Field notes from building a tool that reads agent transcripts — including the measurements it turned out could not be made.
Cursor renamed edit_file to edit_file_v2, the adapter kept exact-match lookup, and 274 sessions were filed as never having tried to write anything.
Eight models cleared the volume floor, giving 28 pairs. Twenty-five never overlapped, two were underpowered, and the one testable pair came out null.
Cursor keeps every conversation in one 4.4 GB SQLite file and records no working directory, so the repository has to be reconstructed from file URIs.
Across 2,364 agent runs, cache-read was 96% of all tokens and output was 0.39%. Eight runs account for half the cache-read; one accounts for 22%.