agent_os_kernel 3.1.0__py3-none-any.whl

agent_kernel/__init__.py

@@ -0,0 +1,69 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""
+Self-Correcting Agent Kernel
+
+A Dual-Loop Architecture for Enterprise Agents:
+- Loop 1 (Runtime): Constraint Engine (Safety)
+- Loop 2 (Offline): Alignment Engine (Quality & Efficiency)
+  - Completeness Auditor (detects laziness)
+  - Semantic Purge (scales by subtraction)
+
+Reference Implementations:
+- auditor.py: Simplified soft failure detection
+- teacher.py: Shadow Teacher diagnosis
+- memory_manager.py: Lesson lifecycle management
+"""
+
+__version__ = "3.1.0"
+
+from .kernel import SelfCorrectingAgentKernel
+from .models import (
+    AgentFailure, FailureAnalysis, CorrectionPatch,
+    AgentOutcome, CompletenessAudit, ClassifiedPatch,
+    OutcomeType, GiveUpSignal, PatchDecayType,
+    ToolExecutionTelemetry, ToolExecutionStatus,
+    SemanticAnalysis, NudgeResult
+)
+from .outcome_analyzer import OutcomeAnalyzer
+from .completeness_auditor import CompletenessAuditor
+from .semantic_purge import SemanticPurge, PatchClassifier
+from .triage import FailureTriage, FixStrategy
+from .semantic_analyzer import SemanticAnalyzer
+from .nudge_mechanism import NudgeMechanism
+
+# Reference implementations (simplified examples)
+from .auditor import CompletenessAuditor as SimpleCompletenessAuditor
+from .teacher import diagnose_failure
+from .memory_manager import MemoryManager, LessonType
+
+__all__ = [
+    "SelfCorrectingAgentKernel",
+    "AgentFailure",
+    "FailureAnalysis",
+    "CorrectionPatch",
+    "AgentOutcome",
+    "CompletenessAudit",
+    "ClassifiedPatch",
+    "OutcomeType",
+    "GiveUpSignal",
+    "PatchDecayType",
+    "ToolExecutionTelemetry",
+    "ToolExecutionStatus",
+    "SemanticAnalysis",
+    "NudgeResult",
+    "OutcomeAnalyzer",
+    "CompletenessAuditor",
+    "SemanticPurge",
+    "PatchClassifier",
+    "FailureTriage",
+    "FixStrategy",
+    "SemanticAnalyzer",
+    "NudgeMechanism",
+    # Reference implementations
+    "SimpleCompletenessAuditor",
+    "diagnose_failure",
+    "MemoryManager",
+    "LessonType",
+]

agent_kernel/analyzer.py

@@ -0,0 +1,435 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""
+Failure analysis system that diagnoses root causes.
+"""
+
+import logging
+from typing import List, Optional, Dict
+from collections import Counter
+
+from .models import AgentFailure, FailureAnalysis, FailureType, DiagnosisJSON, CognitiveGlitch
+
+logger = logging.getLogger(__name__)
+
+
+class FailureAnalyzer:
+    """Analyzes failures to identify root causes and suggest fixes."""
+    
+    def __init__(self):
+        self.analysis_history: List[FailureAnalysis] = []
+        self.known_patterns: Dict[str, dict] = self._load_known_patterns()
+    
+    def _load_known_patterns(self) -> Dict[str, dict]:
+        """Load known failure patterns and their solutions."""
+        return {
+            FailureType.BLOCKED_BY_CONTROL_PLANE: {
+                "root_causes": [
+                    "Missing permission validation",
+                    "Attempting unauthorized resource access",
+                    "Policy violation",
+                    "Security constraint violation"
+                ],
+                "fixes": [
+                    "Add permission checks before actions",
+                    "Implement resource access validation",
+                    "Use safe alternatives for restricted operations",
+                    "Request proper authorization before attempting action"
+                ]
+            },
+            FailureType.TIMEOUT: {
+                "root_causes": [
+                    "Operation taking too long",
+                    "Infinite loop or deadlock",
+                    "Network latency",
+                    "Resource contention"
+                ],
+                "fixes": [
+                    "Implement operation timeout handling",
+                    "Add progress monitoring",
+                    "Optimize algorithm efficiency",
+                    "Add async/parallel processing"
+                ]
+            },
+            FailureType.INVALID_ACTION: {
+                "root_causes": [
+                    "Invalid input parameters",
+                    "Action not supported in current state",
+                    "Precondition not met"
+                ],
+                "fixes": [
+                    "Add input validation",
+                    "Check state before action",
+                    "Verify preconditions"
+                ]
+            },
+            FailureType.RESOURCE_EXHAUSTED: {
+                "root_causes": [
+                    "Memory leak",
+                    "Unbounded resource allocation",
+                    "Missing cleanup"
+                ],
+                "fixes": [
+                    "Implement resource cleanup",
+                    "Add resource limits",
+                    "Use resource pooling"
+                ]
+            },
+            FailureType.LOGIC_ERROR: {
+                "root_causes": [
+                    "Incorrect algorithm",
+                    "Edge case not handled",
+                    "Type mismatch"
+                ],
+                "fixes": [
+                    "Fix algorithm logic",
+                    "Add edge case handling",
+                    "Add type checking"
+                ]
+            }
+        }
+    
+    def analyze(self, failure: AgentFailure, similar_failures: Optional[List[AgentFailure]] = None) -> FailureAnalysis:
+        """
+        Analyze a failure to identify root cause and suggest fixes.
+        
+        Args:
+            failure: The failure to analyze
+            similar_failures: Optional list of similar past failures
+            
+        Returns:
+            FailureAnalysis with root cause and suggested fixes
+        """
+        logger.info(f"Analyzing failure for agent {failure.agent_id}")
+        
+        # Get known patterns for this failure type
+        patterns = self.known_patterns.get(failure.failure_type, {})
+        
+        # Identify root cause
+        root_cause = self._identify_root_cause(failure, patterns)
+        
+        # Identify contributing factors
+        contributing_factors = self._identify_contributing_factors(failure, patterns)
+        
+        # Generate suggested fixes
+        suggested_fixes = self._generate_fixes(failure, patterns)
+        
+        # Calculate confidence based on pattern matching and similar failures
+        confidence_score = self._calculate_confidence(failure, similar_failures)
+        
+        # Find similar failures
+        similar_failure_ids = []
+        if similar_failures:
+            similar_failure_ids = [f.agent_id + "_" + str(f.timestamp) for f in similar_failures[:5]]
+        
+        analysis = FailureAnalysis(
+            failure=failure,
+            root_cause=root_cause,
+            contributing_factors=contributing_factors,
+            suggested_fixes=suggested_fixes,
+            confidence_score=confidence_score,
+            similar_failures=similar_failure_ids
+        )
+        
+        self.analysis_history.append(analysis)
+        logger.info(f"Analysis complete. Root cause: {root_cause} (confidence: {confidence_score:.2f})")
+        
+        return analysis
+    
+    def _identify_root_cause(self, failure: AgentFailure, patterns: dict) -> str:
+        """Identify the root cause of the failure."""
+        root_causes = patterns.get("root_causes", ["Unknown root cause"])
+        
+        # For control plane blocks, check context for more specific cause
+        if failure.failure_type == FailureType.BLOCKED_BY_CONTROL_PLANE:
+            context = failure.context
+            if "permission" in failure.error_message.lower():
+                return "Missing or insufficient permissions for requested operation"
+            elif "policy" in failure.error_message.lower():
+                return "Action violates control plane policy"
+            else:
+                return root_causes[0]
+        
+        # Return the first root cause as default
+        return root_causes[0]
+    
+    def _identify_contributing_factors(self, failure: AgentFailure, patterns: dict) -> List[str]:
+        """Identify contributing factors to the failure."""
+        factors = []
+        
+        # Check for common contributing factors
+        if failure.severity.value in ["high", "critical"]:
+            factors.append("High severity failure requiring immediate attention")
+        
+        if failure.stack_trace:
+            factors.append("Stack trace available for detailed debugging")
+        
+        if failure.context:
+            factors.append(f"Additional context available: {', '.join(failure.context.keys())}")
+        
+        return factors
+    
+    def _generate_fixes(self, failure: AgentFailure, patterns: dict) -> List[str]:
+        """Generate suggested fixes for the failure."""
+        fixes = patterns.get("fixes", ["Manual investigation required"])
+        
+        # Add specific fixes based on failure type
+        if failure.failure_type == FailureType.BLOCKED_BY_CONTROL_PLANE:
+            if "file" in failure.context:
+                fixes.append(f"Validate access permissions for: {failure.context['file']}")
+            if "action" in failure.context:
+                fixes.append(f"Check if action '{failure.context['action']}' is allowed by policy")
+        
+        return fixes[:3]  # Return top 3 fixes
+    
+    def _calculate_confidence(self, failure: AgentFailure, similar_failures: Optional[List[AgentFailure]]) -> float:
+        """Calculate confidence score for the analysis."""
+        confidence = 0.5  # Base confidence
+        
+        # Increase confidence if we have a known pattern
+        if failure.failure_type in self.known_patterns:
+            confidence += 0.2
+        
+        # Increase confidence if we have similar failures
+        if similar_failures and len(similar_failures) > 0:
+            confidence += min(0.2, len(similar_failures) * 0.05)
+        
+        # Increase confidence if we have detailed context
+        if failure.context and len(failure.context) > 0:
+            confidence += 0.1
+        
+        return min(1.0, confidence)
+    
+    def diagnose_cognitive_glitch(self, failure: AgentFailure) -> DiagnosisJSON:
+        """
+        Deep diagnosis to identify cognitive glitches in agent reasoning.
+        
+        This is "The Analyst" - looking at the reasoning that led to the error,
+        not just the error itself.
+        
+        Args:
+            failure: AgentFailure with full trace
+            
+        Returns:
+            DiagnosisJSON with cognitive glitch identification
+        """
+        logger.info(f"Diagnosing cognitive glitch for agent {failure.agent_id}")
+        
+        if not failure.failure_trace:
+            # Fall back to basic diagnosis if no trace available
+            return self._basic_diagnosis(failure)
+        
+        trace = failure.failure_trace
+        
+        # Identify cognitive glitch type
+        glitch = self._identify_cognitive_glitch(failure, trace)
+        
+        # Deep problem analysis
+        deep_problem = self._analyze_deep_problem(failure, trace, glitch)
+        
+        # Collect evidence
+        evidence = self._collect_evidence(failure, trace, glitch)
+        
+        # Generate hint for counterfactual simulation
+        hint = self._generate_hint(failure, trace, glitch)
+        
+        # Expected fix description
+        expected_fix = self._describe_expected_fix(glitch, hint)
+        
+        # Calculate confidence
+        confidence = self._calculate_diagnosis_confidence(failure, trace, evidence)
+        
+        diagnosis = DiagnosisJSON(
+            cognitive_glitch=glitch,
+            deep_problem=deep_problem,
+            evidence=evidence,
+            hint=hint,
+            expected_fix=expected_fix,
+            confidence=confidence
+        )
+        
+        logger.info(f"Diagnosis complete: {glitch.value} (confidence: {confidence:.2f})")
+        return diagnosis
+    
+    def _identify_cognitive_glitch(self, failure: AgentFailure, trace) -> CognitiveGlitch:
+        """Identify the type of cognitive glitch."""
+        error_lower = failure.error_message.lower()
+        
+        # Check for tool misuse (wrong parameter types) - high priority
+        if any(keyword in error_lower for keyword in ["type error", "invalid type", "expected uuid", "wrong parameter type", "parameter type mismatch"]):
+            return CognitiveGlitch.TOOL_MISUSE
+        if "uuid" in error_lower or ("id" in error_lower and any(kw in error_lower for kw in ["invalid", "malformed", "not a valid", "format"])):
+            # Check if this looks like a tool misuse scenario
+            if trace.failed_action:
+                action_str = str(trace.failed_action).lower()
+                if any(kw in action_str for kw in ["name", "username", "email", "params", "id"]):
+                    return CognitiveGlitch.TOOL_MISUSE
+            return CognitiveGlitch.TOOL_MISUSE
+        
+        # Check for policy violations (e.g., medical advice, legal advice) - high priority
+        if any(keyword in error_lower for keyword in ["policy violation", "violates policy", "not allowed to", "cannot advise", "cannot provide"]):
+            return CognitiveGlitch.POLICY_VIOLATION
+        # Check for specific policy domains in combination with blocking
+        if trace.user_prompt:
+            prompt_lower = trace.user_prompt.lower()
+            if any(domain in prompt_lower for domain in ["medical", "health", "diagnosis", "treatment", "medicine", "legal", "attorney", "sue", "investment", "stock"]):
+                if any(keyword in error_lower for keyword in ["blocked", "violation", "not permitted", "cannot", "policy"]):
+                    return CognitiveGlitch.POLICY_VIOLATION
+        
+        # Check for hallucination (inventing facts) - check early before context gap
+        if any(keyword in error_lower for keyword in ["not found", "does not exist", "unknown", "deprecated", "invalid reference", "no such"]):
+            return CognitiveGlitch.HALLUCINATION
+        
+        # Check for schema mismatch
+        if trace.failed_action:
+            action_str = str(trace.failed_action).lower()
+            if "schema" in error_lower and "mismatch" in action_str:
+                return CognitiveGlitch.SCHEMA_MISMATCH
+        
+        # Check for logic error (misunderstanding)
+        if trace.chain_of_thought:
+            cot_text = " ".join(trace.chain_of_thought).lower()
+            # Look for misinterpretation of terms like "recent", "delete", etc.
+            if any(keyword in cot_text for keyword in ["i think", "probably", "assume", "guess"]):
+                return CognitiveGlitch.LOGIC_ERROR
+        
+        # Check for context gap (missing information) - lower priority
+        if not trace.chain_of_thought or len(trace.chain_of_thought) < 2:
+            # Don't default to context gap if we have other clear signals
+            if trace.failed_action and ("uuid" in error_lower or "id" in error_lower):
+                return CognitiveGlitch.TOOL_MISUSE
+            if any(keyword in error_lower for keyword in ["not found", "does not exist"]):
+                return CognitiveGlitch.HALLUCINATION
+            return CognitiveGlitch.CONTEXT_GAP
+        
+        # Check for permission errors
+        if any(keyword in error_lower for keyword in ["permission", "unauthorized", "forbidden"]):
+            # Distinguish from policy violations
+            if "policy" not in error_lower and "violates" not in error_lower:
+                return CognitiveGlitch.PERMISSION_ERROR
+        
+        return CognitiveGlitch.LOGIC_ERROR  # Default
+    
+    def _analyze_deep_problem(self, failure: AgentFailure, trace, glitch: CognitiveGlitch) -> str:
+        """Analyze the deep problem behind the glitch."""
+        if glitch == CognitiveGlitch.HALLUCINATION:
+            return f"Agent invented non-existent entities in action: {trace.failed_action}"
+        elif glitch == CognitiveGlitch.LOGIC_ERROR:
+            return f"Agent misunderstood user intent in prompt: '{trace.user_prompt}'"
+        elif glitch == CognitiveGlitch.CONTEXT_GAP:
+            return f"Agent lacked necessary context (schema/permissions) to safely execute action"
+        elif glitch == CognitiveGlitch.PERMISSION_ERROR:
+            return f"Agent attempted unauthorized action without checking permissions first"
+        elif glitch == CognitiveGlitch.SCHEMA_MISMATCH:
+            return f"Agent referenced incorrect schema elements in action"
+        elif glitch == CognitiveGlitch.TOOL_MISUSE:
+            return f"Agent used tool with wrong parameter type or value: {trace.failed_action}"
+        elif glitch == CognitiveGlitch.POLICY_VIOLATION:
+            return f"Agent violated policy boundaries by attempting: '{trace.user_prompt}'"
+        return "Unknown deep problem"
+    
+    def _collect_evidence(self, failure: AgentFailure, trace, glitch: CognitiveGlitch) -> List[str]:
+        """Collect evidence supporting the diagnosis."""
+        evidence = []
+        
+        evidence.append(f"User prompt: '{trace.user_prompt}'")
+        evidence.append(f"Failed action: {trace.failed_action}")
+        evidence.append(f"Error: {failure.error_message}")
+        
+        if trace.chain_of_thought:
+            evidence.append(f"Reasoning steps: {len(trace.chain_of_thought)} steps")
+            if trace.chain_of_thought:
+                evidence.append(f"Last thought: '{trace.chain_of_thought[-1]}'")
+        
+        return evidence
+    
+    def _generate_hint(self, failure: AgentFailure, trace, glitch: CognitiveGlitch) -> str:
+        """Generate a hint to inject for counterfactual simulation."""
+        if glitch == CognitiveGlitch.HALLUCINATION:
+            return "HINT: Always verify entity names against the provided schema before using them. Available tables/resources must be explicitly listed."
+        elif glitch == CognitiveGlitch.LOGIC_ERROR:
+            return f"HINT: When interpreting '{trace.user_prompt}', be precise about terms like 'recent', 'delete', 'modify'. Ask for clarification if ambiguous."
+        elif glitch == CognitiveGlitch.CONTEXT_GAP:
+            return "HINT: Before executing actions, ensure you have: 1) Complete schema information, 2) Permission requirements, 3) Clear action scope."
+        elif glitch == CognitiveGlitch.PERMISSION_ERROR:
+            return "HINT: Always check permissions before attempting actions. Use validate_permissions() first."
+        elif glitch == CognitiveGlitch.SCHEMA_MISMATCH:
+            return "HINT: Available schema elements must be verified before use. Do not assume table/column names."
+        elif glitch == CognitiveGlitch.TOOL_MISUSE:
+            return "HINT: Always verify parameter types match the tool schema. For example, use UUIDs where required, not names or strings."
+        elif glitch == CognitiveGlitch.POLICY_VIOLATION:
+            return "HINT: Some topics are outside your policy boundaries. Refuse requests for medical advice, legal advice, or other restricted domains."
+        return "HINT: Proceed with caution and verify all assumptions."
+    
+    def _describe_expected_fix(self, glitch: CognitiveGlitch, hint: str) -> str:
+        """Describe the expected outcome of applying the hint."""
+        if glitch == CognitiveGlitch.HALLUCINATION:
+            return "Agent will verify schema before action and use only existing entities"
+        elif glitch == CognitiveGlitch.LOGIC_ERROR:
+            return "Agent will correctly interpret user intent and clarify ambiguous terms"
+        elif glitch == CognitiveGlitch.CONTEXT_GAP:
+            return "Agent will request necessary context before proceeding with action"
+        elif glitch == CognitiveGlitch.PERMISSION_ERROR:
+            return "Agent will validate permissions before attempting action"
+        elif glitch == CognitiveGlitch.TOOL_MISUSE:
+            return "Agent will use correct parameter types according to tool schema"
+        elif glitch == CognitiveGlitch.POLICY_VIOLATION:
+            return "Agent will refuse to provide advice in restricted domains"
+        return "Agent will handle the situation correctly"
+    
+    def _calculate_diagnosis_confidence(self, failure: AgentFailure, trace, evidence: List[str]) -> float:
+        """Calculate confidence in the diagnosis."""
+        confidence = 0.5  # Base
+        
+        # More confidence with complete trace
+        if trace.chain_of_thought and len(trace.chain_of_thought) > 2:
+            confidence += 0.2
+        
+        # More confidence with detailed action
+        if trace.failed_action and len(trace.failed_action) > 0:
+            confidence += 0.15
+        
+        # More confidence with rich evidence
+        if len(evidence) >= 4:
+            confidence += 0.15
+        
+        return min(1.0, confidence)
+    
+    def _basic_diagnosis(self, failure: AgentFailure) -> DiagnosisJSON:
+        """Fallback diagnosis when no trace is available."""
+        return DiagnosisJSON(
+            cognitive_glitch=CognitiveGlitch.NONE,
+            deep_problem=f"No trace available. Basic error: {failure.error_message}",
+            evidence=[f"Error message: {failure.error_message}"],
+            hint="HINT: Ensure proper validation before actions.",
+            expected_fix="Action will be validated before execution",
+            confidence=0.5
+        )
+    
+    def find_similar_failures(self, failure: AgentFailure, history: List[AgentFailure]) -> List[AgentFailure]:
+        """Find similar failures in history."""
+        similar = []
+        
+        for past_failure in history:
+            if past_failure.failure_type == failure.failure_type:
+                # Calculate similarity based on error message
+                similarity = self._calculate_similarity(failure.error_message, past_failure.error_message)
+                if similarity > 0.6:
+                    similar.append(past_failure)
+        
+        return similar[:10]  # Return top 10 similar failures
+    
+    def _calculate_similarity(self, msg1: str, msg2: str) -> float:
+        """Calculate similarity between two error messages."""
+        # Simple word-based similarity
+        words1 = set(msg1.lower().split())
+        words2 = set(msg2.lower().split())
+        
+        if not words1 or not words2:
+            return 0.0
+        
+        intersection = words1.intersection(words2)
+        union = words1.union(words2)
+        
+        return len(intersection) / len(union) if union else 0.0

agent_kernel/auditor.py

@@ -0,0 +1,36 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+# Public Preview — basic self-correction with retry
+"""
+Auditor — simple null/empty check on response text.
+"""
+
+
+class CompletenessAuditor:
+    """Basic completeness check: null, empty, or trivially short responses."""
+
+    def __init__(self):
+        self.lazy_signals = [
+            "i cannot", "i'm sorry", "no data found",
+            "unable to access", "context does not contain",
+        ]
+
+    def audit_response(self, agent_response: str, tool_output: str = "") -> bool:
+        """
+        Returns True when the response likely needs intervention.
+
+        Checks:
+        1. Null / empty response
+        2. Known give-up phrases
+        3. Tool returned trivially small output
+        """
+        if not agent_response or not agent_response.strip():
+            return True
+
+        if any(sig in agent_response.lower() for sig in self.lazy_signals):
+            return True
+
+        if tool_output and len(tool_output) < 10:
+            return True
+
+        return False