Cone-bounded replay.

ConeReplay records a multi-agent trace, then replays a proposed prompt, model, or agent change against that recording. Events outside the causal cone of the change are served from the recording. Events inside the cone run again. The report says whether the outcome diverged.

What gets recorded

A trace is the sequence of LLM calls, tool calls, and cross-agent messages, with parent links. The LangGraph adapter is attach_to_langgraph_app. The published install is pip install 'conereplay[langgraph]' (PyPI 0.1.2).

What the cone limits

Re-executing the whole trace spends tokens and reintroduces sampling noise on turns the change cannot affect. The cone is the set of events downstream of the edit. The rest stays byte-identical to the recording. That is the regression check, not a claim that every production workload saves the same share of tokens. The home page states the measured figures and the corpus they came from.

What you review

conereplay corpus writes a divergence report for a recorded corpus and a modification file. Use that report to decide whether the change did what you intended before you deploy it.