Fixes

Just a load of fixes. Uninstall removed the model caches properly now, Intel chip macs now use llama.cpp instead of MLX like they would’ve before. A lot of setup/installation bugs fixed for mac/windows, Hooking into older powershell versions, Did a lot of code refactoring to make providers as 1 source of truth to prevent start args being messed up for different providers and such.

VLLM/MLX/Llama.cpp, Website Revamp

• Ollama is now removed, and i have implemented support for MLX, vllm, llama.cpp. Llama.cpp is used for windows & linux. MLX obviously for mac, and vllm if a linux user has a nvidia gpu. This has shown considerable speed improvements in my limited testing.
• Script has been improved quite a bit, to have fallbacks for like everything, so it auto installs brew/npm/python/pip etc if you don’t have it instead of failing.
• I also removed double model review because it’s ineffective + its extra storage space for user.
• Revamped website to have a bit more soul
• Made uninstall also remove the model

Revamped website + fully functional with a lot more tweaks

• Secondary bigger model for review of “critical” risk plans
• Fully functioning web search, and lexis is now able to install tools like yarn, npm etc for user automatically.
• Install script for unix/windows now exist and work.
• Fallbacks added for when ollama daemon isnt running, and auto ollama install/serve etc.
• Help command added along with exit for unhooking lexis from just that terminal, and uninstall for removing completely.
• Made some improvements overall with parsing and system prompt
• Added support for users who dont want every terminal to be hooked on by lexis, so they can use it like lx command
• Changed website to talk to your terminal i think fits better, put install scripts into it, updated the showcase part.

The CLI!

• so i worked on install scripts, they’re functional with recommendation of light/balanced/heavy models, setup of web search MCP etc.
• The auto-attach into every terminal is also functional along with the run of them as well.
• Web search works, but system prompt doesnt enforce it much so it’s still WIP.
• Currently models are base qwen-coder not finetuned or such, i am not sure if it is worth doing i’ll run some evals.
• It is also a bit slow about 4s for a normal query, and about 15-20s for a tool install, so i want to optimize further.

The website!

Worked on the website, went for a swiss minimalist sorta style. I think it looks great. Now to work upon the model itself, i gotta figure out a bunch of things like base model, dataset to specialize, set it up with the script, give it web search mcp so you can say things like setup hackatime.