What happens to all the research your AI coding assistant does on your behalf? If you're using tools like Claude Code or Cursor, the answer is usually: it disappears when you close the tab.
Developer Mark Fletcher makes a practical argument for fixing this. When an AI agent helps you debug a production issue or investigate a tricky bug, the reasoning trail, the session transcripts, the distributed traces it pulled up, all of that context is valuable. And right now, most of us throw it away.
Fletcher proposes a structure he calls "Cases" - directories committed to your repo alongside the code changes. Each case contains a human-readable summary, the AI session transcript, any trace results or logs, and whatever artifacts the investigation produced. The idea is simple: investigation records should live where the rest of your work lives, in version control.
He has a concrete reason for caring about this. While working on Groups.io, he hit a recurring calendar feed bug that a previous Claude session had helped him solve. But that session was gone, and he had to re-derive the solution with only partial memory of the original reasoning. Anyone who's used AI coding tools for more than a few weeks has probably hit the same wall.
This won't matter for quick "fix this typo" sessions. But for multi-hour debugging investigations, architectural explorations, or any AI-assisted work where the why matters as much as the what, having that context in your git history could save real time down the road. Fletcher has built an open-source tool called Trellis that gives Claude read-only access to production systems for exactly this kind of structured investigation work.
The overhead is minimal. The payoff comes three months later when you're staring at a commit and can't remember why you made it.