TopOfMind reads these when the /chat page loads. VoicePrint streams live partial transcripts; RunPod records each utterance and transcribes it in one batch call (TOM-side key required).

The pre-flight guard blocks any synthesis / therapy run whose input exceeds the max. Runs use an exact count from the API's count-tokens endpoint when available; the chars/token ratio drives list estimates and re-calibrates itself after every successful run. Blank fields fall back to the .env defaults.