htmlgraph 0.26.23__py3-none-any.whl → 0.26.24__py3-none-any.whl

htmlgraph/analytics/pattern_learning.py

@@ -0,0 +1,771 @@
+"""
+Pattern Learning from Agent Behavior - Phase 2 Feature 2.
+
+Analyzes tool call sequences to identify patterns, anti-patterns, and optimization
+opportunities. Provides actionable recommendations based on historical behavior.
+
+Key Components:
+1. PatternMatcher - Identifies sequences of tool types from event history
+2. InsightGenerator - Converts patterns to actionable recommendations
+3. LearningLoop - Stores patterns and refines based on user feedback
+
+Usage:
+    from htmlgraph.analytics.pattern_learning import PatternLearner
+
+    learner = PatternLearner()
+    patterns = learner.detect_patterns(min_frequency=5)
+    insights = learner.generate_insights()
+    recommendations = learner.get_recommendations()
+"""
+
+import json
+import sqlite3
+from collections import defaultdict
+from dataclasses import dataclass, field
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+
+@dataclass
+class ToolPattern:
+    """
+    Represents a detected tool call sequence pattern.
+
+    Attributes:
+        pattern_id: Unique identifier for the pattern
+        sequence: List of tool names in order (e.g., ["Read", "Grep", "Edit"])
+        frequency: Number of times this pattern occurs
+        success_rate: Percentage of times pattern led to successful outcomes
+        avg_duration_seconds: Average time taken for this pattern
+        last_seen: When this pattern was last observed
+        sessions: List of session IDs where pattern occurred
+        user_feedback: User feedback score (1=helpful, 0=neutral, -1=unhelpful)
+    """
+
+    pattern_id: str
+    sequence: list[str]
+    frequency: int = 0
+    success_rate: float = 0.0
+    avg_duration_seconds: float = 0.0
+    last_seen: datetime | None = None
+    sessions: list[str] = field(default_factory=list)
+    user_feedback: int = 0
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert pattern to dictionary."""
+        return {
+            "pattern_id": self.pattern_id,
+            "sequence": self.sequence,
+            "frequency": self.frequency,
+            "success_rate": self.success_rate,
+            "avg_duration_seconds": self.avg_duration_seconds,
+            "last_seen": self.last_seen.isoformat() if self.last_seen else None,
+            "sessions": self.sessions,
+            "user_feedback": self.user_feedback,
+        }
+
+
+@dataclass
+class PatternInsight:
+    """
+    Actionable insight generated from pattern analysis.
+
+    Attributes:
+        insight_id: Unique identifier
+        insight_type: Type of insight (recommendation, anti-pattern, optimization)
+        title: Human-readable title
+        description: Detailed explanation
+        impact_score: Estimated impact (0-100)
+        patterns: Related pattern IDs
+    """
+
+    insight_id: str
+    insight_type: str  # "recommendation", "anti-pattern", "optimization"
+    title: str
+    description: str
+    impact_score: float
+    patterns: list[str] = field(default_factory=list)
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert insight to dictionary."""
+        return {
+            "insight_id": self.insight_id,
+            "insight_type": self.insight_type,
+            "title": self.title,
+            "description": self.description,
+            "impact_score": self.impact_score,
+            "patterns": self.patterns,
+        }
+
+
+class PatternMatcher:
+    """
+    Identifies sequences of tool types from event history.
+
+    Uses sliding window approach to find common tool call patterns.
+    """
+
+    def __init__(self, db_path: Path | str):
+        """
+        Initialize pattern matcher.
+
+        Args:
+            db_path: Path to HtmlGraph database
+        """
+        self.db_path = Path(db_path)
+
+        if not self.db_path.exists():
+            raise FileNotFoundError(f"Database not found at {self.db_path}")
+
+    def _get_connection(self) -> sqlite3.Connection:
+        """Get database connection."""
+        conn = sqlite3.connect(str(self.db_path))
+        conn.row_factory = sqlite3.Row
+        return conn
+
+    def get_tool_sequences(
+        self, window_size: int = 3, session_id: str | None = None
+    ) -> list[tuple[list[str], str, datetime]]:
+        """
+        Extract tool call sequences from database.
+
+        Args:
+            window_size: Number of consecutive tools in each sequence
+            session_id: Optional session ID to filter by
+
+        Returns:
+            List of (sequence, session_id, timestamp) tuples
+        """
+        conn = self._get_connection()
+        try:
+            # Query tool calls ordered by timestamp
+            query = """
+                SELECT tool_name, session_id, timestamp
+                FROM agent_events
+                WHERE event_type = 'tool_call'
+                AND tool_name IS NOT NULL
+            """
+
+            params: tuple[Any, ...] = ()
+            if session_id:
+                query += " AND session_id = ?"
+                params = (session_id,)
+
+            query += " ORDER BY timestamp ASC"
+
+            cursor = conn.cursor()
+            cursor.execute(query, params)
+
+            # Group by session and extract sequences
+            session_tools: dict[str, list[tuple[str, datetime]]] = defaultdict(list)
+            for row in cursor.fetchall():
+                tool = row["tool_name"]
+                sess_id = row["session_id"]
+                timestamp = (
+                    datetime.fromisoformat(row["timestamp"])
+                    if isinstance(row["timestamp"], str)
+                    else row["timestamp"]
+                )
+                session_tools[sess_id].append((tool, timestamp))
+
+            # Extract sliding windows
+            sequences = []
+            for sess_id, tools in session_tools.items():
+                for i in range(len(tools) - window_size + 1):
+                    sequence = [t[0] for t in tools[i : i + window_size]]
+                    timestamp = tools[i + window_size - 1][
+                        1
+                    ]  # Use last tool's timestamp
+                    sequences.append((sequence, sess_id, timestamp))
+
+            return sequences
+        finally:
+            conn.close()
+
+    def find_patterns(
+        self, window_size: int = 3, min_frequency: int = 5
+    ) -> list[ToolPattern]:
+        """
+        Identify common tool call patterns.
+
+        Args:
+            window_size: Size of tool sequence window (default: 3)
+            min_frequency: Minimum occurrences to be considered a pattern
+
+        Returns:
+            List of detected patterns sorted by frequency
+        """
+        sequences = self.get_tool_sequences(window_size=window_size)
+
+        # Count sequence frequencies
+        sequence_data: dict[tuple[str, ...], list[tuple[str, datetime]]] = defaultdict(
+            list
+        )
+        for seq, sess_id, timestamp in sequences:
+            sequence_data[tuple(seq)].append((sess_id, timestamp))
+
+        # Filter by minimum frequency
+        patterns = []
+        for seq_tuple, occurrences in sequence_data.items():
+            if len(occurrences) >= min_frequency:
+                sequence = list(seq_tuple)
+                pattern_id = self._generate_pattern_id(sequence)
+
+                sessions = [occ[0] for occ in occurrences]
+                last_seen = max(occ[1] for occ in occurrences)
+
+                pattern = ToolPattern(
+                    pattern_id=pattern_id,
+                    sequence=sequence,
+                    frequency=len(occurrences),
+                    last_seen=last_seen,
+                    sessions=sessions,
+                )
+                patterns.append(pattern)
+
+        # Sort by frequency descending
+        patterns.sort(key=lambda p: p.frequency, reverse=True)
+        return patterns
+
+    def _generate_pattern_id(self, sequence: list[str]) -> str:
+        """Generate unique pattern ID from sequence."""
+        seq_str = "->".join(sequence)
+        # Simple hash-based ID
+        import hashlib
+
+        hash_obj = hashlib.md5(seq_str.encode())
+        return f"pat-{hash_obj.hexdigest()[:8]}"
+
+
+class InsightGenerator:
+    """
+    Converts patterns to actionable recommendations.
+
+    Analyzes pattern success rates, costs, and contexts to generate
+    insights about workflow optimization.
+    """
+
+    def __init__(self, db_path: Path | str):
+        """
+        Initialize insight generator.
+
+        Args:
+            db_path: Path to HtmlGraph database
+        """
+        self.db_path = Path(db_path)
+
+        if not self.db_path.exists():
+            raise FileNotFoundError(f"Database not found at {self.db_path}")
+
+    def _get_connection(self) -> sqlite3.Connection:
+        """Get database connection."""
+        conn = sqlite3.connect(str(self.db_path))
+        conn.row_factory = sqlite3.Row
+        return conn
+
+    def calculate_success_rate(self, pattern: ToolPattern) -> float:
+        """
+        Calculate success rate for a pattern.
+
+        Success defined as: pattern followed by passing tests or completion,
+        not followed by error/failure events.
+
+        Args:
+            pattern: Pattern to analyze
+
+        Returns:
+            Success rate as percentage (0-100)
+        """
+        if not pattern.sessions:
+            return 0.0
+
+        conn = self._get_connection()
+        try:
+            successes = 0
+            for session_id in set(pattern.sessions):
+                # Check if session has more successes than failures
+                cursor = conn.cursor()
+                cursor.execute(
+                    """
+                    SELECT
+                        SUM(CASE WHEN event_type = 'completion' THEN 1 ELSE 0 END) as completions,
+                        SUM(CASE WHEN event_type = 'error' THEN 1 ELSE 0 END) as errors
+                    FROM agent_events
+                    WHERE session_id = ?
+                    """,
+                    (session_id,),
+                )
+                row = cursor.fetchone()
+
+                completions = row["completions"] or 0
+                errors = row["errors"] or 0
+
+                # Session is successful if completions > errors
+                if completions > errors:
+                    successes += 1
+
+            return (successes / len(set(pattern.sessions))) * 100
+        finally:
+            conn.close()
+
+    def calculate_avg_duration(self, pattern: ToolPattern) -> float:
+        """
+        Calculate average duration for pattern execution.
+
+        Args:
+            pattern: Pattern to analyze
+
+        Returns:
+            Average duration in seconds
+        """
+        if not pattern.sessions:
+            return 0.0
+
+        conn = self._get_connection()
+        try:
+            cursor = conn.cursor()
+            cursor.execute(
+                """
+                SELECT AVG(execution_duration_seconds) as avg_duration
+                FROM agent_events
+                WHERE session_id IN ({})
+                AND execution_duration_seconds > 0
+                """.format(",".join("?" * len(set(pattern.sessions)))),
+                tuple(set(pattern.sessions)),
+            )
+            row = cursor.fetchone()
+            return row["avg_duration"] or 0.0
+        finally:
+            conn.close()
+
+    def enrich_pattern(self, pattern: ToolPattern) -> ToolPattern:
+        """
+        Enrich pattern with success rate and duration data.
+
+        Args:
+            pattern: Pattern to enrich
+
+        Returns:
+            Enriched pattern with calculated metrics
+        """
+        pattern.success_rate = self.calculate_success_rate(pattern)
+        pattern.avg_duration_seconds = self.calculate_avg_duration(pattern)
+        return pattern
+
+    def generate_insights(self, patterns: list[ToolPattern]) -> list[PatternInsight]:
+        """
+        Generate actionable insights from patterns.
+
+        Args:
+            patterns: List of detected patterns
+
+        Returns:
+            List of insights with recommendations
+        """
+        insights = []
+
+        # Enrich patterns with metrics
+        enriched = [self.enrich_pattern(p) for p in patterns]
+
+        # Find high-success patterns (recommendations)
+        for pattern in enriched:
+            if pattern.success_rate >= 80 and pattern.frequency >= 5:
+                insight = PatternInsight(
+                    insight_id=f"insight-{pattern.pattern_id}",
+                    insight_type="recommendation",
+                    title=f"High Success Pattern: {' → '.join(pattern.sequence)}",
+                    description=(
+                        f"This pattern has a {pattern.success_rate:.1f}% success rate "
+                        f"across {pattern.frequency} occurrences. "
+                        f"Consider using this workflow for similar tasks."
+                    ),
+                    impact_score=pattern.success_rate * (pattern.frequency / 10),
+                    patterns=[pattern.pattern_id],
+                )
+                insights.append(insight)
+
+        # Find low-success patterns (anti-patterns)
+        for pattern in enriched:
+            if pattern.success_rate < 50 and pattern.frequency >= 5:
+                insight = PatternInsight(
+                    insight_id=f"anti-{pattern.pattern_id}",
+                    insight_type="anti-pattern",
+                    title=f"Low Success Pattern: {' → '.join(pattern.sequence)}",
+                    description=(
+                        f"This pattern has only a {pattern.success_rate:.1f}% success rate "
+                        f"across {pattern.frequency} occurrences. "
+                        f"Consider alternative approaches."
+                    ),
+                    impact_score=100 - pattern.success_rate,
+                    patterns=[pattern.pattern_id],
+                )
+                insights.append(insight)
+
+        # Find repeated Read patterns (optimization opportunity)
+        for pattern in enriched:
+            if pattern.sequence.count("Read") >= 2:
+                insight = PatternInsight(
+                    insight_id=f"opt-{pattern.pattern_id}",
+                    insight_type="optimization",
+                    title="Multiple Read Operations Detected",
+                    description=(
+                        f"Pattern '{' → '.join(pattern.sequence)}' contains "
+                        f"{pattern.sequence.count('Read')} Read operations. "
+                        f"Consider delegating exploration to a subagent to reduce context usage."
+                    ),
+                    impact_score=pattern.frequency * 10,
+                    patterns=[pattern.pattern_id],
+                )
+                insights.append(insight)
+
+        # Sort by impact score
+        insights.sort(key=lambda i: i.impact_score, reverse=True)
+        return insights
+
+
+class LearningLoop:
+    """
+    Stores patterns and refines recommendations based on user feedback.
+
+    Maintains a persistent store of patterns and their effectiveness,
+    allowing the system to improve recommendations over time.
+    """
+
+    def __init__(self, db_path: Path | str):
+        """
+        Initialize learning loop.
+
+        Args:
+            db_path: Path to HtmlGraph database
+        """
+        self.db_path = Path(db_path)
+
+        if not self.db_path.exists():
+            raise FileNotFoundError(f"Database not found at {self.db_path}")
+
+        self._ensure_schema()
+
+    def _get_connection(self) -> sqlite3.Connection:
+        """Get database connection."""
+        conn = sqlite3.connect(str(self.db_path))
+        conn.row_factory = sqlite3.Row
+        return conn
+
+    def _ensure_schema(self) -> None:
+        """Create tool_patterns table if it doesn't exist."""
+        conn = self._get_connection()
+        try:
+            cursor = conn.cursor()
+            cursor.execute("""
+                CREATE TABLE IF NOT EXISTS tool_patterns (
+                    pattern_id TEXT PRIMARY KEY,
+                    tool_sequence TEXT NOT NULL,
+                    frequency INTEGER DEFAULT 0,
+                    success_rate REAL DEFAULT 0.0,
+                    avg_duration_seconds REAL DEFAULT 0.0,
+                    last_seen TIMESTAMP,
+                    sessions TEXT,
+                    user_feedback INTEGER DEFAULT 0,
+                    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+                    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+                )
+            """)
+            conn.commit()
+        finally:
+            conn.close()
+
+    def store_pattern(self, pattern: ToolPattern) -> None:
+        """
+        Store or update a pattern in the database.
+
+        Args:
+            pattern: Pattern to store
+        """
+        conn = self._get_connection()
+        try:
+            cursor = conn.cursor()
+            cursor.execute(
+                """
+                INSERT OR REPLACE INTO tool_patterns (
+                    pattern_id,
+                    tool_sequence,
+                    frequency,
+                    success_rate,
+                    avg_duration_seconds,
+                    last_seen,
+                    sessions,
+                    user_feedback,
+                    updated_at
+                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+            """,
+                (
+                    pattern.pattern_id,
+                    "->".join(pattern.sequence),
+                    pattern.frequency,
+                    pattern.success_rate,
+                    pattern.avg_duration_seconds,
+                    pattern.last_seen.isoformat() if pattern.last_seen else None,
+                    json.dumps(pattern.sessions),
+                    pattern.user_feedback,
+                    datetime.now().isoformat(),
+                ),
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+    def get_pattern(self, pattern_id: str) -> ToolPattern | None:
+        """
+        Retrieve a pattern by ID.
+
+        Args:
+            pattern_id: Pattern ID to retrieve
+
+        Returns:
+            Pattern or None if not found
+        """
+        conn = self._get_connection()
+        try:
+            cursor = conn.cursor()
+            cursor.execute(
+                "SELECT * FROM tool_patterns WHERE pattern_id = ?", (pattern_id,)
+            )
+            row = cursor.fetchone()
+
+            if not row:
+                return None
+
+            return ToolPattern(
+                pattern_id=row["pattern_id"],
+                sequence=row["tool_sequence"].split("->"),
+                frequency=row["frequency"],
+                success_rate=row["success_rate"],
+                avg_duration_seconds=row["avg_duration_seconds"],
+                last_seen=datetime.fromisoformat(row["last_seen"])
+                if row["last_seen"]
+                else None,
+                sessions=json.loads(row["sessions"]) if row["sessions"] else [],
+                user_feedback=row["user_feedback"],
+            )
+        finally:
+            conn.close()
+
+    def update_feedback(self, pattern_id: str, feedback: int) -> None:
+        """
+        Update user feedback for a pattern.
+
+        Args:
+            pattern_id: Pattern ID
+            feedback: Feedback score (1=helpful, 0=neutral, -1=unhelpful)
+        """
+        conn = self._get_connection()
+        try:
+            cursor = conn.cursor()
+            cursor.execute(
+                """
+                UPDATE tool_patterns
+                SET user_feedback = ?, updated_at = ?
+                WHERE pattern_id = ?
+            """,
+                (feedback, datetime.now().isoformat(), pattern_id),
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+    def get_all_patterns(self) -> list[ToolPattern]:
+        """
+        Get all stored patterns.
+
+        Returns:
+            List of all patterns
+        """
+        conn = self._get_connection()
+        try:
+            cursor = conn.cursor()
+            cursor.execute("SELECT * FROM tool_patterns ORDER BY frequency DESC")
+
+            patterns = []
+            for row in cursor.fetchall():
+                pattern = ToolPattern(
+                    pattern_id=row["pattern_id"],
+                    sequence=row["tool_sequence"].split("->"),
+                    frequency=row["frequency"],
+                    success_rate=row["success_rate"],
+                    avg_duration_seconds=row["avg_duration_seconds"],
+                    last_seen=datetime.fromisoformat(row["last_seen"])
+                    if row["last_seen"]
+                    else None,
+                    sessions=json.loads(row["sessions"]) if row["sessions"] else [],
+                    user_feedback=row["user_feedback"],
+                )
+                patterns.append(pattern)
+
+            return patterns
+        finally:
+            conn.close()
+
+
+class PatternLearner:
+    """
+    Main interface for pattern learning.
+
+    Combines pattern detection, insight generation, and learning loop
+    into a single API for AI agents.
+
+    Example:
+        >>> learner = PatternLearner()
+        >>> patterns = learner.detect_patterns(min_frequency=5)
+        >>> insights = learner.generate_insights()
+        >>> recommendations = learner.get_recommendations()
+    """
+
+    def __init__(self, graph_dir: Path | None = None):
+        """
+        Initialize pattern learner.
+
+        Args:
+            graph_dir: Root directory for HtmlGraph (defaults to .htmlgraph)
+        """
+        if graph_dir is None:
+            graph_dir = Path.cwd() / ".htmlgraph"
+
+        self.graph_dir = Path(graph_dir)
+        self.db_path = self.graph_dir / "htmlgraph.db"
+
+        # Lazy initialization - only initialize components if database exists
+        # This prevents failures in tests with temporary directories
+        self._matcher: PatternMatcher | None = None
+        self._insight_generator: InsightGenerator | None = None
+        self._learning_loop: LearningLoop | None = None
+
+    @property
+    def matcher(self) -> PatternMatcher:
+        """Lazily initialize matcher."""
+        if self._matcher is None:
+            if not self.db_path.exists():
+                raise FileNotFoundError(f"Database not found at {self.db_path}")
+            self._matcher = PatternMatcher(self.db_path)
+        return self._matcher
+
+    @property
+    def insight_generator(self) -> InsightGenerator:
+        """Lazily initialize insight generator."""
+        if self._insight_generator is None:
+            if not self.db_path.exists():
+                raise FileNotFoundError(f"Database not found at {self.db_path}")
+            self._insight_generator = InsightGenerator(self.db_path)
+        return self._insight_generator
+
+    @property
+    def learning_loop(self) -> LearningLoop:
+        """Lazily initialize learning loop."""
+        if self._learning_loop is None:
+            if not self.db_path.exists():
+                raise FileNotFoundError(f"Database not found at {self.db_path}")
+            self._learning_loop = LearningLoop(self.db_path)
+        return self._learning_loop
+
+    def detect_patterns(
+        self, window_size: int = 3, min_frequency: int = 5
+    ) -> list[ToolPattern]:
+        """
+        Detect tool call patterns from event history.
+
+        Args:
+            window_size: Size of tool sequence window (default: 3)
+            min_frequency: Minimum occurrences to be considered a pattern
+
+        Returns:
+            List of detected patterns
+        """
+        patterns = self.matcher.find_patterns(
+            window_size=window_size, min_frequency=min_frequency
+        )
+
+        # Store patterns in learning loop
+        for pattern in patterns:
+            enriched = self.insight_generator.enrich_pattern(pattern)
+            self.learning_loop.store_pattern(enriched)
+
+        return patterns
+
+    def generate_insights(self) -> list[PatternInsight]:
+        """
+        Generate insights from detected patterns.
+
+        Returns:
+            List of actionable insights
+        """
+        patterns = self.learning_loop.get_all_patterns()
+        return self.insight_generator.generate_insights(patterns)
+
+    def get_recommendations(self, limit: int = 3) -> list[PatternInsight]:
+        """
+        Get top recommendations based on impact.
+
+        Args:
+            limit: Maximum number of recommendations (default: 3)
+
+        Returns:
+            Top recommendations sorted by impact
+        """
+        insights = self.generate_insights()
+        recommendations = [i for i in insights if i.insight_type == "recommendation"]
+        return recommendations[:limit]
+
+    def get_anti_patterns(self) -> list[PatternInsight]:
+        """
+        Get detected anti-patterns.
+
+        Returns:
+            List of anti-pattern insights
+        """
+        insights = self.generate_insights()
+        return [i for i in insights if i.insight_type == "anti-pattern"]
+
+    def export_learnings(self, output_path: Path | str) -> None:
+        """
+        Export learnings to markdown for team sharing.
+
+        Args:
+            output_path: Path to output markdown file
+        """
+        insights = self.generate_insights()
+        patterns = self.learning_loop.get_all_patterns()
+
+        output_path = Path(output_path)
+
+        with open(output_path, "w") as f:
+            f.write("# Pattern Learning Report\n\n")
+            f.write(f"Generated: {datetime.now().isoformat()}\n\n")
+
+            f.write("## Recommendations\n\n")
+            recommendations = [
+                i for i in insights if i.insight_type == "recommendation"
+            ]
+            for insight in recommendations[:5]:
+                f.write(f"### {insight.title}\n\n")
+                f.write(f"{insight.description}\n\n")
+                f.write(f"**Impact Score**: {insight.impact_score:.1f}\n\n")
+
+            f.write("## Anti-Patterns\n\n")
+            anti_patterns = [i for i in insights if i.insight_type == "anti-pattern"]
+            for insight in anti_patterns[:5]:
+                f.write(f"### {insight.title}\n\n")
+                f.write(f"{insight.description}\n\n")
+                f.write(f"**Impact Score**: {insight.impact_score:.1f}\n\n")
+
+            f.write("## Optimization Opportunities\n\n")
+            optimizations = [i for i in insights if i.insight_type == "optimization"]
+            for insight in optimizations[:5]:
+                f.write(f"### {insight.title}\n\n")
+                f.write(f"{insight.description}\n\n")
+                f.write(f"**Impact Score**: {insight.impact_score:.1f}\n\n")
+
+            f.write("## All Detected Patterns\n\n")
+            for pattern in patterns[:20]:
+                f.write(f"- **{' → '.join(pattern.sequence)}** ")
+                f.write(f"(frequency: {pattern.frequency}, ")
+                f.write(f"success: {pattern.success_rate:.1f}%)\n")