Proposal for a git 'command'

[dany-on-demand]

Problem Description

The idea is that forcing the models to use git or any other versioning system will trigger more of the programming training and thus improve logical reasoning? Ideally, the agents would use git history to determine if they're going in the right direction, and possibly undo bad decisions. And the ultimate version would be having the agents be able to interpret git history of existing projects

Proposed Solution

I propose making a 'VersionControl' command with these (hopefully self-explanatory) functions:

• git_init(optional readme_content: string) # I assume the model will like to use the readme as a sort of "description of objective"?
• git_write_and_commit(text: string, mode: (ReWrite,Append), commit_message: string)
• git_cherrypick(commit_hash: string)
• git_history() # show last 10 commits
• git_search_history(filter: string) # show last 10 matching commits, different command as to not confuse the model into using filters when not needed
• git_list_commit(commit_hash: string)
• git_diff_last_commit() # will feed the model peacemeal acc to context size
• git_status()
• git_commit_as(agent_name: string)
• git_switch_branch(branch_name: string) # switches or creates
• git_ignore(regex: string)

git will be local only (no pushing), all commands will git_init if not already initialised

I also assume, this would be primarily used for code & (story)writing. (off-topic, for code, a companion 'CodeTools' command containing stuff like 'cloc' and linters)

Alternatives Considered

No response

Additional Context

No response

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this feature has not been requested yet.
• I have provided enough information for the maintainers to understand and evaluate this request.

[lightningRalf]

This looks interesting.
Can this be somehow be combined with the concepts of recursion and backtracking? / similar to the tree of thoughts paper?

https://arxiv.org/abs/2305.10601

There are two implementations:
Official: https://github.com/princeton-nlp/tree-of-thought-llm
unofficial: https://github.com/kyegomez/tree-of-thoughts / with a monte-carlo approach also.