Instrumenting Claude Code locally gets you the one thing the transcript hides, which is what the agent did between turns and how long each step took. Tool call latency and token spend per step are usually where the surprises show up. Are you capturing tool inputs and outputs as spans too, or only the timings?