Coding adapter

With the prices for tokens rising precipitously I am not a huge fan of how expensive it is to run agents 24/7. I might be getting a lot of work done but those tokens are expensive. In addition the agents commonly make the same mistakes and slow down at the same architectural decisions that deviate from the standard programming lexicon.

To address this I am using a two pronged approach. Many of the tokens are spent on rote work like opening worktrees so I am building Myelin that is a pipeline engine for LLMs. That way regular code can do things like wait on ci, or open worktrees without wasting precious tokens. The second piece of my solution is to train models to use myelin and follow my code architecture from the start.

To build this I am working on a dataset built upon the architecture-checks. I use the architecture checks to break code so it doesn't follow my architecture and then have the agent try multiple times to solve it and then take the best solution. I then do supervised fine-tuning on the eventual dataset to get agents that pass ci on the first try.