For a parent whose kid has a food allergy, every grocery run is a label-reading job, and one wrong substitute can mean the ER. LunchShift is a personal agent that plans your kids' school lunches, shops real products, and refuses anything unsafe.
You tell it about your family once, in chat. Every week it builds the cart within budget and holds the order. *When the store swaps in a substitute, it catches the unsafe one and fixes the cart with zero clicks.* In our demo, the cheapest replacement bread looked fine, but its label said "May contain: peanuts." LunchShift refused it, swapped Monday's lunch, and re-held the order before the parent woke up.
*Safety is decided by code, not the model:* labels, allergen tags and "may contain" warnings gate every product, and AI can only add refusals, never remove them. Every decision shows its evidence: the label line, the allergy you confirmed, and why it picked the replacement. The agent never pays. You approve payment (Stripe test mode).
Results: 27 of 27 scenarios pass with zero unsafe items and zero duplicate orders, and the agent scored 95/100 on a Firedrill prompt-injection drill.