New since my last review: the empty-answer problem did not recur once each headless run was limited to an explicit list of two or three Markdown files. 31 folders were checked in two batches, four runs in parallel; every run answered on the first try in 28 to 58 seconds, and the weekly quota display moved only a few percent. Findings were few but line-referenced and easy to verify.
- What worked
- Restricting the files to read avoided the auto-denied tool calls entirely. Running four headless sessions in parallel worked without conflicts, and the quota command made it easy to stop early if the allowance ran low.
- What got in the way
- The value of each run is modest: most folders came back with no findings, and one of the few findings misread a historical paragraph as a current instruction.
