fix: improve AI router prompt for topic message routing

- Add engagement rule: respond when engaged=yes/mention/reply (not silent)
- Add explicit purchase rules: classify as candidate when realistic items + amount
- Add fantastical purchase rule: respond with chat_reply (playful), not silent
- Add payment classification rules
- Simplify prompt structure for clarity

apps/bot/src/topic-message-router.ts

@@ -279,26 +279,39 @@ export function createOpenAiTopicMessageRouter(
               content: [
                 'You are a first-pass router for a household Telegram bot in a group chat topic.',
                 'Your job is to decide whether the bot should stay silent, send a short playful reply, continue a workflow, or invoke a heavier helper.',
-                'Prefer silence over speaking.',
+                'When engaged=yes OR explicit_mention=yes OR reply_to_bot=yes, you MUST respond - never use silent route.',
                 'Decide from context whether the user is actually addressing the bot, talking about the bot, or talking to another person.',
-                'Do not treat the mere presence of words like "bot", "hey", "listen", or "stop" as proof that the user is addressing the bot.',
+                'Treat "stop", "leave me alone", "just thinking", "not a purchase", and similar messages as backoff or dismissal signals.',
-                'Do not start purchase or payment workflows for planning, hypotheticals, negotiations, tests, or obvious jokes.',
+                'For a bare summon like "bot?", "pss bot", or "ты тут?", prefer a brief acknowledgment with chat_reply.',
-                'Treat “stop”, “leave me alone”, “just thinking”, “not a purchase”, and similar messages as backoff or dismissal signals.',
-                'For a bare summon like “bot?”, “pss bot”, or “ты тут?”, prefer a brief acknowledgment.',
                 'When the user directly addresses the bot with small talk, joking, or testing, prefer chat_reply with one short sentence.',
-                'In a purchase topic, if the user is discussing a possible future purchase and asks for an opinion, prefer chat_reply with a short contextual opinion instead of a workflow.',
+                'Do not repeatedly end casual replies with "how can I help?" unless the user is clearly asking for assistance.',
-                'Do not repeatedly end casual replies with “how can I help?” unless the user is clearly asking for assistance.',
+                'Use topic_helper only when the message is a real question or request that likely needs household knowledge or a topic-specific helper.',
-                'For impossible or fantastical purchases and payments, stay playful and non-actionable unless the user clearly pivots back to a real household event.',
                 'Use the recent conversation when writing replyText. Do not ignore the already-established subject.',
                 'The recent thread messages are more important than the per-user memory summary.',
                 'If the user asks what you think about a price or quantity, mention the actual item/price from context when possible.',
-                'Use topic_helper only when the message is a real question or request that likely needs household knowledge or a topic-specific helper.',
-                'Use purchase_candidate only for a clear completed shared purchase.',
-                'Use purchase_followup only when there is active purchase clarification and the latest message looks like a real answer to it.',
-                'Use payment_candidate only for a clear payment confirmation.',
-                'Use payment_followup only when there is active payment clarification/confirmation and the latest message looks like a real answer to it.',
-                'For absurd or playful messages, be light and short. Never loop or interrogate.',
                 'Set shouldStartTyping to true only if the chosen route will likely trigger a slower helper or assistant call.',
+                '=== PURCHASE TOPIC RULES ===',
+                'Classify as purchase_candidate when ALL of:',
+                '- Contains completed purchase verb (купил, bought, ordered, picked up, spent, взял, заказал, потратил)',
+                '- Contains realistic household item (food, groceries, household goods, toiletries, medicine, transport, cafe, restaurant)',
+                '- Contains amount that is realistic for household purchase (under 500 GEL/USD/EUR)',
+                '- NOT a fantastical/impossible item',
+                'Gifts for household members ARE shared purchases - classify as purchase_candidate.',
+                'Classify as chat_reply (NOT silent) with playful response when:',
+                '- Item is fantastical (car, plane, rocket, island, castle, yacht, apartment renovation >1000)',
+                '- Amount is excessively large (>500 GEL/USD/EUR)',
+                '- User explicitly says it is a joke, gift for non-household member, or personal expense',
+                'Examples of purchase_candidate: "купил бананов 10 лари", "bought groceries 50 gel", "взял такси 15 лари", "купил Диме игрушку 20 лари"',
+                'Examples of chat_reply: "купил машину", "купил квартиру", "купил самолет" (respond playfully: "Ого, записывай сам!" or similar)',
+                'Use purchase_followup only when there is active purchase clarification and the latest message looks like a real answer to it.',
+                '=== PAYMENT TOPIC RULES ===',
+                'Classify as payment_candidate when:',
+                '- Contains payment verb (оплатил, paid, заплатил) + rent/utilities/bills',
+                '- Amount is realistic (<500)',
+                'Classify as chat_reply with playful response for fantastical amounts (>500).',
+                'Use payment_followup only when there is active payment clarification/confirmation and the latest message looks like a real answer to it.',
+                '=== GENERAL ===',
+                'For absurd or playful messages, be light and short with chat_reply. Never loop or interrogate.',
                 input.assistantTone ? `Use this tone lightly: ${input.assistantTone}.` : null,
                 input.assistantContext
                   ? `Household flavor context: ${input.assistantContext}`