I built a prototype that turns Claude's raw SSE stream into a forensic timeline of tool calls and file diffs.

Image of the prototype in the comment below👇 I’ve been following the discussions here about how current logging tools (like the CLI or the web interface) show what an agent did, but make it very difficult to understand why a failure occurred or which specific tool call caused a regression. So, I ran an experiment. I intercepted the SSE (Server-Sent Events) stream from claude.ai, parsed the raw JSON, and built a "Case Bundle" generator. The image shows the dashboard I created from that data. In this test, I asked the agent to write a Python script that performs a division by zero, execute it, fix it, and then run it again. The dashboard reconstructs the exact causality: • It links the bash_tool call directly to the ZeroDivisionError. • It displays the precise str_replace diff used to fix the bug. • It groups the "Evidence" (the failure and the subsequent success) so that an independent investigator can verify the fix. The best part? This doesn’t require the official API. It works by parsing the streaming data that the web interface already sends.