Every Claude Code session is already a rich agent trace sitting on your disk — tool calls, results, errors, the lot. No instrumentation, no export step: pick a session file and analyze it.
JSONL · JSON · NDJSON · CSV · TSV · TXT · GZ · ZIP
~/.claude/projects/<encoded-project-path>/<session-id>.jsonl.ls -lt ~/.claude/projects/*/*.jsonl | headSessions can contain file contents and secrets. Everything is parsed in your browser, and only bounded excerpts of individual traces go to the judging model — but if the code is sensitive, redact first or run GetEvals locally with your own key.
A self-contained HTML report: the failure modes discovered in your own traces (not a generic checklist), a heatmap of which traces hit which mode, the exact quoted evidence behind every verdict, and the judge's own reliability numbers. Plus downloadable evals — a promptfoo config, an OpenAI Evals data file, and a portable judge suite — so the analysis turns into a regression suite you keep.