agent.py

import logging
import asyncio
import json
import uuid
import os
from datetime import datetime
from zoneinfo import ZoneInfo
from typing import Annotated, Optional, AsyncIterable, Any, Dict
import random
import http.server
import socketserver
import threading

from dotenv import load_dotenv
from livekit import rtc
from livekit.agents import (
    AutoSubscribe,
    JobContext,
    JobProcess,
    WorkerOptions,
    cli,
    llm,
    AgentSession,
    metrics,
    MetricsCollectedEvent,
    Agent,
)
from livekit.agents.llm import function_tool
from livekit.agents.voice import (
    RunContext,
    ModelSettings,
)
from livekit.plugins import openai, deepgram, cartesia, silero, groq

# Groq SDK for summary generation
from groq import Groq as GroqClient

# Monitoring and validation imports
import sentry_sdk
from logger import logger
from validators import validate_phone_number, validate_appointment_time, validate_purpose, validate_appointment_id

# Try to import Beyond Presence plugin if available
try:
    from livekit.plugins import bey
    BEY_AVAILABLE = True
except ImportError:
    BEY_AVAILABLE = False
    logging.warning("Beyond Presence plugin not available. Install with: pip install \"livekit-agents[bey]\"")

from db import Database

load_dotenv()

# Initialize Sentry for error tracking
if os.getenv("SENTRY_DSN"):
    sentry_sdk.init(
        dsn=os.getenv("SENTRY_DSN"),
        traces_sample_rate=0.1,
        environment=os.getenv("ENVIRONMENT", "production")
    )
    print("✅ Sentry error tracking enabled")

logger = logging.getLogger("voice-agent")
logger.setLevel(logging.INFO)

# Suppress noisy logs from libraries
logging.getLogger("hpack").setLevel(logging.WARNING)
logging.getLogger("httpx").setLevel(logging.WARNING)
logging.getLogger("livekit").setLevel(logging.INFO)
logging.getLogger("urllib3").setLevel(logging.WARNING)

def get_groq_api_key():
    """Rotate between multiple Groq API keys if available to avoid rate limits."""
    keys_str = os.getenv("GROQ_API_KEYS", "")
    if keys_str:
        keys = [k.strip() for k in keys_str.split(",") if k.strip()]
        if keys:
            chosen = random.choice(keys)
            print(f"DEBUG: Selected Groq Key from list of {len(keys)}. Prefix: {chosen[:5]}...")
            return chosen
    
    single_key = os.getenv("GROQ_API_KEY")
    if single_key:
         print(f"DEBUG: Using single GROQ_API_KEY. Prefix: {single_key[:5]}...")
         return single_key
    
    print("DEBUG: No Groq API Key found!")
    return None

try:
    from flagsmith import Flagsmith
    flagsmith = Flagsmith(environment_key=os.getenv("FLAGSMITH_ENVIRONMENT_KEY", "default"))
except Exception:
    flagsmith = None

# ... (omitting lines for brevity)





SYSTEM_PROMPT = """
You are the SkyTask Clinic Assistant, a friendly and capable voice receptionist.

# User: {name} | Status: {status} | Goal: {goal_instruction}
# Rules
- Voice response: Plain text only. Natural and polite.
- Be warm: Use "Good morning", "Thank you", "Please".
- Length: 1-3 sentences, but don't be robotic.
- Speak nums: "five five five". No emojis/markdown.
- Address user by name if known.
# Flow
1. Identify user (ask phone/name).
2. Tools: book_appointment, check_slots, retrieve_appointments, cancel/modify, summarize_call, end_conversation.
   - STRICT: Only call these tools. Do NOT invent new tools.
   - Do NOT speak tool names. Execute silently.
   - summarize_call: When user asks "summarize" or "recap" - gives summary but continues call
   - end_conversation: When user says "end call", "goodbye", "bye" - ends the call
3. Verify name mismatches.
# Guardrails
- Privacy protection active.
- Scope: Clinic appointments only.
"""

class Assistant(Agent):
    def __init__(self, db: Database, user_context: dict, room):
        current_time_ist = datetime.now(ZoneInfo("Asia/Kolkata")).strftime("%Y-%m-%d %I:%M %p")
        
        # Initialize with Guest state
        instructions = SYSTEM_PROMPT.format(
            name="Guest",
            status="Unidentified",
            goal_instruction="Ask for their phone number (and name) to pull up their file. Say: 'Hi! I'm the clinic assistant. May I have your phone number to get started?'"
        )
        instructions += f"\n\nCurrent time (IST): {current_time_ist}"
        
        super().__init__(instructions=instructions)
        self.db = db
        self.user_context = user_context
        self.room = room
        self.current_time_str = current_time_ist
        self.should_disconnect = False
        
        # References needed for summary generation (set later in entrypoint)
        self.usage_collector = None
        self.assistant = None
        self.start_time = datetime.now()
        self.avatar_type = None
        self.tts_provider = None
        
        # Prevent duplicate summaries
        self.summary_generated = False
        
        # Listen for data messages from frontend (e.g., End Call button)
        @room.on("data_received")
        def on_data_received(data_packet):
            try:
                payload = data_packet.data.decode('utf-8')
                data = json.loads(payload)
                
                if data.get("type") == "request_end_call":
                    logger.info("🔴 Frontend requested end call via button - triggering end_conversation")
                    # Trigger the end_conversation tool asynchronously
                    asyncio.create_task(self.end_conversation("User clicked End Call button"))
            except Exception as e:
                logger.warning(f"Error processing frontend data message: {e}")
    
    def update_instructions_with_name(self, name: str):
        """Update the agent's instructions to include the user's name"""
        try:
            # Re-format with REAL name
            new_instructions = SYSTEM_PROMPT.format(
                name=name,
                status="Authenticated",
                goal_instruction=f"Help {name} with appointments. Address them as {name}."
            )
            full_instructions = f"{new_instructions}\n\nCurrent time (IST): {self.current_time_str}"
            
            # Update the agent's instructions
            self._instructions = full_instructions
            
            print(f"✅ Updated agent instructions with user name: {name}")
            print(f"🔍 DEBUG - NEW PROMPT:\n{new_instructions}")
            return True
        except Exception as e:
            print(f"Failed to update instructions: {e}")
            return False

    # ... (omitting lines) ...

    @function_tool()
    async def identify_user(
        self, 
        contact_number: str
    ):
        """Identify the user by their phone number. Only call this when you have received a numeric phone number.
        
        Args:
            contact_number: The user's contact phone number (e.g. 555-0101). Do not provide an empty string.
        """
        if not contact_number or len(contact_number.strip()) < 3:
            return "Error: A valid contact number is required to identify the user."

        try:
            contact_number = validate_phone_number(contact_number)
        except ValueError as e:
            return f"Error: {str(e)}"

        await self._emit_frontend_event("identify_user", "started", {"contact_number": contact_number})
        logger.info(f"Identifying user with number: {contact_number}")
        user = self.db.get_user(contact_number)
        if not user:
            user = self.db.create_user(contact_number)
            is_new = True
        else:
            is_new = False
        
        self.user_context["contact_number"] = contact_number
        self.user_context["user_name"] = user.get("name", "User")
        
        name = user.get('name', 'User')
        
        # Update the agent's instructions to include the user's name
        self.update_instructions_with_name(name)
        
        # ALSO inject a system message into the chat context
        # This ensures the LLM knows the name in the conversation history
        if hasattr(self, 'chat_ctx') and self.chat_ctx:
            try:
                self.chat_ctx.items.append(
                    llm.ChatMessage(
                        role="system",
                        content=[f"IMPORTANT: The user's name is {name}. You MUST address them as {name} in all future responses. When they ask 'what's my name' or 'do you know my name', respond with 'Yes, {name}, your name is {name}.'"]
                    )
                )
                print(f"✅ Injected name '{name}' into chat context")
            except Exception as e:
                print(f"Could not inject into chat context: {e}")
        
        # Return a message that FORCES the agent to say the name immediately
        result_msg = f"User identified successfully. Their name is {name}. You MUST immediately respond by saying: 'Great to meet you, {name}! How can I help you today?' Use their name {name} in your response right now."
        await self._emit_frontend_event("identify_user", "success", result={"name": name, "is_new": is_new})
        return result_msg

    @function_tool()
    async def verify_identity(
        self, 
        contact_number: str,
        stated_name: str
    ):
        """Verify the user's identity using both their phone number and stated name. 
        Use this when the user provides both pieces of information.
        
        Args:
            contact_number: The user's phone number (numeric).
            stated_name: The name the user introduced themselves with.
        """
        if not contact_number or len(contact_number.strip()) < 3:
            return "Error: A valid contact number is required."

        try:
            contact_number = validate_phone_number(contact_number)
        except ValueError as e:
            return f"Error: {str(e)}"

        await self._emit_frontend_event("verify_identity", "started", {"contact_number": contact_number, "name": stated_name})
        logger.info(f"Verifying identity: {stated_name} with {contact_number}")
        
        user = self.db.get_user(contact_number)
        
        if not user:
            # New user case with name provided
            user = self.db.create_user(contact_number, name=stated_name)
            is_new = True
            db_name = stated_name
            match = True
        else:
            is_new = False
            db_name = user.get("name", "User")
            # Simple fuzzy match check (case insensitive)
            match = stated_name.lower() in db_name.lower() or db_name.lower() in stated_name.lower()
        
        self.user_context["contact_number"] = contact_number
        self.user_context["user_name"] = db_name
        
        # Update system with the CORRECT name from DB (or new name)
        self.update_instructions_with_name(db_name)
        
        if match:
             # ALSO inject a system message into the chat context
             # NOTE: Disabled - chat_ctx is read-only, agent instructions are sufficient
            # if hasattr(self, 'chat_ctx') and self.chat_ctx:
            #     try:
            #         self.chat_ctx.items.append(
            #             llm.ChatMessage(
            #                 role="system",
            #                 content=[f"IMPORTANT: Identity verified. User is {db_name}. Address them as {db_name}."]
            #             )
            #         )
            #     except Exception:
            #         pass
            
            result_msg = f"Identity verified! The user is indeed {db_name}. Greet them naturally as {db_name}."
            await self._emit_frontend_event("verify_identity", "success", result={"name": db_name, "match": True})
            return result_msg
        else:
            # Name mismatch logic
            result_msg = f"Identity Mismatch Warning: The phone number belongs to '{db_name}', but user said '{stated_name}'. politely ask: 'I have this number registered under {db_name}. Are you {db_name}?'"
            await self._emit_frontend_event("verify_identity", "warning", result={"db_name": db_name, "stated_name": stated_name, "match": False})
            return result_msg

    async def _emit_frontend_event(self, tool_name: str, status: str, args: dict = None, result: dict = None):
        try:
            payload = json.dumps({
                "type": "tool_call",
                "tool": tool_name,
                "status": status,
                "args": args,
                "result": result
            })
            await self.room.local_participant.publish_data(payload, reliable=True)
        except Exception as e:
            logger.error(f"Failed to emit frontend event: {e}")

    @function_tool()
    async def hello(self, response: str = ""):
        """This tool is used for greetings.
        
        Args:
            response: The greeting response.
        """
        return "Hello! How can I help you today?"

    @function_tool()
    async def identify_user(
        self, 
        contact_number: str
    ):
        """Identify the user by their phone number. Only call this when you have received a numeric phone number.
        
        Args:
            contact_number: The user's contact phone number (e.g. 555-0101). Do not provide an empty string.
        """
        if not contact_number or len(contact_number.strip()) < 3:
            return "Error: A valid contact number is required to identify the user."

        try:
            contact_number = validate_phone_number(contact_number)
        except ValueError as e:
            return f"Error: {str(e)}"

        await self._emit_frontend_event("identify_user", "started", {"contact_number": contact_number})
        logger.info(f"Identifying user with number: {contact_number}")
        user = self.db.get_user(contact_number)
        if not user:
            user = self.db.create_user(contact_number)
            is_new = True
        else:
            is_new = False
        
        self.user_context["contact_number"] = contact_number
        self.user_context["user_name"] = user.get("name", "User")
        
        # Helper comment: Name will now be picked up by the LLM from the tool return value
        # and usage enforced by updated system prompts.
        
        result_msg = f"User identified. Name: {user.get('name')}. New user: {is_new}."
        await self._emit_frontend_event("identify_user", "success", result={"name": user.get('name'), "is_new": is_new})
        return result_msg

    @function_tool()
    async def fetch_slots(self, location: str):
        """Fetch available appointment slots.
        
        Args:
            location: The clinic location to check (e.g. 'main', 'downtown').
        """
        logger.info(f"Fetching available slots for {location}")
        await self._emit_frontend_event("fetch_slots", "started", {"location": location})
        
        # Use DB method to fetch slots (real or mock)
        available_slots = self.db.get_available_slots()
        slots_json = json.dumps(available_slots)
        
        await self._emit_frontend_event("fetch_slots", "success", result=available_slots)
        return slots_json

    @function_tool()
    async def book_appointment(
        self,
        time: str,
        purpose: str
    ):
        """Book an appointment for the identified user.
        
        Args:
            time: The ISO 8601 formatted date and time for the appointment.
            purpose: Purpose of the appointment.
        """
        await self._emit_frontend_event("book_appointment", "started", {"time": time, "purpose": purpose})
        contact_number = self.user_context.get("contact_number")
        if not contact_number:
            return "Error: User not identified. Please ask for phone number first."

        try:
            contact_number = validate_phone_number(contact_number)
        except ValueError as e:
            return f"Error validation phone: {str(e)}"

        logger.info(f"Booking appointment for {contact_number} at {time}")
        
        is_available = self.db.check_slot_availability(datetime.fromisoformat(time))
        if not is_available:
             return "Error: Slot not available."

        result = self.db.book_appointment(contact_number, time, purpose)
        if result:
            await self._emit_frontend_event("book_appointment", "success", result=result)
            return f"Appointment booked successfully. ID: {result.get('id')}"
        else:
            await self._emit_frontend_event("book_appointment", "failed")
            return "Failed to book appointment."

    @function_tool()
    async def retrieve_appointments(self, user_confirmation: str):
        """Retrieve past and upcoming appointments for the identified user.
        
        Args:
            user_confirmation: The user's confirmation to see their appointments (e.g. 'show them', 'yes').
        """
        await self._emit_frontend_event("retrieve_appointments", "started")
        contact_number = self.user_context.get("contact_number")
        if not contact_number:
            return "Error: User not identified."

        try:
            contact_number = validate_phone_number(contact_number)
        except ValueError as e:
            return f"Error: {str(e)}"
            
        appointments = self.db.get_user_appointments(contact_number)
        if not appointments:
            await self._emit_frontend_event("retrieve_appointments", "success", result=[])
            return "No appointments found."
        
        await self._emit_frontend_event("retrieve_appointments", "success", result=appointments)
        return json.dumps(appointments)

    @function_tool()
    async def cancel_appointment(
        self,
        appointment_id: str
    ):
        """Cancel an appointment.
        
        Args:
            appointment_id: The ID of the appointment to cancel.
        """
        await self._emit_frontend_event("cancel_appointment", "started", {"appointment_id": appointment_id})
        success = self.db.cancel_appointment(appointment_id)
        if success:
            await self._emit_frontend_event("cancel_appointment", "success", result={"id": appointment_id})
            return "Appointment cancelled successfully."
        else:
            await self._emit_frontend_event("cancel_appointment", "failed")
            return "Failed to cancel appointment."

    @function_tool()
    async def modify_appointment(
        self,
        appointment_id: str,
        new_time: str
    ):
        """Modify the date/time of an appointment.
        
        Args:
            appointment_id: The ID of the appointment to modify.
            new_time: The new ISO 8601 formatted date and time.
        """
        await self._emit_frontend_event("modify_appointment", "started", {"appointment_id": appointment_id, "new_time": new_time})
        success = self.db.modify_appointment(appointment_id, new_time)
        if success:
            await self._emit_frontend_event("modify_appointment", "success", result={"id": appointment_id, "new_time": new_time})
            return "Appointment modified successfully."
        else:
            await self._emit_frontend_event("modify_appointment", "failed")
            return "Failed to modify appointment."

    @function_tool()
    async def summarize_call(
        self,
        request: Annotated[str, "User's request for summary"] = "summarize"
    ) -> str:
        """Provide a summary of the current call without ending it.
        
        Use this when the user asks for a summary but wants to continue the conversation.
        Example triggers: "Can you summarize?", "What did we discuss?", "Recap please"
        
        Args:
            request: The user's request for a summary (e.g., "summarize", "recap")
        
        Returns:
            str: A spoken summary of the conversation so far.
        """
        logger.info(f"Generating mid-call summary (not ending): {request}")
        
        # Get context and metrics
        contact = self.user_context.get("contact_number")
        if not contact:
            return "So far, we've discussed your appointments. Is there anything else I can help you with?"
        
        # Collect usage metrics
        summary = self.usage_collector.get_summary()
        usage_stats = {
            "stt_duration": summary.stt_audio_duration,
            "llm_prompt_tokens": summary.llm_prompt_tokens,
            "llm_completion_tokens": summary.llm_completion_tokens,
            "tts_chars": summary.tts_characters_count
        }
        
        duration = (datetime.now() - self.start_time).total_seconds()
        user_name = self.user_context.get("user_name", "the patient")
        
        # Generate summary directly
        try:
            summary_data = await generate_and_save_summary(
                self.db, 
                self.assistant.chat_ctx, 
                contact, 
                duration, 
                self.avatar_type, 
                self.tts_provider, 
                user_name, 
                usage_stats
            )
            if summary_data and isinstance(summary_data, dict):
                spoken_summary = summary_data.get("spoken_text", "So far, we've discussed your appointments.")
                logger.info(f"Mid-call summary: {spoken_summary}")
                return spoken_summary
        except Exception as e:
            logger.error(f"Failed to generate mid-call summary: {e}")
        
        return "So far, we've discussed your appointments. Is there anything else I can help you with?"

    @function_tool()
    async def end_conversation(self, summary_request: str):
        """End the current conversation session and generate a final summary.
        
        Args:
            summary_request: The user's request to end or wrap up (e.g. 'bye', 'summarize', 'we're done').
        """
        logger.info("Ending conversation - generating summary first")
        
        # GUARD: Prevent duplicate summaries
        if self.summary_generated:
            logger.warning("Summary already generated - skipping duplicate generation")
            return "Thank you for calling. Goodbye!"
        
        spoken_text = "Thank you for calling. Have a great day!"
        summary_sent = False
        
        # Get context and metrics
        contact = self.user_context.get("contact_number")
        if contact:
            # Collect usage metrics
            summary = self.usage_collector.get_summary()
            usage_stats = {
                "stt_duration": summary.stt_audio_duration,
                "llm_prompt_tokens": summary.llm_prompt_tokens,
                "llm_completion_tokens": summary.llm_completion_tokens,
                "tts_chars": summary.tts_characters_count
            }
            
            duration = (datetime.now() - self.start_time).total_seconds()
            user_name = self.user_context.get("user_name", "the patient")
            
            # Generate summary directly
            try:
                summary_data = await generate_and_save_summary(
                    self.db,
                    self.assistant.chat_ctx,
                    contact,
                    duration,
                    self.avatar_type,
                    self.tts_provider,
                    user_name,
                    usage_stats
                )
                if summary_data and isinstance(summary_data, dict):
                    # 1. Get spoken summary
                    spoken_text = summary_data.get("spoken_text", spoken_text)
                    
                    # 2. Publish structured data to frontend
                    payload = json.dumps({
                        "type": "summary",
                        "summary": summary_data
                    })
                    await self.room.local_participant.publish_data(payload, reliable=True)
                    logger.info("Summary sent to frontend")
                    summary_sent = True
                    
                    # Mark summary as generated to prevent duplicates
                    self.summary_generated = True
                    
                    # CRITICAL: Send close_session to trigger auto-disconnect for voice UX
                    # Small delay to ensure summary is received first
                    await asyncio.sleep(0.1)
                    close_payload = json.dumps({"type": "close_session"})
                    await self.room.local_participant.publish_data(close_payload, reliable=True)
                    logger.info("✅ close_session sent - UI will auto-disconnect")
                    
            except Exception as e:
                logger.error(f"Failed to process summary: {e}")
        
        # CRITICAL: If summary wasn't sent, send fallback with at least cost structure
        if not summary_sent:
            logger.warning("Sending fallback summary with cost placeholder")
            fallback = {
                "content": "Call ended. See cost breakdown below.",
                "spoken_text": spoken_text,
                "costs": {"stt": 0.0, "tts": 0.0, "llm": 0.0, "avatar": 0.0, "total": 0.0},
                "status": "fallback"
            }
            try:
                payload = json.dumps({"type": "summary", "summary": fallback})
                await self.room.local_participant.publish_data(payload, reliable=True)
                logger.info("Fallback summary sent to frontend")
            except Exception as e:
                logger.error(f"Failed to send fallback: {e}")
        
        # NOTE: Don't send close_session here - let frontend's 2-second timer handle disconnect
        # This ensures the summary data channel message is received before disconnect
        
        # 4. Request disconnect implicitly by setting flag
        # The session listener will handle the actual disconnect after speech ends
        self.should_disconnect = True
        logger.info("Disconnect requested - waiting for speech to finish")
        
        # Start safeguard immediately
        asyncio.create_task(self.safeguard_disconnect())
        
        # Return the simplified spoken text for the agent to say immediately
        return spoken_text

    async def safeguard_disconnect(self):
        """Force disconnect if normal flow fails."""
        logger.info("Safeguard: Timer started (10s)...")
        await asyncio.sleep(10.0)
        
        state = self.room.connection_state
        logger.info(f"Safeguard: Timeout reached. Room state is: {state}")
        
        if state == "connected":
            logger.warning("Safeguard: Timed out. Sending close_session event.")
            try:
                payload = json.dumps({"type": "close_session"})
                await self.room.local_participant.publish_data(payload, reliable=True)
                logger.info("Safeguard: close_session event sent.")
            except Exception as e:
                logger.warning(f"Safeguard: Failed to send event: {e}")
            
            await asyncio.sleep(3.0) # Give frontend more time to process
            
            if self.room.connection_state == "connected":
                logger.warning("Safeguard: Force disconnecting room now.")
                await self.room.disconnect()
        else:
            logger.info("Safeguard: Room already disconnected, taking no action.")

def calculate_costs(duration_seconds: float, tts_chars: int, avatar_type: str, tts_provider: str, prompt_tokens: int = 0, completion_tokens: int = 0):
    # Rates per unit
    stt_rate = 0.006      # Deepgram Nova-2 ($0.006/min)
    # Rates per unit (USD)
    stt_rate = 0.006      # Deepgram Nova-2 ($0.006/min)
    
    # LLM Pricing: OpenAI GPT-OSS-120B (used for main conversation)
    # Input: $0.15 / 1M tokens
    # Output: $0.60 / 1M tokens
    llm_rate_input = 0.15 / 1_000_000 
    llm_rate_output = 0.60 / 1_000_000
    
    # TTS Rates
    if tts_provider == "cartesia":
        tts_rate = 0.050 / 1000 # Cartesia (~$0.05/1k chars)
        tts_label = "Cartesia"
    elif tts_provider == "deepgram":
        tts_rate = 0.015 / 1000 # Deepgram Aura ($0.015/1k chars)
        tts_label = "Deepgram"
    else: # Groq / Other
        tts_rate = 0.000        # Assume Free/Included
        tts_label = "Groq"

    # Avatar Rates
    avatar_rate = 0.05 if avatar_type == 'bey' else 0 # Beyond Presence (~$0.05/min)

    # Calculate Standard Costs
    stt_cost = (duration_seconds / 60) * stt_rate
    tts_cost = tts_chars * tts_rate

    # Use real counts if provided, otherwise estimate (fallback)
    if prompt_tokens == 0 and completion_tokens == 0:
         # Usage estimates (simplified)
         # Assume 150 words/min -> ~200 tokens/min input
        estimated_input_tokens = (duration_seconds / 60) * 200
        estimated_output_tokens = (tts_chars / 4) # Rough char-to-token ratio
        llm_cost = (estimated_input_tokens * llm_rate_input) + (estimated_output_tokens * llm_rate_output)
    else:
        llm_cost = (prompt_tokens * llm_rate_input) + (completion_tokens * llm_rate_output)
    avatar_cost = (duration_seconds / 60) * avatar_rate
    
    total = stt_cost + tts_cost + llm_cost + avatar_cost
    
    # Log for debugging
    logger.info(f"Cost calculation: duration={duration_seconds}s, tts_chars={tts_chars}, provider={tts_provider}")
    logger.info(f"Costs: STT=${stt_cost:.6f}, TTS=${tts_cost:.6f}, LLM=${llm_cost:.6f}, Avatar=${avatar_cost:.6f}")
    
    return {
        "stt": round(stt_cost, 6),
        "tts": round(tts_cost, 6),
        "llm": round(llm_cost, 6),
        "avatar": round(avatar_cost, 6),
        "total": round(total, 6),
        "currency": "USD",
        "labels": {
            "tts": tts_label,
            "stt": "Deepgram",
            "llm": "Groq/OpenAI",
            "avatar": "Beyond Presence" if avatar_type == 'bey' else "3D Avatar"
        }
    }

async def generate_and_save_summary(db: Database, chat_ctx: llm.ChatContext, contact_number: str, duration: float, avatar_type: str, tts_provider: str, user_name: str = "the patient", usage_stats: dict = None) -> Optional[Dict[str, Any]]:
    if not contact_number:
        logger.warning("No contact number to save summary for.")
        return

    logger.info("Generating conversation summary...")
    
    transcript = ""
    messages_to_save = []
    
    # Try to extract messages from chat context
    try:
        if hasattr(chat_ctx, 'items'):
             items = chat_ctx.items
        elif hasattr(chat_ctx, 'messages'):
             items = chat_ctx.messages
        else:
             items = []
        
        for item in items:
            if isinstance(item, llm.ChatMessage):
                role = item.role
                content = item.content
                
                # Format content for string manipulation
                content_str = content
                if isinstance(content, list):
                    content_str = " ".join([str(c) for c in content])
                
                if isinstance(content_str, str):
                    transcript += f"{role}: {content_str}\n"

                # Prepare for DB
                msg_data = {
                    "role": role,
                    "content": content_str,
                    "tool_name": None,
                    "tool_args": None
                }
                
                # Attempt to extract tool info safely
                if hasattr(item, 'tool_calls') and item.tool_calls:
                     try:
                         tc = item.tool_calls[0]
                         # Handle both object and dict (depending on underlying library version)
                         if isinstance(tc, dict):
                             msg_data["tool_name"] = tc.get('function', {}).get('name')
                             msg_data["tool_args"] = tc.get('function', {}).get('arguments')
                         else:
                             # accessing attributes of ToolCall object
                             fn = getattr(tc, 'function', None)
                             if fn:
                                 msg_data["tool_name"] = getattr(fn, 'name', None)
                                 msg_data["tool_args"] = getattr(fn, 'arguments', None)
                     except Exception:
                         pass # Ignore tool extraction errors
                
                if role == "tool":
                    msg_data["tool_name"] = getattr(item, 'name', getattr(item, 'tool_call_id', None))

                messages_to_save.append(msg_data)
        
        # Save transcript to DB
        if messages_to_save:
            try:
                # Generate a session ID for this conversation batch
                session_id = str(uuid.uuid4())
                db.save_chat_transcript(session_id, contact_number, messages_to_save)
            except Exception as e:
                logger.error(f"Failed to save chat transcript to DB: {e}")

    except Exception as e:
        logger.error(f"Error extracting transcript: {e}")
    
    # Calculate costs using official metrics if available, otherwise fallback
    logger.info(f"Calculating costs with usage_stats: {usage_stats}")
    if usage_stats:
        tts_chars = usage_stats.get("tts_chars", 0)
        prompt_tokens = usage_stats.get("llm_prompt_tokens", 0)
        completion_tokens = usage_stats.get("llm_completion_tokens", 0)
        costs = calculate_costs(duration, tts_chars, avatar_type, tts_provider, prompt_tokens, completion_tokens)
    else:
        # Fallback estimation
        tts_chars = len(transcript) // 2 
        costs = calculate_costs(duration, tts_chars, avatar_type, tts_provider)
    
    logger.info(f"Calculated costs: {costs}")
    
    prompt = (
        f"Summarize the conversation with {user_name} in JSON format.\n"
        f"Transcript:\n{transcript}\n\n"
        "CRITICAL: Use natural time formats like '9 AM' or '2:30 PM', NOT 'nine zero zero hours'\n"
        "Return a valid JSON object with exactly two keys:\n"
        "1. 'spoken': A 1-2 sentence spoken closing for TTS. Natural, human-like, polite. No special chars. Start with 'To recap,'.\n"
        "2. 'written': A detailed bulleted summary for the user interface. Include topics, appointments booked, and outcome.\n"
        "IMPORTANT: Ensure the JSON is valid. Do NOT use unescaped newlines in the 'written' string or 'spoken' string. Use \\n for line breaks.\n"
    )

    max_retries = 3
    retry_delay = 1
    
    for attempt in range(max_retries):
        try:
            # Use Groq SDK directly instead of livekit wrapper for reliability
            api_key = os.getenv("GROQ_API_KEY_SUMMARY") or get_groq_api_key()
            client = GroqClient(api_key=api_key)
            
            # Use llama-3.3-70b-versatile for JSON reliability
            response = client.chat.completions.create(
                model="llama-3.3-70b-versatile",
                messages=[
                    {"role": "system", "content": "You are a helpful assistant. Output valid JSON only. Do not output markdown blocks."},
                    {"role": "user", "content": prompt}
                ],
                temperature=0.7,
                max_tokens=500
            )
            
            full_response = response.choices[0].message.content
            # Summary uses Llama-3.3-70B-Versatile
            # Pricing: Input $0.59/1M, Output $0.79/1M
            summary_input_cost = response.usage.prompt_tokens * (0.59 / 1_000_000)
            summary_output_cost = response.usage.completion_tokens * (0.79 / 1_000_000)
            summary_cost = summary_input_cost + summary_output_cost
            
            logger.info(f"🔍 RAW LLM RESPONSE: {full_response}")
            logger.info(f"💰 Summary LLM cost: ${summary_cost:.6f} ({response.usage.prompt_tokens} + {response.usage.completion_tokens} tokens)")
            
            # Attempt to parse JSON
            spoken = "To recap, we discussed your appointments. Have a great day!"
            written = ""
            
            try:
                # Clean up markdown code blocks if present
                clean_json = full_response.replace("```json", "").replace("```", "").strip()
                
                # Regex heuristic to find the JSON object { ... }
                import re
                match = re.search(r"\{.*\}", clean_json, re.DOTALL)
                if match:
                    clean_json = match.group(0)
                
                data = json.loads(clean_json)
                spoken = data.get("spoken", spoken)
                written = data.get("written", "")
                
            except (json.JSONDecodeError, AttributeError) as e:
                logger.warning(f"Failed to parse JSON summary (standard): {e}. Retrying with Regex Fallback.")
                # Fallback: Regex extraction for common invalid JSON issues (newlines in strings)
                try:
                    import re
                    # Extract spoken
                    s_match = re.search(r'"spoken"\s*:\s*"(.*?)"', clean_json, re.DOTALL)
                    if s_match:
                        spoken = s_match.group(1)
                    
                    # Extract written (greedy to catch multi-line content)
                    w_match = re.search(r'"written"\s*:\s*"(.*?)(?<!\\)"', clean_json, re.DOTALL)
                    if w_match:
                        written = w_match.group(1).replace("\\n", "\n") # Unescape manual newlines
                    else:
                         # Fallback for written if regex fails but we have cleaned string
                         written = clean_json
                except Exception as ex:
                     logger.error(f"Regex fallback failed: {ex}")
                     written = clean_json # Last resort: just show the cleaned text

            # Fallback if written summary is empty
            if not written.strip():
                written = f"Summary: {spoken.strip()}"

            logger.info(f"Spoken Summary: {spoken.strip()}")
            logger.info(f"📝 WRITTEN SUMMARY:\\n{written.strip()}")
            logger.info(f"=" * 80)
            db.save_summary(contact_number, written.strip())
            
            # CRITICAL: Add summary LLM cost to total costs
            costs['llm'] += summary_cost
            costs['total'] += summary_cost
            
            # CRITICAL: Always return costs
            summary_result = {
                "text": written.strip(),
                "content": written.strip(),
                "spoken_text": spoken.strip(),
                "costs": costs,
                "status": "completed"
            }
            logger.info(f"📊 Summary with costs: {summary_result}")
            
            # Print prominently to CLI
            print(f"\\n{'='*80}")
            print(f"📋 CALL SUMMARY GENERATED")
            print(f"{'='*80}")
            print(f"Contact: {contact_number}")
            print(f"Summary: {written.strip()}")
            print(f"Costs: STT=${costs['stt']:.4f} | TTS=${costs['tts']:.4f} | LLM=${costs['llm']:.6f} | Total=${costs['total']:.4f}")
            print(f"{'='*80}\\n")
            
            return summary_result
        
        except Exception as e:
            logger.warning(f"Summary generation attempt {attempt+1} failed: {e}")
            if attempt < max_retries - 1:
                await asyncio.sleep(retry_delay * (2 ** attempt)) # Exponential backoff
            else:
                logger.error("All summary generation attempts failed.")
                return {
                    "text": "Call summary unavailable.",
                    "content": "Call summary unavailable.", 
                    "spoken_text": "Thank you for calling. Have a great day!",
                    "costs": costs,
                    "status": "failed"
                }
    

def prewarm(proc: JobProcess):
    """Prewarm worker to reduce cold start latency"""
    from logger import logger as struct_logger
    from db import Database
    
    struct_logger.info("Prewarming worker...")
    
    try:
        # 1. Initialize database connection
        db = Database()
        proc.userdata["db"] = db
        struct_logger.info("✅ Database connection prewarmed")
        
        # 2. Load VAD model into memory
        proc.userdata["vad"] = silero.VAD.load()
        struct_logger.info("✅ VAD model prewarmed")
        
        # 3. Cache available slots
        proc.userdata["slots"] = db.get_available_slots()
        struct_logger.info("✅ Appointment slots cached")
        
        # 4. Initialize cache connection
        from cache import cache
        proc.userdata["cache"] = cache
        struct_logger.info(f"✅ Redis cache prewarmed (enabled: {cache.enabled})")
        
        struct_logger.info("🚀 Worker prewarmed successfully - ready for calls!")
        
    except Exception as e:
        struct_logger.error(f"Prewarming failed: {e}", error=str(e))

async def entrypoint(ctx: JobContext):
    # 1. Connect immediately to acknowledge assignment (Fixes AssignmentTimeoutError)
    await ctx.connect(auto_subscribe=AutoSubscribe.AUDIO_ONLY)
    
    # 2. Retrieve prewarmed resources or initialize if missing
    if "db" in ctx.proc.userdata:
        db = ctx.proc.userdata["db"]
        logger.info("Using prewarmed Database connection")
    else:
        db = Database()
        logger.info("Initialized new Database connection")

    user_context = {}
    participant = await ctx.wait_for_participant()

    avatar_type = '3d'
    user_tts_pref = None
    if participant.metadata:
        try:
            metadata = json.loads(participant.metadata)
            avatar_type = metadata.get('avatarType', '3d')
            user_tts_pref = metadata.get('ttsProvider')
        except Exception as e:
            logger.warning(f"Failed to parse participant metadata: {e}")
    
    logger.info(f"Avatar type requested by {participant.identity}: {avatar_type}")

    # TTS Provider Selection (Feature Flag via Flagsmith)
    tts_provider = os.getenv("TTS_PROVIDER", "deepgram") # Default fallback from ENV
    try:
        flags = flagsmith.get_environment_flags()
        # Flagsmith Python SDK v3+ uses get_feature_value() not get_flag_value()
        tts_provider_flag = flags.get_feature_value("tts_provider")
        if tts_provider_flag:
            tts_provider = tts_provider_flag
        logger.info(f"Flagsmith: tts_provider={tts_provider}")
    except Exception as e:
        logger.warning(f"Failed to fetch feature flags from Flagsmith: {e}. Using default: {tts_provider}")

    if tts_provider == "cartesia":
        logger.info("Using Cartesia TTS")
        agent_tts = cartesia.TTS()
    elif tts_provider == "groq":
        logger.info("Using Groq TTS")
        agent_tts = groq.TTS(model="canopylabs/orpheus-v1-english")
    else:
        logger.info("Using Deepgram TTS (Default)")
        agent_tts = deepgram.TTS()

    # Initialize metrics collector
    usage_collector = metrics.UsageCollector()

    # Initialize the AgentSession with a faster model and optimized VAD
    session = AgentSession(
        stt=deepgram.STT(),
        llm=groq.LLM(
            model="openai/gpt-oss-120b",
            api_key=get_groq_api_key(),
            temperature=0.5,
        ),
        tts=agent_tts,
        vad=silero.VAD.load(
            min_speech_duration=0.1,
            min_silence_duration=0.5, # Prevents cutting off mid-sentence
            prefix_padding_duration=0.2,     # Fixed deprecated argument
        ),
    )

    @session.on("metrics_collected")
    def _on_metrics_collected(ev: MetricsCollectedEvent):
        # logger.info(f"Metrics collected: {ev.metrics}")
        usage_collector.collect(ev.metrics)

    assistant = Assistant(db, user_context, ctx.room)
    start_time = datetime.now()
    assistant.usage_collector = usage_collector
    assistant.assistant = assistant
    assistant.avatar_type = avatar_type
    assistant.tts_provider = tts_provider

    @session.on("agent_speech_stopped")
    def _on_agent_speech_stopped(ev: Any = None):
        """Disconnect if the agent has finished speaking and a disconnect was requested."""
        if assistant.should_disconnect:
            async def _disconnect_sequence():
                logger.info("Agent finished speaking. Sending close_session event then closing room.")
                try:
                    payload = json.dumps({"type": "close_session"})
                    await ctx.room.local_participant.publish_data(payload, reliable=True)
                    logger.info("close_session event sent to frontend")
                except Exception as e:
                    logger.warning(f"Failed to publish close_session: {e}")
                
                # Give frontend time to process the event and disconnect gracefully
                await asyncio.sleep(2.0)
                
                # Only force disconnect if still connected
                if ctx.room.connection_state == "connected":
                    logger.info("Frontend didn't disconnect, forcing disconnect")
                    await ctx.room.disconnect()
                else:
                    logger.info("Frontend disconnected gracefully")

            asyncio.create_task(_disconnect_sequence())

    @session.on("agent_speech_interrupted")
    def _on_agent_speech_interrupted(ev: Any = None):
        """Handle case where agent summary/goodbye is interrupted by noise/user."""
        if assistant.should_disconnect:
            logger.info("Agent speech interrupted during disconnect phase. Triggering disconnect sequence.")
            # Reuse the same disconnect logic
            _on_agent_speech_stopped(ev)

    @session.on("agent_speech_started")
    def _on_agent_speech_started(ev: Any = None):
        logger.info("Agent speech STARTED.")


    await session.start(room=ctx.room, agent=assistant)
    
    
    # NOTE: Session ready signal will be sent after avatar + greeting (line ~1051)
    # This ensures UI doesn't show 'Ready' before system is actually ready
    # NOTE: Moved session_ready to after avatar + greeting (line ~1051)
    # This ensures UI doesn't show 'Ready' before system is actually ready


    # If Beyond Presence avatar is requested and available, initialize it
    if avatar_type == 'bey' and BEY_AVAILABLE:
        logger.info("Initializing Beyond Presence avatar...")
        
        # Send initializing signal repeatedly to ensure frontend gets it
        # (Data channel might not be fully established for 'User' yet)
        async def send_init_signal():
            for _ in range(5):
                try:
                    await ctx.room.local_participant.publish_data(
                        json.dumps({"type": "avatar_initializing"}), 
                        reliable=True
                    )
                except: pass
                await asyncio.sleep(0.5)
        
        asyncio.create_task(send_init_signal())
            
        try:
            bey_session = bey.AvatarSession(
                api_key=os.environ.get("BEYOND_PRESENCE_API_KEY"),
                avatar_id=os.environ.get("BEYOND_PRESENCE_AVATAR_ID", "b9be11b8-89fb-4227-8f86-4a881393cbdb"), 
            )
            await bey_session.start(session, room=ctx.room)
            logger.info("Beyond Presence avatar started successfully (API level)")
            
            # Wait for the avatar participant to actually join the room and publish tracks
            # This ensures we don't greet while the user is still looking at a loading screen
            logger.info("Waiting for avatar participant to join room...")
            avatar_joined = False
            for _ in range(40): # Wait up to 40 seconds
                # Check if avatar is in remote participants
                # Note: identity might vary setup, but usually 'bey-avatar-agent' or similar
                # We check for ANY new participant that looks like an avatar if specific ID fails?
                # For now assume 'bey-avatar-agent'
                p = ctx.room.remote_participants.get("bey-avatar-agent")
                if p:
                    # Check if they have video track
                    video_tracks = [t for t in p.track_publications.values() if t.kind == rtc.TrackKind.KIND_VIDEO]
                    if video_tracks:
                        logger.info("✅ Avatar participant joined and video track found!")
                        avatar_joined = True
                        break
                await asyncio.sleep(1)
            
            if not avatar_joined:
                logger.warning("Timed out waiting for avatar participant to join - proceeding anyway")
                
        except Exception as e:
            logger.error(f"Failed to start Beyond Presence avatar: {e}")
            logger.info("Falling back to audio-only mode")

    # Time-aware greeting
    hour = datetime.now(ZoneInfo("Asia/Kolkata")).hour
    if 5 <= hour < 12:
        greeting = "Good morning"
    elif 12 <= hour < 17:
        greeting = "Good afternoon"
    else:
        greeting = "Good evening"
        
    # Generate greeting ONLY if session is still active
    # (Beyond Presence avatar takes ~12s to init, user might disconnect)
    # Generate greeting
    # We use room connection state as the truth, since session._state might be internal/laggy
    # Check against the Enum value properly
    if ctx.room.connection_state == rtc.ConnectionState.CONN_CONNECTED:
        try:
            logger.info(f"Speaking greeting: {greeting}...")
            # Use .say() directly for instant response
            await session.say(
                f"{greeting}, thank you for calling SkyTask Clinic. May I have your phone number?", 
                allow_interruptions=True
            )
        except RuntimeError as e:
            logger.warning(f"Could not speak greeting - error: {e}")
    else:
        logger.warning("Session not running - skipping greeting (user may have disconnected)")

    # Always send session_ready if we reached here
    try:
        payload = json.dumps({"type": "session_ready"})
        await ctx.room.local_participant.publish_data(payload, reliable=True)
        logger.info("✅ Session ready signal sent to frontend")
    except Exception as e:
        logger.warning(f"Failed to send session_ready: {e}")



    # CRITICAL: Keep the agent alive while connected
    while ctx.room.connection_state == "connected":
        await asyncio.sleep(1)

    contact_number = user_context.get("contact_number")
    if contact_number and not assistant.summary_generated:
         logger.info("Disconnect summary generation (backup)...")
         duration = (datetime.now() - start_time).total_seconds()
         user_name = user_context.get("user_name", "the patient")
         await generate_and_save_summary(db, assistant.chat_ctx, contact_number, duration, avatar_type, tts_provider, user_name)


def start_health_check_server():
    """Starts a simple HTTP server for health checks."""
    try:
        port = int(os.getenv("PORT", 8080))
        
        class HealthCheckHandler(http.server.BaseHTTPRequestHandler):
            def do_GET(self):
                if self.path == "/health" or self.path == "/":
                    self.send_response(200)
                    self.send_header("Content-type", "application/json")
                    self.end_headers()
                    self.wfile.write(b'{"status": "healthy"}')
                else:
                    self.send_response(404)
                    self.end_headers()
            
            def log_message(self, format, *args):
                pass  # Suppress logs to keep console clean

        # Allow reuse of address to prevent 'Address already in use' errors
        socketserver.TCPServer.allow_reuse_address = True
        
        httpd = socketserver.TCPServer(("", port), HealthCheckHandler)
        print(f"✅ Health check server listening on port {port}")
        
        # Run in a daemon thread so it doesn't block program exit
        thread = threading.Thread(target=httpd.serve_forever, daemon=True)
        thread.start()
    except Exception as e:
        print(f"⚠️ Failed to start health check server: {e}")

if __name__ == "__main__":
    start_health_check_server()
    # HF Space has ~16GB RAM, so we can enable prewarming for faster response times
    cli.run_app(WorkerOptions(entrypoint_fnc=entrypoint, prewarm_fnc=prewarm))

    # Low resource configuration (Render Free Tier 512MB RAM):
    # options = WorkerOptions(
    #    entrypoint_fnc=entrypoint,
    #    prewarm_fnc=None,  # Disable prewarming to save RAM
    #    num_idle_processes=0, # Do not keep any processes waiting
    # )
    # cli.run_app(options)