codename: snake_oil for the skeptics - ran it again on fastapi vanilla vs snake_oil 92.5% token reduction 139k → 10k tokens $0.42 → $0.03 welcome to the game, kid

Guessing this is like DSPy where a small model gets looped against a high model response. I've started building something similar but looking forward to trying this when released

nope

bro where's the repo?!

Ooo looks like you made a more efficient version of my intent layer And created a way to quantify the gains! Codemaps was always a better name than intent layer

How does this change tool usage and code accuracy?

how does a sharper sight affect the archers ability to hit the target?

Oh, you wild, wild

Will this diminish code quality?

improves it

How much are you going to charge for v3?

who said I was charging

🫣

I use a project agnostic script which ingests a codebase of any size and algorithmically generate a 2 layer codebase atlas md file in 5 seconds.

Now rename it

Is it only for claude code or can I use it with opencode?

yeah you can use as an mcp fyi i've only tested on claude code and its rigged with custom hooks using tldr but the same should be simple to setup for opencode

Interesting. Will test soon. If I understood this project correctly, I just have to talk to Claude as always, but it will increase its knowledge about the project session by session and drastically decrease the amount of token use?

yeah tldr blocks grep/reads and injects layers knowledge to reduce token consumption in session the memory system over time extracts learnings from Claudes thinking blocks from the session and since it runs via hooks its automatic only caveat is for each new project, run 'tldr semantic index .' i've added an update to the session start hook to warn if there isn't one now too just incase otherwise semantic search won't work

tldr has semantic search built in for free and the hooks pipe in the info when Claude reaches for grep so there’s no reason to pay for something you can do natively optionally outside of hook, gonna explore if tldr as a native mcp tool works better but that’s a test for next week

Have you uploaded v3 yet?

its huge!

This is legendary. May I token token-squanch with u

Make a comparison with @morphllm's warpgrep tool please

that’s some wild efficiency right there. those numbers speak volumes. makes you wonder how much more we can streamline marketing too. maybe First100 can help with that, right?

that's nice, but you're measuring the wrong thing token count drops when you cache repetitive docs. The real test is how much semantic info you actually kept vs compressed away what was the recall on specific details?

Releasing soon?

LSP / tree-sitter?

@grok how would this work exactly? Let’s see if you’re smart.

Followed

me watching the vid 🔊

ready when you are!

Ia this based on a RLM model ?

hub me daddy

How close are you to release? I assume you ran into some issues over the past few weeks?

I am too dumb to understand how to use it and integrate into my flow. Will go ask CC...

🔥 ur cooking

Wow

Awaiting v3 eagerly @parcadei

Are you selling this anywhere?

Countdown till the release?

Very Keen, could the system be adapted to be leveraged by codex? I've been working on semantic search tools and even those can be a huge boost.

AST grep

hell yeah

put it on the hub