diff --git a/traceroot/ai/__init__.py b/traceroot/ai/__init__.py
new file mode 100644
index 00000000..3b9a4e5a
--- /dev/null
+++ b/traceroot/ai/__init__.py
@@ -0,0 +1,3 @@
+from .context import ContextAggregator, DebugContext
+
+__all__ = ['ContextAggregator', 'DebugContext']
diff --git a/traceroot/ai/context/__init__.py b/traceroot/ai/context/__init__.py
new file mode 100644
index 00000000..05f17417
--- /dev/null
+++ b/traceroot/ai/context/__init__.py
@@ -0,0 +1,5 @@
+from .models import ContextNode, DebugContext
+from .aggregator import ContextAggregator
+from .utils import flatten_tree, count_nodes
+
+__all__ = ['ContextNode', 'DebugContext', 'ContextAggregator', 'flatten_tree', 'count_nodes']
diff --git a/traceroot/ai/traceroot/ai/context/models.py b/traceroot/ai/traceroot/ai/context/models.py
new file mode 100644
index 00000000..b282669a
--- /dev/null
+++ b/traceroot/ai/traceroot/ai/context/models.py
@@ -0,0 +1,20 @@
+from dataclasses import dataclass, field
+from typing import Dict, List, Any, Optional
+
+@dataclass
+class ContextNode:
+    """Node in the heterogeneous tree (span or log)"""
+    node_type: str  # 'span' or 'log'
+    data: Dict[str, Any]
+    children: List['ContextNode'] = field(default_factory=list)
+    parent: Optional['ContextNode'] = None
+    timestamp: Optional[float] = None
+    duration: Optional[float] = None
+
+@dataclass
+class DebugContext:
+    trace_id: str
+    root_node: ContextNode
+    github_context: Optional[Dict[str, Any]] = None
+    error_patterns: List[str] = field(default_factory=list)
+    performance_metrics: Dict[str, float] = field(default_factory=dict)
diff --git a/traceroot/ai/traceroot/ai/context/traceroot/ai/context/aggregator.py b/traceroot/ai/traceroot/ai/context/traceroot/ai/context/aggregator.py
new file mode 100644
index 00000000..604b87b8
--- /dev/null
+++ b/traceroot/ai/traceroot/ai/context/traceroot/ai/context/aggregator.py
@@ -0,0 +1,147 @@
+import asyncio
+from typing import Dict, List, Any, Optional
+from .models import ContextNode, DebugContext
+from .utils import flatten_tree, count_nodes
+
+class ContextAggregator:
+    """AI-powered, TraceRoot-compliant context aggregation for debugging"""
+
+    def __init__(self, llm_client):
+        self.llm_client = llm_client
+
+    async def aggregate_context(self, trace_data: Dict[str, Any]) -> DebugContext:
+        logger = traceroot.get_logger()
+        # Step 1: Build heterogeneous tree (spans and logs)
+        root_node = self._build_tree(trace_data)
+        # Step 2: LLM feature filter
+        filtered_tree = await self._llm_feature_filter(root_node)
+        # Step 3: LLM structure filter (optional human-in-the-loop)
+        pruned_tree = await self._llm_structure_filter(filtered_tree)
+        # Step 4: Hierarchical/temporal encoding
+        self._encode_tree(pruned_tree)
+        # Step 5: Analyze error patterns using AI
+        error_patterns = await self._analyze_error_patterns(pruned_tree)
+        # Step 6: Calculate performance metrics
+        performance_metrics = self._calculate_performance_metrics(pruned_tree)
+        # Step 7: Get GitHub context if available
+        github_context = await self._fetch_github_context(trace_data.get('github_info'))
+
+        logger.info({
+            "trace_id": trace_data['trace_id'],
+            "tree_nodes": count_nodes(pruned_tree),
+            "error_patterns_found": len(error_patterns)
+        }, f"Context aggregated for trace {trace_data['trace_id']}")
+
+        return DebugContext(
+            trace_id=trace_data['trace_id'],
+            root_node=pruned_tree,
+            github_context=github_context,
+            error_patterns=error_patterns,
+            performance_metrics=performance_metrics
+        )
+
+    def _build_tree(self, trace_data: Dict[str, Any]) -> ContextNode:
+        """Constructs a heterogeneous tree from spans and logs."""
+        spans = trace_data.get('spans', [])
+        logs = trace_data.get('logs', [])
+        # Build root node
+        root = ContextNode(node_type='root', data={'trace_id': trace_data['trace_id']})
+        # Build span nodes
+        span_nodes = []
+        for span in spans:
+            node = ContextNode(
+                node_type='span',
+                data=span,
+                timestamp=span.get('start_time'),
+                duration=span.get('duration')
+            )
+            span_nodes.append(node)
+        # Build log nodes and attach to spans (if possible)
+        log_nodes = []
+        for log in logs:
+            node = ContextNode(
+                node_type='log',
+                data=log,
+                timestamp=log.get('timestamp')
+            )
+            log_nodes.append(node)
+            # Attach log to its span if span_id exists
+            if 'span_id' in log and log['span_id']:
+                for span_node in span_nodes:
+                    if span_node.data.get('span_id') == log['span_id']:
+                        span_node.children.append(node)
+                        node.parent = span_node
+                        break
+                else:
+                    root.children.append(node)
+                    node.parent = root
+            else:
+                root.children.append(node)
+                node.parent = root
+        # Attach spans to root
+        for node in span_nodes:
+            root.children.append(node)
+            node.parent = root
+        return root
+
+    async def _llm_feature_filter(self, root_node: ContextNode) -> ContextNode:
+        """LLM-based feature filtering: remove useless fields from nodes."""
+        # Flatten tree for feature filtering
+        all_nodes = flatten_tree(root_node)
+        for node in all_nodes:
+            prompt = f"Given this node data of type {node.node_type}: {node.data}\nList only the most essential features for root cause debugging. Remove irrelevant fields."
+            features = await self.llm_client.generate(prompt)
+            if features:
+                # Keep only the features returned by LLM
+                essential = [x.strip() for x in features.split('\n') if x.strip()]
+                node.data = {k: v for k, v in node.data.items() if k in essential}
+        return root_node
+
+    async def _llm_structure_filter(self, root_node: ContextNode) -> ContextNode:
+        """LLM-based structure filtering: prune useless nodes."""
+        all_nodes = flatten_tree(root_node)
+        for node in all_nodes:
+            prompt = f"Should this node of type {node.node_type} with data {node.data} be kept for debugging context? Reply ONLY 'yes' or 'no'."
+            keep = await self.llm_client.generate(prompt)
+            if keep.strip().lower() != 'yes':
+                if node.parent:
+                    node.parent.children.remove(node)
+        return root_node
+
+    def _encode_tree(self, root_node: ContextNode):
+        """Hierarchical and temporal encoding (may be extended)."""
+        # This can be as simple as traversing the tree and setting hierarchy or more advanced encodings
+        pass
+
+    async def _analyze_error_patterns(self, root_node: ContextNode) -> List[str]:
+        """AI-powered error pattern detection from tree."""
+        errors = []
+        for node in flatten_tree(root_node):
+            if node.node_type == 'span' and node.data.get('status', {}).get('code') == 'ERROR':
+                errors.append(f"Span error: {node.data.get('status', {}).get('message', '')}")
+            if node.node_type == 'log' and node.data.get('level') in ['ERROR', 'FATAL']:
+                errors.append(f"Log error: {node.data.get('message', '')}")
+        if not errors:
+            return []
+        prompt = f"Analyze these errors and identify common patterns for debugging:\n{errors}\nFocus on failure modes, bottlenecks, and integration issues."
+        patterns = await self.llm_client.generate(prompt)
+        return [x.strip() for x in patterns.split('\n') if x.strip()] if patterns else []
+
+    def _calculate_performance_metrics(self, root_node: ContextNode) -> Dict[str, float]:
+        """Aggregate performance metrics from spans."""
+        durations = []
+        for node in flatten_tree(root_node):
+            if node.node_type == 'span' and node.duration:
+                durations.append(node.duration)
+        return {
+            "avg_duration": sum(durations) / len(durations) if durations else 0.0,
+            "max_duration": max(durations) if durations else 0.0,
+            "min_duration": min(durations) if durations else 0.0,
+        }
+
+    async def _fetch_github_context(self, github_info: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
+        """Fetch GitHub context (stub example)."""
+        if not github_info:
+            return None
+        # Example: fetch PR or issue details from GitHub API if needed
+        return github_info
diff --git a/traceroot/ai/traceroot/ai/context/traceroot/ai/context/raceroot/ai/context/traceroot/ai/context/__init__.py b/traceroot/ai/traceroot/ai/context/traceroot/ai/context/raceroot/ai/context/traceroot/ai/context/__init__.py
new file mode 100644
index 00000000..05f17417
--- /dev/null
+++ b/traceroot/ai/traceroot/ai/context/traceroot/ai/context/raceroot/ai/context/traceroot/ai/context/__init__.py
@@ -0,0 +1,5 @@
+from .models import ContextNode, DebugContext
+from .aggregator import ContextAggregator
+from .utils import flatten_tree, count_nodes
+
+__all__ = ['ContextNode', 'DebugContext', 'ContextAggregator', 'flatten_tree', 'count_nodes']
diff --git a/traceroot/ai/traceroot/ai/context/traceroot/ai/context/raceroot/ai/context/utils.py b/traceroot/ai/traceroot/ai/context/traceroot/ai/context/raceroot/ai/context/utils.py
new file mode 100644
index 00000000..472ada2e
--- /dev/null
+++ b/traceroot/ai/traceroot/ai/context/traceroot/ai/context/raceroot/ai/context/utils.py
@@ -0,0 +1,16 @@
+from typing import List
+from .models import ContextNode
+
+def flatten_tree(root: ContextNode) -> List[ContextNode]:
+    """Utility: returns all nodes in the tree, depth-first."""
+    result = []
+    stack = [root]
+    while stack:
+        node = stack.pop()
+        result.append(node)
+        stack.extend(node.children)
+    return result
+
+def count_nodes(root: ContextNode) -> int:
+    """Count all nodes in the tree."""
+    return len(flatten_tree(root))