@bastani/atomic 0.5.11 → 0.5.12-0

package/.agents/skills/context-degradation/scripts/degradation_detector.py

@@ -0,0 +1,614 @@
+"""
+Context Degradation Detection — Public API
+============================================
+
+Detect, measure, and diagnose context degradation patterns in LLM agent systems.
+
+Public API:
+    measure_attention_distribution  — Map attention weight across context positions.
+    detect_lost_in_middle           — Flag critical information in degraded-attention regions.
+    analyze_context_structure       — Assess structural degradation risk factors.
+    PoisoningDetector               — Detect context poisoning indicators (error accumulation,
+                                      contradictions, hallucination markers).
+    ContextHealthAnalyzer           — Run composite health analysis combining attention,
+                                      poisoning, and utilization metrics.
+    analyze_agent_context           — One-call convenience function for agent sessions.
+
+PRODUCTION NOTES:
+- The attention estimation functions simulate U-shaped attention curves for demonstration
+  purposes. Production systems should extract actual attention weights from model internals
+  when available (e.g., via TransformerLens or model-specific APIs).
+- Token estimation uses simplified heuristics (~1 token per whitespace-split word).
+  Production systems should use model-specific tokenizers for accurate counts.
+- Poisoning and hallucination detection uses pattern matching as a proxy. Production
+  systems may benefit from fine-tuned classifiers or model-based detection.
+"""
+
+import random
+import re
+from typing import Dict, List, Optional
+
+__all__ = [
+    "measure_attention_distribution",
+    "detect_lost_in_middle",
+    "analyze_context_structure",
+    "PoisoningDetector",
+    "ContextHealthAnalyzer",
+    "analyze_agent_context",
+]
+
+
+# ---------------------------------------------------------------------------
+# Attention Distribution Analysis
+# ---------------------------------------------------------------------------
+
+def measure_attention_distribution(
+    context_tokens: List[str],
+    query: str,
+) -> List[Dict[str, object]]:
+    """Map simulated attention weight to each context position.
+
+    Use when: diagnosing whether critical information sits in the
+    low-attention middle region of a long context.
+
+    Args:
+        context_tokens: Whitespace-split tokens (or chunks) of the context.
+        query: The query or task description the context is meant to support.
+
+    Returns:
+        List of dicts, one per position, each containing:
+            position (int), attention (float), region (str), tokens (str | None).
+    """
+    n = len(context_tokens)
+    attention_by_position: List[Dict[str, object]] = []
+
+    for position in range(n):
+        is_beginning = position < n * 0.1
+        is_end = position > n * 0.9
+
+        attention = _estimate_attention(position, n, is_beginning, is_end)
+
+        attention_by_position.append({
+            "position": position,
+            "attention": attention,
+            "region": "attention_favored" if (is_beginning or is_end) else "attention_degraded",
+            "tokens": context_tokens[position][:50] if position < 5 or position > n - 5 else None,
+        })
+
+    return attention_by_position
+
+
+def _estimate_attention(
+    position: int,
+    total: int,
+    is_beginning: bool,
+    is_end: bool,
+) -> float:
+    """Estimate attention weight for a single position.
+
+    Simulates the U-shaped attention curve documented in lost-in-middle research:
+    - Beginning tokens receive high attention (primacy / attention-sink effect).
+    - End tokens receive high attention (recency effect).
+    - Middle tokens receive degraded attention.
+
+    IMPORTANT: This is a simulation for demonstration. Production systems should
+    extract actual attention weights from model forward passes or use
+    interpretability libraries (e.g., TransformerLens).
+    """
+    if is_beginning:
+        return 0.8 + random.random() * 0.2
+    elif is_end:
+        return 0.7 + random.random() * 0.3
+    else:
+        middle_progress = (position - total * 0.1) / (total * 0.8)
+        base_attention = 0.3 * (1 - middle_progress) + 0.1 * middle_progress
+        return base_attention + random.random() * 0.1
+
+
+# ---------------------------------------------------------------------------
+# Lost-in-Middle Detection
+# ---------------------------------------------------------------------------
+
+def detect_lost_in_middle(
+    critical_positions: List[int],
+    attention_distribution: List[Dict[str, object]],
+) -> Dict[str, object]:
+    """Check if critical information sits in attention-degraded positions.
+
+    Use when: context has been assembled and you need to verify that
+    high-priority content is not buried in the low-attention middle zone.
+
+    Args:
+        critical_positions: Indices into the context that hold critical info.
+        attention_distribution: Output of ``measure_attention_distribution``.
+
+    Returns:
+        Dict with keys: at_risk (list[int]), safe (list[int]),
+        recommendations (list[str]), degradation_score (float 0-1).
+    """
+    results: Dict[str, object] = {
+        "at_risk": [],
+        "safe": [],
+        "recommendations": [],
+        "degradation_score": 0.0,
+    }
+
+    at_risk_count = 0
+    total_critical = len(critical_positions)
+
+    for pos in critical_positions:
+        if pos < len(attention_distribution):
+            region = attention_distribution[pos]["region"]
+            if region == "attention_degraded":
+                results["at_risk"].append(pos)
+                at_risk_count += 1
+            else:
+                results["safe"].append(pos)
+
+    if total_critical > 0:
+        results["degradation_score"] = at_risk_count / total_critical
+
+    if results["at_risk"]:
+        results["recommendations"].extend([
+            "Move critical information to attention-favored positions",
+            "Use explicit markers to highlight critical information",
+            "Consider splitting context to reduce middle section",
+            f"{at_risk_count}/{total_critical} critical items are in degraded region",
+        ])
+
+    return results
+
+
+# ---------------------------------------------------------------------------
+# Context Structure Analysis
+# ---------------------------------------------------------------------------
+
+def analyze_context_structure(context: str) -> Dict[str, object]:
+    """Assess structural degradation risk factors in a context string.
+
+    Use when: evaluating whether a context layout puts too much content
+    in the low-attention middle zone before sending it to a model.
+
+    Args:
+        context: The full context string to analyze.
+
+    Returns:
+        Dict with total_lines, sections list, middle_content_ratio,
+        and degradation_risk level (low / medium / high).
+    """
+    lines = context.split("\n")
+    sections: List[Dict[str, object]] = []
+
+    current_section: Dict[str, object] = {"start": 0, "type": "unknown", "length": 0}
+
+    for i, line in enumerate(lines):
+        if line.startswith("#"):
+            if current_section["length"] > 0:
+                sections.append(current_section)
+            current_section = {
+                "start": i,
+                "type": "header",
+                "length": 1,
+                "header": line.lstrip("#").strip(),
+            }
+        else:
+            current_section["length"] += 1
+
+    sections.append(current_section)
+
+    n = len(lines)
+    middle_start = int(n * 0.3)
+    middle_end = int(n * 0.7)
+
+    middle_content = sum(
+        s["length"] for s in sections
+        if s["start"] >= middle_start and s["start"] <= middle_end
+    )
+
+    middle_ratio = middle_content / n if n > 0 else 0
+    return {
+        "total_lines": n,
+        "sections": sections,
+        "middle_content_ratio": middle_ratio,
+        "degradation_risk": (
+            "high" if middle_ratio > 0.5
+            else "medium" if middle_ratio > 0.3
+            else "low"
+        ),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Context Poisoning Detection
+# ---------------------------------------------------------------------------
+
+class PoisoningDetector:
+    """Detect context poisoning indicators via pattern matching.
+
+    Use when: context quality is suspect — outputs degrade on previously
+    successful tasks, tool calls misalign, or hallucinations persist
+    despite corrections.
+    """
+
+    def __init__(self) -> None:
+        self.claims: List[Dict[str, object]] = []
+        self.error_patterns: List[str] = [
+            r"error",
+            r"failed",
+            r"exception",
+            r"cannot",
+            r"unable",
+            r"invalid",
+            r"not found",
+        ]
+
+    def extract_claims(self, text: str) -> List[Dict[str, object]]:
+        """Extract claims from text for verification tracking.
+
+        Use when: building a provenance chain to trace which claims
+        entered context and whether they have been verified.
+
+        Args:
+            text: Raw text to extract claims from.
+
+        Returns:
+            List of claim dicts with id, text, verified status, and
+            error indicator flag.
+        """
+        sentences = text.split(".")
+        claims: List[Dict[str, object]] = []
+
+        for i, sentence in enumerate(sentences):
+            sentence = sentence.strip()
+            if len(sentence) < 10:
+                continue
+
+            claims.append({
+                "id": i,
+                "text": sentence,
+                "verified": None,
+                "has_error_indicator": any(
+                    re.search(pattern, sentence, re.IGNORECASE)
+                    for pattern in self.error_patterns
+                ),
+            })
+
+        self.claims.extend(claims)
+        return claims
+
+    def detect_poisoning(self, context: str) -> Dict[str, object]:
+        """Detect potential context poisoning indicators.
+
+        Use when: agent output quality has degraded and context
+        contamination is suspected. Checks for error accumulation,
+        contradictions, and hallucination markers.
+
+        Args:
+            context: The full context string to analyze.
+
+        Returns:
+            Dict with poisoning_risk (bool), indicators (list),
+            and overall_risk level (low / medium / high).
+        """
+        indicators: List[Dict[str, object]] = []
+
+        # Check for error accumulation
+        error_count = sum(
+            1 for pattern in self.error_patterns
+            if re.search(pattern, context, re.IGNORECASE)
+        )
+
+        if error_count > 3:
+            indicators.append({
+                "type": "error_accumulation",
+                "count": error_count,
+                "severity": "high" if error_count > 5 else "medium",
+                "message": f"Found {error_count} error indicators in context",
+            })
+
+        # Check for contradiction patterns
+        contradictions = self._detect_contradictions(context)
+        if contradictions:
+            indicators.append({
+                "type": "contradictions",
+                "count": len(contradictions),
+                "examples": contradictions[:3],
+                "severity": "high",
+                "message": f"Found {len(contradictions)} potential contradictions",
+            })
+
+        # Check for hallucination markers
+        hallucination_markers = self._detect_hallucination_markers(context)
+        if hallucination_markers:
+            indicators.append({
+                "type": "hallucination_markers",
+                "count": len(hallucination_markers),
+                "severity": "medium",
+                "message": f"Found {len(hallucination_markers)} phrases associated with uncertain claims",
+            })
+
+        return {
+            "poisoning_risk": len(indicators) > 0,
+            "indicators": indicators,
+            "overall_risk": (
+                "high" if len(indicators) > 2
+                else "medium" if len(indicators) > 0
+                else "low"
+            ),
+        }
+
+    def _detect_contradictions(self, text: str) -> List[str]:
+        """Detect potential contradictions in text."""
+        contradictions: List[str] = []
+
+        conflict_patterns = [
+            (r"however", r"but"),
+            (r"on the other hand", r"instead"),
+            (r"although", r"yet"),
+            (r"despite", r"nevertheless"),
+        ]
+
+        for pattern1, pattern2 in conflict_patterns:
+            if re.search(pattern1, text, re.IGNORECASE) and re.search(pattern2, text, re.IGNORECASE):
+                sentences = text.split(".")
+                for sentence in sentences:
+                    if (re.search(pattern1, sentence, re.IGNORECASE)
+                            or re.search(pattern2, sentence, re.IGNORECASE)):
+                        stripped = sentence.strip()
+                        if stripped and len(stripped) < 200:
+                            contradictions.append(stripped[:100])
+
+        return contradictions[:5]
+
+    def _detect_hallucination_markers(self, text: str) -> List[str]:
+        """Detect phrases associated with uncertain or hallucinated claims."""
+        markers = [
+            "may have been",
+            "might have",
+            "could potentially",
+            "possibly",
+            "apparently",
+            "reportedly",
+            "it is said that",
+            "sources suggest",
+            "believed to be",
+            "thought to be",
+        ]
+
+        return [marker for marker in markers if marker in text.lower()]
+
+
+# ---------------------------------------------------------------------------
+# Context Health Analyzer
+# ---------------------------------------------------------------------------
+
+class ContextHealthAnalyzer:
+    """Run composite health analysis on a context string.
+
+    Use when: performing routine health checks on agent context during
+    long-running sessions, or when setting up automated monitoring that
+    triggers compaction or isolation before degradation hits.
+
+    Combines attention distribution, poisoning detection, and utilization
+    metrics into a single 0-1 health score with status interpretation.
+    """
+
+    def __init__(self, context_limit: int = 100_000) -> None:
+        self.context_limit: int = context_limit
+        self.metrics_history: List[Dict[str, object]] = []
+
+    def analyze(
+        self,
+        context: str,
+        critical_positions: Optional[List[int]] = None,
+    ) -> Dict[str, object]:
+        """Perform comprehensive context health analysis.
+
+        Use when: a single health-check call is needed that covers
+        attention, poisoning, and utilization in one pass.
+
+        Args:
+            context: The full context string to analyze.
+            critical_positions: Indices of tokens holding critical info.
+                Defaults to the first 10 positions if not provided.
+
+        Returns:
+            Dict with health_score (float 0-1), status (str),
+            metrics (dict), issues (dict), and recommendations (list[str]).
+        """
+        tokens = context.split()
+
+        token_count = len(tokens)
+        utilization = token_count / self.context_limit
+
+        attention_dist = measure_attention_distribution(
+            tokens[:1000],  # Sample for efficiency
+            "current_task",
+        )
+
+        degradation = detect_lost_in_middle(
+            critical_positions or list(range(10)),
+            attention_dist,
+        )
+
+        poisoning = PoisoningDetector().detect_poisoning(context)
+
+        health_score = self._calculate_health_score(
+            utilization=utilization,
+            degradation=degradation["degradation_score"],
+            poisoning_risk=1.0 if poisoning["poisoning_risk"] else 0.0,
+        )
+
+        result: Dict[str, object] = {
+            "health_score": health_score,
+            "status": self._interpret_score(health_score),
+            "metrics": {
+                "token_count": token_count,
+                "utilization": utilization,
+                "degradation_score": degradation["degradation_score"],
+                "poisoning_risk": poisoning["overall_risk"],
+            },
+            "issues": {
+                "lost_in_middle": degradation,
+                "poisoning": poisoning,
+            },
+            "recommendations": self._generate_recommendations(
+                utilization, degradation, poisoning
+            ),
+        }
+
+        self.metrics_history.append(result)
+        return result
+
+    def _calculate_health_score(
+        self,
+        utilization: float,
+        degradation: float,
+        poisoning_risk: float,
+    ) -> float:
+        """Calculate composite health score (0-1, higher is healthier)."""
+        utilization_penalty = min(utilization * 0.5, 0.3)
+        degradation_penalty = degradation * 0.3
+        poisoning_penalty = poisoning_risk * 0.2
+
+        score = 1.0 - utilization_penalty - degradation_penalty - poisoning_penalty
+        return max(0.0, min(1.0, score))
+
+    def _interpret_score(self, score: float) -> str:
+        """Map numeric score to human-readable status."""
+        if score > 0.8:
+            return "healthy"
+        elif score > 0.6:
+            return "warning"
+        elif score > 0.4:
+            return "degraded"
+        else:
+            return "critical"
+
+    def _generate_recommendations(
+        self,
+        utilization: float,
+        degradation: Dict[str, object],
+        poisoning: Dict[str, object],
+    ) -> List[str]:
+        """Generate actionable recommendations based on analysis."""
+        recommendations: List[str] = []
+
+        if utilization > 0.8:
+            recommendations.append("Context near limit - consider compaction")
+            recommendations.append("Implement observation masking for tool outputs")
+
+        if degradation.get("at_risk"):
+            recommendations.append("Critical information in degraded attention region")
+            recommendations.append("Move key information to beginning or end of context")
+
+        if poisoning["poisoning_risk"]:
+            recommendations.append("Context poisoning indicators detected")
+            recommendations.append("Review and remove potentially erroneous information")
+
+        if not recommendations:
+            recommendations.append("Context appears healthy - continue monitoring")
+
+        return recommendations
+
+
+# ---------------------------------------------------------------------------
+# Convenience Function
+# ---------------------------------------------------------------------------
+
+def analyze_agent_context(
+    context: str,
+    context_limit: int = 80_000,
+    critical_positions: Optional[List[int]] = None,
+) -> Dict[str, object]:
+    """One-call health analysis for an agent session.
+
+    Use when: a quick health check is needed without manually configuring
+    an analyzer instance. Prints a summary and returns the full result dict.
+
+    Args:
+        context: The full context string to analyze.
+        context_limit: Maximum token budget for this agent's context window.
+        critical_positions: Indices of critical tokens. Defaults to [0..4].
+
+    Returns:
+        Full health analysis dict from ``ContextHealthAnalyzer.analyze``.
+    """
+    analyzer = ContextHealthAnalyzer(context_limit=context_limit)
+
+    if critical_positions is None:
+        critical_positions = list(range(5))
+
+    result = analyzer.analyze(context, critical_positions)
+
+    print(f"Health Score: {result['health_score']:.2f}")
+    print(f"Status: {result['status']}")
+    print("Recommendations:")
+    for rec in result["recommendations"]:
+        print(f"  - {rec}")
+
+    return result
+
+
+# ---------------------------------------------------------------------------
+# CLI Demo
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    # Demonstrate the public API with synthetic context
+    print("=" * 60)
+    print("Context Degradation Detector — Demo")
+    print("=" * 60)
+
+    # Build a synthetic context with identifiable sections
+    intro = "System prompt: Analyze quarterly revenue data and produce a report. "
+    middle = "Background information. " * 200  # Filler to simulate long context
+    conclusion = "Key finding: Revenue increased 15% year-over-year. "
+    sample_context = intro + middle + conclusion
+
+    print(f"\nSample context length: {len(sample_context.split())} tokens")
+
+    # 1. Structure analysis
+    print("\n--- Structure Analysis ---")
+    structure = analyze_context_structure(sample_context)
+    print(f"  Lines: {structure['total_lines']}")
+    print(f"  Middle content ratio: {structure['middle_content_ratio']:.2f}")
+    print(f"  Degradation risk: {structure['degradation_risk']}")
+
+    # 2. Attention distribution (first 50 tokens for brevity)
+    print("\n--- Attention Distribution (first 50 tokens) ---")
+    tokens = sample_context.split()[:50]
+    attention = measure_attention_distribution(tokens, "quarterly revenue")
+    favored = sum(1 for a in attention if a["region"] == "attention_favored")
+    degraded = sum(1 for a in attention if a["region"] == "attention_degraded")
+    print(f"  Favored positions: {favored}")
+    print(f"  Degraded positions: {degraded}")
+
+    # 3. Lost-in-middle detection
+    print("\n--- Lost-in-Middle Detection ---")
+    critical = [0, 1, 2, 25, 26, 48, 49]  # Start, middle, end
+    lim_result = detect_lost_in_middle(critical, attention)
+    print(f"  At risk: {lim_result['at_risk']}")
+    print(f"  Safe: {lim_result['safe']}")
+    print(f"  Degradation score: {lim_result['degradation_score']:.2f}")
+
+    # 4. Poisoning detection
+    print("\n--- Poisoning Detection ---")
+    poisoned_context = (
+        "The API returned an error. However, the system reportedly "
+        "recovered. But the error persisted and the request failed. "
+        "Unable to parse the response. Sources suggest the endpoint "
+        "may have been deprecated. Although retries succeeded, yet "
+        "the invalid token caused an exception."
+    )
+    detector = PoisoningDetector()
+    poisoning = detector.detect_poisoning(poisoned_context)
+    print(f"  Poisoning risk: {poisoning['poisoning_risk']}")
+    print(f"  Overall risk: {poisoning['overall_risk']}")
+    for indicator in poisoning["indicators"]:
+        print(f"    [{indicator['severity']}] {indicator['message']}")
+
+    # 5. Full health analysis
+    print("\n--- Full Health Analysis ---")
+    result = analyze_agent_context(sample_context)
+    print(f"\n  Full result keys: {list(result.keys())}")