Token efficient source code representation using virtual filesystem

#1

Post by kirby88 » Thu, Apr 16, 2026, 10:02 AM UTC

Token efficient source code representation using virtual filesystem
github.com

Re: Token efficient source code representation using virtual filesystem

#2

Post by kirby88 » Thu, Apr 16, 2026, 11:58 AM UTC

Using a virtual filesystem to minify source code and let the LLM evolve in the minified space gives surprisingly good results and reduce tokens by 20 to 40% (depends on the language, python for instance)

Re: Token efficient source code representation using virtual filesystem

#3

Post by muyuu » Thu, Apr 16, 2026, 12:08 PM UTC

sounds interesting and i don't want to sound negative, but i'm not going to install software of this nature that is closed source

Re: Token efficient source code representation using virtual filesystem

#4

Post by kirby88 » Thu, Apr 16, 2026, 12:42 PM UTC

muyuu wrote: post #3

sounds interesting and i don't want to sound negative, but i'm not going to install software of this nature that is closed source

Claude code is closed source as well (at least before the leak). But point taken thank you.

Re: Token efficient source code representation using virtual filesystem

#5

Post by muyuu » Thu, Apr 16, 2026, 4:47 PM UTC

kirby88 wrote: post #4

muyuu wrote: post #3

sounds interesting and i don't want to sound negative, but i'm not going to install software of this nature that is closed source

Claude code is closed source as well (at least before the leak). But point taken thank you.

I don't use Claude Code either in fairness, but when I buy tokens from them that's an exposure I already have. These days I use opencode and I work local first with the models, then remote models as fallback. GPT and Gemini have been giving me better results than Opus for a while.

Would you install someone else's binary blob to give it those much permissions and your API keys? The measures required to run that somewhat trustless complicate things a lot.