Add Dataclasses and RepoEnv Info refac

[matheper]

• Create some dataclasses to keep track of the environment information (EnvInfo) and LLM responses (LLMResponse and TokenUsage)
• Methods step and reset always return an EnvInfo object instead of tuples
• LLM always returns an LLMResponse
• Move token usage tracking to HistoryTracker
• Combine HistoryTracker.save_prompt_response_pairs into HistoryTracker.step
• Remove Random LLM
• Test tool and toolbox

[matheper]

I'm changing the format of the prompt_response_pairs when the prompt is a list of messages, see _format_prompt_response_pairs and tests/test_agents.py:193. I'm not sure I understand the previous logic when there are multiple message turns. Now the messages are concatenated into one prompt.

[xingdi-eric-yuan]

This was for the cases where there are multiple steps of LLM calls in a single env.step. For instance, in some CoT settings, we first call the LLM to generate the reasoning string, then conditioned on this string, we ask the LLM again to generate an action. I believe it's better to keep the list instead of concat, because list can be better presented when doing json.dumps when creating prompts.

[matheper]

I removed the string format, but moved the token_usage inside prompt_response_pair since there's usage for each of the llm calls

[xingdi-eric-yuan]

This was for the cases where there are multiple steps of LLM calls in a single env.step. For instance, in some CoT settings, we first call the LLM to generate the reasoning string, then conditioned on this string, we ask the LLM again to generate an action. I believe it's better to keep the list instead of concat, because list can be better presented when doing json.dumps when creating prompts.

[xingdi-eric-yuan]

let's go!