I wanted to understand some recent commits I'd seen in a very large Rust project that I have a day job interest in keeping track of. It's a blockchain node (aka server) and I believe is one of the largest Rust projects currently. The code originated at Facebook, subsequently enhanced and maintained by various other groups. Specifically the task was to understand the purpose of the "dag" code within the consensus module. I'd heard that possibly a new optimized consensus algorithm might be somehow using that code, but initially wasn't sure and also didn't have a deep understanding of the consensus code in general.
I began this exercise without considering the use of LLM tools. I just opened the files in the consensus/dag subdirectory and read them. I expected that this would tell me what I wanted to know. After about 15 minutes it became apparent that although I'd looked at all the source files, and generally they made sense, its clear purpose still wasn't revealed.
So after cloning the repository and firing up Claude Code, I asked for its take. Since it's common to see online comments that LLMs are only useful for simple Web UI and CRUD code, I thought it was worthwhile posting the session as an example of an LLM providing useful analysis of code that's about as polar opposite to React and CRUD as it's possible to get!
I used this utility writen by Simon Wilson that extracts Claude code session content through its API. This is how the session transcript was generated. I hand-edited the resulting HTML to include a link to this page.