What actually goes wrong in your agent sessions?
When a session goes badly, it’s tempting to blame the model. Often it’s the floor: a missing command, a shell option, a flag the installed CLI doesn’t accept. A smarter model doesn’t fix any of those. This page reads your session transcripts and counts, in code, which failures keep coming back, how many come from your environment, what the sessions cost, and whether the problems you fixed stay fixed.
Your sessions
Everything stays on your machine. Transcripts hold your source code and maybe secrets, so this page reads them in this tab and never sends them anywhere. There’s no server and no model involved.
Drop session transcripts or whole folders here
Files are read in your browser. Nothing is uploaded.
.jsonl file per session in ~/.claude/projects/. Drop that
folder, or one project inside it. Subagent transcripts in a subagents folder count toward the session they sit beside.Or try a made-up set of sessions
Paste a transcript instead
Drop your transcripts, or pick one of the made-up sets above, and the audit starts with a question for you.