learning_hub/reflector.py

# learning_hub/reflector.py
# (محدث بالكامل - V2 - تمرير الأخبار للتعلم السريع)
import json
import traceback
from typing import Dict, Any, TYPE_CHECKING
from .schemas import TraceLog, ReflectorOutput
from .memory_store import MemoryStore

# (استخدام TYPE_CHECKING لتجنب الاستيراد الدائري الفعلي)
if TYPE_CHECKING:
    from LLM import LLMService

class Reflector:
    def __init__(self, llm_service: 'LLMService', memory_store: MemoryStore):
        self.llm_service = llm_service
        self.memory_store = memory_store
        print("✅ Learning Hub Module: Reflector (Fast-Learner) loaded")

    async def analyze_trade_outcome(self, trade_object: Dict[str, Any], close_reason: str):
        """
        Analyzes the trade outcome using LLM to generate a learning 'Delta' (rule).
        (Implements Point 2 & 4 of the 16-point plan)
        """
        try:
            # 1. Create the TraceLog
            # (Note: We assume TradeManager now saves 'market_context_at_decision',
            # 'indicators_at_decision', 'news_text', and 'news_score'
            # inside 'decision_data' when opening the trade)
            
            decision_data = trade_object.get('decision_data', {})
            
            trace_log = TraceLog(
                decision_context=decision_data,
                market_context_at_decision=decision_data.get('market_context_at_decision', {}),
                indicators_at_decision=decision_data.get('indicators_at_decision', {}),
                closed_trade_object=trade_object,
                actual_outcome_reason=close_reason
            )
            
            # 2. Create the Reflector Prompt (Now in English)
            # 🔴 --- START OF CHANGE (V2 - News Learning) --- 🔴
            # (تمرير بيانات الأخبار (التي يفترض أنها في decision_data) إلى الـ prompt)
            news_text_at_decision = decision_data.get('news_text', 'No news data available at decision time.')
            news_score_at_decision = decision_data.get('news_score', 0.0) # (VADER raw score)
            
            prompt = self._create_reflector_prompt(
                trace_log, 
                news_text_at_decision, 
                news_score_at_decision
            )
            # 🔴 --- END OF CHANGE --- 🔴
            
            # 3. Call the LLM
            response_text = await self.llm_service._call_llm(prompt)
            
            if not response_text:
                raise ValueError("Reflector LLM call returned no response.")

            # 4. Parse the response
            # (We use the enhanced parser from LLM.py which handles JSON)
            reflector_json = self.llm_service._parse_llm_response_enhanced(
                response_text, 
                fallback_strategy="reflection", 
                symbol=trade_object.get('symbol', 'N/A')
            )
            
            if not reflector_json:
                raise ValueError(f"Failed to parse Reflector LLM response: {response_text}")

            # (Validate against the strict schema from schemas.py)
            reflector_output = ReflectorOutput(**reflector_json)
            
            # 5. Determine the 'Domain' for the Delta
            strategy = trade_object.get('strategy', 'general')
            domain = self._determine_domain(strategy, reflector_output.error_mode)
            
            # 6. Save the suggested 'Delta' to the Memory Store
            # (MemoryStore will use PolicyEngine to decide on auto-approval)
            await self.memory_store.save_new_delta(
                reflector_output=reflector_output,
                trade_object=trade_object,
                domain=domain
            )
            
            print(f"✅ [Reflector] Successfully analyzed {trade_object.get('symbol')}. New Delta created.")
            
        except Exception as e:
            print(f"❌ [Reflector] Failed to analyze trade outcome for {trade_object.get('symbol')}: {e}")
            traceback.print_exc()

    def _determine_domain(self, strategy: str, error_mode: str) -> str:
        """Determines the domain the suggested Delta belongs to."""
        error_mode = error_mode.lower()
        if "pattern" in error_mode or "triangle" in error_mode or "flag" in error_mode:
            return "pattern"
        if "indicator" in error_mode or "rsi" in error_mode or "macd" in error_mode:
            return "indicator"
        if "monte_carlo" in error_mode or "garch" in error_mode or "simulation" in error_mode:
            return "monte_carlo"
        
        # 🔴 --- START OF CHANGE (V2 - News Learning) --- 🔴
        if "news" in error_mode or "sentiment" in error_mode or "sec" in error_mode:
            return "general" # (أو يمكننا إنشاء مجال "news" جديد)
        # 🔴 --- END OF CHANGE --- 🔴
        
        if "strategy" in error_mode or "exit" in error_mode or "entry" in error_mode:
            return "strategy"
        
        # Default to the strategy's domain
        if strategy in ["trend_following", "mean_reversion", "breakout_momentum"]:
             return "strategy"
             
        return "general"

    # 🔴 --- START OF CHANGE (V2 - News Learning) --- 🔴
    def _create_reflector_prompt(
        self, 
        trace_log: TraceLog, 
        news_text: str, 
        news_score: float
    ) -> str:
    # 🔴 --- END OF CHANGE --- 🔴
        """
        Creates the (English-only) prompt for the LLM to act as a Reflector.
        (Implements Point 4 - Reflector prompt)
        """
        
        trade = trace_log.closed_trade_object
        pnl_percent = trade.get('pnl_percent', 0)
        
        # Determine initial success
        is_success = pnl_percent > 0.1 # (Consider any small profit a success)
        
        # 🔴 --- START OF CHANGE (V2 - News Learning) --- 🔴
        # (إضافة قسم الأخبار إلى الـ prompt)
        news_context_section = f"""
4.  **News Context (at entry):**
    * VADER Score (Raw): {news_score:.4f}
    * News Text: {news_text}
"""
        # 🔴 --- END OF CHANGE --- 🔴
        
        prompt = f"""
SYSTEM: You are an expert trading analyst Reflector. Your task is to analyze a completed trade "Trace" and determine the cause of success or failure. You must suggest a concise "Rule" (Delta) (max 25 words) to improve future performance.

--- TRACE LOG START ---

1.  **Original Decision Context (What we decided):**
    * Strategy Used: {trade.get('strategy', 'N/A')}
    * Exit Profile: {trade.get('decision_data', {}).get('exit_profile', 'N/A')}
    * Reasoning (at entry): {trade.get('decision_data', {}).get('reasoning', 'N/A')[:200]}...
    * Entry Price: {trade.get('entry_price')}
    * Initial Stop Loss: {trade.get('stop_loss')}
    * Initial Take Profit: {trade.get('take_profit')}

2.  **Environment Context (When we decided):**
    * Market Context: {json.dumps(trace_log.market_context_at_decision)}
    * Key Indicators: {json.dumps(trace_log.indicators_at_decision)}

3.  **Actual Outcome (What happened):**
    * Close Price: {trade.get('close_price')}
    * Final PnL: {pnl_percent:+.2f}%
    * Close Reason: {trace_log.actual_outcome_reason}
    * Trade Duration: {trade.get('hold_duration_minutes', 'N/A')} minutes

{news_context_section}
--- TRACE LOG END ---

TASK: Analyze the Trace above.
1.  Compare the "Actual Outcome" with the "Original Decision Context".
2.  **Crucially, review the "News Context".** Did the market react as the VADER score predicted? Did the news text contain critical information that was missed?
3.  Identify the primary "Error Mode" (e.g., 'ignored_negative_news', 'premature_exit') or "Success Factor" (e.g., 'correct_pattern_identification').
4.  Suggest ONE concise "Rule" (Delta) (max 25 words) to improve performance. If the news was the cause, the rule MUST mention news.

OUTPUT FORMAT (JSON Only - Adhere strictly to this schema):
{{
  "success": {str(is_success).lower()},
  "score": 0.0,
  "error_mode": "Short description of the error mode (e.g., 'ignored_negative_news_SEC_investigation').",
  "suggested_rule": "The concise 25-word rule (e.g., 'If news contains 'SEC' or 'investigation', do not BUY regardless of technicals.').",
  "confidence": 0.0 
}}
"""
        return prompt