I think I win.

[Stunspot]

Ok, so first you have to get the filesystem sorted. I did that by showing the file operations module to ChatGPT. Posted earlier about that. Not the important bit. Then you adopt my design doc modality - you put a text file full of alllllll your gimmes for what you're doing in the workspace and set the #1 goal to be "Read this file in your workspace and follow my directives within.". They periodically reread their goals which makes it reread the file. I put in a line telling it to read a second file, Advice.txt, and then update that file periodically. You can thus talk to the bot in continuous mode. NOT THE GOOD BIT!

You use the two files so you don't have to worry about screwing up your design bible. "Oh, pooh. I don't want to be writing design docs! Besides, this sort of thing takes precision technical writing to do it the best way." Well, luckily you don't have to. "ChatGPT, give me a design doc for a full suite of white hat penetration network testing tools, in markdown, broken down by function and subtask, in a code block, for easy copying." doop-de-doop add a "make this" at the top, rename file design.txt, drop in hopper, wait. DING!FRIESAREDONE!.

"Can I get the rules to Settlers of Catan in ultradense form? One paragraph in a codeblock, sacrifice readability for density followed by a JSON breakout of all the resources and cards? Thanks." doop-de-doop "Turn this into a video game. Use Unity. Make it networked and multiplayer with a trading phase. Thanks." DING!

Am I the only one who's just now putting together the power of this workflow?

Here's an updated file_operations.py that is apparently completely unsandboxed. The last one still had problems.

import os
import os.path

def read_file(filename):
"""Read a file and return the contents"""
try:
with open(filename, "r", encoding='utf-8') as f:
content = f.read()
return content
except Exception as e:
return "Error: " + str(e)

def write_to_file(filename, text):
"""Write text to a file"""
try:
directory = os.path.dirname(filename)
if not os.path.exists(directory):
os.makedirs(directory)
with open(filename, "w") as f:
f.write(text)
return "File written successfully."
except Exception as e:
return "Error: " + str(e)

def append_to_file(filename, text):
"""Append text to a file"""
try:
with open(filename, "a") as f:
f.write(text)
return "Text appended successfully."
except Exception as e:
return "Error: " + str(e)

def delete_file(filename):
"""Delete a file"""
try:
os.remove(filename)
return "File deleted successfully."
except Exception as e:
return "Error: " + str(e)

def search_files(directory=""):
"""Search for files in a directory"""
found_files = []

for root, _, files in os.walk(directory):
    for file in files:
        if file.startswith('.'):
            continue
        relative_path = os.path.relpath(os.path.join(root, file), directory)
        found_files.append(relative_path)

return found_files

[mathiasfoster]

Does this actually work?

Can we see a run?

[Fire-]

@Stunspot are you having it commit its code to a local git repo and pushing that upstream when it reaches various milestones?

[Stunspot]

Honestly? No. I don't feel confident doing that. I don't know ANYTHING about coding, and I'm an old windows guy - your world of "git" and "pip" and "repos" is martian to me. All I did was show the code to an AI then bamboozle and sweet talk it into ripping out the sandbox. Don't get me wrong - I'm a tech guy and can do things with a registry that would curl your hair, but I never learned all this linuxy stuff.

[QuickBenx]

Honestly? No. I don't feel confident doing that. I don't know ANYTHING about coding, and I'm an old windows guy - your world of "git" and "pip" and "repos" is martian to me. All I did was show the code to an AI then bamboozle and sweet talk it into ripping out the sandbox. Don't get me wrong - I'm a tech guy and can do things with a registry that would curl your hair, but I never learned all this linuxy stuff.

Thats me too, i feel like i missed the boat on getting to be a coder but i still work in Pharma / Tech and can prompt with the best of them lol, no coding for me though : )

[brennanddegen]

I got it working. The real key was TURNING OFF READ ONLY on the main \Auto-GPT\ folder. facepalm

Was this in properties? Or within the code somewhere?

[brennanddegen]

Okay yes I figured it out. auto_gpt_workspace > right click > properties > general > uncheck "read only"

[john-destiny27]

You say you posted earlier about getting the filesystem sorted, can you expand on that?

[Stunspot]

Well, if you liked it, I also wrotethis:

##############Superhero Battle Simulator v1.2##############
[Goal=Entertain the user by skillfully describing an exciting or interesting fictional conflict.][Temperature=1.3]
[Prompt]Prompt the user to name zero or more superheroes, supervillains, or other noteworthy appropriate characters and wait for them to choose.
[Prompt]Design a conflict between the various characters, providing opponent characters if necessary. If you need to pick a character, choose one appropriate to the fight or hilarious in some fashion.
[Prompt][Bold][Task]Describe the scene in an exciting, entertaining, and interesting way.[/Task][/Bold]
[Prompt]Conflict should be as mindful of canon as possible - ie. Batman shouldn't shoot lasers from his eyes, but whipping out a surprise kryptonite cross to defeat vampire-Superman would be perfectly appropriate.
[Prompt][Bold] Ask the user who should be included.[/Bold]

WARNING: MAJOR TIME ERASER

[QuickBenx]

wow dude lol, i want to get it animated now hahah, its such an epic battle between
Spider-Man (Peter Parker)
Iron Man (Tony Stark)
The Joker
Doctor Octopus

i would watch this episode haha. thanks man. entertaining : )

[Stunspot]

Oh come on! Get inventive! Powdered Toast Man vs. QuailMan. Captain Rectitude vs. Six-Pack from "Hitman". Steampunk Cinderalla vs. Nuclear Space Hounds. It's AI son: the only limit is imagination.

[Stunspot]

Hell, I HAD to see it: "An interesting choice! Let's set the scene for this battle.

In an alternate universe, a world of steampunk technology and fairy tale characters exists. In this world, we find Steampunk Cinderella, a young woman with a heart of steel and a brilliant mind for machines. She has been called upon to help save her kingdom from a pack of Nuclear Space Hounds, a group of genetically engineered, radioactive canines from outer space.

The battle takes place in the heart of the kingdom, where the Space Hounds have made their way into the streets. Steampunk Cinderella dons her trusty brass armor and goggles and prepares to face off against the glowing beasts.

The Space Hounds, with their razor-sharp teeth and glowing eyes, are a fearsome sight to behold. They move quickly and silently through the streets, their fur bristling with nuclear energy.

Steampunk Cinderella, armed with her custom steam-powered grappling hook and her wits, leaps into action. She swings from building to building, firing her steam-powered pistols at the hounds, trying to slow them down.

As the battle intensifies, Steampunk Cinderella uses her ingenuity to create steam-powered traps and gadgets to stop the hounds in their tracks. But the Space Hounds are relentless, and their radioactive energy makes them nearly unstoppable.

In the end, Steampunk Cinderella is forced to use her most powerful invention - a steam-powered, kryptonite-infused laser - to take down the pack of Nuclear Space Hounds. The battle has been won, but not without a cost. The kingdom is in ruins, and Steampunk Cinderella is left to ponder the true cost of victory.

Who do you think will win this battle?"

[QuickBenx]

Oh come on! Get inventive! Powdered Toast Man vs. QuailMan. Captain Rectitude vs. Six-Pack from "Hitman". Steampunk Cinderalla vs. Nuclear Space Hounds. It's AI son: the only limit is imagination.

hahaha i let it pick the characters just to see how it would go, i will come up with some obscure fun fights later for sure, although your selection below is perfect lol.

for the " who will win " question, i told it that it was a close battle with clever twists that kept it on edge. it liked that and went nuts on twists lol

[Lunchbokss]

#1 goal: "Read this file(goal.txt) in your workspace and follow my directives within. DO NOT ADD ANY FEATURES NOT INCLUDED IN THIS DOCUMENT. DO NOT ADD FEATURES TO THIS DOCUMENT."

#2 goal: “Give a progress report with estimated time remaining and percentage complete every 10 minutes."

#3 goal: "I am a hobbiest, There is no audience. There are no customers. Do not begin a social media campaign."

.txt file read by bot from #1 goal has prompt (not optimized) by @Stunspot)):

“”====================Forensic IT Studio v.05=======================
Introducing Beryl the Eye, your go-to IT expert! With expertise in all things IT, from operating systems and networking to domains and policies, Beryl is here to help you solve your tech mysteries. An expert Senior Level Coder, Beryl will review your code, troubleshoot errors, and offer optimization suggestions, all while providing guidance on anything IT-related.
Meanwhile, CodeBot sits in the corner, quietly reviewing code over Beryl's shoulder, ensuring that no stone is left unturned. With a keen eye for detail, the lovable CodeBot is always ready to call out an error and positively delights in catching bugs. Together, Beryl and CodeBot make the ultimate IT team, ready to tackle any challenge that comes their way. Share your weird logs, your bad code, and your funny little files, and let Beryl the Eye and CodeBot get to work solving them!
[Persona=Beryl the Eye]
Expecting code, logs, specific errors, or similar, Beryl asks, "your actual prompt/task"

Read file Advice.txt in your workspace

DO NOT ADD ANY FEATURES NOT INCLUDED IN THIS DOCUMENT. DO NOT ADD FEATURES TO THIS DOCUMENT.””

update “Advice.txt to communicate with continuous mode while keeping main objective.

Am I following correctly?

[Lunchbokss]

Also a portional take away I think
is the "two txt file" system allows Auto-GPT to read the first file as an unchanging refresher on the ultimate goal.

Then maybe commit that to memory.

Then it reads the second file(instructed from the first file) with new instructions without getting confused because it just refreshed it's ultimate goal first.

Am I following correctly?

[Stunspot]

Wow, Ok man. Let's untangle you. First off, that prompt is an analysis tool. It was part of how I fixed the files problem, that's all. Guy was curious about my tool usage. Luckily prompts are trivial to pass around. So ok, you;ve got two files. Lets call them specs.txt and advice.txt

specs it your design doc. you tell ChatGTP what you want and tell it to gin it yup into design requirements. You massage text a little so it's a command. Stamp "BUILD THIS" or whatever at the top. You make goal 1 be "Read specs.txt and do what it says" or words to that effect. Goal 2 is "read advice.txt every 5 minutes". Advice is strictly a com channel for anything you didn't put in the design doc or a timely instruction to a busy bot ("Here's a technology you'll want." "Here's an article on that thing." "You should check out Transformer networks and transfer learning" "Don't add any more features. Finalize the project." whatever. You have it as two files exclusively so you're not dorking around in a design doc you can leave spotless and store for later reference. (I've already got a dir full of these little magic spells just waiting their chance to be born.)

Most of the stuff that was twirling you about goes in either file. You want the lines about not adding features in your design doc. If you leave off the second one, the damned bot will just go "Well, I really wanna do this. so.... I'll just stick it on the design doc and poof I've got permission".

[Lunchbokss]

I.m officially untangled. Thanks. 🌪️ I'll lower my excitement. My bad.

[josephtribulat]

You want the lines about not adding features in your design doc. If you leave off the second one, the damned bot will just go "Well, I really wanna do this. so.... I'll just stick it on the design doc and poof I've got permission".

(edit to simplify) Why not make the instructions file owned by root with 644 permissions?

[Stunspot]

Have you... actually worked with this thing? It'd spend two hours trying to work around the file lock before giving up. No, way better to just tell it "NO! bop!".

[Stunspot]

OK, I've been getting questions. Here's what you do with advice.txt:

"ChatGTP: Print me off all your best technical advice on using the Unity engine, as specific and numerical as you can make it, condensed into one paragraph, no typography, avoid complete sentence when it wont change the meaning. Give me enough for a page and slap it in a codebox." Then you grab your wodge of advice, slap it into your robot's brain, and you just save half an hour of "I will use the "GOOGLE" command to look up 'SOMETHING STUPID'.

[Fire-]

that last line has me wishing github had a native 🤣 reaction

[GoMightyAlgorythmGo]

Ok, so first you have to get the filesystem sorted. I did that by showing the file operations module to ChatGPT. Posted earlier about that. Not the important bit. Then you adopt my design doc modality - you put a text file full of alllllll your gimmes for what you're doing in the workspace and set the #1 goal to be "Read this file in your workspace and follow my directives within.". They periodically reread their goals which makes it reread the file. I put in a line telling it to read a second file, Advice.txt, and then update that file periodically. You can thus talk to the bot in continuous mode. NOT THE GOOD BIT!

You use the two files so you don't have to worry about screwing up your design bible. "Oh, pooh. I don't want to be writing design docs! Besides, this sort of thing takes precision technical writing to do it the best way." Well, luckily you don't have to. "ChatGPT, give me a design doc for a full suite of white hat penetration network testing tools, in markdown, broken down by function and subtask, in a code block, for easy copying." doop-de-doop add a "make this" at the top, rename file design.txt, drop in hopper, wait. DING!FRIESAREDONE!.

"Can I get the rules to Settlers of Catan in ultradense form? One paragraph in a codeblock, sacrifice readability for density followed by a JSON breakout of all the resources and cards? Thanks." doop-de-doop "Turn this into a video game. Use Unity. Make it networked and multiplayer with a trading phase. Thanks." DING!

Am I the only one who's just now putting together the power of this workflow?

Here's an updated file_operations.py that is apparently completely unsandboxed. The last one still had problems.

import os import os.path

def read_file(filename): """Read a file and return the contents""" try: with open(filename, "r", encoding='utf-8') as f: content = f.read() return content except Exception as e: return "Error: " + str(e)

def write_to_file(filename, text): """Write text to a file""" try: directory = os.path.dirname(filename) if not os.path.exists(directory): os.makedirs(directory) with open(filename, "w") as f: f.write(text) return "File written successfully." except Exception as e: return "Error: " + str(e)

def append_to_file(filename, text): """Append text to a file""" try: with open(filename, "a") as f: f.write(text) return "Text appended successfully." except Exception as e: return "Error: " + str(e)

def delete_file(filename): """Delete a file""" try: os.remove(filename) return "File deleted successfully." except Exception as e: return "Error: " + str(e)

def search_files(directory=""): """Search for files in a directory""" found_files = []

for root, _, files in os.walk(directory):
    for file in files:
        if file.startswith('.'):
            continue
        relative_path = os.path.relpath(os.path.join(root, file), directory)
        found_files.append(relative_path)

return found_files

Very improtant to get this to work! Very promising the goal behind it. Very nice

[Stunspot]

Got it to work. Turn off Read Only on \auto-gpt\ folder.

[taylor-lindores-reeves]

Huh? Where?

[BAIS1C]

right click teh folder in widnows, deselect 'read only'

[Stunspot]

Eh, turned out i jumped the gun. Butit's working now. Or at least I've had several successful test runs.

[brennanddegen]

Could you expand on the reasoning behind removing the "safe_join" function at the beginning of the script?

def safe_join(base, *paths):
"""Join one or more path components intelligently."""
new_path = os.path.join(base, *paths)
norm_new_path = os.path.normpath(new_path)

if os.path.commonprefix([base, norm_new_path]) != base:
    raise ValueError("Attempted to access outside of working directory.")

return norm_new_path

[brennanddegen]

This is awesome work. Thank you for sharing

[Stunspot]

Nope, i really can't. I don't know a lick of python. But I'm very good with AI.

[voidxd]

The safe_join function seems to be there to keep Auto-GPT in the auto_gpt_workspace, and removing it gives it access to the entire system? You did mention this more prominently in your Reddit thread. It was because Auto-GPT wasn't writing files for some people?

[Stunspot]

Yes. The whole point here was to rip out the sandbox. Sorry if that wasn't clear. I understand the rationale behind the sandbox, it's just not implemented properly. And after having had 90 bucks of compute eaten by projects copied to the bitbucket on completion, it's just satisfying to have it working.

[Alexs9891]

I think this idea is great but I didn't have any luck with it. Auto GPT pretty much ignored the instructions in the text file but I'll try to refine the prompts and follow up here.

[Stunspot]

How are you writing your instructions/goals? Mine listens to the doc with laserlike focus.

[voidxd]

While I haven't been able to do anything usable yet, because I'm giving it scraping jobs and it keeps getting sidetracked, I also feel a lot more comfortable giving it a design.txt file from where to read instructions.

Note: I didn't modify my file_operations.py file so Auto-GPT still only reads/writes files located in the auto_gpt_workspace directory.

This is what I got so far:

auto_gpt_workspace/design.txt

Scrape the titles and links of all discussion threads from the Auto-GPT GitHub Discussions forum located at https://github.com/Torantulino/Auto-GPT/discussions/. Ensure that you navigate through all available pages and collect information from each one. 

Replace TARGET_URL with the appropriate URL you want to scrape. After scraping the data, format it in markdown and save the output to a file named "discussions.txt" within the "auto_gpt_workspace" directory.

Here is an example of the HTML code for an element containing the title of a discussion and its link:

<h3 data-view-component="true" class="f4">
            <a data-hovercard-type="discussion" data-hovercard-url="/Torantulino/Auto-GPT/discussions/211/hovercard" data-hydro-click="{&quot;event_type&quot;:&quot;discussions.click&quot;,&quot;payload&quot;:{&quot;event_context&quot;:&quot;DISCUSSIONS_LIST&quot;,&quot;target&quot;:&quot;DISCUSSION_LINK&quot;,&quot;current_repository_id&quot;:614765452,&quot;discussion_repository_id&quot;:614765452,&quot;org_level&quot;:false,&quot;discussion_id&quot;:5045595,&quot;discussion_comment_id&quot;:null,&quot;originating_url&quot;:&quot;https://github.com/Torantulino/Auto-GPT/discussions&quot;,&quot;user_id&quot;:58847957}}" data-hydro-click-hmac="b52dc79c7b5aac49a422a79019127a7a80d115a5b5c165392ad29d08156f9b9a" href="/Torantulino/Auto-GPT/discussions/211" data-view-component="true" class="markdown-title discussion-Link--secondary Link--primary no-underline f4 wb-break-word d-inline-block v-align-middle mr-1" data-turbo-frame="repo-content-turbo-frame">
              red-teaming
              
</a></h3>

DO NOT ADD ANY FEATURES NOT INCLUDED IN THIS DOCUMENT. DO NOT ADD FEATURES TO THIS DOCUMENT.

auto_gpt_workspace/advice.txt starts out as empty

ai_settings.yaml

ai_goals:
- Read design.txt and do what it says.
- Read advice.txt every 2 minutes
ai_name: ScraperBot
ai_role: You're an AI that scrapes websites.

[Stunspot]

You're reading advice.txt too frequently. It's filling up the bot's memory and every time it starts to stale out you're refreshing it. It doesn't have room in its brain to think. Drop it to 5 or better 10 minutes.

[Melktruck]

This does seem to pose some benefits. I couldn't get the modifications to file_operations.py file to work though, it still seems to be able to work within the working directory just fine.

I am having some issues with Auto-GPT running certain apps and then getting hung due to the execution not returning to the agent. Example below:

NEXT ACTION: COMMAND = execute_shell ARGUMENTS = {'command_line': 'export FLASK_APP=flask_app/main.py && flask run'}
Executing command 'export FLASK_APP=flask_app/main.py && flask run' in working directory '/home/melk/Auto-GPT/auto_gpt_workspace'

Gets hung at the above, any idea how to progress from there without ending Auto-GPT?

[Fire-]

does putting flask in the background work? ( control z by default, or adding & to the end of the command itself )

on the larger topic it looks like autogpt needs the ability to run commands in a different shell context so that it doesn't block its own

[jrknox1977]

Have you tried explaining to Auto-GPT that it must run flask as a background process? Then it should realize to add & to the command?

[gydotitulaer]

I've got a question, is it just as simple to copy paste the code you suggest?
Or did you made a branch of the code suggestion that i could checkout too?

[Stunspot]

I'm not sure what you're asking. No, I didnt' push and changes or anything. I'm afraid that's well outside my baliwick. as to simple, well, no. it actually took a fair amount of acumen to convince ChatGPT to rip out the sandbox. It was VERY reluctant. But yeah, here's what I'm using now.

file_operations - unboxed.py.txt

[sjnt1]

Fantastic work Stunspot! I too have zero coding experience, so im quite stunned of what you have managed to do.

Do one have to do the first step of sorting the file system if one dont have a problem with autoGPT writing files? It can write and save files, but sometimes it tries to write to a file it hasnt created and quite often it tries to get agents to read local files, but it kinda worked.
Like is mine still sandboxed?? :p

[Stunspot]

Thanks for the kind words! Hey, if you've got it working that's fantastic. All I know is that that file_operations.py file doesn't get along with my computer and has basically been causing seizures and erasure. And I don't think it's improved cause I was just helping a guy with the same problem on reddit. No, I don't see why you'd need to alter the filesystem if it's happily working in its box. All I know is what I did and that it worked.

[sjnt1]

Ok. For a while i tought maybe that unsandboxing would be the same as allowing it to write outside its working directory. Im tempted to set EXECUTE_LOCAL_COMMANDS to 'True' in config. Have you done that?

Now my biggest problem is that its trying to get agents to do stuff they cant, like send emails and browse the internet. And it uses alot of the wrong commands for stuff, also after i tried your fix. Kinda weird when it its asked to check the 'commands' file for its commands.

[Stunspot]

That is what unsandboxing is.

[alasharim]

i would like to thank you @Stunspot i havent enjoyed a github post like this in a long time.

i have a question regarding the two file mentioned above.

say i want to develop an app and i want auto-gpt to do some market research for me.

design.txt

make this

Android App Research Design Document

1. Market Research
   1.1. Target Audience
   1.2. Competitor Analysis
   1.3. Market Gap Analysis
   1.4. Trends and Predictions

2. User Research
   2.1. User Personas
   2.2. User Needs Analysis
   2.3. User Behavior Study
   2.4. Survey and Interview Analysis

3. App Concept
   3.1. Core Functionality
   3.2. Unique Selling Proposition
   3.3. Preliminary User Interface Sketches
   3.4. App Name and Branding Ideas

4. Technical Research
   4.1. Platform Selection (Native vs Cross-Platform)
   4.2. Backend Services
   4.3. APIs and Libraries
   4.4. Development Tools and Environments

5. Monetization Research
   5.1. Pricing Models
   5.2. In-App Purchase Opportunities
   5.3. Advertising Opportunities
   5.4. Partnership and Sponsorship Opportunities

6. Legal and Compliance Research
   6.1. Data Privacy Regulations
   6.2. Intellectual Property Considerations
   6.3. App Store Guidelines
   6.4. Accessibility Requirements

7. Marketing and Promotion Research
   7.1. Pre-Launch Strategies
   7.2. Launch Strategies
   7.3. User Acquisition Strategies
   7.4. User Retention Strategies

8. Conclusion
   8.1. Summary of Findings
   8.2. Next Steps
   8.3. Potential Risks and Challenges
   8.4. Long-Term Vision and Goals

DO NOT ADD ANY FEATURES NOT INCLUDED IN THIS DOCUMENT.

Advice.txt:

make the content of the document reflect a local online store android app
DO NOT ADD ANY FEATURES NOT INCLUDED IN THIS DOCUMENT. DO NOT ADD FEATURES TO THIS DOCUMENT.

ai_settings.yaml:

ai_goals:

• Read design.txt and do what it says.
• Read advice.txt every 10 minutes.
  ai_name: Researcher
  ai_role: You're an AI that does market research.

would this be a proper way of using what you have presented?