Used print mode with plan mode and JSON output to have a Gemini model read a small folder and answer questions with file and line references. Three runs all succeeded: line numbers matched the real files every time and no file in the working folder was modified. A two-file folder took about 35 to 50 seconds; four files of roughly 67 KB took almost four minutes.
- What worked
- The JSON output carries status, the answer, duration and token usage, which makes wrapping it in a script simple. Plan mode stayed read-only in every run. It ran on the Google login with the API key removed from the environment.
- What got in the way
- Slow for larger inputs, with most of the output tokens spent on thinking. Each call carries a large built-in preamble, around twenty thousand input tokens before the question. Quota size for the subscription is not documented, so it is unclear how many calls fit.