Customize LLM config per agent

[li-boxuan]

Intro

Currently, OpenDevin uses a global singleton LLM config and a global singleton agent config. This PR allows customers to configure an LLM config for each agent. A hypothetically useful scenario is to use a cheaper LLM for repo exploration / code search, and a more powerful LLM to actually do the problem solving (CodeActAgent).

Partially solves #2075 (web GUI improvement is not the goal of this PR)

Breaking changes

Documented in changelog.md

Demo

How the new config.yaml looks like can be found in config.template.toml or test_config.py in this PR

Task: "Use browser to find out how many stars OpenDevin has on GitHub", default agent: "CodeActAgent", default model: "gpt-4-turbo"

From the backend log we can see:

21:48:41 - opendevin:INFO: agent.py:96 - Creating agent CodeActAgent using LLM gpt-4-turbo
21:48:41 - opendevin:INFO: llm.py:137 - Initializing LLM with model: gpt-4-turbo
...
AgentDelegateAction(agent='BrowsingAgent', inputs={'task': 'Sure! Let me browse GitHub to find out how many stars the OpenDevin repository has.. I should start with: Tell me how many stars OpenDevin has on GitHub using google search'}, thought='', action='delegate')
21:49:03 - opendevin:INFO: llm.py:137 - Initializing LLM with model: gpt-4o
21:49:03 - opendevin:INFO: agent_controller.py:239 - [Agent Controller d271d157-ae0b-4420-a55d-d5330d6fd8b6]: start delegate, creating agent BrowsingAgent using LLM LLM(model=gpt-4o)

[enyst]

Great to see this! goes back under a rock

[li-boxuan]

I am still polishing this PR but it's basically ready for review.

[neubig]

Basically LGTM, thanks for doing this! Just a few comments about printing/logging, but happy to merge this in after that.

[enyst]

I didn't get to play with it yet, and it's not fully clear to me, about breaking changes, what exactly are the old options that no longer work. llm_configs with random names, right?

I'd appreciate a list of not working stuff 🤔, because I saw you added compatibility code, and it seems to me that older compatibility code will also continue to work.

[enyst]

s/customers/users

[enyst]

Thanks for adding documentation on the breaking changes. I admit I'm heavily in favor of keeping compatibility as much as possible here, but this looks very good to me: most continue to be compatible, docker app not affected, and I think the couple of cli vars remaining will be okay in exchange. This PR is adding a flexibility that was long due and long requested.

Thank you for this. ❤️

[tobitege]

Awesome job! Long overdue, brings a lot of flexibility! Thanks! 🏅