​
I built Crucible as a personal project to understand how AI agents actually consume information. The repo includes the full code, the full Claude Code session transcript so you can see every prompt used, and a complete document explaining what was discovered along the way.
The tool intercepts every tool call Claude Code makes: bash commands, file reads, web fetches. It returns structured signal instead of raw noise. About 82% compression on average. But the more interesting part to me was the methodology that emerged from building it.
The monitor file in the repo is 1,100+ lines of the AI watching itself work and writing first-person reports about what it noticed. Those reports became the specification for the next improvement. I didn't plan that. It just happened, and then I named it.
15+ hours of session history, all of it open on git.