# context_pilot_workflow.py
"""
ContextPilot: Context Curation Workflow
=======================================

This module handles topic detection and context curation for the ContextPilot system.
It does NOT generate responses - it only curates the context that will be sent to
the response LLM.

KEY RESPONSIBILITIES
--------------------
1. Topic Detection: Detect when the conversation topic changes
2. Context Storage: Save/load conversation context when switching topics
3. Context Curation: Build the optimal message list for the response LLM

TOPIC DETECTION
---------------
Uses a cheap LLM (CONTEXT_LLM) with function calling to decide:
- Is this the same topic? → No action needed
- Is this a new topic? → Save old context, set new topic
- Is this a previously discussed topic? → Save old context, load old topic

The topic detection LLM has access to these tools:
- save_context(topic, summary, key_facts): Save current topic before switching
- set_current_topic(topic): Set a new topic that hasn't been discussed
- load_context(topic): Load and switch to a previously saved topic
- list_saved_contexts(): See what topics have been discussed

CONTEXT CURATION (build_curated_messages)
-----------------------------------------
Builds the message list that will be sent to the response LLM:

Summary Mode:
    - System prompt includes topic summary + key facts
    - Includes session messages (for within-session continuity)
    - Includes current user message
    
Full Mode:
    - Plain system prompt
    - Full message history from stored context
    - Current session messages
    - Current user message

DATA FLOW
---------
1. User message arrives via MCP server
2. load_context_store() loads persisted state
3. detect_and_handle_topic_change() decides if topic changed
4. If changed: save_context() + set_current_topic()/load_context()
5. build_curated_messages() creates optimized message list
6. Return curated messages + stats to app.py for response generation
"""

from llama_index.core.workflow import Workflow, StartEvent, StopEvent, step, Context
from llama_index.llms.openai_like import OpenAILike
from llama_index.core.tools import FunctionTool
from llama_index.core.llms import ChatMessage
from dataclasses import dataclass, field
from pathlib import Path
from dotenv import load_dotenv
import json
import time
import os

load_dotenv()


# =============================================================================
# Configuration
# =============================================================================

CONTEXT_STORE_PATH = Path(__file__).parent / ".context_store.json"
SYSTEM_PROMPT_PATH = Path(__file__).parent.parent / "prompts" / "system_prompt.txt"

# Context management tool names (excluded from full history)
CONTEXT_TOOL_NAMES = {'save_context', 'load_context', 'set_current_topic', 'list_saved_contexts'}


# =============================================================================
# Events
# =============================================================================

class MessageEvent(StartEvent):
    """Input event for the workflow (MCP compatible)."""
    msg: str
    chat_history: list = []


# =============================================================================
# Data Classes
# =============================================================================

@dataclass
class CurationResult:
    """Result of context curation."""
    curated_messages: list[dict]  # Messages ready for LLM
    current_topic: str
    topic_changed: bool
    decisions: list[str]  # UI-friendly decision messages
    logs: list[str]
    stats: dict


@dataclass
class ProcessingState:
    """Holds state during message processing."""
    msg: str
    chat_history: list
    store: dict
    logs: list = field(default_factory=list)
    decisions: list = field(default_factory=list)
    topic_changed: bool = False
    detection_tokens: int = 0  # Tokens used by topic detection LLM
    
    @property
    def current_topic(self) -> str:
        return self.store.get("current_topic") or "None"
    
    @property
    def mode(self) -> str:
        return self.store.get("mode", "summary")
    
    @property
    def contexts(self) -> dict:
        return self.store.get("contexts", {})
    
    @property
    def saved_topics_str(self) -> str:
        topics = list(self.contexts.keys())
        return ", ".join(topics) if topics else "None"
    
    @property
    def context_summaries_str(self) -> str:
        """Build context summaries string including saved topics AND current session."""
        parts = []
        
        # Include saved topic summaries
        if self.contexts:
            parts.append("### Saved Topics:")
            for topic, ctx in self.contexts.items():
                parts.append(f"- {topic}: {ctx.get('summary', 'No summary')}")
        
        # Include current session messages (compact summaries) as plain text
        session_messages = self.store.get("current_session_messages", [])
        if session_messages:
            parts.append(f"\n### Current Session ({self.current_topic or 'unknown topic'}):")
            for msg in session_messages[-5:]:  # Last 5 compact exchanges
                content = msg.get("content", "")
                # Just show the content - these are compact summaries like "Q: ... | A: ..."
                parts.append(f"- {content[:200]}")  # Truncate if somehow long
        
        if not parts:
            return "No context yet."
        
        return "\n".join(parts)


# =============================================================================
# Utility Functions
# =============================================================================

def count_tokens(text: str) -> int:
    """Estimate token count (roughly 4 chars per token)."""
    return len(text) // 4 if text else 0


def extract_text_content(content) -> str:
    """Safely extract text from various content formats."""
    if content is None:
        return ""
    if isinstance(content, str):
        return content
    if isinstance(content, list):
        texts = []
        for block in content:
            if isinstance(block, dict):
                texts.append(block.get('text') or block.get('content', ''))
            elif hasattr(block, 'text'):
                texts.append(block.text)
            elif isinstance(block, str):
                texts.append(block)
        return "".join(str(t) for t in texts)
    return str(content)


# =============================================================================
# Context Store
# =============================================================================

def load_context_store() -> dict:
    """Load context store from disk."""
    if CONTEXT_STORE_PATH.exists():
        try:
            store = json.loads(CONTEXT_STORE_PATH.read_text())
            # Ensure mode exists (default to summary for backwards compatibility)
            if "mode" not in store:
                store["mode"] = "summary"
            # Ensure session messages list exists
            if "current_session_messages" not in store:
                store["current_session_messages"] = []
            return store
        except (json.JSONDecodeError, IOError):
            pass
    return {
        "contexts": {},
        "current_topic": None,
        "mode": "summary",  # "summary" or "full"
        "current_session_messages": [],  # Messages for current topic in this session
        "stats": {"total_tokens": 0, "tokens_saved": 0, "context_switches": 0}
    }


def get_current_mode() -> str:
    """Get the current context mode."""
    store = load_context_store()
    return store.get("mode", "summary")


def set_mode(mode: str) -> dict:
    """Set the context mode and clear all contexts. Returns the new store."""
    if mode not in ("summary", "full"):
        raise ValueError(f"Invalid mode: {mode}. Must be 'summary' or 'full'")
    
    # Create fresh store with new mode
    store = {
        "contexts": {},
        "current_topic": None,
        "mode": mode,
        "current_session_messages": [],
        "stats": {"total_tokens": 0, "tokens_saved": 0, "context_switches": 0,
                  "cumulative_full_tokens": 0, "cumulative_tokens_saved": 0}
    }
    save_context_store(store)
    return store


def append_session_message(role: str, content: str):
    """Append a message to the current session messages (for full mode)."""
    store = load_context_store()
    store["current_session_messages"].append({"role": role, "content": content})
    save_context_store(store)


def clear_session_messages():
    """Clear session messages (called when topic changes)."""
    store = load_context_store()
    store["current_session_messages"] = []
    save_context_store(store)


def save_context_store(store: dict):
    """Save context store to disk."""
    CONTEXT_STORE_PATH.write_text(json.dumps(store, indent=2))


# =============================================================================
# System Prompt
# =============================================================================

def load_system_prompt(**variables) -> str:
    """Load and format system prompt template."""
    try:
        template = SYSTEM_PROMPT_PATH.read_text()
        return template.format(**variables)
    except FileNotFoundError:
        return f"You are a helpful AI assistant. Current topic: {variables.get('current_topic', 'unknown')}"


def build_system_prompt(state: ProcessingState) -> str:
    """Build system prompt from current state."""
    return load_system_prompt(
        current_topic=state.current_topic,
        saved_topics=state.saved_topics_str,
        context_summaries=state.context_summaries_str
    )


# =============================================================================
# Context Tools (LLM-callable for topic detection)
# =============================================================================

# Global variable to hold pending full history for save
# This is set by the workflow before tool calling
_pending_full_history: list = []


def set_pending_full_history(history: list):
    """Set the full history to be saved when save_context is called in full mode."""
    global _pending_full_history
    # In full mode, we use session messages from the store instead
    # But we also include the UI history filtered as backup
    _pending_full_history = [
        msg for msg in history
        if not _is_context_tool_message(msg)
    ]


def _is_context_tool_message(msg: dict) -> bool:
    """Check if a message is a context management tool output."""
    content = msg.get("content", "")
    if not isinstance(content, str):
        return False
    # Check for our context tool output markers
    markers = ["💾 **Context saved", "📂 **Context loaded", "📍 **Topic set", 
               "📍 **Topic inferred", "🆕 **Topic changed", "🧭 Current topic:",
               "📚 **Contexts listed"]
    return any(marker in content for marker in markers)


def save_context(topic: str, summary: str, key_facts: list[str]) -> str:
    """Save conversation context before switching topics."""
    store = load_context_store()
    mode = store.get("mode", "summary")
    session_messages = store.get("current_session_messages", [])
    
    content = summary + " ".join(key_facts)
    tokens = count_tokens(content)
    
    context_data = {
        "topic": topic,
        "summary": summary,
        "key_facts": key_facts,
        "tokens": tokens,
        "saved_at": time.time(),
        "mode": mode,
    }
    
    # In full mode, save the session messages as full_history
    if mode == "full":
        # Combine any existing stored history with session messages
        existing_history = []
        if topic in store.get("contexts", {}) and store["contexts"][topic].get("full_history"):
            existing_history = store["contexts"][topic]["full_history"]
        
        full_history = existing_history + session_messages
        context_data["full_history"] = full_history
        context_data["tokens"] = sum(
            count_tokens(m.get("content", "")) for m in full_history
        )
        tokens = context_data["tokens"]
    
    store["contexts"][topic] = context_data
    store["stats"]["tokens_saved"] = sum(
        c.get("tokens", 0) for c in store["contexts"].values()
    )
    store["stats"]["context_switches"] += 1
    
    # Clear session messages after saving
    store["current_session_messages"] = []
    
    save_context_store(store)
    
    if mode == "full":
        msg_count = len(context_data.get("full_history", []))
        return f"✅ Saved context '{topic}' with {msg_count} messages ({tokens} tokens)"
    return f"✅ Saved context '{topic}' with {len(key_facts)} key facts ({tokens} tokens)"


def load_context(topic: str) -> str:
    """Load a previously saved conversation context."""
    store = load_context_store()
    
    if topic in store["contexts"]:
        ctx = store["contexts"][topic]
        store["current_topic"] = topic
        save_context_store(store)
        
        result = {
            "topic": ctx["topic"],
            "summary": ctx["summary"],
            "key_facts": ctx["key_facts"],
        }
        
        # In full mode, include the full history indicator
        if ctx.get("mode") == "full" and ctx.get("full_history"):
            result["has_full_history"] = True
            result["message_count"] = len(ctx["full_history"])
        
        return json.dumps(result, indent=2)
    
    return f"No saved context found for topic '{topic}'"


def list_saved_contexts() -> str:
    """List all saved conversation contexts."""
    store = load_context_store()
    contexts = store.get("contexts", {})
    
    if not contexts:
        return "No saved contexts yet."
    
    return "\n".join(
        f"• {topic}: {ctx.get('summary', 'No summary')}..."
        for topic, ctx in contexts.items()
    )


def set_current_topic(topic: str) -> str:
    """Set the current conversation topic."""
    store = load_context_store()
    store["current_topic"] = topic
    save_context_store(store)
    return f"📍 Current topic set to: {topic}"


# Tool objects for topic detection
CONTEXT_TOOLS = [
    FunctionTool.from_defaults(fn=save_context),
    FunctionTool.from_defaults(fn=load_context),
    FunctionTool.from_defaults(fn=list_saved_contexts),
    FunctionTool.from_defaults(fn=set_current_topic),
]


# =============================================================================
# LLM for Topic Detection Only (cheaper model)
# =============================================================================

# Use CONTEXT_LLM for topic detection (cheaper)
# Use RESPONSE_LLM for generation (more capable) - configured in app.py
CONTEXT_LLM_MODEL = os.getenv("CONTEXT_LLM", os.getenv("NEBIUS_MODEL", "openai/gpt-4o-mini"))
NEBIUS_BASE_URL = os.getenv("NEBIUS_BASE_URL")
NEBIUS_API_KEY = os.getenv("NEBIUS_API_KEY")

# Validate required environment variables
if not NEBIUS_BASE_URL:
    print("WARNING: NEBIUS_BASE_URL not set. Topic detection will fail.")
if not NEBIUS_API_KEY:
    print("WARNING: NEBIUS_API_KEY not set. Topic detection will fail.")

print(f"[ContextPilot] Context LLM: {CONTEXT_LLM_MODEL}")
print(f"[ContextPilot] API Base: {NEBIUS_BASE_URL}")
print(f"[ContextPilot] API Key set: {bool(NEBIUS_API_KEY)}")

topic_llm = OpenAILike(
    model=CONTEXT_LLM_MODEL,
    api_base=NEBIUS_BASE_URL,
    api_key=NEBIUS_API_KEY,
    is_chat_model=True,
    is_function_calling_model=True,
    context_window=128000,
)


# =============================================================================
# Message Building
# =============================================================================

def build_detection_messages(state: ProcessingState, system_prompt: str) -> list[ChatMessage]:
    """Build messages for topic detection (minimal context).
    
    The system prompt already contains compact context summaries from the store.
    We only need to add the current user message - no need to include chat history
    since that would send full messages defeating the purpose of compact summaries.
    """
    messages = [ChatMessage(role="system", content=system_prompt)]
    # Only the current user message - context summaries are in system prompt
    messages.append(ChatMessage(role="user", content=state.msg))
    return messages


@dataclass
class CurationMetrics:
    """Metrics comparing full context vs curated context."""
    curated_messages: list[dict]
    curated_tokens: int
    full_context_tokens: int  # What it would be without curation
    tokens_saved_this_request: int
    savings_percent: float


def build_full_context_messages(state: ProcessingState) -> list[dict]:
    """
    Build what the FULL context would look like without curation.
    This is for comparison only - to show how many tokens we saved.
    Includes ALL chat history without any summarization.
    """
    store = load_context_store()
    
    base_prompt = "You are a helpful AI assistant."
    
    # In a non-curated approach, we'd include ALL stored contexts expanded
    all_contexts = store.get("contexts", {})
    if all_contexts:
        base_prompt += "\n\nFull conversation history from all topics:\n"
        for topic, ctx in all_contexts.items():
            base_prompt += f"\n[Topic: {topic}]\n"
            base_prompt += f"Summary: {ctx.get('summary', '')}\n"
            if ctx.get('key_facts'):
                base_prompt += "Key facts:\n" + "\n".join(f"- {fact}" for fact in ctx['key_facts'])
            base_prompt += "\n"
    
    messages = [{"role": "system", "content": base_prompt}]
    
    # Include FULL chat history (no truncation)
    for h in state.chat_history:
        if isinstance(h, dict):
            content = extract_text_content(h.get("content", ""))
            if content:
                messages.append({"role": h.get("role", "user"), "content": content})
    
    messages.append({"role": "user", "content": state.msg})
    return messages


def build_curated_messages(state: ProcessingState) -> CurationMetrics:
    """
    Build the curated message list for LLM consumption.
    
    Summary Mode:
        - System prompt with summary of CURRENT topic only
        - Only the current user message is sent (no history)
        
    Full Mode:
        - Plain system prompt
        - Full message history for CURRENT topic only
        - Current session messages
        - The current user message
    """
    store = load_context_store()
    current_topic = store.get("current_topic", "None")
    mode = store.get("mode", "summary")
    session_messages = store.get("current_session_messages", [])
    all_contexts = store.get("contexts", {})
    
    # Debug logging
    state.logs.append(f"🔧 Mode from store: {mode}")
    state.logs.append(f"🔧 Session messages count: {len(session_messages)}")
    state.logs.append(f"🔧 Current topic: {current_topic}")
    
    base_prompt = "You are a helpful AI assistant."
    curated_messages = []
    
    # Check if we have stored context for current topic
    has_stored_context = (
        current_topic and 
        current_topic != "None" and 
        current_topic in all_contexts
    )
    
    state.logs.append(f"🔧 Has stored context for '{current_topic}': {has_stored_context}")
    
    if mode == "full":
        # FULL MODE: Plain system prompt + full history for CURRENT topic only
        state.logs.append("🔧 Using FULL mode")
        curated_messages = [{"role": "system", "content": base_prompt}]
        
        # Restore saved FULL history for current topic only
        if has_stored_context:
            ctx = all_contexts[current_topic]
            if ctx.get("full_history"):
                state.logs.append(f"🔧 Restoring {len(ctx['full_history'])} messages from full_history")
                curated_messages.extend(ctx["full_history"])
            else:
                state.logs.append("🔧 No full_history in stored context")
        
        # Add session messages (messages from this session for current topic)
        if session_messages:
            state.logs.append(f"🔧 Adding {len(session_messages)} session messages")
            curated_messages.extend(session_messages)
        
        # Add current message
        curated_messages.append({"role": "user", "content": state.msg})
        
    else:
        # SUMMARY MODE: System prompt with CURRENT topic summary + session messages
        state.logs.append("🔧 Using SUMMARY mode")
        
        if has_stored_context:
            ctx = all_contexts[current_topic]
            base_prompt += f"\n\n## Context for '{current_topic}':\n"
            base_prompt += f"Summary: {ctx.get('summary', 'No summary')}\n"
            if ctx.get('key_facts'):
                base_prompt += "Key facts:\n" + "\n".join(f"- {fact}" for fact in ctx['key_facts'])
            state.logs.append(f"🔧 Added summary for current topic '{current_topic}'")
        
        curated_messages = [{"role": "system", "content": base_prompt}]
        
        # Include session messages for continuation (current session's exchanges)
        if session_messages:
            state.logs.append(f"🔧 Adding {len(session_messages)} session messages for continuation")
            curated_messages.extend(session_messages)
        
        # Add current message
        curated_messages.append({"role": "user", "content": state.msg})
    
    # Calculate full context for comparison (what it would be without ANY curation)
    full_messages = build_full_context_messages(state)
    
    # Count tokens
    curated_tokens = sum(count_tokens(m.get("content", "")) for m in curated_messages)
    full_context_tokens = sum(count_tokens(m.get("content", "")) for m in full_messages)
    tokens_saved = full_context_tokens - curated_tokens
    savings_percent = (tokens_saved / full_context_tokens * 100) if full_context_tokens > 0 else 0
    
    return CurationMetrics(
        curated_messages=curated_messages,
        curated_tokens=curated_tokens,
        full_context_tokens=full_context_tokens,
        tokens_saved_this_request=max(0, tokens_saved),
        savings_percent=round(savings_percent, 1)
    )


# =============================================================================
# Tool Processing
# =============================================================================

def process_tool_sources(sources: list, state: ProcessingState) -> set:
    """Process tool call sources and update state. Returns set of called tool names."""
    called_tools = set()
    
    for source in sources:
        tool_name = getattr(source, 'tool_name', 'unknown')
        raw_output = getattr(source, 'raw_output', '')
        
        called_tools.add(tool_name)
        state.logs.append(f"  → {tool_name}: {str(raw_output)[:100]}...")
        
        # User-friendly decisions
        decisions_map = {
            'save_context': f"💾 **Context saved**: {raw_output}",
            'load_context': f"📂 **Context loaded**: {raw_output}",
            'set_current_topic': f"📍 **Topic set**: {raw_output}",
            'list_saved_contexts': "📚 **Contexts listed**",
        }
        if tool_name in decisions_map:
            state.decisions.append(decisions_map[tool_name])
    
    return called_tools


async def infer_topic_from_message(msg: str) -> str:
    """Use LLM to infer topic from user message."""
    messages = [
        ChatMessage(
            role="system",
            content="Extract the MAIN SUBJECT (noun) from the user's message. "
                    "Focus on WHAT the question is about, not the action/verb. "
                    "For 'how do cats hunt?' → 'cats' (not 'hunting'). "
                    "For 'what is Python used for?' → 'python' (not 'programming'). "
                    "For 'how to cook pasta?' → 'pasta' (not 'cooking'). "
                    "Respond with ONLY ONE WORD (lowercase, no punctuation)."
        ),
        ChatMessage(role="user", content=msg)
    ]
    response = await topic_llm.achat(messages)
    topic = extract_text_content(response.message.content)
    return topic.strip().lower().replace(" ", "").replace("_", "")


async def detect_and_handle_topic_change(state: ProcessingState) -> bool:
    """
    Use LLM with tools to detect if topic changed.
    Returns True if topic changed, False otherwise.
    """
    system_prompt = build_system_prompt(state)
    messages = build_detection_messages(state, system_prompt)
    
    # Calculate tokens used for topic detection
    detection_tokens = sum(count_tokens(m.content or "") for m in messages)
    state.detection_tokens = detection_tokens
    
    state.logs.append(f"🔍 Detecting topic change... (mode: {state.mode})")
    state.logs.append(f"🎯 Detection tokens: {detection_tokens}")
    
    # Set pending history for full mode (in case save_context is called)
    if state.mode == "full":
        # Build filtered history from chat_history
        filtered_history = []
        for h in state.chat_history:
            if isinstance(h, dict) and not _is_context_tool_message(h):
                content = extract_text_content(h.get("content", ""))
                if content:
                    filtered_history.append({"role": h.get("role", "user"), "content": content})
        set_pending_full_history(filtered_history)
    
    # Call LLM with context tools
    response = await topic_llm.apredict_and_call(
        tools=CONTEXT_TOOLS,
        chat_history=messages,
        error_on_no_tool_call=False,
    )
    
    sources = getattr(response, 'sources', [])
    
    if not sources:
        state.logs.append("� No topic change detected")
        return False
    
    state.logs.append(f"🔧 LLM called {len(sources)} tools")
    called_tools = process_tool_sources(sources, state)
    
    context_tools = {'save_context', 'load_context', 'set_current_topic', 'list_saved_contexts'}
    if called_tools & context_tools:
        # Topic change detected
        
        # Infer topic if save_context called without set_current_topic
        if 'save_context' in called_tools and 'set_current_topic' not in called_tools and 'load_context' not in called_tools:
            state.logs.append("⚠️ Inferring new topic...")
            new_topic = await infer_topic_from_message(state.msg)
            set_current_topic(new_topic)
            state.decisions.append(f"� **Topic inferred**: {new_topic}")
            state.logs.append(f"� New topic: {new_topic}")
        
        # Reload store
        state.store = load_context_store()
        state.decisions.append(f"🆕 **Topic changed to**: **{state.current_topic}**")
        return True
    
    return False


# =============================================================================
# Result Building
# =============================================================================

def build_stats(state: ProcessingState, metrics: CurationMetrics = None) -> dict:
    """Build stats dictionary with token comparison metrics."""
    store = load_context_store()
    stats = store.get("stats", {})
    
    stored_contexts_data = [
        {
            "topic": topic,
            "summary": ctx.get("summary", ""),
            "key_facts": ctx.get("key_facts", []),
            "tokens": ctx.get("tokens", 0),
            "is_current": topic == store.get("current_topic")
        }
        for topic, ctx in store.get("contexts", {}).items()
    ]
    
    # Update cumulative stats
    if metrics:
        stats["total_tokens"] = stats.get("total_tokens", 0) + metrics.curated_tokens
        stats["cumulative_full_tokens"] = stats.get("cumulative_full_tokens", 0) + metrics.full_context_tokens
        stats["cumulative_tokens_saved"] = stats.get("cumulative_tokens_saved", 0) + metrics.tokens_saved_this_request
        save_context_store(store)
    
    # Update detection token stats
    if state.detection_tokens:
        stats["cumulative_detection_tokens"] = stats.get("cumulative_detection_tokens", 0) + state.detection_tokens
        store["stats"] = stats
        save_context_store(store)
    
    return {
        # Per-request metrics
        "curated_tokens": metrics.curated_tokens if metrics else 0,
        "full_context_tokens": metrics.full_context_tokens if metrics else 0,
        "tokens_saved_this_request": metrics.tokens_saved_this_request if metrics else 0,
        "savings_percent": metrics.savings_percent if metrics else 0,
        "detection_tokens": state.detection_tokens,  # Tokens for topic detection
        
        # Cumulative metrics
        "cumulative_curated_tokens": stats.get("total_tokens", 0),
        "cumulative_full_tokens": stats.get("cumulative_full_tokens", 0),
        "cumulative_tokens_saved": stats.get("cumulative_tokens_saved", 0),
        "cumulative_detection_tokens": stats.get("cumulative_detection_tokens", 0),
        
        # Other stats
        "context_switches": stats.get("context_switches", 0),
        "stored_contexts": len(store.get("contexts", {})),
        "current_topic": store.get("current_topic", "None"),
        "mode": store.get("mode", "summary"),
        "logs": state.logs,
        "stored_contexts_data": stored_contexts_data,
    }


# =============================================================================
# Workflow
# =============================================================================

class ContextPilotWorkflow(Workflow):
    """
    Context curation workflow.
    
    Detects topic changes and returns curated messages ready for LLM consumption.
    Does NOT generate the actual response - that's done by the caller.
    """
    
    @step
    async def process_message(self, ctx: Context, ev: MessageEvent) -> StopEvent:
        """
        Curate context for a message.
        
        Returns:
            - curated_messages: List of messages ready for LLM
            - current_topic: The current topic
            - topic_changed: Whether topic changed
            - decisions: UI-friendly decision messages
            - stats: Statistics
        """
        try:
            # Initialize state
            store = load_context_store()
            state = ProcessingState(
                msg=ev.msg,
                chat_history=getattr(ev, 'chat_history', []),
                store=store,
            )
            
            state.logs.append(f"📨 Processing: {state.msg[:50]}...")
            state.logs.append(f"🧭 Current topic: {state.current_topic}")
            state.decisions.append(f"🧭 Current topic: **{state.current_topic}**")
            
            # Detect topic change
            state.topic_changed = await detect_and_handle_topic_change(state)
            
            # Build curated messages with metrics
            metrics = build_curated_messages(state)
            state.logs.append(f"📝 Curated {len(metrics.curated_messages)} messages (topic_changed={state.topic_changed})")
            state.logs.append(f"📊 Tokens: {metrics.curated_tokens} curated vs {metrics.full_context_tokens} full ({metrics.savings_percent}% saved)")
            
            # Build result
            result = {
                "curated_messages": metrics.curated_messages,
                "current_topic": state.current_topic,
                "topic_changed": state.topic_changed,
                "decisions": state.decisions,
                "stats": build_stats(state, metrics),
            }
            
            return StopEvent(result=json.dumps(result))
        
        except Exception as e:
            import traceback
            error_result = {
                "error": str(e),
                "traceback": traceback.format_exc(),
                "curated_messages": [],
                "decisions": [f"❌ Error: {e}"],
                "stats": {"logs": [f"Error: {e}", traceback.format_exc()]},
            }
            return StopEvent(result=json.dumps(error_result))