[Bug]: opendevin seems donot do Substantive work ， just print out "All done! What's next on the agenda?"

[kinigsonstar]

Is there an existing issue for the same bug?

• I have checked the troubleshooting document at https://github.com/OpenDevin/OpenDevin/blob/main/docs/guides/Troubleshooting.md
• I have checked the existing issues.

Describe the bug

opendevin seems donot do Substantive work ， just print out "All done! What's next on the agenda?"

Current Version

INFO:     Started server process [7216]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://127.0.0.1:3000 (Press CTRL+C to quit)
Connection to localhost (127.0.0.1) 3000 port [tcp/*] succeeded!
Backend started successfully.
Starting frontend with npm...

> opendevin-frontend@0.1.0 start
> vite --port 3001 --host


  VITE v5.2.8  ready in 6945 ms

  ➜  Local:   http://localhost:3001/
  ➜  Network: http://172.16.2.230:3001/
  ➜  press h + enter to show help

Installation and Configuration

make build; make run

Model and Agent

LLM_MODEL="gpt-3.5-turbo"
LLM_API_KEY="sk-..."
LLM_EMBEDDING_MODEL=""
WORKSPACE_BASE="./workspace"
SANDBOX_TYPE="exec"

Reproduction Steps

No response

Logs, Errors, Screenshots, and Additional Context

No response

[enyst]

@kinigsonstar I modified your post to remove the API key. Please make sure to cancel it, if it was real, and make another.

On the issue: I've seen too that GPT-3.5 is sometimes doing that. You can restart the app, or at least the session, if necessary, it would get it back on track. It must be related to the history it got up to that point, so if you start a new session it should be cleared.
I think the LLM sometimes decides to answer with 'finish' every time, and if that is so, there may or may not be a lot we can do, we can look into it.

[kinigsonstar]

thanks enyst, the API key was not real,
Based on your preliminary analysis， the issue may be related to the LLM?

[enyst]

the issue may be related to the LLM?

It can be, FWIW, GPT-4 is better, including on not "forgetting" what it was told, and on interpreting its task, and I believe I haven't seen this particular issue on GPT-4, but what you're seeing here is not just LLM.

For one, recently, we have changed the way it worked after finishing a task, so that it resets when it's done. I think you're on an older version. So if you upgrade, it might not repeat itself here, but: I think it will not be able to answer your question anyway, precisely because it "forgot" the previous task.

Also, currently the monologue agent works better if you give it a task to do. It's not so good as a chatbot. I think of it this way: it "thinks" it got a task from the user, so it tries to do what you say, not just answering questions, not just chatting.

That may change soon, and the behavior will be better, allowing for both another task, and some chat with the user about the previous task.

[rbren]

Which agent are you using? monologue?

[kinigsonstar]

thanks enyst
I will get latest version and have a try with GPT-4

thanks rbren,
“Which agent are you using? monologue?” I am new to opendevin , how can get which agent I used?

[rbren]

You can see it by clicking the settings wheel in the bottom right (or left, on older versions)

Monologue is default, so that's probably it

[li-boxuan]

I think there's a gap. From your screenshot, clearly you are hoping OpenDevin can understand & remember the context of previous task. I noticed you used wording like "this program", which OpenDevin has no idea about, because you gave OpenDevin a new task without telling it the context. It's very intuitive for one who is familiar with chatbots to have that mindset, but that doesn't perfectly suit here (today).

That being said, OpenDevin does have some "context" as in it keeps the current workspace, so it can read whatever files created by previous tasks. Could you please try rephrasing your task from

how to run this program in linux

to

how to run 'hello_world.cpp' in linux

That might make a difference.

[rbren]

Going to close this one--I think it was fixed by a recent PR to reset task state properly