Ask a structural question about how an agent worked and get an exact answer over the whole trajectory dataset, not a sample. ProcGrep runs server-side: it ingests the dataset, canonicalizes it into an action vocabulary, and scans every trace. No model call.
These are public agent-rollout corpora on Hugging Face (SWE-agent and OpenHands scaffolds solving SWE-bench-family GitHub issues). ProcGrep canonicalizes each trajectory into an action spine; nothing is re-hosted.
The diff strip names what separates two groups: the procedures that most distinguish them, their action-mix distance, and the length and reasoning gaps. No model call.
A real Claude Code session reduced to its action structure: each human prompt opens a turn, and the agent's tool calls become the turn's atoms. Pick it in the sidebar, replay it, or run an atom pattern over it — the same grammar as the eval query.
These are sample traces. Run procgrep locally to explore your own sessions:
python examples/procgrep_view.py