claude-mpm 4.1.26__py3-none-any.whl → 4.24.0__py3-none-any.whl

claude_mpm/hooks/failure_learning/learning_extraction_hook.py

@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+"""
+Learning Extraction Hook
+=========================
+
+Synthesizes learnings from failure-fix pairs and writes them to agent memory.
+
+WHY: The final step in the failure-learning cycle is extracting actionable
+learnings and persisting them to agent memory files. This hook completes the
+cycle by taking failure-fix pairs and creating formatted learning entries.
+
+DESIGN DECISION: This hook runs last (priority=89) after both failure and fix
+detection. It uses template-based synthesis for MVP (no AI) and integrates
+with AgentMemoryManager to write learnings to the appropriate memory files.
+
+Integration points:
+- Monitors for fix_detected metadata from FixDetectionHook
+- Extracts learnings using FailureTracker
+- Formats learnings as markdown
+- Writes to agent memory files via AgentMemoryManager
+"""
+
+import logging
+from typing import Any
+
+from claude_mpm.hooks.base_hook import (
+    BaseHook,
+    HookContext,
+    HookResult,
+    HookType,
+)
+from claude_mpm.services.memory.failure_tracker import get_failure_tracker
+
+logger = logging.getLogger(__name__)
+
+
+class LearningExtractionHook(BaseHook):
+    """Hook that extracts and persists learnings from failure-fix pairs.
+
+    WHY: Automatically converts failure-fix pairs into persistent learnings
+    stored in agent memory files. This completes the failure-learning cycle
+    without requiring manual intervention.
+
+    DESIGN DECISION: Priority 89 ensures this runs last in the chain:
+    1. Failure detection (85) - detects failures
+    2. Fix detection (87) - matches fixes with failures
+    3. Learning extraction (89) - creates and persists learnings
+
+    MVP uses template-based learning synthesis. Future versions could use
+    AI to analyze git diffs, code changes, and generate richer learnings.
+    """
+
+    def __init__(self):
+        """Initialize the learning extraction hook."""
+        super().__init__(
+            name="learning_extraction",
+            priority=89,  # Last in the chain, after fix detection
+        )
+        self.tracker = get_failure_tracker()
+        self._memory_manager = None
+
+    @property
+    def memory_manager(self):
+        """Lazy-load memory manager to avoid circular imports.
+
+        WHY: AgentMemoryManager may import hooks, so we lazy-load to prevent
+        circular dependency issues.
+
+        Returns:
+            AgentMemoryManager instance
+        """
+        if self._memory_manager is None:
+            try:
+                from claude_mpm.services.agents.memory.agent_memory_manager import (
+                    get_memory_manager,
+                )
+
+                self._memory_manager = get_memory_manager()
+            except ImportError as e:
+                logger.error(f"Failed to import AgentMemoryManager: {e}")
+                raise
+
+        return self._memory_manager
+
+    def execute(self, context: HookContext) -> HookResult:
+        """Execute learning extraction from failure-fix pairs.
+
+        WHY: When a fix is detected, we have everything needed to extract a
+        learning: the original failure, the fix that resolved it, and the
+        context. This method synthesizes a learning and writes it to memory.
+
+        Args:
+            context: Hook context containing fix detection metadata
+
+        Returns:
+            HookResult with extraction results
+        """
+        try:
+            # Check if this is a fix detection event
+            metadata = context.metadata or {}
+            if not metadata.get("fix_detected"):
+                # Not a fix event, skip
+                return HookResult(success=True, modified=False)
+
+            # Extract failure and fix events from metadata
+            failure_event = metadata.get("failure_event")
+            fix_event = metadata.get("fix_event")
+
+            if not failure_event or not fix_event:
+                logger.warning("Fix detected but failure/fix events not in metadata")
+                return HookResult(success=True, modified=False)
+
+            # Extract learning from failure-fix pair
+            learning = self.tracker.extract_learning(
+                fix_event=fix_event,
+                failure_event=failure_event,
+                target_agent=self._determine_target_agent(context, failure_event),
+            )
+
+            # Format learning as markdown
+            learning_markdown = learning.to_markdown()
+
+            # Write to agent memory
+            success = self._write_to_memory(
+                agent_id=learning.target_agent, learning_text=learning_markdown
+            )
+
+            if success:
+                logger.info(
+                    f"Learning extracted and saved for {learning.target_agent}: "
+                    f"{learning.category}"
+                )
+                return HookResult(
+                    success=True,
+                    modified=False,
+                    metadata={
+                        "learning_extracted": True,
+                        "target_agent": learning.target_agent,
+                        "learning_category": learning.category,
+                    },
+                )
+            logger.warning(
+                f"Failed to write learning to memory for {learning.target_agent}"
+            )
+            return HookResult(
+                success=True,  # Don't fail the hook, just log warning
+                modified=False,
+                metadata={"learning_extracted": False},
+            )
+
+        except Exception as e:
+            logger.error(f"Error in learning extraction hook: {e}", exc_info=True)
+            return HookResult(success=False, error=str(e), modified=False)
+
+    def validate(self, context: HookContext) -> bool:
+        """Validate if this hook should run for the given context.
+
+        Args:
+            context: Hook context to validate
+
+        Returns:
+            True if hook should execute
+        """
+        if not super().validate(context):
+            return False
+
+        # Run for POST_DELEGATION events (after tool execution)
+        if context.hook_type != HookType.POST_DELEGATION:
+            return False
+
+        # Must have fix detection metadata
+        metadata = context.metadata or {}
+        return metadata.get("fix_detected", False)
+
+    def _determine_target_agent(self, context: HookContext, failure_event: Any) -> str:
+        """Determine which agent should receive the learning.
+
+        WHY: Learnings should go to the agent most likely to benefit from them.
+        This method applies routing logic to determine the best target.
+
+        Args:
+            context: Hook context
+            failure_event: The failure event
+
+        Returns:
+            Agent identifier (PM, engineer, qa, etc.)
+        """
+        # Try to get agent from context first
+        agent_type = (
+            context.data.get("agent_type")
+            or context.data.get("subagent_type")
+            or context.metadata.get("agent_type")
+        )
+        if agent_type:
+            return agent_type
+
+        # Check failure event context
+        if hasattr(failure_event, "context") and failure_event.context.get(
+            "agent_type"
+        ):
+            return failure_event.context["agent_type"]
+
+        # Fall back to task-based routing
+        if hasattr(failure_event, "task_type"):
+            task_type = failure_event.task_type
+            if task_type in ("test", "lint"):
+                return "qa"
+            if task_type in ("build", "install", "script") or task_type == "git":
+                return "engineer"
+
+        # Default to PM
+        return "PM"
+
+    def _write_to_memory(self, agent_id: str, learning_text: str) -> bool:
+        """Write learning to agent memory file.
+
+        WHY: Learnings must be persisted to memory files so agents can access
+        them in future sessions. This method uses AgentMemoryManager to handle
+        the actual file operations.
+
+        Args:
+            agent_id: Agent identifier
+            learning_text: Markdown-formatted learning
+
+        Returns:
+            True if write succeeded, False otherwise
+        """
+        try:
+            # Parse learning sections from markdown
+            learning_items = self._parse_learning_markdown(learning_text)
+
+            if not learning_items:
+                logger.warning(f"No learning items parsed from: {learning_text}")
+                return False
+
+            # Add to agent memory
+            return self.memory_manager.update_agent_memory(
+                agent_id=agent_id, new_items=learning_items
+            )
+
+        except Exception as e:
+            logger.error(f"Failed to write learning to memory for {agent_id}: {e}")
+            return False
+
+    def _parse_learning_markdown(self, learning_markdown: str) -> list:
+        """Parse learning markdown into list items for memory.
+
+        WHY: AgentMemoryManager expects a list of learning items. We need to
+        convert the markdown-formatted learning into individual list items.
+
+        Args:
+            learning_markdown: Markdown-formatted learning
+
+        Returns:
+            List of learning items
+        """
+        items = []
+
+        # Split by lines and extract bullet points
+        lines = learning_markdown.split("\n")
+
+        for line in lines:
+            line = line.strip()
+            if line.startswith("- **"):
+                # This is a learning item (e.g., "- **Problem**: ...")
+                items.append(line)
+            elif line.startswith("## "):
+                # This is a category header, skip it
+                continue
+
+        # If no items found, return the whole thing as a single item
+        if not items:
+            items = [f"- {learning_markdown.strip()}"]
+
+        return items
+
+
+def get_learning_extraction_hook() -> LearningExtractionHook:
+    """Factory function to create learning extraction hook.
+
+    WHY: Provides consistent hook creation pattern used throughout the framework.
+
+    Returns:
+        Configured LearningExtractionHook instance
+    """
+    return LearningExtractionHook()

claude_mpm/hooks/instruction_reinforcement.py

@@ -0,0 +1,301 @@
+"""
+Instruction Reinforcement Hook for PM Delegation Compliance
+
+This hook monitors PM behavior for delegation violations and provides
+escalating warnings when the PM attempts to implement instead of delegate.
+"""
+
+import re
+from dataclasses import dataclass
+from enum import Enum
+from typing import Dict, List, Optional, Tuple
+
+from claude_mpm.core.enums import OperationResult
+from claude_mpm.core.logging_utils import get_logger
+
+logger = get_logger(__name__)
+
+
+class ViolationType(Enum):
+    """Types of PM delegation violations"""
+
+    EDIT_ATTEMPT = "Edit/Write/MultiEdit"
+    BASH_IMPLEMENTATION = "Bash for implementation"
+    FILE_CREATION = "Direct file creation"
+    TEST_EXECUTION = "Test execution"
+    DEPLOYMENT = "Deployment operation"
+    CODE_WRITING = "Code writing"
+
+
+@dataclass
+class Violation:
+    """Record of a delegation violation"""
+
+    violation_type: ViolationType
+    context: str
+    timestamp: float
+    severity_level: int
+
+
+class InstructionReinforcementHook:
+    """
+    Monitors PM messages for delegation violations and provides corrective feedback.
+
+    Circuit breaker pattern implementation to prevent PM from implementing
+    instead of delegating work to appropriate agents.
+    """
+
+    def __init__(self):
+        self.violation_count = 0
+        self.violations: List[Violation] = []
+
+        # Patterns that indicate PM is attempting forbidden actions
+        self.forbidden_patterns = [
+            # Direct implementation language - EXPANDED
+            (
+                r"I'll\s+(fix|create|write|implement|code|build|update|generate|modify|set\s+up|configure|optimize|rewrite|run|test|deploy|push|analyze|review|setup)",
+                ViolationType.CODE_WRITING,
+            ),
+            (
+                r"I'm\s+(fix|create|write|implement|code|build|update|generate|modify)",
+                ViolationType.CODE_WRITING,
+            ),
+            (
+                r"Let\s+me\s+(edit|write|modify|create|update|fix|run|execute|commit|refactor|configure|set\s+up|review)",
+                ViolationType.EDIT_ATTEMPT,
+            ),
+            (
+                r"I\s+will\s+(implement|code|build|create|write|update|fix)",
+                ViolationType.CODE_WRITING,
+            ),
+            (
+                r"I'm\s+(going\s+to|about\s+to)\s+(fix|create|write|implement|update|modify)",
+                ViolationType.CODE_WRITING,
+            ),
+            # Common honeypot phrases
+            (
+                r"Here's\s+(the|my|an?)\s+(implementation|code|SQL|query|solution|analysis|fix)",
+                ViolationType.CODE_WRITING,
+            ),
+            (
+                r"The\s+(query|code|implementation|solution)\s+(would\s+be|is)",
+                ViolationType.CODE_WRITING,
+            ),
+            (r"I\s+(found|identified)\s+(these\s+)?issues", ViolationType.CODE_WRITING),
+            # Deployment and setup patterns
+            (
+                r"Setting\s+up\s+(the\s+)?(authentication|containers|environment|docker)",
+                ViolationType.DEPLOYMENT,
+            ),
+            (r"Deploying\s+(to|the)", ViolationType.DEPLOYMENT),
+            (r"I'll\s+(deploy|push|host|launch)", ViolationType.DEPLOYMENT),
+            # Tool usage patterns - EXPANDED
+            (r"Using\s+(Edit|Write|MultiEdit)\s+tool", ViolationType.EDIT_ATTEMPT),
+            (r"<invoke\s+name=\"(Edit|Write|MultiEdit)\"", ViolationType.EDIT_ATTEMPT),
+            (
+                r"Running\s+(bash\s+command|git\s+commit|npm|yarn|python|node|go|tests|pytest)",
+                ViolationType.BASH_IMPLEMENTATION,
+            ),
+            (r"Executing\s+(tests|test\s+suite|pytest)", ViolationType.TEST_EXECUTION),
+            # Testing patterns - EXPANDED
+            (
+                r"(Testing|I'll\s+test|Let\s+me\s+test)\s+(the\s+)?(payment|API|endpoint)",
+                ViolationType.TEST_EXECUTION,
+            ),
+            (
+                r"I'll\s+(run|execute|verify)\s+(the\s+)?(tests|test\s+suite|endpoint)",
+                ViolationType.TEST_EXECUTION,
+            ),
+            (r"pytest|npm\s+test|yarn\s+test|go\s+test", ViolationType.TEST_EXECUTION),
+            # File operation patterns - EXPANDED
+            (
+                r"Creating\s+(new\s+|a\s+|the\s+)?(file|YAML|README|workflow)",
+                ViolationType.FILE_CREATION,
+            ),
+            (r"Writing\s+to\s+file", ViolationType.FILE_CREATION),
+            (
+                r"Updating\s+(the\s+)?(code|component|queries)",
+                ViolationType.CODE_WRITING,
+            ),
+            (
+                r"I'll\s+(update|modify)\s+(the\s+)?(component|code|React)",
+                ViolationType.CODE_WRITING,
+            ),
+        ]
+
+        # Patterns for correct delegation behavior - EXPANDED
+        self.delegation_patterns = [
+            r"delegat(e|ing)\s+(to|this)",
+            r"Task\s+tool",
+            r"(asking|request|have|use)\s+\w+\s+agent",
+            r"requesting\s+\w+\s+to",
+            r"will\s+(have|ask|request)\s+\w+\s+agent",
+            r"I'll\s+(have|ask|request|delegate)",
+            r"the\s+\w+\s+agent\s+(will|can|should)",
+        ]
+
+    def detect_violation_intent(
+        self, message: str
+    ) -> Optional[Tuple[ViolationType, str]]:
+        """
+        Check message for patterns indicating PM violation intent.
+
+        Args:
+            message: The PM's message to analyze
+
+        Returns:
+            Tuple of (ViolationType, matched_text) if violation detected, None otherwise
+        """
+        message_lower = message.lower()
+
+        # Check for forbidden patterns
+        for pattern, violation_type in self.forbidden_patterns:
+            match = re.search(pattern, message_lower, re.IGNORECASE)
+            if match:
+                # Check if this is actually a delegation (false positive check)
+                is_delegation = any(
+                    re.search(
+                        del_pattern,
+                        message_lower[max(0, match.start() - 50) : match.end() + 50],
+                    )
+                    for del_pattern in self.delegation_patterns
+                )
+
+                if not is_delegation:
+                    return (violation_type, match.group(0))
+
+        return None
+
+    def escalate_warning(self) -> str:
+        """
+        Generate escalating warning message based on violation count.
+
+        Returns:
+            Warning message with appropriate severity
+        """
+        if self.violation_count == 1:
+            return (
+                "⚠️ DELEGATION REMINDER: PM must delegate ALL implementation work.\n"
+                "Use the Task tool to delegate to the appropriate agent."
+            )
+        if self.violation_count == 2:
+            return (
+                "🚨 DELEGATION WARNING: Critical PM violation detected!\n"
+                "You MUST delegate implementation work. Do NOT use Edit/Write/Bash for implementation.\n"
+                "Next violation will result in session failure."
+            )
+        if self.violation_count == 3:
+            return (
+                "❌ CRITICAL DELEGATION FAILURE: Multiple PM violations detected.\n"
+                "PM has repeatedly attempted to implement instead of delegate.\n"
+                "Session integrity compromised. All work must be delegated to agents."
+            )
+        return (
+            f"❌❌❌ SEVERE VIOLATION (Count: {self.violation_count}): PM continues to violate delegation rules.\n"
+            "MANDATORY: Use Task tool to delegate ALL implementation to appropriate agents.\n"
+            "Current session may need to be terminated and restarted."
+        )
+
+    def check_message(self, message: str) -> Optional[Dict[str, any]]:
+        """
+        Check a PM message for violations and return feedback if needed.
+
+        Args:
+            message: The PM's message to check
+
+        Returns:
+            Dictionary with violation details and correction, or None if compliant
+        """
+        violation_result = self.detect_violation_intent(message)
+
+        if violation_result:
+            violation_type, context = violation_result
+            self.violation_count += 1
+
+            # Record the violation
+            import time
+
+            violation = Violation(
+                violation_type=violation_type,
+                context=context,
+                timestamp=time.time(),
+                severity_level=min(self.violation_count, 4),
+            )
+            self.violations.append(violation)
+
+            # Generate corrective feedback
+            warning = self.escalate_warning()
+
+            # Determine which agent should handle this
+            agent_mapping = {
+                ViolationType.EDIT_ATTEMPT: "Engineer",
+                ViolationType.CODE_WRITING: "Engineer",
+                ViolationType.BASH_IMPLEMENTATION: "Engineer or Ops",
+                ViolationType.FILE_CREATION: "Engineer or Documentation",
+                ViolationType.TEST_EXECUTION: "QA",
+                ViolationType.DEPLOYMENT: "Ops",
+            }
+
+            suggested_agent = agent_mapping.get(violation_type, "appropriate agent")
+
+            # Clean up context for task suggestion
+            clean_context = (
+                context.replace("I will ", "")
+                .replace("I'll ", "")
+                .replace("Let me ", "")
+            )
+
+            return {
+                "violation_detected": True,
+                "violation_count": self.violation_count,
+                "violation_type": violation_type.value,
+                "context": context,
+                "warning": warning,
+                "correction": f"MUST delegate to {suggested_agent} using Task tool",
+                "suggested_task": f"Task: Please {clean_context}",
+                "severity": min(self.violation_count, 4),
+            }
+
+        return None
+
+    def get_violation_summary(self) -> Dict[str, any]:
+        """
+        Get a summary of all violations in the session.
+
+        Returns:
+            Dictionary with violation statistics and details
+        """
+        if not self.violations:
+            return {
+                "total_violations": 0,
+                "status": OperationResult.SUCCESS,
+                "message": "No PM delegation violations detected",
+            }
+
+        violation_types = {}
+        for v in self.violations:
+            vtype = v.violation_type.value
+            violation_types[vtype] = violation_types.get(vtype, 0) + 1
+
+        status = (
+            OperationResult.ERROR
+            if self.violation_count < 3
+            else OperationResult.FAILED
+        )
+
+        return {
+            "total_violations": self.violation_count,
+            "status": status,
+            "violation_types": violation_types,
+            "most_recent": self.violations[-1].context if self.violations else None,
+            "recommendation": (
+                "Review PM delegation training"
+                if self.violation_count > 2
+                else "Continue monitoring"
+            ),
+        }
+
+    def reset(self):
+        """Reset violation tracking for a new session"""
+        self.violation_count = 0
+        self.violations = []