attune-ai 2.0.0__py3-none-any.whl

attune/orchestration/meta_orchestrator.py

@@ -0,0 +1,1168 @@
+"""Meta-orchestrator for intelligent agent composition.
+
+This module implements the core orchestration logic that analyzes tasks,
+selects appropriate agents, and chooses composition patterns.
+
+Security:
+    - All inputs validated before processing
+    - No eval() or exec() usage
+    - Agent selection based on whitelisted templates
+
+Example:
+    >>> orchestrator = MetaOrchestrator()
+    >>> plan = orchestrator.analyze_and_compose(
+    ...     task="Boost test coverage to 90%",
+    ...     context={"current_coverage": 75}
+    ... )
+    >>> print(plan.strategy)
+    sequential
+    >>> print([a.role for a in plan.agents])
+    ['Test Coverage Expert', 'Test Generation Specialist', 'Quality Assurance Validator']
+"""
+
+import logging
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Any
+
+from .agent_templates import AgentTemplate, get_template, get_templates_by_capability
+
+logger = logging.getLogger(__name__)
+
+
+class TaskComplexity(Enum):
+    """Task complexity classification."""
+
+    SIMPLE = "simple"  # Single agent, straightforward
+    MODERATE = "moderate"  # 2-3 agents, some coordination
+    COMPLEX = "complex"  # 4+ agents, multi-phase execution
+
+
+class TaskDomain(Enum):
+    """Task domain classification."""
+
+    TESTING = "testing"
+    SECURITY = "security"
+    CODE_QUALITY = "code_quality"
+    DOCUMENTATION = "documentation"
+    PERFORMANCE = "performance"
+    ARCHITECTURE = "architecture"
+    REFACTORING = "refactoring"
+    GENERAL = "general"
+
+
+class CompositionPattern(Enum):
+    """Available composition patterns (grammar rules)."""
+
+    # Original 7 patterns
+    SEQUENTIAL = "sequential"  # A → B → C
+    PARALLEL = "parallel"  # A || B || C
+    DEBATE = "debate"  # A ⇄ B ⇄ C → Synthesis
+    TEACHING = "teaching"  # Junior → Expert validation
+    REFINEMENT = "refinement"  # Draft → Review → Polish
+    ADAPTIVE = "adaptive"  # Classifier → Specialist
+    CONDITIONAL = "conditional"  # If-then-else routing
+
+    # Anthropic-inspired patterns (Patterns 8-10)
+    TOOL_ENHANCED = "tool_enhanced"  # Single agent with tools
+    PROMPT_CACHED_SEQUENTIAL = "prompt_cached_sequential"  # Shared cached context
+    DELEGATION_CHAIN = "delegation_chain"  # Hierarchical delegation (≤3 levels)
+
+
+@dataclass
+class TaskRequirements:
+    """Extracted requirements from task analysis.
+
+    Attributes:
+        complexity: Task complexity level
+        domain: Primary task domain
+        capabilities_needed: List of capabilities required
+        parallelizable: Whether task can be parallelized
+        quality_gates: Quality thresholds to enforce
+        context: Additional context for customization
+    """
+
+    complexity: TaskComplexity
+    domain: TaskDomain
+    capabilities_needed: list[str]
+    parallelizable: bool = False
+    quality_gates: dict[str, Any] = field(default_factory=dict)
+    context: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ExecutionPlan:
+    """Plan for agent execution.
+
+    Attributes:
+        agents: List of agents to execute
+        strategy: Composition pattern to use
+        quality_gates: Quality thresholds to enforce
+        estimated_cost: Estimated execution cost
+        estimated_duration: Estimated time in seconds
+    """
+
+    agents: list[AgentTemplate]
+    strategy: CompositionPattern
+    quality_gates: dict[str, Any] = field(default_factory=dict)
+    estimated_cost: float = 0.0
+    estimated_duration: int = 0
+
+
+class MetaOrchestrator:
+    """Intelligent task analyzer and agent composition engine.
+
+    The meta-orchestrator analyzes tasks to determine requirements,
+    selects appropriate agents, and chooses optimal composition patterns.
+
+    Example:
+        >>> orchestrator = MetaOrchestrator()
+        >>> plan = orchestrator.analyze_and_compose(
+        ...     task="Prepare for v3.12.0 release",
+        ...     context={"version": "3.12.0"}
+        ... )
+    """
+
+    # Keyword patterns for task analysis
+    COMPLEXITY_KEYWORDS = {
+        TaskComplexity.SIMPLE: [
+            "format",
+            "lint",
+            "check",
+            "validate",
+            "document",
+        ],
+        TaskComplexity.MODERATE: [
+            "improve",
+            "refactor",
+            "optimize",
+            "test",
+            "review",
+        ],
+        TaskComplexity.COMPLEX: [
+            "release",
+            "migrate",
+            "redesign",
+            "architecture",
+            "prepare",
+        ],
+    }
+
+    DOMAIN_KEYWORDS = {
+        TaskDomain.TESTING: [
+            "test",
+            "coverage",
+            "pytest",
+            "unit test",
+            "integration test",
+        ],
+        TaskDomain.SECURITY: [
+            "security",
+            "vulnerability",
+            "audit",
+            "penetration",
+            "threat",
+        ],
+        TaskDomain.CODE_QUALITY: [
+            "quality",
+            "code review",
+            "lint",
+            "best practices",
+            "maintainability",
+        ],
+        TaskDomain.DOCUMENTATION: [
+            "docs",
+            "documentation",
+            "readme",
+            "guide",
+            "tutorial",
+        ],
+        TaskDomain.PERFORMANCE: [
+            "performance",
+            "optimize",
+            "speed",
+            "benchmark",
+            "profile",
+        ],
+        TaskDomain.ARCHITECTURE: [
+            "architecture",
+            "design",
+            "structure",
+            "pattern",
+            "dependency",
+        ],
+        TaskDomain.REFACTORING: [
+            "refactor",
+            "cleanup",
+            "simplify",
+            "restructure",
+            "debt",
+        ],
+    }
+
+    # Capability mapping by domain
+    DOMAIN_CAPABILITIES = {
+        TaskDomain.TESTING: [
+            "analyze_gaps",
+            "suggest_tests",
+            "validate_coverage",
+        ],
+        TaskDomain.SECURITY: [
+            "vulnerability_scan",
+            "threat_modeling",
+            "compliance_check",
+        ],
+        TaskDomain.CODE_QUALITY: [
+            "code_review",
+            "quality_assessment",
+            "best_practices_check",
+        ],
+        TaskDomain.DOCUMENTATION: [
+            "generate_docs",
+            "check_completeness",
+            "update_examples",
+        ],
+        TaskDomain.PERFORMANCE: [
+            "profile_code",
+            "identify_bottlenecks",
+            "suggest_optimizations",
+        ],
+        TaskDomain.ARCHITECTURE: [
+            "analyze_architecture",
+            "identify_patterns",
+            "suggest_improvements",
+        ],
+        TaskDomain.REFACTORING: [
+            "identify_code_smells",
+            "suggest_refactorings",
+            "validate_changes",
+        ],
+    }
+
+    def __init__(self):
+        """Initialize meta-orchestrator."""
+        logger.info("MetaOrchestrator initialized")
+
+    def analyze_task(self, task: str, context: dict[str, Any] | None = None) -> TaskRequirements:
+        """Analyze task to extract requirements (public wrapper for testing).
+
+        Args:
+            task: Task description (e.g., "Boost test coverage to 90%")
+            context: Optional context dictionary
+
+        Returns:
+            TaskRequirements with extracted information
+
+        Raises:
+            ValueError: If task is invalid
+
+        Example:
+            >>> orchestrator = MetaOrchestrator()
+            >>> requirements = orchestrator.analyze_task(
+            ...     task="Improve test coverage",
+            ...     context={"current_coverage": 75}
+            ... )
+            >>> print(requirements.domain)
+            TaskDomain.TESTING
+        """
+        if not task or not isinstance(task, str):
+            raise ValueError("task must be a non-empty string")
+
+        context = context or {}
+        return self._analyze_task(task, context)
+
+    def create_execution_plan(
+        self,
+        requirements: TaskRequirements,
+        agents: list[AgentTemplate],
+        strategy: CompositionPattern,
+    ) -> ExecutionPlan:
+        """Create execution plan from components (extracted for testing).
+
+        Args:
+            requirements: Task requirements with quality gates
+            agents: Selected agents for execution
+            strategy: Composition pattern to use
+
+        Returns:
+            ExecutionPlan with all components configured
+
+        Example:
+            >>> orchestrator = MetaOrchestrator()
+            >>> requirements = TaskRequirements(
+            ...     complexity=TaskComplexity.MODERATE,
+            ...     domain=TaskDomain.TESTING,
+            ...     capabilities_needed=["analyze_gaps"],
+            ...     quality_gates={"min_coverage": 80}
+            ... )
+            >>> agents = [get_template("test_coverage_analyzer")]
+            >>> strategy = CompositionPattern.SEQUENTIAL
+            >>> plan = orchestrator.create_execution_plan(requirements, agents, strategy)
+            >>> print(plan.strategy)
+            CompositionPattern.SEQUENTIAL
+        """
+        return ExecutionPlan(
+            agents=agents,
+            strategy=strategy,
+            quality_gates=requirements.quality_gates,
+            estimated_cost=self._estimate_cost(agents),
+            estimated_duration=self._estimate_duration(agents, strategy),
+        )
+
+    def analyze_and_compose(
+        self, task: str, context: dict[str, Any] | None = None, interactive: bool = False
+    ) -> ExecutionPlan:
+        """Analyze task and create execution plan.
+
+        This is the main entry point for the meta-orchestrator.
+
+        Args:
+            task: Task description (e.g., "Boost test coverage to 90%")
+            context: Optional context dictionary
+            interactive: If True, prompts user for low-confidence cases (default: False)
+
+        Returns:
+            ExecutionPlan with agents and strategy
+
+        Raises:
+            ValueError: If task is invalid
+
+        Example:
+            >>> orchestrator = MetaOrchestrator()
+            >>> plan = orchestrator.analyze_and_compose(
+            ...     task="Improve test coverage",
+            ...     context={"current_coverage": 75}
+            ... )
+        """
+        if not task or not isinstance(task, str):
+            raise ValueError("task must be a non-empty string")
+
+        context = context or {}
+
+        # Use interactive mode if requested
+        if interactive:
+            return self.analyze_and_compose_interactive(task, context)
+
+        logger.info(f"Analyzing task: {task}")
+
+        # Step 1: Analyze task requirements
+        requirements = self._analyze_task(task, context)
+        logger.info(
+            f"Task analysis: complexity={requirements.complexity.value}, "
+            f"domain={requirements.domain.value}, "
+            f"capabilities={requirements.capabilities_needed}"
+        )
+
+        # Step 2: Select appropriate agents
+        agents = self._select_agents(requirements)
+        logger.info(f"Selected {len(agents)} agents: {[a.id for a in agents]}")
+
+        # Step 3: Choose composition pattern
+        strategy = self._choose_composition_pattern(requirements, agents)
+        logger.info(f"Selected strategy: {strategy.value}")
+
+        # Step 4: Create execution plan (using extracted public method)
+        plan = self.create_execution_plan(requirements, agents, strategy)
+
+        return plan
+
+    def analyze_and_compose_interactive(
+        self, task: str, context: dict[str, Any] | None = None
+    ) -> ExecutionPlan:
+        """Analyze task with user confirmation for ambiguous cases.
+
+        This method uses confidence scoring to determine when to ask the user
+        for input. High-confidence selections proceed automatically, while
+        low-confidence cases prompt the user to choose.
+
+        Args:
+            task: Task description
+            context: Optional context dictionary
+
+        Returns:
+            ExecutionPlan with agents and strategy
+
+        Raises:
+            ValueError: If task is invalid
+            ImportError: If AskUserQuestion tool is not available
+
+        Example:
+            >>> orchestrator = MetaOrchestrator()
+            >>> plan = orchestrator.analyze_and_compose_interactive(
+            ...     task="Complex architectural redesign",
+            ...     context={}
+            ... )
+            # User may be prompted to choose approach if confidence is low
+        """
+        if not task or not isinstance(task, str):
+            raise ValueError("task must be a non-empty string")
+
+        context = context or {}
+        logger.info(f"Analyzing task interactively: {task}")
+
+        # Step 1: Analyze task requirements
+        requirements = self._analyze_task(task, context)
+        logger.info(
+            f"Task analysis: complexity={requirements.complexity.value}, "
+            f"domain={requirements.domain.value}"
+        )
+
+        # Step 2: Select agents
+        agents = self._select_agents(requirements)
+        logger.info(f"Selected {len(agents)} agents: {[a.id for a in agents]}")
+
+        # Step 3: Choose pattern
+        recommended_pattern = self._choose_composition_pattern(requirements, agents)
+        logger.info(f"Recommended strategy: {recommended_pattern.value}")
+
+        # Step 4: Calculate confidence in recommendation
+        confidence = self._calculate_confidence(requirements, agents, recommended_pattern)
+        logger.info(f"Confidence score: {confidence:.2f}")
+
+        # Step 5: Branch based on confidence
+        if confidence >= 0.8:
+            # High confidence → automatic execution
+            logger.info("High confidence - proceeding automatically")
+            return self.create_execution_plan(requirements, agents, recommended_pattern)
+
+        else:
+            # Low confidence → ask user
+            logger.info("Low confidence - prompting user for choice")
+            return self._prompt_user_for_approach(
+                requirements, agents, recommended_pattern, confidence
+            )
+
+    def _calculate_confidence(
+        self,
+        requirements: TaskRequirements,
+        agents: list[AgentTemplate],
+        pattern: CompositionPattern,
+    ) -> float:
+        """Calculate confidence in automatic pattern selection.
+
+        Confidence scoring considers:
+        - Domain clarity (GENERAL domain reduces confidence)
+        - Agent count (many agents = complex coordination)
+        - Task complexity (complex tasks have multiple valid approaches)
+        - Pattern specificity (Anthropic patterns have clear heuristics)
+
+        Args:
+            requirements: Task requirements
+            agents: Selected agents
+            pattern: Recommended composition pattern
+
+        Returns:
+            Confidence score between 0.0 and 1.0
+
+        Example:
+            >>> confidence = orchestrator._calculate_confidence(
+            ...     requirements=TaskRequirements(
+            ...         complexity=TaskComplexity.SIMPLE,
+            ...         domain=TaskDomain.TESTING,
+            ...         capabilities_needed=["analyze_gaps"]
+            ...     ),
+            ...     agents=[test_agent],
+            ...     pattern=CompositionPattern.SEQUENTIAL
+            ... )
+            >>> confidence >= 0.8  # High confidence for simple, clear task
+            True
+        """
+        confidence = 1.0
+
+        # Reduce confidence for ambiguous cases
+        if requirements.domain == TaskDomain.GENERAL:
+            confidence *= 0.7  # Generic tasks are less clear
+
+        if len(agents) > 5:
+            confidence *= 0.8  # Many agents → complex coordination
+
+        if requirements.complexity == TaskComplexity.COMPLEX:
+            confidence *= 0.85  # Complex → multiple valid approaches
+
+        # Increase confidence for clear patterns
+        if pattern in [
+            CompositionPattern.TOOL_ENHANCED,
+            CompositionPattern.DELEGATION_CHAIN,
+            CompositionPattern.PROMPT_CACHED_SEQUENTIAL,
+        ]:
+            confidence *= 1.1  # New Anthropic patterns have clear heuristics
+
+        # Specific domain patterns also get confidence boost
+        if pattern in [
+            CompositionPattern.TEACHING,
+            CompositionPattern.REFINEMENT,
+        ] and requirements.domain in [TaskDomain.DOCUMENTATION, TaskDomain.REFACTORING]:
+            confidence *= 1.05  # Domain-specific pattern match
+
+        return min(confidence, 1.0)
+
+    def _prompt_user_for_approach(
+        self,
+        requirements: TaskRequirements,
+        agents: list[AgentTemplate],
+        recommended_pattern: CompositionPattern,
+        confidence: float,
+    ) -> ExecutionPlan:
+        """Prompt user to choose approach when confidence is low.
+
+        Presents three options:
+        1. Use recommended pattern (with confidence score)
+        2. Customize team composition
+        3. Show all patterns and choose
+
+        Args:
+            requirements: Task requirements
+            agents: Selected agents
+            recommended_pattern: Recommended pattern
+            confidence: Confidence score (0.0-1.0)
+
+        Returns:
+            ExecutionPlan based on user choice
+
+        Raises:
+            ImportError: If AskUserQuestion tool not available
+        """
+        try:
+            # Import here to avoid circular dependency and allow graceful degradation
+            from attune.tools import AskUserQuestion
+        except ImportError as e:
+            logger.warning(f"AskUserQuestion not available: {e}")
+            logger.info("Falling back to automatic selection")
+            return self.create_execution_plan(requirements, agents, recommended_pattern)
+
+        # Format agent list for display
+        agent_summary = ", ".join([a.role for a in agents])
+
+        # Ask user for approach
+        response = AskUserQuestion(
+            questions=[
+                {
+                    "header": "Approach",
+                    "question": "How would you like to create the agent team?",
+                    "multiSelect": False,
+                    "options": [
+                        {
+                            "label": f"Use recommended: {recommended_pattern.value} (Recommended)",
+                            "description": f"Auto-selected based on task analysis. "
+                            f"{len(agents)} agents: {agent_summary}. "
+                            f"Confidence: {confidence:.0%}",
+                        },
+                        {
+                            "label": "Customize team composition",
+                            "description": "Choose specific agents and pattern manually",
+                        },
+                        {
+                            "label": "Show all 10 patterns",
+                            "description": "Learn about patterns and select one",
+                        },
+                    ],
+                }
+            ]
+        )
+
+        # Handle user response
+        user_choice = response.get("Approach", "")
+
+        if "Use recommended" in user_choice:
+            logger.info("User accepted recommended approach")
+            return self.create_execution_plan(requirements, agents, recommended_pattern)
+
+        elif "Customize" in user_choice:
+            logger.info("User chose to customize team")
+            return self._interactive_team_builder(requirements, agents, recommended_pattern)
+
+        else:  # Show patterns
+            logger.info("User chose to explore patterns")
+            return self._pattern_chooser_wizard(requirements, agents)
+
+    def _interactive_team_builder(
+        self,
+        requirements: TaskRequirements,
+        suggested_agents: list[AgentTemplate],
+        suggested_pattern: CompositionPattern,
+    ) -> ExecutionPlan:
+        """Interactive team builder for manual customization.
+
+        Allows user to:
+        1. Review suggested agents and modify selection
+        2. Choose composition pattern
+        3. Configure quality gates
+
+        Args:
+            requirements: Task requirements
+            suggested_agents: Auto-selected agents
+            suggested_pattern: Auto-selected pattern
+
+        Returns:
+            ExecutionPlan with user-customized configuration
+        """
+        try:
+            from attune.tools import AskUserQuestion
+        except ImportError:
+            logger.warning("AskUserQuestion not available, using defaults")
+            return self.create_execution_plan(requirements, suggested_agents, suggested_pattern)
+
+        # Step 1: Agent selection
+        agent_response = AskUserQuestion(
+            questions=[
+                {
+                    "header": "Agents",
+                    "question": "Which agents should be included in the team?",
+                    "multiSelect": True,
+                    "options": [
+                        {
+                            "label": agent.role,
+                            "description": f"{agent.id} - {', '.join(agent.capabilities[:3])}",
+                        }
+                        for agent in suggested_agents
+                    ],
+                }
+            ]
+        )
+
+        # Filter agents based on user selection
+        selected_agent_roles = agent_response.get("Agents", [])
+        if not isinstance(selected_agent_roles, list):
+            selected_agent_roles = [selected_agent_roles]
+
+        selected_agents = [a for a in suggested_agents if a.role in selected_agent_roles]
+        if not selected_agents:
+            # User deselected all - use defaults
+            selected_agents = suggested_agents
+
+        # Step 2: Pattern selection
+        pattern_response = AskUserQuestion(
+            questions=[
+                {
+                    "header": "Pattern",
+                    "question": "Which composition pattern should be used?",
+                    "multiSelect": False,
+                    "options": [
+                        {
+                            "label": f"{suggested_pattern.value} (Recommended)",
+                            "description": self._get_pattern_description(suggested_pattern),
+                        },
+                        {
+                            "label": "sequential",
+                            "description": "Execute agents one after another (A → B → C)",
+                        },
+                        {
+                            "label": "parallel",
+                            "description": "Execute agents simultaneously (A || B || C)",
+                        },
+                        {
+                            "label": "tool_enhanced",
+                            "description": "Single agent with comprehensive tool access",
+                        },
+                    ],
+                }
+            ]
+        )
+
+        # Parse pattern choice
+        pattern_choice = pattern_response.get("Pattern", suggested_pattern.value)
+        if "(Recommended)" in pattern_choice:
+            selected_pattern = suggested_pattern
+        else:
+            # Extract pattern name
+            pattern_name = pattern_choice.split()[0]
+            try:
+                selected_pattern = CompositionPattern(pattern_name)
+            except ValueError:
+                logger.warning(f"Invalid pattern: {pattern_name}, using suggested")
+                selected_pattern = suggested_pattern
+
+        # Create execution plan with user selections
+        return self.create_execution_plan(requirements, selected_agents, selected_pattern)
+
+    def _pattern_chooser_wizard(
+        self,
+        requirements: TaskRequirements,
+        suggested_agents: list[AgentTemplate],
+    ) -> ExecutionPlan:
+        """Interactive pattern chooser with educational previews.
+
+        Shows all 10 composition patterns with:
+        - Description and when to use
+        - Visual preview of agent flow
+        - Estimated cost and duration
+        - Examples of similar tasks
+
+        Args:
+            requirements: Task requirements
+            suggested_agents: Auto-selected agents
+
+        Returns:
+            ExecutionPlan with user-selected pattern
+        """
+        try:
+            from attune.tools import AskUserQuestion
+        except ImportError:
+            logger.warning("AskUserQuestion not available, using defaults")
+            suggested_pattern = self._choose_composition_pattern(requirements, suggested_agents)
+            return self.create_execution_plan(
+                requirements, suggested_agents, suggested_pattern
+            )
+
+        # Present all patterns with descriptions
+        pattern_response = AskUserQuestion(
+            questions=[
+                {
+                    "header": "Pattern",
+                    "question": "Choose a composition pattern (with preview):",
+                    "multiSelect": False,
+                    "options": [
+                        {
+                            "label": "sequential",
+                            "description": "A → B → C | Step-by-step pipeline | "
+                            "Example: Parse → Analyze → Report",
+                        },
+                        {
+                            "label": "parallel",
+                            "description": "A || B || C | Independent tasks | "
+                            "Example: Security + Quality + Performance audits",
+                        },
+                        {
+                            "label": "debate",
+                            "description": "A ⇄ B ⇄ C → Synthesis | Multiple perspectives | "
+                            "Example: 3 reviewers discuss approach",
+                        },
+                        {
+                            "label": "teaching",
+                            "description": "Junior → Expert validation | Draft + review | "
+                            "Example: Cheap model drafts, expert validates",
+                        },
+                        {
+                            "label": "refinement",
+                            "description": "Draft → Review → Polish | Iterative improvement | "
+                            "Example: Code → Review → Refine",
+                        },
+                        {
+                            "label": "adaptive",
+                            "description": "Classifier → Specialist | Dynamic routing | "
+                            "Example: Analyze task type → Route to expert",
+                        },
+                        {
+                            "label": "tool_enhanced (NEW)",
+                            "description": "Single agent + tools | Most efficient | "
+                            "Example: File reader with analysis tools",
+                        },
+                        {
+                            "label": "prompt_cached_sequential (NEW)",
+                            "description": "Shared large context | Cost-optimized | "
+                            "Example: 3 agents using same codebase docs",
+                        },
+                        {
+                            "label": "delegation_chain (NEW)",
+                            "description": "Coordinator → Specialists | Hierarchical | "
+                            "Example: Task planner delegates to architects",
+                        },
+                    ],
+                }
+            ]
+        )
+
+        # Parse pattern choice
+        pattern_choice = pattern_response.get("Pattern", "sequential")
+        pattern_name = pattern_choice.split()[0]  # Extract name before any annotations
+
+        try:
+            selected_pattern = CompositionPattern(pattern_name)
+        except ValueError:
+            logger.warning(f"Invalid pattern: {pattern_name}, using sequential")
+            selected_pattern = CompositionPattern.SEQUENTIAL
+
+        logger.info(f"User selected pattern: {selected_pattern.value}")
+
+        # Create execution plan with user-selected pattern
+        return self.create_execution_plan(requirements, suggested_agents, selected_pattern)
+
+    def _get_pattern_description(self, pattern: CompositionPattern) -> str:
+        """Get human-readable description of a pattern.
+
+        Args:
+            pattern: Composition pattern
+
+        Returns:
+            Description string
+        """
+        descriptions = {
+            CompositionPattern.SEQUENTIAL: "Execute agents one after another (A → B → C)",
+            CompositionPattern.PARALLEL: "Execute agents simultaneously (A || B || C)",
+            CompositionPattern.DEBATE: "Multiple agents discuss and synthesize (A ⇄ B → Result)",
+            CompositionPattern.TEACHING: "Junior agent with expert validation (Draft → Review)",
+            CompositionPattern.REFINEMENT: "Iterative improvement (Draft → Review → Polish)",
+            CompositionPattern.ADAPTIVE: "Dynamic routing based on classification",
+            CompositionPattern.CONDITIONAL: "If-then-else branching logic",
+            CompositionPattern.TOOL_ENHANCED: "Single agent with comprehensive tool access",
+            CompositionPattern.PROMPT_CACHED_SEQUENTIAL: "Sequential with shared cached context",
+            CompositionPattern.DELEGATION_CHAIN: "Hierarchical coordinator → specialists",
+        }
+        return descriptions.get(pattern, "Custom composition pattern")
+
+    def _analyze_task(self, task: str, context: dict[str, Any]) -> TaskRequirements:
+        """Analyze task to extract requirements.
+
+        Args:
+            task: Task description
+            context: Context dictionary
+
+        Returns:
+            TaskRequirements with extracted information
+        """
+        task_lower = task.lower()
+
+        # Determine complexity
+        complexity = self._classify_complexity(task_lower)
+
+        # Determine domain
+        domain = self._classify_domain(task_lower)
+
+        # Extract needed capabilities
+        capabilities = self._extract_capabilities(domain, context)
+
+        # Determine if parallelizable
+        parallelizable = self._is_parallelizable(task_lower, complexity)
+
+        # Extract quality gates from context
+        quality_gates = context.get("quality_gates", {})
+
+        return TaskRequirements(
+            complexity=complexity,
+            domain=domain,
+            capabilities_needed=capabilities,
+            parallelizable=parallelizable,
+            quality_gates=quality_gates,
+            context=context,
+        )
+
+    def _classify_complexity(self, task_lower: str) -> TaskComplexity:
+        """Classify task complexity based on keywords.
+
+        Args:
+            task_lower: Lowercase task description
+
+        Returns:
+            TaskComplexity classification
+        """
+        # Check for complex keywords first (most specific)
+        for keyword in self.COMPLEXITY_KEYWORDS[TaskComplexity.COMPLEX]:
+            if keyword in task_lower:
+                return TaskComplexity.COMPLEX
+
+        # Check for moderate keywords
+        for keyword in self.COMPLEXITY_KEYWORDS[TaskComplexity.MODERATE]:
+            if keyword in task_lower:
+                return TaskComplexity.MODERATE
+
+        # Check for simple keywords
+        for keyword in self.COMPLEXITY_KEYWORDS[TaskComplexity.SIMPLE]:
+            if keyword in task_lower:
+                return TaskComplexity.SIMPLE
+
+        # Default to moderate if no keywords match
+        return TaskComplexity.MODERATE
+
+    def _classify_domain(self, task_lower: str) -> TaskDomain:
+        """Classify task domain based on keywords.
+
+        Args:
+            task_lower: Lowercase task description
+
+        Returns:
+            TaskDomain classification
+        """
+        # Score each domain based on keyword matches
+        domain_scores: dict[TaskDomain, int] = dict.fromkeys(TaskDomain, 0)
+
+        for domain, keywords in self.DOMAIN_KEYWORDS.items():
+            for keyword in keywords:
+                if keyword in task_lower:
+                    domain_scores[domain] += 1
+
+        # Return domain with highest score
+        max_score = max(domain_scores.values())
+        if max_score > 0:
+            for domain, score in domain_scores.items():
+                if score == max_score:
+                    return domain
+
+        # Default to general if no keywords match
+        return TaskDomain.GENERAL
+
+    def _extract_capabilities(self, domain: TaskDomain, context: dict[str, Any]) -> list[str]:
+        """Extract needed capabilities based on domain.
+
+        Args:
+            domain: Task domain
+            context: Context dictionary
+
+        Returns:
+            List of capability names
+        """
+        # Get default capabilities for domain
+        capabilities = self.DOMAIN_CAPABILITIES.get(domain, []).copy()
+
+        # Add capabilities from context if provided
+        if "capabilities" in context:
+            additional = context["capabilities"]
+            if isinstance(additional, list):
+                capabilities.extend(additional)
+
+        return capabilities
+
+    def _is_parallelizable(self, task_lower: str, complexity: TaskComplexity) -> bool:
+        """Determine if task can be parallelized.
+
+        Args:
+            task_lower: Lowercase task description
+            complexity: Task complexity
+
+        Returns:
+            True if task can be parallelized
+        """
+        # Keywords indicating parallel execution
+        parallel_keywords = [
+            "release",
+            "audit",
+            "check",
+            "validate",
+            "review",
+        ]
+
+        # Keywords indicating sequential execution
+        sequential_keywords = [
+            "migrate",
+            "refactor",
+            "generate",
+            "create",
+        ]
+
+        # Check for sequential keywords first (higher precedence)
+        for keyword in sequential_keywords:
+            if keyword in task_lower:
+                return False
+
+        # Check for parallel keywords
+        for keyword in parallel_keywords:
+            if keyword in task_lower:
+                return True
+
+        # Complex tasks often benefit from parallel execution
+        return complexity == TaskComplexity.COMPLEX
+
+    def _select_agents(self, requirements: TaskRequirements) -> list[AgentTemplate]:
+        """Select appropriate agents based on requirements.
+
+        Args:
+            requirements: Task requirements
+
+        Returns:
+            List of agent templates
+
+        Raises:
+            ValueError: If no agents match requirements
+        """
+        agents: list[AgentTemplate] = []
+
+        # Select agents based on needed capabilities
+        for capability in requirements.capabilities_needed:
+            templates = get_templates_by_capability(capability)
+            if templates:
+                # Pick the first template with this capability
+                # In future: could rank by success rate, cost, etc.
+                agent = templates[0]
+                if agent not in agents:
+                    agents.append(agent)
+
+        # If no agents found, use domain-appropriate default
+        if not agents:
+            agents = self._get_default_agents(requirements.domain)
+
+        if not agents:
+            raise ValueError(f"No agents available for domain: {requirements.domain.value}")
+
+        return agents
+
+    def _get_default_agents(self, domain: TaskDomain) -> list[AgentTemplate]:
+        """Get default agents for a domain.
+
+        Args:
+            domain: Task domain
+
+        Returns:
+            List of default agent templates
+        """
+        defaults = {
+            TaskDomain.TESTING: ["test_coverage_analyzer"],
+            TaskDomain.SECURITY: ["security_auditor"],
+            TaskDomain.CODE_QUALITY: ["code_reviewer"],
+            TaskDomain.DOCUMENTATION: ["documentation_writer"],
+            TaskDomain.PERFORMANCE: ["performance_optimizer"],
+            TaskDomain.ARCHITECTURE: ["architecture_analyst"],
+            TaskDomain.REFACTORING: ["refactoring_specialist"],
+        }
+
+        template_ids = defaults.get(domain, ["code_reviewer"])
+        agents = []
+        for template_id in template_ids:
+            template = get_template(template_id)
+            if template:
+                agents.append(template)
+
+        return agents
+
+    def _choose_composition_pattern(
+        self, requirements: TaskRequirements, agents: list[AgentTemplate]
+    ) -> CompositionPattern:
+        """Choose optimal composition pattern.
+
+        Args:
+            requirements: Task requirements
+            agents: Selected agents
+
+        Returns:
+            CompositionPattern to use
+        """
+        num_agents = len(agents)
+        context = requirements.context
+
+        # Anthropic Pattern 8: Tool-Enhanced (single agent + tools preferred)
+        if num_agents == 1 and context.get("tools"):
+            return CompositionPattern.TOOL_ENHANCED
+
+        # Anthropic Pattern 10: Delegation Chain (hierarchical coordination)
+        # Use when: Complex task + coordinator pattern + 2+ specialists
+        has_coordinator = any("coordinator" in agent.role.lower() for agent in agents)
+        if (
+            requirements.complexity == TaskComplexity.COMPLEX
+            and has_coordinator
+            and num_agents >= 2
+        ):
+            return CompositionPattern.DELEGATION_CHAIN
+
+        # Anthropic Pattern 9: Prompt-Cached Sequential (large shared context)
+        # Use when: 3+ agents need same large context (>2000 tokens)
+        large_context = context.get("cached_context") or context.get("shared_knowledge")
+        if num_agents >= 3 and large_context and len(str(large_context)) > 2000:
+            return CompositionPattern.PROMPT_CACHED_SEQUENTIAL
+
+        # Parallelizable tasks: use parallel strategy (check before single agent)
+        if requirements.parallelizable:
+            return CompositionPattern.PARALLEL
+
+        # Security/architecture: benefit from multiple perspectives (even with 1 agent)
+        if requirements.domain in [TaskDomain.SECURITY, TaskDomain.ARCHITECTURE]:
+            return CompositionPattern.PARALLEL
+
+        # Documentation: teaching pattern (cheap → validate → expert if needed)
+        if requirements.domain == TaskDomain.DOCUMENTATION:
+            return CompositionPattern.TEACHING
+
+        # Refactoring: refinement pattern (identify → refactor → validate)
+        if requirements.domain == TaskDomain.REFACTORING:
+            return CompositionPattern.REFINEMENT
+
+        # Single agent: sequential (after domain-specific patterns)
+        if num_agents == 1:
+            return CompositionPattern.SEQUENTIAL
+
+        # Multiple agents with same capability: debate/consensus
+        capabilities = [cap for agent in agents for cap in agent.capabilities]
+        if len(capabilities) != len(set(capabilities)):
+            # Duplicate capabilities detected → debate
+            return CompositionPattern.DEBATE
+
+        # Testing domain: typically sequential (analyze → generate → validate)
+        if requirements.domain == TaskDomain.TESTING:
+            return CompositionPattern.SEQUENTIAL
+
+        # Complex tasks: adaptive routing
+        if requirements.complexity == TaskComplexity.COMPLEX:
+            return CompositionPattern.ADAPTIVE
+
+        # Default: sequential
+        return CompositionPattern.SEQUENTIAL
+
+    def _estimate_cost(self, agents: list[AgentTemplate]) -> float:
+        """Estimate execution cost based on agent tiers.
+
+        Args:
+            agents: List of agents
+
+        Returns:
+            Estimated cost in arbitrary units
+        """
+        tier_costs = {
+            "CHEAP": 1.0,
+            "CAPABLE": 3.0,
+            "PREMIUM": 10.0,
+        }
+
+        total_cost = 0.0
+        for agent in agents:
+            total_cost += tier_costs.get(agent.tier_preference, 3.0)
+
+        return total_cost
+
+    def _estimate_duration(self, agents: list[AgentTemplate], strategy: CompositionPattern) -> int:
+        """Estimate execution duration in seconds.
+
+        Args:
+            agents: List of agents
+            strategy: Composition pattern
+
+        Returns:
+            Estimated duration in seconds
+        """
+        # Get max timeout from agents
+        max_timeout = max(
+            (agent.resource_requirements.timeout_seconds for agent in agents),
+            default=300,
+        )
+
+        # Sequential: sum of timeouts
+        if strategy == CompositionPattern.SEQUENTIAL:
+            return sum(agent.resource_requirements.timeout_seconds for agent in agents)
+
+        # Parallel: max timeout
+        if strategy == CompositionPattern.PARALLEL:
+            return max_timeout
+
+        # Debate: multiple rounds, estimate 2x max timeout
+        if strategy == CompositionPattern.DEBATE:
+            return max_timeout * 2
+
+        # Teaching: initial attempt + possible expert review
+        if strategy == CompositionPattern.TEACHING:
+            return int(max_timeout * 1.5)
+
+        # Refinement: 3 passes (draft → review → polish)
+        if strategy == CompositionPattern.REFINEMENT:
+            return max_timeout * 3
+
+        # Adaptive: classification + specialist
+        if strategy == CompositionPattern.ADAPTIVE:
+            return int(max_timeout * 1.2)
+
+        # Anthropic Pattern 8: Tool-Enhanced (single agent with tools, efficient)
+        if strategy == CompositionPattern.TOOL_ENHANCED:
+            return max_timeout  # Similar to sequential for single agent
+
+        # Anthropic Pattern 9: Prompt-Cached Sequential (faster with cache hits)
+        if strategy == CompositionPattern.PROMPT_CACHED_SEQUENTIAL:
+            # Sequential but 20% faster due to cached context reducing token processing
+            total = sum(agent.resource_requirements.timeout_seconds for agent in agents)
+            return int(total * 0.8)
+
+        # Anthropic Pattern 10: Delegation Chain (coordinator + specialists in sequence)
+        if strategy == CompositionPattern.DELEGATION_CHAIN:
+            # Coordinator analyzes, then specialists execute (sequential-like)
+            return sum(agent.resource_requirements.timeout_seconds for agent in agents)
+
+        # Conditional: branch evaluation + selected path
+        if strategy == CompositionPattern.CONDITIONAL:
+            return int(max_timeout * 1.1)
+
+        # Default: max timeout
+        return max_timeout