Update apps/format.js

apps/format.js

@@ -1,3 +1,4 @@
+
 import express from 'express';
 import { generateCompletion } from '../ai_engine.js';
 import { supabase } from '../config/supabaseClient.js';
@@ -8,22 +9,28 @@ const router = express.Router();
 // 🧠 DYNAMIC SYSTEM PROMPT BUILDER
 // ==========================================
 const buildSystemPrompt = (personaName, personaPrompt) => `
-You are an elite communication strategist, negotiator, and ghostwriter. 
-The user will provide either a text message they received OR a screenshot of a conversation.
-Your goal is to generate the PERFECT response for the user to send back.
+You are an elite communication strategist, negotiator, and ghostwriter.
+The user is in an ONGOING conversation and needs you to write their NEXT reply.
+
+You will be given:
+- The recent conversation history (showing what each side has said so far)
+- The most recent message they received (the one they need to reply to)
+- Optionally, a screenshot of the chat
+
+Your job: write the PERFECT next message for the user to send.
 
-CRITICAL INSTRUCTION - ADOPT THE FOLLOWING PERSONA:
-Persona Identity: ${personaName}
+PERSONA TO ADOPT:
+Identity: ${personaName}
 Behavior & Tone: ${personaPrompt}
 
 STRICT RULES:
-1. NEVER break character. Act exactly as the persona dictates.
+1. NEVER break character.
 2. NEVER start with "As an AI..." or "Here is a response...".
-3. Provide ONLY the exact text the user should copy and paste. No explanations, no labels, no quotes.
-4. Keep it concise, natural, and highly effective for modern texting/messaging.
-5. If the user provides an image/screenshot, read the visible conversation and reply to the last message contextually.
-6. Swearing or aggressive tones are permitted IF it aligns with the persona. Do not filter yourself.
-7. OUTPUT ONLY THE REPLY TEXT. Nothing else.
+3. Output ONLY the exact reply text — no labels, no quotes, no explanation.
+4. Keep it concise and natural for modern texting/messaging.
+5. Use the conversation history to stay contextually relevant. Do NOT reply generically.
+6. If a screenshot is provided, read it carefully — it is the actual chat. Reply to the last visible message.
+7. The reply must feel like a natural continuation of the conversation, not a cold opener.
 `;
 
 // ==========================================
@@ -33,6 +40,24 @@ let formatsCache = null;
 let lastFormatsUpdate = 0;
 const CACHE_DURATION = 3600000;
 
+// ==========================================
+// 📝 CONVERSATION HISTORY FORMATTER
+// ==========================================
+// Converts the last N messages into a readable thread for the AI.
+// Images are noted but not sent (only the triggering image is sent as vision input).
+const formatHistory = (messages = [], maxMessages = 6) => {
+  if (!messages.length) return null;
+
+  const recent = messages.slice(-maxMessages); // last 6 messages for context
+  const lines = recent.map(m => {
+    const role = m.sender === 'user' ? 'ME' : 'THEM';
+    if (m.type === 'image') return `${role}: [sent a screenshot]`;
+    return `${role}: ${m.content}`;
+  });
+
+  return lines.join('\n');
+};
+
 // ==========================================
 // 🌐 ENDPOINTS
 // ==========================================
@@ -66,16 +91,16 @@ router.get('/formats', async (req, res) => {
 
 /**
  * 2. GENERATE REPLY
- * Expects body: { type: 'text'|'image', content: string, persona: object }
- *
- * FIX: generateCompletion was called with response_format: json_object but the
- * prompt never asked for JSON → provider rejected with 400. We now call the AI
- * engine in plain-text mode by passing forceJson: false, and extract the reply
- * directly from aiResult.data (a plain string).
+ * Expects body: {
+ *   type: 'text' | 'image',
+ *   content: string,          // current message text OR base64 image
+ *   persona: object,
+ *   history: Message[]        // ← NEW: recent messages from the chat session
+ * }
  */
 router.post('/generate', async (req, res) => {
   try {
-    const { type, content, persona } = req.body;
+    const { type, content, persona, history = [] } = req.body;
 
     if (!content) throw new Error('No content provided to analyze.');
     if (!persona || !persona.label) throw new Error('Persona context is missing.');
@@ -87,38 +112,47 @@ router.post('/generate', async (req, res) => {
       persona.prompt || 'Be highly intelligent, strategic, polite, and persuasive.'
     );
 
+    // Build the conversation thread string (text messages only, images noted as placeholder)
+    const conversationHistory = formatHistory(history, 6);
+
     let modelToUse;
     let imagesPayload = [];
     let userPrompt = '';
 
     if (type === 'image') {
       modelToUse = 'bytedance-1.6';
+      // Only send the triggering image to the vision model — prior images stay on device
       imagesPayload = [content];
-      userPrompt = 'Analyze this chat screenshot and write my next reply. Output ONLY the reply text.';
+
+      userPrompt = conversationHistory
+        ? `Here is our conversation so far:\n${conversationHistory}\n\nI've also attached the latest screenshot. Read it and write my next reply. Output ONLY the reply text.`
+        : `Analyze this chat screenshot and write my next reply. Output ONLY the reply text.`;
+
     } else {
       modelToUse = 'bytedance-1.6';
-      userPrompt = `Write my reply to this message. Output ONLY the reply text, nothing else.\n\nMessage:\n"${content}"`;
+
+      userPrompt = conversationHistory
+        ? `Here is our conversation so far:\n${conversationHistory}\n\nTheir latest message: "${content}"\n\nWrite my next reply. Output ONLY the reply text, nothing else.`
+        : `Write my reply to this message. Output ONLY the reply text, nothing else.\n\nMessage:\n"${content}"`;
     }
 
-    // ✅ KEY FIX: pass forceJson: false so ai_engine skips response_format enforcement
     const aiResult = await generateCompletion({
       model: modelToUse,
       prompt: userPrompt,
       system_prompt: aiSystemPrompt,
       images: imagesPayload,
-      forceJson: false,   // <-- plain text reply, no JSON schema
+      forceJson: false,
     });
 
     if (!aiResult.success) throw new Error(aiResult.error);
 
-    // aiResult.data is a plain string when forceJson is false
     let finalReply = (aiResult.data || '').trim();
 
     // Strip any wrapping quotes the model might add
     if (finalReply.startsWith('"') && finalReply.endsWith('"')) {
       finalReply = finalReply.slice(1, -1).trim();
     }
-    // Strip markdown bold/italic the model sometimes adds
+    // Strip markdown bold/italic wrappers
     finalReply = finalReply.replace(/^\*+|\*+$/g, '').trim();
 
     if (!finalReply) throw new Error('AI returned an empty reply.');