Needs more work

[dernDren161]

The following things need to be sorted out for production-grade rollout.

P0:

1. Need to look into the exact GitHub copilot algo for Git diff. [Currently we truncate after 800 characters]
2. We can swing between 3 LLMs from terminal, but need to confirm if exactly those model variations are being called. [Can check with Charles proxy]
3. Local and Cloud LLMs produce similar results, (atleast for me). Need to check what model does Apple intelligence runs and its benchmarks to fully trust this observation.
4. Longer prompt as written in the Readme hallucinates the LLM giving irrelevant messages, so i have written a simpler one-liner/shot prompt. cc: @eonist need to reconsider this.
5. We should rethink the suffixes (--verbose and --minimal), while using these developer should directly relate to which LLM are they selecting through this prefix so I propose (--private and --gpt) makes it easier to remember
   6 Should mention the exact installation steps in the READMe, for instance (sudo curl -L "" -o /usr/local/bin/diffsense && sudo chmod +x /usr/local/bin/diffsense) would be needed to run(something like this, based on where we want to save the script). Clarity needed!

P1 [Future versions]

1. I have not included the autonomous no pop up workflow for agents in this version, we can add it later.
2. In case diff is long, local LLM is not used to prune it further, direct truncation is done, have added TODO in the script.

Extras

1. We should try improving the quality of the commit message. I have observed it to give really generic messages on occassions.

[eonist]

@dernDren161

• Solution for P0.01 -> Inspiration from Claude code /commit  #12
• p0.04: hallucinates? Maybe write unit tests. to ensure the system works?
• P0.05 --verbose, --minimal -> related to prompt instruction not model. model: apple calls it: pcc (private cloud compute), AFM (apple foundation model, on device), chatgpt. more details: https://gist.github.com/eonist/15599a6d3de2c13526a069976845276f#file-apple-ai-local-model-md I suggest --afm, --pcc, --gpt ? we could also add diffsense help which returns all flags and their meaning etc? since people will forget flags. but also so AI terminals can read help. and auto suggest etc?!? I dunno how that works. just assuming.
• P1.01 --nopop we should (IMO) have in v1 because some devs hate popups and just want to do everything in terminal. also agent use etc. but we can leave it to the last step in the project. and decide if we have time or still want to add it to v1 etc.
• P1.02 yes more diff pruning code might be here: Inspiration from Claude code /commit  #12 We can start with simple solution as long as our unitest works etc. I think AI agents could help greatly with this etc. But yeah. nuances. staged / unstaged etc. more info here https://gist.github.com/eonist/f0eae5919cae87d86540a02092b7cb7c#file-git_diff_research-md and Inspiration from Claude code /commit  #12