The three things it looks at
All three are read locally, used to answer one question — is an agent working right now — and then discarded.
- The process list. Names, full command lines, and CPU usage, the same thing
psshows you. Worth knowing: a command line includes any prompt text passed on it. - Session file names and timestamps. The names and modification times of agent session files, so it can tell that an agent waiting on a model reply is still working even at zero CPU. The contents of those files are never read.
- On-screen text — only if you grant Accessibility. Optional and off by default. When on, the app reads text from the agent windows you enable and matches it against short status phrases in memory. For a terminal or code editor that text can include your code. It is never written to disk, never enters the log file, and never leaves the Mac. Without the grant this input simply does not run.