Current state of Stable branch; Docker images.

[GECORegulatory]

Current State of AutoGPT Stable Branch and Docker Images

Introduction

As some of you may be aware, AutoGPT is currently undergoing a major architectural overhaul. Unfortunately, this means that bug reports are starting to pile up, and the state of our stable branch and Docker images is not as robust as we would like. I'm making a personal commitment to help remedy this situation, but I could definitely use some assistance from the community.

Current task

Figure out which bugs exist in the stable branch
Find the new highest-severity bug
Track the presence of bugs until the bug count lowers, and until there is little discrepancy between the official docker image and one built from source

Workaround

There is a large discrepancy between bugs in the stable branch, and in the docker images. Because of this, I recommend building your own images from source. I've provided a script for this. Please note that you must first clone the repository, cd and edit your .env with your API key. You cannot build the image without setting it.

Script (bash/zsh)

cp .env.template .env
vim .env
#Edit your .env with your API keys
docker-compose up --build
docker build -t my-autogpt-image .
docker run -it --rm my-autogpt-image

Bugs Overview

Here's a brief rundown of the most pressing issues. By not reproducible, I mean that I wasn't able to reproduce it or find a procedure to reproduce it.:

High Severity Bugs

• Pasting source code into the human prompt results in a crash #4956: There's a potential high-severity crash involving a bad API parameter (max_tokens) sent to the OpenAI API. It's not reproducible and needs more info. This is potentially the current highest severity bug report.
• AutoGPT crashes when processing a prompt related to reading a file #4966: This issue involves severe undefined behavior from the read_file command (an issue I personally reported). It's also reproducible. Workaround currently solves this.

Medium Severity Bugs

• Utilities.py NoneType object has no attribute 'startswith'  #4871: Fatal parsing error with backticks; this is reproducible and the highest severity bug we have right now. Testingasrequested #4986 is working to fix this. See workaround
• Plugin config is invalid, continuing without plugins. Error: stat: path should be string, bytes, os.PathLike or integer, not NoneType #4964: This is a medium severity issue with parsing the plugins file. It has not enough information and I have not personally tested it, but it appears reproducible. is reproducible, usually without any special procedure.
• Agent trying to use goals_accomplished, can't steer away from command despite agent acknowledging goals not accomplished #4951: There's an issue with the model attempting to use deprecated methods.
• browse_website: Config needs to be passed to nested summarize calls #4949: This issue pertains to a NoneType error that needs a PR to fix. The issue is already traced, this is a great start to anyone to jump in on. Reproducability is unclear

Low Severity Bugs

• Unable to use AutoGPT with voice #4926: Low severity bug with ElevenLabs API; potential re-emergence of old bug --Speak is broken TypeError: say_text() missing 1 required positional argument: 'config' #4794. It needs more info and is not reproducible. A potential workaround exists in i think we need this #4934 involving editing config.py manually, but the solution lacks detail.
• Unknown Command "speak"  #4897: Low severity bug affecting TTS.
• SYSTEM: Command execute_python_file returned: sh: npx: not found sh: npm: not found #4963: Low severity potential junk bug report. The user likely does not have npm environment set up correctly, but it's worth checking the Docker image.
• Run Auto-GPT Error  #4800: Non-reproducible 'module not found' error.
• There are several low quality reports, such as gpt running files and stuff #4967, ImportError: Bad git executable. #4939, Auto-GPT does not work #4920, and Agent kept running command not exist #4886. These are either lacking in information or are not particularly constructive. (Edited due to typo)

Fixed

• Replaced Fictitious color name Fore.ORANGE #4972: My PR regarding a nonexistent color (ORANGE) in Colorama; it might be worth investigating how this didn't fail a test, since it causes a fatal error if run.sh is run without having an OpenAI API key set up.
• Speak mode causing crash on Mac #4923: This seems to be a potential OSX-specific crash related to ElevenLabs; it doesn't seem to be a dupe of Unable to use AutoGPT with voice #4926 but worth checking. As I do not own a MacBook, I cannot reproduce it, and we would appreciate help from a MacOS user.

Moving Forward

I'd like to propose the following steps to improve our situation:

1. Allow self-assigning of bugs and severity reports.
2. Clearly mark a finite number of issues as the most severe for visibility.
3. Implement GitHub Actions to filter out low-quality reports that don't follow the proper format.
4. Automatically close stale issues to keep the queue manageable.
5. Review our testing procedures to ensure comprehensive coverage.
6. Revise constructive bug report procedures to help users provide all necessary details.
7. Take a holistic look at bug fixing, from discovery, triage, fix, testing all the way patch deployment, to see if this can be improved. It is sad to see so many bug reports and so few replies to them other than "me too".

As a starting point, I will begin working on the most severe issue (#4871). Any help would be greatly appreciated!

[Yaski]

Hello, I don't understand a little bit

Utilities.py NoneType object has no attribute 'startswith' #4871: Fatal parsing error with backticks; this is reproducible and the highest severity bug we have right now. #4986 is working to fix this. See workaround

Is workaround to build autogpt from source? Should I use stable or master branch? I tried both and still got this error.

docker-compose up --build

This command doesn't work for me. I used docker-compose run --build --rm auto-gpt

[Yaski]

docker-compose up --build auto-gpt

This command worked