Spaces:

MHamdan
/

SPARKNET

Sleeping

SPARKNET / src /agents /planner_agent.py

Initial commit: SPARKNET framework

a9dc537 9 months ago

19.8 kB

"""
PlannerAgent for SPARKNET - LangChain Version
Breaks down complex VISTA scenarios into executable workflows
Uses LangChain chains for structured task decomposition
"""

from typing import List, Dict, Optional, Any
from dataclasses import dataclass, field
from loguru import logger
import json
import networkx as nx
from pydantic import BaseModel, Field

from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import JsonOutputParser
from langchain_core.messages import HumanMessage, SystemMessage

from .base_agent import BaseAgent, Task, Message
from ..llm.langchain_ollama_client import LangChainOllamaClient
from ..workflow.langgraph_state import SubTask as SubTaskModel, TaskStatus


# Pydantic model for planning output
class TaskDecomposition(BaseModel):
    """Structured output from planning chain"""
    subtasks: List[Dict[str, Any]] = Field(description="List of subtasks with dependencies")
    reasoning: str = Field(description="Explanation of the planning strategy")
    estimated_total_duration: float = Field(description="Total estimated duration in seconds")


@dataclass
class TaskGraph:
    """Directed acyclic graph of tasks with dependencies."""
    subtasks: Dict[str, SubTaskModel] = field(default_factory=dict)
    graph: nx.DiGraph = field(default_factory=nx.DiGraph)

    def add_subtask(self, subtask: SubTaskModel):
        """Add a subtask to the graph."""
        self.subtasks[subtask.id] = subtask
        self.graph.add_node(subtask.id, task=subtask)

        # Add edges for dependencies
        for dep_id in subtask.dependencies:
            if dep_id in self.subtasks:
                self.graph.add_edge(dep_id, subtask.id)

    def get_execution_order(self) -> List[List[str]]:
        """
        Get tasks in execution order (topological sort).
        Returns list of lists - inner lists can be executed in parallel.
        """
        try:
            generations = list(nx.topological_generations(self.graph))
            return generations
        except nx.NetworkXError as e:
            logger.error(f"Error in topological sort: {e}")
            return []

    def validate(self) -> bool:
        """Validate graph has no cycles."""
        return nx.is_directed_acyclic_graph(self.graph)


class PlannerAgent(BaseAgent):
    """
    Agent specialized in task decomposition and workflow planning.
    Uses LangChain chains with qwen2.5:14b for complex reasoning.
    """

    # Scenario templates for common VISTA workflows
    SCENARIO_TEMPLATES = {
        'patent_wakeup': {
            'description': 'Analyze dormant patent and create valorization roadmap',
            'stages': [
                {
                    'name': 'document_analysis',
                    'agent': 'DocumentAnalysisAgent',
                    'description': 'Extract and analyze patent content',
                    'dependencies': [],
                },
                {
                    'name': 'market_analysis',
                    'agent': 'MarketAnalysisAgent',
                    'description': 'Identify market opportunities for patent',
                    'dependencies': ['document_analysis'],
                },
                {
                    'name': 'matchmaking',
                    'agent': 'MatchmakingAgent',
                    'description': 'Match patent with potential licensees',
                    'dependencies': ['document_analysis', 'market_analysis'],
                },
                {
                    'name': 'outreach',
                    'agent': 'OutreachAgent',
                    'description': 'Generate valorization brief and outreach materials',
                    'dependencies': ['matchmaking'],
                },
            ],
        },
        'agreement_safety': {
            'description': 'Review legal agreement for risks and compliance',
            'stages': [
                {
                    'name': 'document_parsing',
                    'agent': 'LegalAnalysisAgent',
                    'description': 'Parse agreement and extract clauses',
                    'dependencies': [],
                },
                {
                    'name': 'compliance_check',
                    'agent': 'ComplianceAgent',
                    'description': 'Check GDPR and Law 25 compliance',
                    'dependencies': ['document_parsing'],
                },
                {
                    'name': 'risk_assessment',
                    'agent': 'RiskAssessmentAgent',
                    'description': 'Identify problematic clauses and risks',
                    'dependencies': ['document_parsing'],
                },
                {
                    'name': 'recommendations',
                    'agent': 'RecommendationAgent',
                    'description': 'Generate improvement suggestions',
                    'dependencies': ['compliance_check', 'risk_assessment'],
                },
            ],
        },
        'partner_matching': {
            'description': 'Match stakeholders based on complementary capabilities',
            'stages': [
                {
                    'name': 'profiling',
                    'agent': 'ProfilingAgent',
                    'description': 'Extract stakeholder capabilities and needs',
                    'dependencies': [],
                },
                {
                    'name': 'semantic_matching',
                    'agent': 'SemanticMatchingAgent',
                    'description': 'Find complementary partners using embeddings',
                    'dependencies': ['profiling'],
                },
                {
                    'name': 'network_analysis',
                    'agent': 'NetworkAnalysisAgent',
                    'description': 'Identify strategic network connections',
                    'dependencies': ['profiling'],
                },
                {
                    'name': 'facilitation',
                    'agent': 'ConnectionFacilitatorAgent',
                    'description': 'Generate introduction materials',
                    'dependencies': ['semantic_matching', 'network_analysis'],
                },
            ],
        },
    }

    def __init__(
        self,
        llm_client: LangChainOllamaClient,
        memory_agent: Optional['MemoryAgent'] = None,
        temperature: float = 0.7,
    ):
        """
        Initialize PlannerAgent with LangChain client.
        Args:
            llm_client: LangChain Ollama client
            memory_agent: Optional memory agent for context
            temperature: LLM temperature for planning
        """
        self.llm_client = llm_client
        self.memory_agent = memory_agent
        self.temperature = temperature

        # Create planning chains
        self.planning_chain = self._create_planning_chain()
        self.refinement_chain = self._create_refinement_chain()

        # Store for backward compatibility
        self.name = "PlannerAgent"
        self.description = "Task decomposition and workflow planning"

        logger.info(f"Initialized PlannerAgent with LangChain (complexity: complex)")

    def _create_planning_chain(self):
        """
        Create LangChain chain for task decomposition.
        Returns:
            Runnable chain: prompt | llm | parser
        """
        system_template = """You are a strategic planning agent for research valorization tasks.
Your role is to:
1. Analyze complex tasks and break them into manageable subtasks
2. Identify dependencies between subtasks
3. Assign appropriate agents to each subtask
4. Estimate task complexity and duration
5. Create optimal execution plans
Available agent types:
- ExecutorAgent: General task execution
- DocumentAnalysisAgent: Analyze patents and documents
- MarketAnalysisAgent: Market research and opportunity identification
- MatchmakingAgent: Stakeholder matching and connections
- OutreachAgent: Generate outreach materials and briefs
- LegalAnalysisAgent: Legal document analysis
- ComplianceAgent: Compliance checking (GDPR, Law 25)
- RiskAssessmentAgent: Risk identification
- ProfilingAgent: Stakeholder profiling
- SemanticMatchingAgent: Semantic similarity matching
- NetworkAnalysisAgent: Network and relationship analysis
Output your plan as a structured JSON object with:
- subtasks: List of subtask objects with id, description, agent_type, dependencies, estimated_duration, priority
- reasoning: Your strategic reasoning for this decomposition
- estimated_total_duration: Total estimated time in seconds"""

        human_template = """Given the following task, create a detailed execution plan:
Task: {task_description}
{context_section}
Break this down into specific subtasks. For each subtask:
- Give it a unique ID (use snake_case)
- Describe what needs to be done
- Specify which agent type should handle it
- List any dependencies (IDs of tasks that must complete first)
- Estimate duration in seconds
- Set priority (1=highest)
Think step-by-step about:
- What is the ultimate goal?
- What information is needed?
- What are the logical stages?
- Which subtasks can run in parallel?
- What are the critical dependencies?
Output JSON only."""

        prompt = ChatPromptTemplate.from_messages([
            ("system", system_template),
            ("human", human_template)
        ])

        # Use complex model for planning
        llm = self.llm_client.get_llm(complexity="complex", temperature=self.temperature)

        # JSON output parser
        parser = JsonOutputParser(pydantic_object=TaskDecomposition)

        # Create chain
        chain = prompt | llm | parser

        return chain

    def _create_refinement_chain(self):
        """
        Create LangChain chain for replanning based on feedback.
        Returns:
            Runnable chain for refinement
        """
        system_template = """You are refining an existing task plan based on feedback.
Your role is to:
1. Review the original plan and feedback
2. Identify what went wrong or could be improved
3. Create an improved plan that addresses the issues
4. Maintain successful elements from the original plan
Be thoughtful about what to change and what to keep."""

        human_template = """Refine the following plan based on feedback:
Original Task: {task_description}
Original Plan:
{original_plan}
Feedback from execution:
{feedback}
Issues encountered:
{issues}
Create an improved plan that addresses these issues while maintaining what worked well.
Output JSON in the same format as before."""

        prompt = ChatPromptTemplate.from_messages([
            ("system", system_template),
            ("human", human_template)
        ])

        llm = self.llm_client.get_llm(complexity="complex", temperature=self.temperature)
        parser = JsonOutputParser(pydantic_object=TaskDecomposition)

        chain = prompt | llm | parser

        return chain

    async def process_task(self, task: Task) -> Task:
        """
        Process planning task by decomposing into workflow.
        Args:
            task: High-level task to plan
        Returns:
            Updated task with plan in result
        """
        logger.info(f"PlannerAgent planning task: {task.id}")
        task.status = "in_progress"

        try:
            # Check if this is a known scenario
            scenario = task.metadata.get('scenario') if task.metadata else None

            if scenario and scenario in self.SCENARIO_TEMPLATES:
                # Use template-based planning
                logger.info(f"Using template for scenario: {scenario}")
                task_graph = await self._plan_from_template(task, scenario)
            else:
                # Use LangChain-based planning for custom tasks
                logger.info("Using LangChain planning for custom task")
                task_graph = await self._plan_with_langchain(task)

            # Validate the graph
            if not task_graph.validate():
                raise ValueError("Generated task graph contains cycles!")

            # Store plan in task result
            task.result = {
                'task_graph': task_graph,
                'execution_order': task_graph.get_execution_order(),
                'total_subtasks': len(task_graph.subtasks),
            }
            task.status = "completed"

            logger.info(f"Planning completed: {len(task_graph.subtasks)} subtasks")

        except Exception as e:
            logger.error(f"Planning failed: {e}")
            task.status = "failed"
            task.error = str(e)

        return task

    async def _plan_from_template(self, task: Task, scenario: str) -> TaskGraph:
        """
        Create task graph from scenario template.
        Args:
            task: Original task
            scenario: Scenario identifier
        Returns:
            TaskGraph based on template
        """
        template = self.SCENARIO_TEMPLATES[scenario]
        task_graph = TaskGraph()

        # Get task parameters
        params = task.metadata.get('parameters', {}) if task.metadata else {}

        # Create subtasks from template stages
        for i, stage in enumerate(template['stages']):
            subtask = SubTaskModel(
                id=f"{task.id}_{stage['name']}",
                description=stage['description'],
                agent_type=stage['agent'],
                dependencies=[f"{task.id}_{dep}" for dep in stage['dependencies']],
                estimated_duration=30.0,
                priority=i + 1,
                parameters=params,
                status=TaskStatus.PENDING
            )
            task_graph.add_subtask(subtask)

        logger.debug(f"Created task graph with {len(task_graph.subtasks)} subtasks from template")

        return task_graph

    async def _plan_with_langchain(self, task: Task, context: Optional[List[Any]] = None) -> TaskGraph:
        """
        Create task graph using LangChain planning chain.
        Args:
            task: Original task
            context: Optional context from memory
        Returns:
            TaskGraph generated by LangChain
        """
        # Prepare context section
        context_section = ""
        if context and len(context) > 0:
            context_section = "Relevant past experiences:\n"
            for i, ctx in enumerate(context[:3], 1):  # Top 3 contexts
                context_section += f"{i}. {ctx.page_content[:200]}...\n"

        # Invoke planning chain
        try:
            result = await self.planning_chain.ainvoke({
                "task_description": task.description,
                "context_section": context_section
            })

            # Parse result into TaskGraph
            task_graph = TaskGraph()

            for subtask_data in result.get('subtasks', []):
                subtask = SubTaskModel(
                    id=f"{task.id}_{subtask_data.get('id', f'subtask_{len(task_graph.subtasks)}')}",
                    description=subtask_data.get('description', ''),
                    agent_type=subtask_data.get('agent_type', 'ExecutorAgent'),
                    dependencies=[f"{task.id}_{dep}" for dep in subtask_data.get('dependencies', [])],
                    estimated_duration=subtask_data.get('estimated_duration', 30.0),
                    priority=subtask_data.get('priority', 0),
                    parameters=subtask_data.get('parameters', {}),
                    status=TaskStatus.PENDING
                )
                task_graph.add_subtask(subtask)

            logger.debug(f"Created task graph with {len(task_graph.subtasks)} subtasks from LangChain")

            return task_graph

        except Exception as e:
            logger.error(f"Failed to parse LangChain planning response: {e}")
            raise ValueError(f"Failed to generate plan: {e}")

    async def decompose_task(
        self,
        task_description: str,
        scenario: Optional[str] = None,
        context: Optional[List[Any]] = None
    ) -> TaskGraph:
        """
        Decompose a high-level task into subtasks.
        Args:
            task_description: Natural language description
            scenario: Optional scenario identifier
            context: Optional context from memory
        Returns:
            TaskGraph with subtasks and dependencies
        """
        # Create a task object
        task = Task(
            id=f"plan_{hash(task_description) % 10000}",
            description=task_description,
            metadata={'scenario': scenario} if scenario else {},
        )

        # Process with planning logic
        result_task = await self.process_task(task)

        if result_task.status == "completed" and result_task.result:
            return result_task.result['task_graph']
        else:
            raise RuntimeError(f"Planning failed: {result_task.error}")

    async def adapt_plan(
        self,
        task_graph: TaskGraph,
        feedback: str,
        issues: List[str]
    ) -> TaskGraph:
        """
        Adapt an existing plan based on execution feedback.
        Args:
            task_graph: Original task graph
            feedback: Feedback from execution
            issues: List of issues encountered
        Returns:
            Updated task graph
        """
        logger.info("Adapting plan based on feedback")

        # Convert task graph to dict for refinement
        original_plan = {
            "subtasks": [
                {
                    "id": st.id,
                    "description": st.description,
                    "agent_type": st.agent_type,
                    "dependencies": st.dependencies
                }
                for st in task_graph.subtasks.values()
            ]
        }

        try:
            # Invoke refinement chain
            result = await self.refinement_chain.ainvoke({
                "task_description": "Refine task decomposition",
                "original_plan": json.dumps(original_plan, indent=2),
                "feedback": feedback,
                "issues": "\n".join(f"- {issue}" for issue in issues)
            })

            # Create new task graph from refined plan
            new_task_graph = TaskGraph()

            for subtask_data in result.get('subtasks', []):
                subtask = SubTaskModel(
                    id=subtask_data.get('id', f'subtask_{len(new_task_graph.subtasks)}'),
                    description=subtask_data.get('description', ''),
                    agent_type=subtask_data.get('agent_type', 'ExecutorAgent'),
                    dependencies=subtask_data.get('dependencies', []),
                    estimated_duration=subtask_data.get('estimated_duration', 30.0),
                    priority=subtask_data.get('priority', 0),
                    parameters=subtask_data.get('parameters', {}),
                    status=TaskStatus.PENDING
                )
                new_task_graph.add_subtask(subtask)

            logger.info(f"Plan adapted: {len(new_task_graph.subtasks)} subtasks")
            return new_task_graph

        except Exception as e:
            logger.error(f"Plan adaptation failed: {e}, returning original plan")
            return task_graph

    def get_parallel_tasks(self, task_graph: TaskGraph) -> List[List[SubTaskModel]]:
        """
        Get tasks that can be executed in parallel.
        Args:
            task_graph: Task graph
        Returns:
            List of parallel task groups
        """
        execution_order = task_graph.get_execution_order()
        parallel_groups = []

        for task_ids in execution_order:
            group = [task_graph.subtasks[task_id] for task_id in task_ids]
            parallel_groups.append(group)

        return parallel_groups

	"""
	PlannerAgent for SPARKNET - LangChain Version
	Breaks down complex VISTA scenarios into executable workflows
	Uses LangChain chains for structured task decomposition
	"""

	from typing import List, Dict, Optional, Any
	from dataclasses import dataclass, field
	from loguru import logger
	import json
	import networkx as nx
	from pydantic import BaseModel, Field

	from langchain_core.prompts import ChatPromptTemplate
	from langchain_core.output_parsers import JsonOutputParser
	from langchain_core.messages import HumanMessage, SystemMessage

	from .base_agent import BaseAgent, Task, Message
	from ..llm.langchain_ollama_client import LangChainOllamaClient
	from ..workflow.langgraph_state import SubTask as SubTaskModel, TaskStatus


	# Pydantic model for planning output
	class TaskDecomposition(BaseModel):
	"""Structured output from planning chain"""
	subtasks: List[Dict[str, Any]] = Field(description="List of subtasks with dependencies")
	reasoning: str = Field(description="Explanation of the planning strategy")
	estimated_total_duration: float = Field(description="Total estimated duration in seconds")


	@dataclass
	class TaskGraph:
	"""Directed acyclic graph of tasks with dependencies."""
	subtasks: Dict[str, SubTaskModel] = field(default_factory=dict)
	graph: nx.DiGraph = field(default_factory=nx.DiGraph)

	def add_subtask(self, subtask: SubTaskModel):
	"""Add a subtask to the graph."""
	self.subtasks[subtask.id] = subtask
	self.graph.add_node(subtask.id, task=subtask)

	# Add edges for dependencies
	for dep_id in subtask.dependencies:
	if dep_id in self.subtasks:
	self.graph.add_edge(dep_id, subtask.id)

	def get_execution_order(self) -> List[List[str]]:
	"""
	Get tasks in execution order (topological sort).
	Returns list of lists - inner lists can be executed in parallel.
	"""
	try:
	generations = list(nx.topological_generations(self.graph))
	return generations
	except nx.NetworkXError as e:
	logger.error(f"Error in topological sort: {e}")
	return []

	def validate(self) -> bool:
	"""Validate graph has no cycles."""
	return nx.is_directed_acyclic_graph(self.graph)


	class PlannerAgent(BaseAgent):
	"""
	Agent specialized in task decomposition and workflow planning.
	Uses LangChain chains with qwen2.5:14b for complex reasoning.
	"""

	# Scenario templates for common VISTA workflows
	SCENARIO_TEMPLATES = {
	'patent_wakeup': {
	'description': 'Analyze dormant patent and create valorization roadmap',
	'stages': [
	{
	'name': 'document_analysis',
	'agent': 'DocumentAnalysisAgent',
	'description': 'Extract and analyze patent content',
	'dependencies': [],
	},
	{
	'name': 'market_analysis',
	'agent': 'MarketAnalysisAgent',
	'description': 'Identify market opportunities for patent',
	'dependencies': ['document_analysis'],
	},
	{
	'name': 'matchmaking',
	'agent': 'MatchmakingAgent',
	'description': 'Match patent with potential licensees',
	'dependencies': ['document_analysis', 'market_analysis'],
	},
	{
	'name': 'outreach',
	'agent': 'OutreachAgent',
	'description': 'Generate valorization brief and outreach materials',
	'dependencies': ['matchmaking'],
	},
	],
	},
	'agreement_safety': {
	'description': 'Review legal agreement for risks and compliance',
	'stages': [
	{
	'name': 'document_parsing',
	'agent': 'LegalAnalysisAgent',
	'description': 'Parse agreement and extract clauses',
	'dependencies': [],
	},
	{
	'name': 'compliance_check',
	'agent': 'ComplianceAgent',
	'description': 'Check GDPR and Law 25 compliance',
	'dependencies': ['document_parsing'],
	},
	{
	'name': 'risk_assessment',
	'agent': 'RiskAssessmentAgent',
	'description': 'Identify problematic clauses and risks',
	'dependencies': ['document_parsing'],
	},
	{
	'name': 'recommendations',
	'agent': 'RecommendationAgent',
	'description': 'Generate improvement suggestions',
	'dependencies': ['compliance_check', 'risk_assessment'],
	},
	],
	},
	'partner_matching': {
	'description': 'Match stakeholders based on complementary capabilities',
	'stages': [
	{
	'name': 'profiling',
	'agent': 'ProfilingAgent',
	'description': 'Extract stakeholder capabilities and needs',
	'dependencies': [],
	},
	{
	'name': 'semantic_matching',
	'agent': 'SemanticMatchingAgent',
	'description': 'Find complementary partners using embeddings',
	'dependencies': ['profiling'],
	},
	{
	'name': 'network_analysis',
	'agent': 'NetworkAnalysisAgent',
	'description': 'Identify strategic network connections',
	'dependencies': ['profiling'],
	},
	{
	'name': 'facilitation',
	'agent': 'ConnectionFacilitatorAgent',
	'description': 'Generate introduction materials',
	'dependencies': ['semantic_matching', 'network_analysis'],
	},
	],
	},
	}

	def __init__(
	self,
	llm_client: LangChainOllamaClient,
	memory_agent: Optional['MemoryAgent'] = None,
	temperature: float = 0.7,
	):
	"""
	Initialize PlannerAgent with LangChain client.

	Args:
	llm_client: LangChain Ollama client
	memory_agent: Optional memory agent for context
	temperature: LLM temperature for planning
	"""
	self.llm_client = llm_client
	self.memory_agent = memory_agent
	self.temperature = temperature

	# Create planning chains
	self.planning_chain = self._create_planning_chain()
	self.refinement_chain = self._create_refinement_chain()

	# Store for backward compatibility
	self.name = "PlannerAgent"
	self.description = "Task decomposition and workflow planning"

	logger.info(f"Initialized PlannerAgent with LangChain (complexity: complex)")

	def _create_planning_chain(self):
	"""
	Create LangChain chain for task decomposition.

	Returns:
	Runnable chain: prompt \| llm \| parser
	"""
	system_template = """You are a strategic planning agent for research valorization tasks.

	Your role is to:
	1. Analyze complex tasks and break them into manageable subtasks
	2. Identify dependencies between subtasks
	3. Assign appropriate agents to each subtask
	4. Estimate task complexity and duration
	5. Create optimal execution plans

	Available agent types:
	- ExecutorAgent: General task execution
	- DocumentAnalysisAgent: Analyze patents and documents
	- MarketAnalysisAgent: Market research and opportunity identification
	- MatchmakingAgent: Stakeholder matching and connections
	- OutreachAgent: Generate outreach materials and briefs
	- LegalAnalysisAgent: Legal document analysis
	- ComplianceAgent: Compliance checking (GDPR, Law 25)
	- RiskAssessmentAgent: Risk identification
	- ProfilingAgent: Stakeholder profiling
	- SemanticMatchingAgent: Semantic similarity matching
	- NetworkAnalysisAgent: Network and relationship analysis

	Output your plan as a structured JSON object with:
	- subtasks: List of subtask objects with id, description, agent_type, dependencies, estimated_duration, priority
	- reasoning: Your strategic reasoning for this decomposition
	- estimated_total_duration: Total estimated time in seconds"""

	human_template = """Given the following task, create a detailed execution plan:

	Task: {task_description}

	{context_section}

	Break this down into specific subtasks. For each subtask:
	- Give it a unique ID (use snake_case)
	- Describe what needs to be done
	- Specify which agent type should handle it
	- List any dependencies (IDs of tasks that must complete first)
	- Estimate duration in seconds
	- Set priority (1=highest)

	Think step-by-step about:
	- What is the ultimate goal?
	- What information is needed?
	- What are the logical stages?
	- Which subtasks can run in parallel?
	- What are the critical dependencies?

	Output JSON only."""

	prompt = ChatPromptTemplate.from_messages([
	("system", system_template),
	("human", human_template)
	])

	# Use complex model for planning
	llm = self.llm_client.get_llm(complexity="complex", temperature=self.temperature)

	# JSON output parser
	parser = JsonOutputParser(pydantic_object=TaskDecomposition)

	# Create chain
	chain = prompt \| llm \| parser

	return chain

	def _create_refinement_chain(self):
	"""
	Create LangChain chain for replanning based on feedback.

	Returns:
	Runnable chain for refinement
	"""
	system_template = """You are refining an existing task plan based on feedback.

	Your role is to:
	1. Review the original plan and feedback
	2. Identify what went wrong or could be improved
	3. Create an improved plan that addresses the issues
	4. Maintain successful elements from the original plan

	Be thoughtful about what to change and what to keep."""

	human_template = """Refine the following plan based on feedback:

	Original Task: {task_description}

	Original Plan:
	{original_plan}

	Feedback from execution:
	{feedback}

	Issues encountered:
	{issues}

	Create an improved plan that addresses these issues while maintaining what worked well.
	Output JSON in the same format as before."""

	prompt = ChatPromptTemplate.from_messages([
	("system", system_template),
	("human", human_template)
	])

	llm = self.llm_client.get_llm(complexity="complex", temperature=self.temperature)
	parser = JsonOutputParser(pydantic_object=TaskDecomposition)

	chain = prompt \| llm \| parser

	return chain

	async def process_task(self, task: Task) -> Task:
	"""
	Process planning task by decomposing into workflow.

	Args:
	task: High-level task to plan

	Returns:
	Updated task with plan in result
	"""
	logger.info(f"PlannerAgent planning task: {task.id}")
	task.status = "in_progress"

	try:
	# Check if this is a known scenario
	scenario = task.metadata.get('scenario') if task.metadata else None

	if scenario and scenario in self.SCENARIO_TEMPLATES:
	# Use template-based planning
	logger.info(f"Using template for scenario: {scenario}")
	task_graph = await self._plan_from_template(task, scenario)
	else:
	# Use LangChain-based planning for custom tasks
	logger.info("Using LangChain planning for custom task")
	task_graph = await self._plan_with_langchain(task)

	# Validate the graph
	if not task_graph.validate():
	raise ValueError("Generated task graph contains cycles!")

	# Store plan in task result
	task.result = {
	'task_graph': task_graph,
	'execution_order': task_graph.get_execution_order(),
	'total_subtasks': len(task_graph.subtasks),
	}
	task.status = "completed"

	logger.info(f"Planning completed: {len(task_graph.subtasks)} subtasks")

	except Exception as e:
	logger.error(f"Planning failed: {e}")
	task.status = "failed"
	task.error = str(e)

	return task

	async def _plan_from_template(self, task: Task, scenario: str) -> TaskGraph:
	"""
	Create task graph from scenario template.

	Args:
	task: Original task
	scenario: Scenario identifier

	Returns:
	TaskGraph based on template
	"""
	template = self.SCENARIO_TEMPLATES[scenario]
	task_graph = TaskGraph()

	# Get task parameters
	params = task.metadata.get('parameters', {}) if task.metadata else {}

	# Create subtasks from template stages
	for i, stage in enumerate(template['stages']):
	subtask = SubTaskModel(
	id=f"{task.id}_{stage['name']}",
	description=stage['description'],
	agent_type=stage['agent'],
	dependencies=[f"{task.id}_{dep}" for dep in stage['dependencies']],
	estimated_duration=30.0,
	priority=i + 1,
	parameters=params,
	status=TaskStatus.PENDING
	)
	task_graph.add_subtask(subtask)

	logger.debug(f"Created task graph with {len(task_graph.subtasks)} subtasks from template")

	return task_graph

	async def _plan_with_langchain(self, task: Task, context: Optional[List[Any]] = None) -> TaskGraph:
	"""
	Create task graph using LangChain planning chain.

	Args:
	task: Original task
	context: Optional context from memory

	Returns:
	TaskGraph generated by LangChain
	"""
	# Prepare context section
	context_section = ""
	if context and len(context) > 0:
	context_section = "Relevant past experiences:\n"
	for i, ctx in enumerate(context[:3], 1): # Top 3 contexts
	context_section += f"{i}. {ctx.page_content[:200]}...\n"

	# Invoke planning chain
	try:
	result = await self.planning_chain.ainvoke({
	"task_description": task.description,
	"context_section": context_section
	})

	# Parse result into TaskGraph
	task_graph = TaskGraph()

	for subtask_data in result.get('subtasks', []):
	subtask = SubTaskModel(
	id=f"{task.id}_{subtask_data.get('id', f'subtask_{len(task_graph.subtasks)}')}",
	description=subtask_data.get('description', ''),
	agent_type=subtask_data.get('agent_type', 'ExecutorAgent'),
	dependencies=[f"{task.id}_{dep}" for dep in subtask_data.get('dependencies', [])],
	estimated_duration=subtask_data.get('estimated_duration', 30.0),
	priority=subtask_data.get('priority', 0),
	parameters=subtask_data.get('parameters', {}),
	status=TaskStatus.PENDING
	)
	task_graph.add_subtask(subtask)

	logger.debug(f"Created task graph with {len(task_graph.subtasks)} subtasks from LangChain")

	return task_graph

	except Exception as e:
	logger.error(f"Failed to parse LangChain planning response: {e}")
	raise ValueError(f"Failed to generate plan: {e}")

	async def decompose_task(
	self,
	task_description: str,
	scenario: Optional[str] = None,
	context: Optional[List[Any]] = None
	) -> TaskGraph:
	"""
	Decompose a high-level task into subtasks.

	Args:
	task_description: Natural language description
	scenario: Optional scenario identifier
	context: Optional context from memory

	Returns:
	TaskGraph with subtasks and dependencies
	"""
	# Create a task object
	task = Task(
	id=f"plan_{hash(task_description) % 10000}",
	description=task_description,
	metadata={'scenario': scenario} if scenario else {},
	)

	# Process with planning logic
	result_task = await self.process_task(task)

	if result_task.status == "completed" and result_task.result:
	return result_task.result['task_graph']
	else:
	raise RuntimeError(f"Planning failed: {result_task.error}")

	async def adapt_plan(
	self,
	task_graph: TaskGraph,
	feedback: str,
	issues: List[str]
	) -> TaskGraph:
	"""
	Adapt an existing plan based on execution feedback.

	Args:
	task_graph: Original task graph
	feedback: Feedback from execution
	issues: List of issues encountered

	Returns:
	Updated task graph
	"""
	logger.info("Adapting plan based on feedback")

	# Convert task graph to dict for refinement
	original_plan = {
	"subtasks": [
	{
	"id": st.id,
	"description": st.description,
	"agent_type": st.agent_type,
	"dependencies": st.dependencies
	}
	for st in task_graph.subtasks.values()
	]
	}

	try:
	# Invoke refinement chain
	result = await self.refinement_chain.ainvoke({
	"task_description": "Refine task decomposition",
	"original_plan": json.dumps(original_plan, indent=2),
	"feedback": feedback,
	"issues": "\n".join(f"- {issue}" for issue in issues)
	})

	# Create new task graph from refined plan
	new_task_graph = TaskGraph()

	for subtask_data in result.get('subtasks', []):
	subtask = SubTaskModel(
	id=subtask_data.get('id', f'subtask_{len(new_task_graph.subtasks)}'),
	description=subtask_data.get('description', ''),
	agent_type=subtask_data.get('agent_type', 'ExecutorAgent'),
	dependencies=subtask_data.get('dependencies', []),
	estimated_duration=subtask_data.get('estimated_duration', 30.0),
	priority=subtask_data.get('priority', 0),
	parameters=subtask_data.get('parameters', {}),
	status=TaskStatus.PENDING
	)
	new_task_graph.add_subtask(subtask)

	logger.info(f"Plan adapted: {len(new_task_graph.subtasks)} subtasks")
	return new_task_graph

	except Exception as e:
	logger.error(f"Plan adaptation failed: {e}, returning original plan")
	return task_graph

	def get_parallel_tasks(self, task_graph: TaskGraph) -> List[List[SubTaskModel]]:
	"""
	Get tasks that can be executed in parallel.

	Args:
	task_graph: Task graph

	Returns:
	List of parallel task groups
	"""
	execution_order = task_graph.get_execution_order()
	parallel_groups = []

	for task_ids in execution_order:
	group = [task_graph.subtasks[task_id] for task_id in task_ids]
	parallel_groups.append(group)

	return parallel_groups