feat: integrate Claude API via zunef gateway with JSON parsing

- Add Claude API support via requests library (bypass Cloudflare blocks)
- Implement manual JSON parsing for Claude models (no structured output support)
- Fix system prompt extraction and injection in AnthropicWrapper
- Update classifier/stylist prompts with JSON format requirements
- Fix graph.py to handle both dict and object outputs from LLM
- Add debug logging for raw API responses
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/agent/lead_stage_agent/graph.py

@@ -228,15 +228,13 @@ class LeadStageGraph:
 
         # AI 1: Classifier — structured output, NHẸ (streaming=False)
         _classifier_base = create_llm(model_name=self.model_name, streaming=False)
-        self.classifier_llm = _classifier_base.with_structured_output(
-            ClassifierOutput, method="function_calling"
-        )
+        self.classifier_llm = _classifier_base
+        # Note: Claude models bypass structured output, parse JSON manually
 
         # AI 2: Stylist — structured output, NẶNG (streaming=False vì structured)
         _stylist_base = create_llm(model_name=self.model_name, streaming=False)
-        self.stylist_llm = _stylist_base.with_structured_output(
-            StylistOutput, method="function_calling"
-        )
+        self.stylist_llm = _stylist_base
+        # Note: Claude models bypass structured output, parse JSON manually
 
         self._compiled = None
         logger.info(f"✅ LeadStageGraph initialized | model: {self.model_name}")
@@ -347,9 +345,33 @@ class LeadStageGraph:
 
         start = time.time()
         try:
-            output: ClassifierOutput = await self.classifier_llm.ainvoke(
+            response = await self.classifier_llm.ainvoke(
                 classifier_messages, config=config
             )
+            # Parse JSON from text response (Claude models return text, not structured output)
+            import json as json_lib
+            text = response.content if hasattr(response, "content") else str(response)
+            logger.info(f"[CLASSIFIER] Raw response repr: {repr(text[:500])}")
+            # Extract JSON from text (may have extra text)
+            text = text.strip()
+            if text.startswith('```json'):
+                text = text[7:]
+            if text.startswith('```'):
+                text = text[3:]
+            if text.endswith('```'):
+                text = text[:-3]
+            text = text.strip()
+
+            try:
+                output = json_lib.loads(text)
+            except json_lib.JSONDecodeError as je:
+                logger.error(f"[CLASSIFIER] JSON parse error: {je}. Text was: {repr(text[:500])}")
+                raise
+            # Convert dict to object-like access
+            class DictToObj:
+                def __init__(self, data):
+                    self.__dict__.update(data)
+            output = DictToObj(output)
         except Exception as e:
             logger.error(f"❌ Classifier error: {e}")
             return {
@@ -364,7 +386,13 @@ class LeadStageGraph:
 
         # ── Resolve tool_args: typed lead_search_args hoac raw dict ──
         if tool_name == "lead_search_tool" and output.lead_search_args:
-            tool_args = output.lead_search_args.model_dump(exclude_none=False)
+            # Handle both dict and Pydantic model cases
+            if isinstance(output.lead_search_args, dict):
+                tool_args = output.lead_search_args
+            elif hasattr(output.lead_search_args, 'model_dump'):
+                tool_args = output.lead_search_args.model_dump(exclude_none=False)
+            else:
+                tool_args = dict(output.lead_search_args) if output.lead_search_args else {}
         elif tool_name == "lead_search_tool" and output.tool_args:
             # fallback: model dung tool_args thay vi lead_search_args
             tool_args = output.tool_args
@@ -512,15 +540,39 @@ class LeadStageGraph:
                 HumanMessage(content="=== KHÔNG CÓ TOOL RESULT === (Khách chào hỏi hoặc câu hỏi chung)")
             )
 
-        # ── LLM Call: Structured Output ──
+        # ── LLM Call: Parse JSON from text (Claude returns text, not structured) ──
         start = time.time()
         try:
-            output: StylistOutput = await self.stylist_llm.ainvoke(
+            response = await self.stylist_llm.ainvoke(
                 stylist_messages, config=config
             )
+            # Parse JSON from text response
+            text = response.content if hasattr(response, "content") else str(response)
+            logger.info(f"[STYLIST] Raw response text (first 500 chars): {repr(text[:500])}")
+            text = text.strip()
+            if text.startswith('```json'):
+                text = text[7:]
+            if text.startswith('```'):
+                text = text[3:]
+            if text.endswith('```'):
+                text = text[:-3]
+            text = text.strip()
+
+            output = json.loads(text)
+            # Convert dict to object-like access
+            class DictToObj:
+                def __init__(self, data):
+                    self.__dict__.update(data)
+                def __getattr__(self, name):
+                    return self.__dict__.get(name)
+            output = DictToObj(output)
+
+            # Handle nested user_insight dict
+            if hasattr(output, "user_insight") and isinstance(output.user_insight, dict):
+                output.user_insight = DictToObj(output.user_insight)
+
         except Exception as e:
-            logger.error(f"❌ Stylist structured output error: {e}")
-            # Fallback: trả response text thuần
+            logger.error(f"❌ Stylist JSON parse error: {e}")
             fallback = "Dạ bạn cho mình hỏi thêm để tư vấn chính xác hơn nhé!"
             return {
                 "messages": [AIMessage(content=fallback)],
@@ -533,8 +585,15 @@ class LeadStageGraph:
 
         elapsed_ms = (time.time() - start) * 1000
 
-        # Extract insight
-        insight_dict = output.user_insight.model_dump()
+        # Extract insight - handle both dict and object cases
+        insight_raw = getattr(output, "user_insight", None)
+        if isinstance(insight_raw, dict):
+            insight_dict = insight_raw
+        elif hasattr(insight_raw, "__dict__"):
+            insight_dict = insight_raw.__dict__
+        else:
+            insight_dict = {}
+
         lead_stage = {
             "stage": insight_dict.get("STAGE_NUM", 1),
             "stage_name": insight_dict.get("STAGE", "BROWSE"),
@@ -543,25 +602,28 @@ class LeadStageGraph:
         }
         stage_injection = format_insight_injection(insight_dict)
 
+        ai_response = getattr(output, "ai_response", "Dạ mình chưa hiểu, bạn nói lại nhé!")
+        product_ids = getattr(output, "product_ids", []) or []
+
         logger.info(
             f"💬 Stylist: stage={insight_dict.get('STAGE')} | goal={insight_dict.get('GOAL')} | "
-            f"response_len={len(output.ai_response)} | products={output.product_ids} | {elapsed_ms:.0f}ms"
+            f"response_len={len(ai_response)} | products={product_ids} | {elapsed_ms:.0f}ms"
         )
 
         diag = {
             "step": "stylist",
             "label": "💬 Stylist (Response + Insight)",
-            "content": output.ai_response[:500],
+            "content": ai_response[:500],
             "elapsed_ms": round(elapsed_ms),
-            "raw_json": json.dumps(output.model_dump(), ensure_ascii=False, indent=2),
+            "raw_json": json.dumps(insight_dict, ensure_ascii=False, indent=2),
         }
 
         return {
-            "messages": [AIMessage(content=output.ai_response)],
+            "messages": [AIMessage(content=ai_response)],
             "updated_insight": insight_dict,
             "lead_stage": lead_stage,
             "stage_injection": stage_injection,
-            "product_ids": output.product_ids or [],
+            "product_ids": product_ids,
             "diagnostics": [diag],
         }
 

backend/agent/lead_stage_agent/prompts.py

@@ -225,6 +225,18 @@ Query: "combo gì cho ngày 2/9"
 → inferred.tags: ["Màu đỏ", "Đi chơi / dạo phố"]
 → inferred.keywords: ["quốc khánh", "2/9", "lễ lớn"]
 ```
+
+**QUAN TRỌNG: BẮT BUỘC OUTPUT FORMAT**
+Bạn PHẢI trả về DUY NHẤT một JSON object hợp lệ, KHÔNG text thêm thắt.
+JSON schema:
+{
+  "reasoning": "lý luận ngắn gọn",
+  "tool_name": null hoặc "lead_search_tool",
+  "lead_search_args": {"literal": {"raw_text": "..."}, "inferred": {"product_line_vn": [], "gender_by_product": null, ...}, "magento_ref_code": null, "reasoning": "..."},
+  "tool_args": null,
+  "ai_response": null hoặc "câu chào hỏi",
+  "product_ids": []
+}
 """
 
 # ═══════════════════════════════════════════════
@@ -389,6 +401,15 @@ Cập nhật 12 trường dưới đây sau mỗi turn. Cộng dồn thông tin.
 - KHÔNG dán link URL
 - KHÔNG tự bịa tính năng không có trong tool results
 - Mở đầu bằng 1 câu đồng cảm dựa trên GOAL
+
+**QUAN TRỌNG: BẮT BUỘC OUTPUT FORMAT**
+Bạn PHẢI trả về DUY NHẤT một JSON object hợp lệ, KHÔNG text thêm thắt.
+JSON schema:
+{
+  "ai_response": "câu trả lời cho khách (max 250 từ)",
+  "product_ids": [],
+  "user_insight": {"USER": "...", "TARGET": "...", "GOAL": "...", ...}
+}
 """
 
 # ═══════════════════════════════════════════════

backend/config.py

@@ -81,7 +81,14 @@ OPENAI_API_KEY: str | None = os.getenv("OPENAI_API_KEY")
 GOOGLE_API_KEY: str | None = os.getenv("GOOGLE_API_KEY")
 GROQ_API_KEY: str | None = os.getenv("GROQ_API_KEY")
 
-DEFAULT_MODEL: str = os.getenv("DEFAULT_MODEL", "gemini-3.1-flash-lite-preview")
+# DS2API (DeepSeek via OpenAI-compatible wrapper)
+DS2API_BASE_URL: str | None = os.getenv("DS2API_BASE_URL")
+DS2API_API_KEY: str | None = os.getenv("DS2API_API_KEY")
+
+# Claude API (via zunef gateway)
+CLAUDE_BASE_URL: str | None = os.getenv("CLAUDE_BASE_URL", "https://claude-api.zunef.com/v1/ai")
+
+DEFAULT_MODEL: str = os.getenv("DEFAULT_MODEL", "claude-sonnet-4-6")
 # DEFAULT_MODEL: str = os.getenv("DEFAULT_MODEL", "gpt-5.1-codex-mini")
 
 # ====================== JWT CONFIGURATION ======================