claude-mpm 4.8.0__py3-none-any.whl → 4.8.3__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()