You’re on the right track with the prompt, but the issue is that the model still reasons in UTC internally. The reliable fix is to enforce the timezone at the tool/backend level rather than trusting the prompt. Convert all incoming times to Asia/Calcutta (Asia/Kolkata, Kolkata) on your server and only pass fully formatted ISO timestamps with the +05:30 offset to check_availability. This ensures the assistant can’t accidentally fall back to UTC even if it ignores the prompt instructions. I can help you @choudhary9549