The runner now avoids several repeated steps when starting an agent session. It reuses local tool binaries and Pi skill copies where applicable, mounts drives in parallel, and reuses a harness capability probe.

When its trace is ready, a completed Pi turn no longer waits for the full trace-drain polling window. Organization owners and admins can test their connections from a project that does not use them. Conversation replay no longer cuts through a character and leaves a chat unable to continue.