" makes it seem like it's a path to a file already in the system.
If you want to close as "won't do", that's okay. But I don't think it's a duplicate.
Ah, thanks for elaborating! I think this is something we could add without too much effort. The tricky thing is to properly sandbox it, in a way equivalent to execute_python_file.
this has more to do with in-memory execution of code that isn't written to disk, I suppose ? If so, that's related to the API discovery idea #56
Yeah, it is somewhat related.
I think this issue might supersede that one/that issue might supersede this one depending on how it's implemented. Although this one is a bit more general (e.g. the Agent might spin up a python instance just to do some calculations, so nothing necessarily related to an API).
Edit: As a matter of fact, now that I think about it, I think these should be separate tasks. Meaning, a search_for_api task, followed by a write_python_code task, which would use that knowledge. That would keep the system more general, but fulfil the capabilities of #56.
search_for_api would be a specialization of a do_research (crawl) command, where as api could be either a classical API or a networking API
Some of us have succeeded getting Agent-GPT to write code by exploring API docs already - my recent experiments made it download the github API docs and come up with a CLI tool to filter PRs based on excluding those that are touching the same paths/files: https://github.com/Significant-Gravitas/Auto-GPT/compare/master...Boostrix:Auto-GPT:topic/PRHelper.py
While this is trivial in nature, it can already be pretty helpful to identify PRs that can be easily reviewed/integrated, because they're not stepping on anyone's toes. And it would be easy to extend as well.
The point being, having some sort of search API / extend yourself mechanism is exactly what many folks here are suggesting when it comes to "self-improving", in its simplest form: adding features without having to write much/any code.
So, thinking about it, I am inclined to think that commands should be based on classes that can be extended - a research command would be based on a crawler/spider class [http requests: #2730), and a find_api command would be based on the research command [class]
That way, you can have your cake and eat it, while also ensuring that the underlying functionality (searching/exploring the solution space), is available for other use-cases - like the idea of hooking up the agent to a research paper server (#826 ) or making it process pdf files (#1353 )
Commands in their current form worked, but to support scaling and reduce code rot, it would make sense to identify overlapping functionality and then use a layered approach for common building blocks.
The "API explorer" you mentioned could also be API based itself, so there is no need to go through HTML scraping - but some folks may need exactly that, so a scraping mechanism would be a higher-level implementation of a crawler #2730
Related talks collated here: https://github.com/Significant-Gravitas/Auto-GPT/issues/514#issuecomment-1534224349
This issue was closed automatically because it has been stale for 10 days with no activity.
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}