clustertrace is a local-first debugger for LLM agents. Instead of scrolling through individual run logs, it groups failing runs by execution pattern — the order of functions and tool calls each run made — so two underlying bugs surface as two patterns instead of twelve scattered traces. Patterns below are ordered by how often they occur; the ones with red fail-rates are where the bugs are.
The longest sequence of functions that every failing run starts with. When failures come from different agents the shared path is empty — check the per-agent breakdown below.