Build a Trace of an Agent Run

The problem, from first principles

A single agent run might call an LLM, then a tool, then the LLM again, then another tool -- and when something goes wrong (or costs too much, or takes too long), the question is always "which step, and when." A trace answers that: a timeline of every step in a run, in order, with when it started and how long it took.

The task

Write build_trace(events) that takes a list of (name, kind, duration) events run one after another with no overlap, and returns one dict per event with its name, kind, start_offset (seconds since the run began), and duration.