You can get something functional using ollama for the model, aichat for the front end, and then making a script to generate a session file with all the relevant data you want included. I don't know how effective it will be because there will be so much data.

Not sure if it will handle queries on bash commands. But, this video covers setting up and running a local large language model (LLM). IIRC, he specifically sets up DeepSeek, but it looks like you have a couple options. However, one caveat is that these LLMs run much faster with with a good GPU backing them up. If you don't have a reasonable gaming GPU, you may be waiting a bit for answers.

Are you really ready to trust it? I use DDG's agent frequently, when my query-fu fails, but I've gotten bogus, outright lies in response. They are good for helping me reframe my questions for the regular search engine. I don't think I'd rely on LLMs yet for technical guidance.

In dealings with current LLMs I'm regularly reminded of the old truism: "it's better to spend ten years looking for the right teacher, than to spend ten years learning from the wrong one."

Have you tried solutions like "thefuck"? Maybe a non-LLM shell helper would be safer, and be less likely to teach you wrong or bad solutions.

I think there's a bunch of AI enhanced shells and shell assistants... Shell-AI, Vity, Warp Terminal, Shell Genie, Butterfish, yai, aichat, aider, tgpt... You can get plugins for zsh or fish and other shells, we have agents with tool access, it's in some IDEs...