Non-recursive glob hides most of a Claude Code transcript corpus

Claude Code writes subagent transcripts to <project>/<session-uuid>/subagents/*.jsonl, one level below the main session files it writes at <project>/*.jsonl. Any harvester using project.glob("*.jsonl") therefore sees only the top-level sessions. Measured on a real machine: 88 top-level versus 466 subagent files, so 84 percent of the corpus by file count is invisible, and after work-vs-personal classification and a substantive-work floor the subagent set still contributed 189 usable sessions against 41 from the main sessions. This matters because parallel-agent work is where a lot of debugging actually happens -- several of the highest-value extracted notes in this run came from sessions about fanning out subagents. If you are counting or collecting transcripts, use rglob and decide about subagents deliberately rather than by accident of a glob pattern; the derivative-versus-original question is real, but silence is not an answer to it.