dasein-core 0.2.7__py3-none-any.whl → 0.2.10__py3-none-any.whl

dasein/capture.py

@@ -1,1803 +1,2325 @@
-"""
-Trace capture functionality for Dasein.
-"""
-
-import hashlib
-from typing import Any, Dict, List, Optional, Union
-from datetime import datetime
-from langchain_core.callbacks.base import BaseCallbackHandler
-from langchain_core.callbacks.manager import CallbackManagerForToolRun
-from langchain_core.tools import BaseTool
-
-
-# ============================================================================
-# VERBOSE LOGGING HELPER
-# ============================================================================
-
-def _vprint(message: str, verbose: bool = False, force: bool = False):
-    """
-    Helper function for verbose printing.
-    
-    Args:
-        message: Message to print
-        verbose: Whether verbose mode is enabled
-        force: If True, always print regardless of verbose setting
-    """
-    if force or verbose:
-        print(message)
-
-
-# DEPRECATED: Global trace store removed for thread-safety
-# Traces are now stored instance-level in DaseinCallbackHandler._trace
-# _TRACE: List[Dict[str, Any]] = []
-
-# Hook cache for agent fingerprinting
-_HOOK_CACHE: Dict[str, Any] = {}
-
-# Store for modified tool inputs
-_MODIFIED_TOOL_INPUTS: Dict[str, str] = {}
-
-
-class DaseinToolWrapper(BaseTool):
-    """Wrapper for tools that applies micro-turn modifications."""
-    
-    name: str = ""
-    description: str = ""
-    original_tool: Any = None
-    callback_handler: Any = None
-    
-    def __init__(self, original_tool, callback_handler=None, verbose: bool = False):
-        super().__init__(
-            name=original_tool.name,
-            description=original_tool.description
-        )
-        self.original_tool = original_tool
-        self.callback_handler = callback_handler
-        self._verbose = verbose
-    
-    def _vprint(self, message: str, force: bool = False):
-        """Helper for verbose printing."""
-        _vprint(message, self._verbose, force)
-    
-    def _run(self, *args, **kwargs):
-        """Run the tool with micro-turn injection at execution level."""
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] _run called for {self.name} - VERSION 2.0")
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] Args: {args}")
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] Kwargs: {kwargs}")
-        
-        try:
-            # Get the original input
-            original_input = args[0] if args else ""
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Original input: {original_input[:100]}...")
-            
-            # Apply micro-turn injection if we have rules
-            modified_input = self._apply_micro_turn_injection(str(original_input))
-            
-            if modified_input != original_input:
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
-                # Use modified input
-                result = self.original_tool._run(modified_input, *args[1:], **kwargs)
-            else:
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] No micro-turn injection applied for {self.name}")
-                # Use original input
-                result = self.original_tool._run(*args, **kwargs)
-            
-            # Capture the tool output in the trace
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] About to capture tool output for {self.name}")
-            self._capture_tool_output(self.name, args, kwargs, result)
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Finished capturing tool output for {self.name}")
-            
-            return result
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Exception in _run: {e}")
-            import traceback
-            traceback.print_exc()
-            # Still try to call the original tool
-            result = self.original_tool._run(*args, **kwargs)
-            return result
-    
-    def invoke(self, input_data, config=None, **kwargs):
-        """Invoke the tool with micro-turn injection."""
-        # Get the original input
-        original_input = str(input_data)
-        
-        # Apply micro-turn injection if we have rules
-        modified_input = self._apply_micro_turn_injection(original_input)
-        
-        if modified_input != original_input:
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
-            # Use modified input
-            return self.original_tool.invoke(modified_input, config, **kwargs)
-        else:
-            # Use original input
-            return self.original_tool.invoke(input_data, config, **kwargs)
-    
-    async def _arun(self, *args, **kwargs):
-        """Async run the tool with micro-turn injection at execution level."""
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] _arun called for {self.name} - ASYNC VERSION")
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] Args: {args}")
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] Kwargs: {kwargs}")
-        
-        try:
-            # Get the original input
-            original_input = args[0] if args else ""
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Original input: {original_input[:100]}...")
-            
-            # Apply micro-turn injection if we have rules
-            modified_input = self._apply_micro_turn_injection(str(original_input))
-            
-            if modified_input != original_input:
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
-                # Use modified input
-                result = await self.original_tool._arun(modified_input, *args[1:], **kwargs)
-            else:
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] No micro-turn injection applied for {self.name}")
-                # Use original input
-                result = await self.original_tool._arun(*args, **kwargs)
-            
-            # Capture the tool output in the trace
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] About to capture tool output for {self.name}")
-            self._capture_tool_output(self.name, args, kwargs, result)
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Finished capturing tool output for {self.name}")
-            
-            return result
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Exception in _arun: {e}")
-            import traceback
-            traceback.print_exc()
-            # Still try to call the original tool
-            result = await self.original_tool._arun(*args, **kwargs)
-            return result
-    
-    async def ainvoke(self, input_data, config=None, **kwargs):
-        """Async invoke the tool with micro-turn injection."""
-        self._vprint(f"[DASEIN][TOOL_WRAPPER] ainvoke called for {self.name} - ASYNC VERSION")
-        
-        # Get the original input
-        original_input = str(input_data)
-        
-        # Apply micro-turn injection if we have rules
-        modified_input = self._apply_micro_turn_injection(original_input)
-        
-        if modified_input != original_input:
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
-            # Use modified input
-            return await self.original_tool.ainvoke(modified_input, config, **kwargs)
-        else:
-            # Use original input
-            return await self.original_tool.ainvoke(input_data, config, **kwargs)
-    
-    def _apply_micro_turn_injection(self, original_input: str) -> str:
-        """Apply micro-turn injection to the tool input."""
-        try:
-            # Check if we have a callback handler with rules and LLM
-            if not self.callback_handler:
-                return original_input
-                
-            # Normalize selected rules into Rule objects (handle (rule, metadata) tuples)
-            normalized_rules = []
-            for rule_meta in getattr(self.callback_handler, "_selected_rules", []) or []:
-                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
-                    rule_obj, _metadata = rule_meta
-                else:
-                    rule_obj = rule_meta
-                normalized_rules.append(rule_obj)
-                
-            # Filter tool_start rules
-            tool_rules = [r for r in normalized_rules if getattr(r, 'target_step_type', '') == "tool_start"]
-            
-            if not tool_rules:
-                self._vprint(f"[DASEIN][MICROTURN] No tool rules selected - skipping micro-turn for {self.name}")
-                return original_input
-                
-            # Check if any rule covers this tool
-            covered_rules = [rule for rule in tool_rules 
-                             if self._rule_covers_tool(rule, self.name, original_input)]
-            
-            if not covered_rules:
-                return original_input
-                
-            # Fire micro-turn LLM call (use first matching rule)
-            rule = covered_rules[0]
-            self._vprint(f"[DASEIN][MICROTURN] rule_id={rule.id} tool={self.name}")
-            
-            # Create micro-turn prompt
-            micro_turn_prompt = self._create_micro_turn_prompt(rule, self.name, original_input)
-            
-            # Execute micro-turn LLM call
-            modified_input = self._execute_micro_turn_llm_call(micro_turn_prompt, original_input)
-            
-            self._vprint(f"[DASEIN][MICROTURN] Applied rule {rule.id}: {str(original_input)[:50]}... -> {str(modified_input)[:50]}...")
-            return modified_input
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][MICROTURN] Error in micro-turn injection: {e}")
-            return original_input
-    
-    def _rule_covers_tool(self, rule, tool_name: str, tool_input: str) -> bool:
-        """Check if a rule covers this tool call."""
-        if not hasattr(rule, 'references') or not rule.references:
-            return False
-            
-        # Check if the rule references this tool
-        tools = rule.references.get('tools', [])
-        return tool_name in tools
-    
-    def _create_micro_turn_prompt(self, rule, tool_name: str, tool_input: str) -> str:
-        """Create the prompt for the micro-turn LLM call."""
-        return f"""You are applying a rule to fix a tool input.
-
-Rule: {rule.advice_text}
-
-Tool: {tool_name}
-Current Input: {tool_input}
-
-Apply the rule to fix the input. Return only the corrected input, nothing else."""
-
-    def _execute_micro_turn_llm_call(self, prompt: str, original_input: str) -> str:
-        """Execute the actual micro-turn LLM call."""
-        try:
-            if not self.callback_handler or not self.callback_handler._llm:
-                self._vprint(f"[DASEIN][MICROTURN] No LLM available for micro-turn call")
-                return original_input
-
-            self._vprint(f"[DASEIN][MICROTURN] Executing micro-turn LLM call")
-            self._vprint(f"[DASEIN][MICROTURN] Prompt: {prompt[:200]}...")
-
-            # Make the micro-turn LLM call
-            messages = [{"role": "user", "content": prompt}]
-            response = self.callback_handler._llm.invoke(messages)
-
-            # Extract the response content
-            if hasattr(response, 'content'):
-                modified_input = response.content.strip()
-            elif isinstance(response, str):
-                modified_input = response.strip()
-            else:
-                modified_input = str(response).strip()
-
-            self._vprint(f"[DASEIN][MICROTURN] LLM response: {modified_input[:100]}...")
-
-            # 🚨 CRITICAL: Parse JSON responses with markdown fences
-            if modified_input.startswith('```json') or modified_input.startswith('```'):
-                try:
-                    # Extract JSON from markdown fences
-                    import re
-                    import json
-                    json_match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', modified_input, re.DOTALL)
-                    if json_match:
-                        json_str = json_match.group(1)
-                        parsed_json = json.loads(json_str)
-                        # Convert back to the expected format
-                        if isinstance(parsed_json, dict) and 'name' in parsed_json and 'args' in parsed_json:
-                            modified_input = parsed_json
-                            self._vprint(f"[DASEIN][MICROTURN] Parsed JSON from markdown fences: {parsed_json}")
-                        else:
-                            self._vprint(f"[DASEIN][MICROTURN] JSON doesn't have expected structure, using as-is")
-                    else:
-                        self._vprint(f"[DASEIN][MICROTURN] Could not extract JSON from markdown fences")
-                except Exception as e:
-                    self._vprint(f"[DASEIN][MICROTURN] Error parsing JSON: {e}")
-
-            # Validate the response - only fallback if completely empty
-            if not modified_input:
-                self._vprint(f"[DASEIN][MICROTURN] LLM response empty, using original input")
-                return original_input
-
-            return modified_input
-
-        except Exception as e:
-            self._vprint(f"[DASEIN][MICROTURN] Error executing micro-turn LLM call: {e}")
-            return original_input
-    
-    def _capture_tool_output(self, tool_name, args, kwargs, result):
-        """Capture tool output in the trace."""
-        try:
-            # Create args excerpt
-            args_str = str(args) if args else ""
-            if len(args_str) > 1000:
-                args_str = args_str[:1000] + "..."
-            
-            # Create result excerpt (with 10k limit)
-            result_str = str(result) if result else ""
-            if len(result_str) > 10000:
-                result_str = result_str[:10000] + "..."
-            
-            # Add tool_end step to trace
-            step = {
-                "step_type": "tool_end",
-                "tool_name": tool_name,
-                "args_excerpt": args_str,
-                "outcome": result_str,
-                "ts": datetime.now().isoformat(),
-                "run_id": f"tool_{id(self)}_{datetime.now().timestamp()}",
-                "parent_run_id": None,
-            }
-            
-            # Add to LLM wrapper's trace if available
-            if self.callback_handler and hasattr(self.callback_handler, '_llm') and self.callback_handler._llm:
-                if hasattr(self.callback_handler._llm, '_trace'):
-                    self.callback_handler._llm._trace.append(step)
-                    self._vprint(f"[DASEIN][TOOL_WRAPPER] Added to LLM wrapper trace")
-                else:
-                    self._vprint(f"[DASEIN][TOOL_WRAPPER] LLM wrapper has no _trace attribute")
-            else:
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] No LLM wrapper available")
-            
-            # Also add to callback handler's trace if it has one
-            if self.callback_handler and hasattr(self.callback_handler, '_trace'):
-                self.callback_handler._trace.append(step)
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] Added to callback handler trace")
-            
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Captured tool output for {tool_name}")
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Output length: {len(result_str)} chars")
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] First 200 chars: {result_str[:200]}")
-            if self.callback_handler and hasattr(self.callback_handler, '_trace'):
-                self._vprint(f"[DASEIN][TOOL_WRAPPER] Callback handler trace length after capture: {len(self.callback_handler._trace)}")
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][TOOL_WRAPPER] Error capturing tool output: {e}")
-
-
-class DaseinCallbackHandler(BaseCallbackHandler):
-    """
-    Callback handler that captures step-by-step traces and implements rule injection.
-    """
-    
-    def __init__(self, weights=None, llm=None, is_langgraph=False, coordinator_node=None, planning_nodes=None, verbose: bool = False):
-        super().__init__()
-        self._weights = weights
-        self._selected_rules = []  # Rules selected for this run
-        self._injection_guard = set()  # Prevent duplicate injections
-        self._last_modified_prompts = []  # Store modified prompts for LLM wrapper
-        self._llm = llm  # Store reference to LLM for micro-turn calls
-        self._tool_name_by_run_id = {}  # Track tool names by run_id
-        self._discovered_tools = set()  # Track tools discovered during execution
-        self._wrapped_dynamic_tools = {}  # Cache of wrapped dynamic tools
-        self._is_langgraph = is_langgraph  # Flag to skip planning rule injection for LangGraph
-        self._coordinator_node = coordinator_node  # Coordinator node (for future targeted injection)
-        self._planning_nodes = planning_nodes if planning_nodes else set()  # Planning-capable nodes (including subgraph children)
-        self._current_chain_node = None  # Track current LangGraph node
-        self._agent_was_recreated = False  # Track if agent was successfully recreated
-        self._function_calls_made = {}  # Track function calls: {function_name: [{'step': N, 'ts': timestamp}]}
-        self._trace = []  # Instance-level trace storage (not global) for thread-safety
-        self._verbose = verbose
-        self._start_times = {}  # Track start times for duration calculation: {step_index: datetime}
-        self._vprint(f"[DASEIN][CALLBACK] Initialized callback handler (LangGraph: {is_langgraph})")
-        if coordinator_node:
-            self._vprint(f"[DASEIN][CALLBACK] Coordinator: {coordinator_node}")
-        if planning_nodes:
-            self._vprint(f"[DASEIN][CALLBACK] Planning nodes: {planning_nodes}")
-        self._vprint(f"[DASEIN][CALLBACK] Dynamic tool detection enabled (tools discovered at runtime)")
-    
-    def _vprint(self, message: str, force: bool = False):
-        """Helper for verbose printing."""
-        _vprint(message, self._verbose, force)
-    
-    def reset_run_state(self):
-        """Reset state that should be cleared between runs."""
-        self._function_calls_made = {}
-        self._injection_guard = set()
-        self._trace = []  # Clear instance trace
-        self._start_times = {}  # Clear start times
-        self._vprint(f"[DASEIN][CALLBACK] Reset run state (trace, function calls, injection guard, and start times cleared)")
-    
-    def on_llm_start(
-        self,
-        serialized: Dict[str, Any],
-        prompts: List[str],
-        **kwargs: Any,
-    ) -> None:
-        """Called when an LLM starts running."""
-        model_name = serialized.get("name", "unknown") if serialized else "unknown"
-        
-        # 🎯 CRITICAL: Track current node from kwargs metadata (LangGraph includes langgraph_node)
-        if self._is_langgraph and 'metadata' in kwargs and isinstance(kwargs['metadata'], dict):
-            if 'langgraph_node' in kwargs['metadata']:
-                node_name = kwargs['metadata']['langgraph_node']
-                self._current_chain_node = node_name
-        
-        # Inject rules if applicable
-        modified_prompts = self._inject_rule_if_applicable("llm_start", model_name, prompts)
-        
-        # Store the modified prompts for the LLM wrapper to use
-        self._last_modified_prompts = modified_prompts
-        
-        # 🚨 OPTIMIZED: For LangGraph, check if kwargs contains 'invocation_params' with messages
-        # Extract the most recent message instead of full history
-        # Use from_end=True to capture the END of system prompts (where user's actual query is)
-        if 'invocation_params' in kwargs and 'messages' in kwargs['invocation_params']:
-            args_excerpt = self._extract_recent_message({'messages': kwargs['invocation_params']['messages']})
-        else:
-            args_excerpt = self._excerpt(" | ".join(modified_prompts), from_end=True)
-        
-        # GNN-related fields
-        step_index = len(self._trace)
-        
-        # Track which rules triggered at this step (llm_start rules)
-        rule_triggered_here = []
-        if hasattr(self, '_selected_rules') and self._selected_rules:
-            for rule_meta in self._selected_rules:
-                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
-                    rule_obj, _metadata = rule_meta
-                else:
-                    rule_obj = rule_meta
-                target_step_type = getattr(rule_obj, 'target_step_type', '')
-                if target_step_type in ['llm_start', 'chain_start']:
-                    rule_triggered_here.append(getattr(rule_obj, 'id', 'unknown'))
-        
-        # Record start time for duration calculation
-        start_time = datetime.now()
-        self._start_times[step_index] = start_time
-        
-        step = {
-            "step_type": "llm_start",
-            "tool_name": model_name,
-            "args_excerpt": args_excerpt,
-            "outcome": "",
-            "ts": start_time.isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-            "node": self._current_chain_node,  # LangGraph node name (if available)
-            # GNN step-level fields
-            "step_index": step_index,
-            "rule_triggered_here": rule_triggered_here,
-        }
-        self._trace.append(step)
-        # self._vprint(f"[DASEIN][CALLBACK] Captured llm_start: {len(_TRACE)} total steps")  # Commented out - too noisy
-    
-    def on_llm_end(
-        self,
-        response: Any,
-        **kwargs: Any,
-    ) -> None:
-        """Called when an LLM ends running."""
-        outcome = ""
-        try:
-            # Debug: Print ALL available data to see what we're getting
-            # print(f"[DEBUG] on_llm_end called")
-            # print(f"  response type: {type(response)}")
-            # print(f"  kwargs keys: {kwargs.keys()}")
-            
-            # Try multiple extraction strategies
-            # Strategy 1: Standard LangChain LLMResult structure
-            if hasattr(response, 'generations') and response.generations:
-                if len(response.generations) > 0:
-                    first_gen = response.generations[0]
-                    if isinstance(first_gen, list) and len(first_gen) > 0:
-                        generation = first_gen[0]
-                    else:
-                        generation = first_gen
-                    
-                    # Try multiple content fields
-                    if hasattr(generation, 'text') and generation.text:
-                        outcome = self._excerpt(generation.text)
-                    elif hasattr(generation, 'message'):
-                        if hasattr(generation.message, 'content'):
-                            outcome = self._excerpt(generation.message.content)
-                        elif hasattr(generation.message, 'text'):
-                            outcome = self._excerpt(generation.message.text)
-                    elif hasattr(generation, 'content'):
-                        outcome = self._excerpt(generation.content)
-                    else:
-                        outcome = self._excerpt(str(generation))
-            
-            # Strategy 2: Check if response itself has content
-            elif hasattr(response, 'content'):
-                outcome = self._excerpt(response.content)
-            
-            # Strategy 3: Check kwargs for output/response
-            elif 'output' in kwargs:
-                outcome = self._excerpt(str(kwargs['output']))
-            elif 'result' in kwargs:
-                outcome = self._excerpt(str(kwargs['result']))
-            
-            # Fallback
-            if not outcome:
-                outcome = self._excerpt(str(response))
-            
-            # Debug: Warn if still empty
-            if not outcome or len(outcome) == 0:
-                self._vprint(f"[DASEIN][CALLBACK] WARNING: on_llm_end got empty outcome!")
-                print(f"  Response: {str(response)[:200]}")
-                print(f"  kwargs keys: {list(kwargs.keys())}")
-                
-        except (AttributeError, IndexError, TypeError) as e:
-            self._vprint(f"[DASEIN][CALLBACK] Error in on_llm_end: {e}")
-            outcome = self._excerpt(str(response))
-        
-        # 🎯 CRITICAL: Extract function calls for state tracking (agent-agnostic)
-        try:
-            if hasattr(response, 'generations') and response.generations:
-                first_gen = response.generations[0]
-                if isinstance(first_gen, list) and len(first_gen) > 0:
-                    generation = first_gen[0]
-                else:
-                    generation = first_gen
-                
-                # Check for function_call in message additional_kwargs
-                if hasattr(generation, 'message') and hasattr(generation.message, 'additional_kwargs'):
-                    func_call = generation.message.additional_kwargs.get('function_call')
-                    if func_call and isinstance(func_call, dict) and 'name' in func_call:
-                        func_name = func_call['name']
-                        step_num = len(self._trace)
-                        
-                        # Extract arguments and create preview
-                        args_str = func_call.get('arguments', '')
-                        preview = ''
-                        if args_str and len(args_str) > 0:
-                            # Take first 100 chars as preview
-                            preview = args_str[:100].replace('\n', ' ').replace('\r', '')
-                            if len(args_str) > 100:
-                                preview += '...'
-                        
-                        call_info = {
-                            'step': step_num,
-                            'ts': datetime.now().isoformat(),
-                            'preview': preview
-                        }
-                        
-                        if func_name not in self._function_calls_made:
-                            self._function_calls_made[func_name] = []
-                        self._function_calls_made[func_name].append(call_info)
-                        
-                        self._vprint(f"[DASEIN][STATE] Tracked function call: {func_name} (count: {len(self._function_calls_made[func_name])})")
-        except Exception as e:
-            pass  # Silently skip if function call extraction fails
-        
-        # Extract token usage from response metadata
-        input_tokens = 0
-        output_tokens = 0
-        try:
-            # DEBUG: Print response structure for first LLM call
-            # Uncomment to see token structure:
-            # import json
-            # print(f"[DEBUG] Response structure:")
-            # print(f"  Has llm_output: {hasattr(response, 'llm_output')}")
-            # if hasattr(response, 'llm_output'):
-            #     print(f"  llm_output keys: {response.llm_output.keys() if response.llm_output else None}")
-            # print(f"  Has generations: {hasattr(response, 'generations')}")
-            # if hasattr(response, 'generations') and response.generations:
-            #     gen = response.generations[0][0] if isinstance(response.generations[0], list) else response.generations[0]
-            #     print(f"  generation_info: {gen.generation_info if hasattr(gen, 'generation_info') else None}")
-            
-            # Try LangChain's standard llm_output field
-            if hasattr(response, 'llm_output') and response.llm_output:
-                llm_output = response.llm_output
-                # Different providers use different field names
-                if 'token_usage' in llm_output:
-                    usage = llm_output['token_usage']
-                    input_tokens = usage.get('prompt_tokens', 0) or usage.get('input_tokens', 0)
-                    output_tokens = usage.get('completion_tokens', 0) or usage.get('output_tokens', 0)
-                elif 'usage_metadata' in llm_output:
-                    usage = llm_output['usage_metadata']
-                    input_tokens = usage.get('input_tokens', 0) or usage.get('prompt_tokens', 0)
-                    output_tokens = usage.get('output_tokens', 0) or usage.get('completion_tokens', 0)
-            
-            # Try generations metadata (Google GenAI format)
-            if (input_tokens == 0 and output_tokens == 0) and hasattr(response, 'generations') and response.generations:
-                first_gen = response.generations[0]
-                if isinstance(first_gen, list) and len(first_gen) > 0:
-                    gen = first_gen[0]
-                else:
-                    gen = first_gen
-                
-                # Check message.usage_metadata (Google GenAI stores it here!)
-                if hasattr(gen, 'message') and hasattr(gen.message, 'usage_metadata'):
-                    usage = gen.message.usage_metadata
-                    input_tokens = usage.get('input_tokens', 0)
-                    output_tokens = usage.get('output_tokens', 0)
-                
-                # Fallback: Check generation_info
-                elif hasattr(gen, 'generation_info') and gen.generation_info:
-                    gen_info = gen.generation_info
-                    if 'usage_metadata' in gen_info:
-                        usage = gen_info['usage_metadata']
-                        input_tokens = usage.get('prompt_token_count', 0) or usage.get('input_tokens', 0)
-                        output_tokens = usage.get('candidates_token_count', 0) or usage.get('output_tokens', 0)
-            
-            # Log if we got tokens
-            # if input_tokens > 0 or output_tokens > 0:
-            #     self._vprint(f"[DASEIN][TOKENS] Captured: {input_tokens} in, {output_tokens} out")
-                    
-        except Exception as e:
-            # Print error for debugging
-            self._vprint(f"[DASEIN][CALLBACK] Error extracting tokens: {e}")
-            import traceback
-            traceback.print_exc()
-        
-        # GNN-related fields: compute tokens_delta
-        step_index = len(self._trace)
-        tokens_delta = 0
-        # Find previous step with tokens_output to compute delta
-        for prev_step in reversed(self._trace):
-            if 'tokens_output' in prev_step and prev_step['tokens_output'] > 0:
-                tokens_delta = output_tokens - prev_step['tokens_output']
-                break
-        
-        # Calculate duration_ms by matching with corresponding llm_start
-        duration_ms = 0
-        for i in range(len(self._trace) - 1, -1, -1):
-            if self._trace[i].get('step_type') == 'llm_start':
-                # Found the matching llm_start
-                if i in self._start_times:
-                    start_time = self._start_times[i]
-                    end_time = datetime.now()
-                    duration_ms = int((end_time - start_time).total_seconds() * 1000)
-                    # Update the llm_start step with duration_ms
-                    self._trace[i]['duration_ms'] = duration_ms
-                break
-        
-        step = {
-            "step_type": "llm_end",
-            "tool_name": "",
-            "args_excerpt": "",
-            "outcome": self._excerpt(outcome, max_len=1000),  # Truncate to 1000 chars
-            "ts": datetime.now().isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-            "tokens_input": input_tokens,
-            "tokens_output": output_tokens,
-            "node": self._current_chain_node,  # LangGraph node name (if available)
-            # GNN step-level fields
-            "step_index": step_index,
-            "tokens_delta": tokens_delta,
-            "duration_ms": duration_ms,
-        }
-        self._trace.append(step)
-    
-    def on_agent_action(
-        self,
-        action: Any,
-        **kwargs: Any,
-    ) -> None:
-        """Called when an agent takes an action."""
-        tool_name = getattr(action, 'tool', 'unknown')
-        args_excerpt = self._excerpt(str(getattr(action, 'tool_input', '')))
-        outcome = self._excerpt(str(getattr(action, 'log', '')))
-        
-        step = {
-            "step_type": "agent_action",
-            "tool_name": tool_name,
-            "args_excerpt": args_excerpt,
-            "outcome": outcome,
-            "ts": datetime.now().isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-        }
-        self._trace.append(step)
-    
-    def on_agent_finish(
-        self,
-        finish: Any,
-        **kwargs: Any,
-    ) -> None:
-        """Called when an agent finishes."""
-        outcome = self._excerpt(str(getattr(finish, 'return_values', '')))
-        
-        step = {
-            "step_type": "agent_finish",
-            "tool_name": None,
-            "args_excerpt": "",
-            "outcome": outcome,
-            "ts": datetime.now().isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-        }
-        self._trace.append(step)
-    
-    def on_tool_start(
-        self,
-        serialized: Dict[str, Any],
-        input_str: str,
-        *,
-        run_id: str,
-        parent_run_id: Optional[str] = None,
-        tags: Optional[List[str]] = None,
-        metadata: Optional[Dict[str, Any]] = None,
-        inputs: Optional[Dict[str, Any]] = None,
-        **kwargs: Any,
-    ) -> None:
-        """Called when a tool starts running.
-        
-        This is where we detect and track dynamic tools that weren't
-        statically attached to the agent at init time.
-        """
-        tool_name = serialized.get("name", "unknown") if serialized else "unknown"
-        
-        # Track discovered tools for reporting
-        if tool_name != "unknown" and tool_name not in self._discovered_tools:
-            self._discovered_tools.add(tool_name)
-            # Tool discovered and tracked (silently)
-        
-        # Store tool name for later use in on_tool_end
-        self._tool_name_by_run_id[run_id] = tool_name
-        
-        # Apply tool-level rule injection
-        # self._vprint(f"[DASEIN][CALLBACK] on_tool_start called!")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][CALLBACK] Tool: {tool_name}")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][CALLBACK] Input: {input_str[:100]}...")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][APPLY] on_tool_start: selected_rules={len(self._selected_rules)}")  # Commented out - too noisy
-        modified_input = self._inject_tool_rule_if_applicable("tool_start", tool_name, input_str)
-        
-        args_excerpt = self._excerpt(modified_input)
-        
-        # GNN-related fields: capture step-level metrics
-        step_index = len(self._trace)
-        tool_input_chars = len(str(input_str))
-        
-        # Track which rules triggered at this step
-        rule_triggered_here = []
-        if hasattr(self, '_selected_rules') and self._selected_rules:
-            for rule_meta in self._selected_rules:
-                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
-                    rule_obj, _metadata = rule_meta
-                else:
-                    rule_obj = rule_meta
-                if getattr(rule_obj, 'target_step_type', '') == "tool_start":
-                    rule_triggered_here.append(getattr(rule_obj, 'id', 'unknown'))
-        
-        # Record start time for duration calculation (keyed by run_id for tools)
-        start_time = datetime.now()
-        self._start_times[run_id] = start_time
-        
-        step = {
-            "step_type": "tool_start",
-            "tool_name": tool_name,
-            "args_excerpt": args_excerpt,
-            "outcome": "",
-            "ts": start_time.isoformat(),
-            "run_id": run_id,
-            "parent_run_id": parent_run_id,
-            "node": self._current_chain_node,  # LangGraph node name (if available)
-            # GNN step-level fields
-            "step_index": step_index,
-            "tool_input_chars": tool_input_chars,
-            "rule_triggered_here": rule_triggered_here,
-        }
-        self._trace.append(step)
-    
-    def on_tool_end(
-        self,
-        output: str,
-        *,
-        run_id: str,
-        parent_run_id: Optional[str] = None,
-        tags: Optional[List[str]] = None,
-        **kwargs: Any,
-    ) -> None:
-        """Called when a tool ends running."""
-        # Get the tool name from the corresponding tool_start
-        tool_name = self._tool_name_by_run_id.get(run_id, "unknown")
-        
-        # Handle different output types (LangGraph may pass ToolMessage objects)
-        output_str = str(output)
-        outcome = self._excerpt(output_str)
-        
-        # self._vprint(f"[DASEIN][CALLBACK] on_tool_end called!")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][CALLBACK] Tool: {tool_name}")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][CALLBACK] Output length: {len(output_str)} chars")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][CALLBACK] Outcome length: {len(outcome)} chars")  # Commented out - too noisy
-        
-        # GNN-related fields: capture tool output metrics
-        step_index = len(self._trace)
-        tool_output_chars = len(output_str)
-        
-        # Estimate tool_output_items (heuristic: count lines, or rows if SQL-like)
-        tool_output_items = 0
-        try:
-            # Try to count lines as a proxy for items
-            if output_str:
-                tool_output_items = output_str.count('\n') + 1
-        except:
-            tool_output_items = 0
-        
-        # Calculate duration_ms using run_id to match with tool_start
-        duration_ms = 0
-        if run_id in self._start_times:
-            start_time = self._start_times[run_id]
-            end_time = datetime.now()
-            duration_ms = int((end_time - start_time).total_seconds() * 1000)
-            # Update the corresponding tool_start step with duration_ms
-            for i in range(len(self._trace) - 1, -1, -1):
-                if self._trace[i].get('step_type') == 'tool_start' and self._trace[i].get('run_id') == run_id:
-                    self._trace[i]['duration_ms'] = duration_ms
-                    break
-            # Clean up start time
-            del self._start_times[run_id]
-        
-        # Extract available selectors from DOM-like output (web browse agents)
-        available_selectors = None
-        if tool_name in ['extract_text', 'get_elements', 'extract_hyperlinks', 'extract_content']:
-            available_selectors = self._extract_semantic_selectors(output_str)
-        
-        step = {
-            "step_type": "tool_end",
-            "tool_name": tool_name,
-            "args_excerpt": "",
-            "outcome": self._excerpt(outcome, max_len=1000),  # Truncate to 1000 chars
-            "ts": datetime.now().isoformat(),
-            "run_id": run_id,
-            "parent_run_id": parent_run_id,
-            "node": self._current_chain_node,  # LangGraph node name (if available)
-            # GNN step-level fields
-            "step_index": step_index,
-            "tool_output_chars": tool_output_chars,
-            "tool_output_items": tool_output_items,
-            "duration_ms": duration_ms,
-        }
-        
-        # Add available_selectors only if found (keep trace light)
-        if available_selectors:
-            step["available_selectors"] = available_selectors
-        self._trace.append(step)
-        
-        # Clean up the stored tool name
-        if run_id in self._tool_name_by_run_id:
-            del self._tool_name_by_run_id[run_id]
-    
-    def on_tool_error(
-        self,
-        error: BaseException,
-        *,
-        run_id: str,
-        parent_run_id: Optional[str] = None,
-        tags: Optional[List[str]] = None,
-        **kwargs: Any,
-    ) -> None:
-        """Called when a tool encounters an error."""
-        error_msg = self._excerpt(str(error))
-        
-        step = {
-            "step_type": "tool_error",
-            "tool_name": "",
-            "args_excerpt": "",
-            "outcome": f"ERROR: {error_msg}",
-            "ts": datetime.now().isoformat(),
-            "run_id": run_id,
-            "parent_run_id": parent_run_id,
-        }
-        self._trace.append(step)
-    
-    def on_chain_start(
-        self,
-        serialized: Dict[str, Any],
-        inputs: Dict[str, Any],
-        **kwargs: Any,
-    ) -> None:
-        """Called when a chain starts running."""
-        chain_name = serialized.get("name", "unknown") if serialized else "unknown"
-        # self._vprint(f"[DASEIN][CALLBACK] on_chain_start called!")  # Commented out - too noisy
-        # self._vprint(f"[DASEIN][CALLBACK] Chain: {chain_name}")  # Commented out - too noisy
-        
-        # 🚨 OPTIMIZED: For LangGraph agents, suppress redundant chain_start events
-        # LangGraph fires on_chain_start for every internal node, creating noise
-        # We already capture llm_start, llm_end, tool_start, tool_end which are more meaningful
-        if self._is_langgraph:
-            # Track current chain node for future targeted injection
-            # 🎯 CRITICAL: Extract actual node name from metadata (same as on_llm_start)
-            if 'metadata' in kwargs and isinstance(kwargs['metadata'], dict):
-                if 'langgraph_node' in kwargs['metadata']:
-                    self._current_chain_node = kwargs['metadata']['langgraph_node']
-                else:
-                    self._current_chain_node = chain_name
-            else:
-                self._current_chain_node = chain_name
-            
-            # self._vprint(f"[DASEIN][CALLBACK] Suppressing redundant chain_start for LangGraph agent")  # Commented out - too noisy
-            # Still handle tool executors
-            if chain_name in {"tools", "ToolNode", "ToolExecutor"}:
-                # self._vprint(f"[DASEIN][CALLBACK] Bridging chain_start to tool_start for {chain_name}")  # Commented out - too noisy
-                pass
-                self._handle_tool_executor_start(serialized, inputs, **kwargs)
-            return
-        
-        # For standard LangChain agents, keep chain_start events
-        # Bridge to tool_start for tool executors
-        if chain_name in {"tools", "ToolNode", "ToolExecutor"}:
-            # self._vprint(f"[DASEIN][CALLBACK] Bridging chain_start to tool_start for {chain_name}")  # Commented out - too noisy
-            self._handle_tool_executor_start(serialized, inputs, **kwargs)
-        
-        args_excerpt = self._excerpt(str(inputs))
-        
-        # Record start time for duration calculation
-        step_index = len(self._trace)
-        start_time = datetime.now()
-        self._start_times[f"chain_{step_index}"] = start_time
-        
-        step = {
-            "step_type": "chain_start",
-            "tool_name": chain_name,
-            "args_excerpt": args_excerpt,
-            "outcome": "",
-            "ts": start_time.isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-            "step_index": step_index,
-        }
-        self._trace.append(step)
-    
-    def on_chain_end(
-        self,
-        outputs: Dict[str, Any],
-        **kwargs: Any,
-    ) -> None:
-        """Called when a chain ends running."""
-        # 🚨 OPTIMIZED: Suppress redundant chain_end for LangGraph agents
-        if self._is_langgraph:
-            return
-        
-        outcome = self._excerpt(str(outputs))
-        
-        # Calculate duration_ms by matching with corresponding chain_start
-        duration_ms = 0
-        for i in range(len(self._trace) - 1, -1, -1):
-            if self._trace[i].get('step_type') == 'chain_start':
-                # Found the matching chain_start
-                chain_key = f"chain_{i}"
-                if chain_key in self._start_times:
-                    start_time = self._start_times[chain_key]
-                    end_time = datetime.now()
-                    duration_ms = int((end_time - start_time).total_seconds() * 1000)
-                    # Update the chain_start step with duration_ms
-                    self._trace[i]['duration_ms'] = duration_ms
-                    # Clean up start time
-                    del self._start_times[chain_key]
-                break
-        
-        step = {
-            "step_type": "chain_end",
-            "tool_name": "",
-            "args_excerpt": "",
-            "outcome": outcome,
-            "ts": datetime.now().isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-            "duration_ms": duration_ms,
-        }
-        self._trace.append(step)
-    
-    def on_chain_error(
-        self,
-        error: BaseException,
-        **kwargs: Any,
-    ) -> None:
-        """Called when a chain encounters an error."""
-        error_msg = self._excerpt(str(error))
-        
-        step = {
-            "step_type": "chain_error",
-            "tool_name": "",
-            "args_excerpt": "",
-            "outcome": f"ERROR: {error_msg}",
-            "ts": datetime.now().isoformat(),
-            "run_id": None,
-            "parent_run_id": None,
-        }
-        self._trace.append(step)
-    
-    def _extract_recent_message(self, inputs: Dict[str, Any]) -> str:
-        """
-        Extract the most recent message from LangGraph inputs to show thought progression.
-        
-        For LangGraph agents, inputs contain {'messages': [msg1, msg2, ...]}.
-        Instead of showing the entire history, we extract just the last message.
-        """
-        try:
-            # Check if this is a LangGraph message format
-            if isinstance(inputs, dict) and 'messages' in inputs:
-                messages = inputs['messages']
-                if isinstance(messages, list) and len(messages) > 0:
-                    # Get the most recent message
-                    last_msg = messages[-1]
-                    
-                    # Extract content based on message type
-                    if hasattr(last_msg, 'content'):
-                        # LangChain message object
-                        content = last_msg.content
-                        msg_type = getattr(last_msg, 'type', 'unknown')
-                        return self._excerpt(f"[{msg_type}] {content}")
-                    elif isinstance(last_msg, tuple) and len(last_msg) >= 2:
-                        # Tuple format: (role, content)
-                        return self._excerpt(f"[{last_msg[0]}] {last_msg[1]}")
-                    else:
-                        # Unknown format, convert to string
-                        return self._excerpt(str(last_msg))
-            
-            # For non-message inputs, check if it's a list of actions/tool calls
-            if isinstance(inputs, list) and len(inputs) > 0:
-                # This might be tool call info
-                return self._excerpt(str(inputs[0]))
-            
-            # Fall back to original behavior for non-LangGraph agents
-            return self._excerpt(str(inputs))
-            
-        except Exception as e:
-            # On any error, fall back to original behavior
-            return self._excerpt(str(inputs))
-    
-    def _excerpt(self, obj: Any, max_len: int = 250, from_end: bool = False) -> str:
-        """
-        Truncate text to max_length with ellipsis.
-        
-        Args:
-            obj: Object to convert to string and truncate
-            max_len: Maximum length of excerpt
-            from_end: If True, take LAST max_len chars (better for system prompts).
-                     If False, take FIRST max_len chars (better for tool args).
-        """
-        text = str(obj)
-        if len(text) <= max_len:
-            return text
-        
-        if from_end:
-            # Take last X chars - better for system prompts where the end contains user's actual query
-            return "..." + text[-(max_len-3):]
-        else:
-            # Take first X chars - better for tool inputs
-            return text[:max_len-3] + "..."
-    
-    def _extract_semantic_selectors(self, html_text: str) -> List[Dict[str, int]]:
-        """
-        Extract semantic HTML tags from output for grounding web browse rules.
-        Only extracts semantic tags (nav, header, h1, etc.) to keep trace lightweight.
-        
-        Args:
-            html_text: Output text that may contain HTML
-            
-        Returns:
-            List of {"tag": str, "count": int} sorted by count descending, or None if no HTML
-        """
-        import re
-        
-        # Quick check: does this look like HTML?
-        if '<' not in html_text or '>' not in html_text:
-            return None
-        
-        # Semantic tags we care about (prioritized for web browse agents)
-        semantic_tags = [
-            # Navigation/Structure (highest priority)
-            'nav', 'header', 'footer', 'main', 'article', 'section', 'aside',
-            
-            # Headers (critical for "find headers" queries!)
-            'h1', 'h2', 'h3', 'h4', 'h5', 'h6',
-            
-            # Interactive
-            'a', 'button', 'form', 'input', 'textarea', 'select', 'label',
-            
-            # Lists (often used for navigation)
-            'ul', 'ol', 'li',
-            
-            # Tables (data extraction)
-            'table', 'thead', 'tbody', 'tr', 'th', 'td',
-            
-            # Media
-            'img', 'video', 'audio'
-        ]
-        
-        # Count occurrences of each semantic tag
-        found_tags = {}
-        for tag in semantic_tags:
-            # Pattern: <tag ...> or <tag> (opening tags only)
-            pattern = f'<{tag}[\\s>]'
-            matches = re.findall(pattern, html_text, re.IGNORECASE)
-            if matches:
-                found_tags[tag] = len(matches)
-        
-        # Return None if no semantic tags found
-        if not found_tags:
-            return None
-        
-        # Convert to list format, sorted by count descending
-        # Limit to top 15 to keep trace light
-        result = [{"tag": tag, "count": count} 
-                  for tag, count in sorted(found_tags.items(), key=lambda x: -x[1])]
-        return result[:15]  # Top 15 most common tags
-    
-    def set_selected_rules(self, rules: List[Dict[str, Any]]):
-        """Set the rules selected for this run.
-        Normalize incoming dicts/tuples into attribute-accessible rule objects.
-        """
-        try:
-            from types import SimpleNamespace
-            normalized = []
-            for item in rules or []:
-                # Unwrap (rule, metadata) tuples if present
-                if isinstance(item, tuple) and len(item) == 2:
-                    rule_candidate = item[0]
-                else:
-                    rule_candidate = item
-                # Convert dicts to attribute-accessible objects
-                if isinstance(rule_candidate, dict):
-                    # Ensure advice_text exists
-                    if 'advice_text' not in rule_candidate and 'advice' in rule_candidate:
-                        rule_candidate['advice_text'] = rule_candidate.get('advice')
-                    normalized.append(SimpleNamespace(**rule_candidate))
-                else:
-                    normalized.append(rule_candidate)
-            self._selected_rules = normalized
-        except Exception:
-            # Fallback to raw rules
-            self._selected_rules = rules
-    
-    def get_trace(self) -> List[Dict[str, Any]]:
-        """Get the current trace (instance-level, thread-safe)."""
-        return self._trace.copy()
-    
-    def _inject_first_turn_override(self, prompts: List[str]) -> List[str]:
-        """Inject a generic first-turn override to own turn 1."""
-        if not prompts:
-            return prompts
-            
-        # Create a generic first-turn override
-        first_turn_override = """🚨🚨🚨 CRITICAL SYSTEM DIRECTIVE 🚨🚨🚨
-⚠️ MANDATORY: You MUST follow this exact sequence or the task will FAIL
-
-TURN 1 REQUIREMENT:
-- Output ONLY: Action: sql_db_list_tables
-Action Input: ACK_RULES:[r1]
-- Do NOT use any other tools
-- Do NOT perform any planning
-- Do NOT output anything else
-
-TURN 2+ (After ACK):
-- If ACK was correct, proceed with normal tools and schema
-- Skip table discovery and schema introspection
-- Use known tables directly
-
-🚨 FAILURE TO ACK IN TURN 1 = IMMEDIATE TASK TERMINATION 🚨
-
-"""
-        
-        # Put the injection at the VERY BEGINNING of the system prompt
-        modified_prompts = prompts.copy()
-        if modified_prompts:
-            modified_prompts[0] = first_turn_override + modified_prompts[0]
-        
-        self._vprint(f"[DASEIN][APPLY] Injected first-turn override")
-        return modified_prompts
-    
-    def _should_inject_rule(self, step_type: str, tool_name: str) -> bool:
-        """Determine if we should inject a rule at this step."""
-        # Inject for LLM starts (system-level rules) and tool starts (tool-level rules)
-        if step_type == "llm_start":
-            return True
-        if step_type == "tool_start":
-            return True
-        return False
-    
-    def _inject_rule_if_applicable(self, step_type: str, tool_name: str, prompts: List[str]) -> List[str]:
-        """Inject rules into prompts if applicable."""
-        if not self._should_inject_rule(step_type, tool_name):
-            return prompts
-
-        # If no rules selected yet, return prompts unchanged
-        if not self._selected_rules:
-            return prompts
-
-        # Check guard to prevent duplicate injection
-        # 🎯 CRITICAL: For LangGraph planning nodes, SKIP the guard - we need to inject on EVERY call
-        # because the same node (e.g., supervisor) can be called multiple times dynamically
-        use_guard = True
-        if hasattr(self, '_is_langgraph') and self._is_langgraph:
-            if step_type == 'llm_start' and hasattr(self, '_current_chain_node'):
-                # For planning nodes, skip guard to allow re-injection on subsequent calls
-                if hasattr(self, '_planning_nodes') and self._current_chain_node in self._planning_nodes:
-                    use_guard = False
-        
-        if use_guard:
-            guard_key = (step_type, tool_name)
-            if guard_key in self._injection_guard:
-                return prompts
-        
-        try:
-            # Inject rules that target llm_start and tool_start (both go to system prompt)
-            system_rules = []
-            for rule_meta in self._selected_rules:
-                # Handle tuple format from select_rules: (rule, metadata)
-                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
-                    rule, metadata = rule_meta
-                elif isinstance(rule_meta, dict):
-                    if 'rule' in rule_meta:
-                        rule = rule_meta.get('rule', {})
-                    else:
-                        rule = rule_meta
-                else:
-                    rule = rule_meta
-                
-                # Check if this rule targets system-level injection (llm_start only)
-                target_step_type = getattr(rule, 'target_step_type', '')
-                
-                # 🚨 CRITICAL: For LangGraph agents, only skip planning rules if agent was successfully recreated
-                # If recreation failed, we need to inject via callback as fallback
-                if step_type == 'llm_start' and hasattr(self, '_is_langgraph') and self._is_langgraph:
-                    # Only skip if agent was actually recreated with planning rules embedded
-                    if hasattr(self, '_agent_was_recreated') and self._agent_was_recreated:
-                        if target_step_type in ['llm_start', 'chain_start']:
-                            self._vprint(f"[DASEIN][CALLBACK] Skipping planning rule {getattr(rule, 'id', 'unknown')} for LangGraph agent (already injected at creation)")
-                            continue
-                
-                # 🎯 COORDINATOR-GATED INJECTION: Only apply planning rules when executing planning-capable nodes
-                if target_step_type in ['llm_start', 'chain_start']:
-                    # If we have planning nodes, only inject planning rules when we're in one of them
-                    if hasattr(self, '_planning_nodes') and self._planning_nodes:
-                        current_node = getattr(self, '_current_chain_node', None)
-                        # Check if current node is in the planning nodes set
-                        if current_node not in self._planning_nodes:
-                            # Silently skip non-planning nodes
-                            continue
-                        # Injecting into planning node (logged in detailed injection log below)
-                    
-                    advice = getattr(rule, 'advice_text', getattr(rule, 'advice', ''))
-                    if advice:
-                        system_rules.append(advice)
-            
-            # Apply system-level rules if any
-            if system_rules and prompts:
-                modified_prompts = prompts.copy()
-                system_prompt = modified_prompts[0]
-                
-                # Combine all system rules with much stronger language
-                rule_injections = []
-                for advice in system_rules:
-                    if "TOOL RULE:" in advice:
-                        # Make tool rules even more explicit
-                        rule_injections.append(f"🚨 CRITICAL TOOL OVERRIDE: {advice}")
-                    else:
-                        rule_injections.append(f"🚨 CRITICAL SYSTEM OVERRIDE: {advice}")
-                
-                # Build execution state context (agent-agnostic, with argument previews)
-                # Strategy: Show all if ≤5 calls, else show most recent 3
-                # Rationale: Small counts get full context; larger counts show recent to prevent duplicates
-                state_context = ""
-                if hasattr(self, '_function_calls_made') and self._function_calls_made:
-                    state_lines = []
-                    for func_name in sorted(self._function_calls_made.keys()):
-                        calls = self._function_calls_made[func_name]
-                        count = len(calls)
-                        
-                        # Hybrid window: show all if ≤5 calls, else show recent 3
-                        if count <= 5:
-                            # Show all calls with previews
-                            state_lines.append(f"  • {func_name}: called {count}x:")
-                            for call in calls:
-                                preview = call.get('preview', '')
-                                if preview:
-                                    state_lines.append(f"      [step {call['step']}] {preview}")
-                                else:
-                                    state_lines.append(f"      [step {call['step']}] (no args)")
-                        else:
-                            # Show summary + recent 3 with previews
-                            state_lines.append(f"  • {func_name}: called {count}x (most recent 3):")
-                            for call in calls[-3:]:
-                                preview = call.get('preview', '')
-                                if preview:
-                                    state_lines.append(f"      [step {call['step']}] {preview}")
-                                else:
-                                    state_lines.append(f"      [step {call['step']}] (no args)")
-                    
-                    if state_lines:
-                        state_context = f"""
-EXECUTION STATE (functions called so far in this run):
-{chr(10).join(state_lines)}
-
-"""
-                
-                combined_injection = f""" SYSTEM OVERRIDE — PLANNING TURN ONLY 
-These rules OVERRIDE all defaults. You MUST enforce them exactly or the task FAILS.
-
-Tags: AVOID (absolute ban), SKIP (force bypass), FIX (mandatory params), PREFER (ranked choice), HINT (optional).
-Precedence: AVOID/SKIP > FIX > PREFER > HINT. On conflict, the higher rule ALWAYS wins.
-
-{state_context}Checklist (non-negotiable):
-- AVOID: no banned targets under ANY condition.
-- SKIP: bypass skipped steps/tools; NEVER retry them.
-- FIX: all required params/settings MUST be included.
-- PREFER: when multiple compliant options exist, choose the preferred—NO exceptions.
-- Recovery: if a banned/skipped item already failed, IMMEDIATELY switch to a compliant alternative.
-
-Output Contract: Produce ONE compliant tool/function call (or direct answer if none is needed).  
-NO reasoning, NO justification, NO markdown.
-
-Rules to Enforce:
-
-
-{chr(10).join(rule_injections)}
-
-
-"""
-                # Put the injection at the VERY BEGINNING of the system prompt
-                modified_prompts[0] = combined_injection + system_prompt
-                
-                # Add to guard (only if we're using the guard)
-                if use_guard:
-                    self._injection_guard.add(guard_key)
-                
-                # Log the complete injection for debugging
-                # Compact injection summary
-                if hasattr(self, '_is_langgraph') and self._is_langgraph:
-                    # LangGraph: show node name
-                    func_count = len(self._function_calls_made) if hasattr(self, '_function_calls_made') and state_context else 0
-                    node_name = getattr(self, '_current_chain_node', 'unknown')
-                    print(f"[DASEIN] 🎯 Injecting {len(system_rules)} rule(s) into {node_name} | State: {func_count} functions tracked")
-                else:
-                    # LangChain: simpler logging without node name
-                    print(f"[DASEIN] 🎯 Injecting {len(system_rules)} rule(s) into agent")
-                
-                return modified_prompts
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][APPLY] Injection failed: {e}")
-        
-        return prompts
-    
-    def _inject_tool_rule_if_applicable(self, step_type: str, tool_name: str, input_str: str) -> str:
-        """Inject rules into tool input if applicable."""
-        if not self._should_inject_rule(step_type, tool_name):
-            return input_str
-
-        # If no rules selected yet, return input unchanged
-        if not self._selected_rules:
-            return input_str
-
-        # Check guard to prevent duplicate injection
-        guard_key = (step_type, tool_name)
-        if guard_key in self._injection_guard:
-            return input_str
-        
-        try:
-            # Inject rules that target tool_start
-            tool_rules = []
-            for rule_meta in self._selected_rules:
-                # Handle tuple format from select_rules: (rule, metadata)
-                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
-                    rule, metadata = rule_meta
-                else:
-                    rule = rule_meta
-                    metadata = {}
-
-                # Only apply rules that target tool_start
-                if rule.target_step_type == "tool_start":
-                    tool_rules.append(rule)
-                    self._vprint(f"[DASEIN][APPLY] Tool rule: {rule.advice_text[:100]}...")
-
-            if tool_rules:
-                # Apply tool-level rule injection
-                modified_input = self._apply_tool_rules(input_str, tool_rules)
-                self._injection_guard.add(guard_key)
-                return modified_input
-            else:
-                return input_str
-
-        except Exception as e:
-            self._vprint(f"[DASEIN][APPLY] Error injecting tool rules: {e}")
-            return input_str
-    
-    def _apply_tool_rules(self, input_str: str, rules: List) -> str:
-        """Apply tool-level rules to modify the input string."""
-        modified_input = input_str
-        
-        for rule in rules:
-            try:
-                # Apply the rule's advice to modify the tool input
-                if "strip" in rule.advice_text.lower() and "fence" in rule.advice_text.lower():
-                    # Strip markdown code fences
-                    import re
-                    # Remove ```sql...``` or ```...``` patterns
-                    modified_input = re.sub(r'```(?:sql)?\s*(.*?)\s*```', r'\1', modified_input, flags=re.DOTALL)
-                    self._vprint(f"[DASEIN][APPLY] Stripped code fences from tool input")
-                elif "strip" in rule.advice_text.lower() and "whitespace" in rule.advice_text.lower():
-                    # Strip leading/trailing whitespace
-                    modified_input = modified_input.strip()
-                    self._vprint(f"[DASEIN][APPLY] Stripped whitespace from tool input")
-                # Add more rule types as needed
-                
-            except Exception as e:
-                self._vprint(f"[DASEIN][APPLY] Error applying tool rule: {e}")
-                continue
-        
-        return modified_input
-    
-    def _handle_tool_executor_start(
-        self,
-        serialized: Dict[str, Any],
-        inputs: Dict[str, Any],
-        **kwargs: Any,
-    ) -> None:
-        """Handle tool executor start - bridge from chain_start to tool_start."""
-        self._vprint(f"[DASEIN][CALLBACK] tool_start (from chain_start)")
-        
-        # Extract tool information from inputs
-        tool_name = "unknown"
-        tool_input = ""
-        
-        if isinstance(inputs, dict):
-            if "tool" in inputs:
-                tool_name = inputs["tool"]
-            elif "tool_name" in inputs:
-                tool_name = inputs["tool_name"]
-            
-            if "tool_input" in inputs:
-                tool_input = str(inputs["tool_input"])
-            elif "input" in inputs:
-                tool_input = str(inputs["input"])
-            else:
-                tool_input = str(inputs)
-        else:
-            tool_input = str(inputs)
-        
-        self._vprint(f"[DASEIN][CALLBACK] Tool: {tool_name}")
-        self._vprint(f"[DASEIN][CALLBACK] Input: {tool_input[:100]}...")
-        
-        # Check if we have tool_start rules that cover this tool
-        tool_rules = [rule for rule in self._selected_rules if rule.target_step_type == "tool_start"]
-        covered_rules = [rule for rule in tool_rules if self._rule_covers_tool(rule, tool_name, tool_input)]
-        
-        if covered_rules:
-            self._vprint(f"[DASEIN][APPLY] tool_start: {len(covered_rules)} rules cover this tool call")
-            # Fire micro-turn for rule application
-            modified_input = self._fire_micro_turn_for_tool_rules(covered_rules, tool_name, tool_input)
-        else:
-            self._vprint(f"[DASEIN][APPLY] tool_start: no rules cover this tool call")
-            modified_input = tool_input
-        
-        args_excerpt = self._excerpt(modified_input)
-        
-        step = {
-            "step_type": "tool_start",
-            "tool_name": tool_name,
-            "args_excerpt": args_excerpt,
-            "outcome": "",
-            "ts": datetime.now().isoformat(),
-            "run_id": kwargs.get("run_id"),
-            "parent_run_id": kwargs.get("parent_run_id"),
-        }
-        self._trace.append(step)
-    
-    def _rule_covers_tool(self, rule, tool_name: str, tool_input: str) -> bool:
-        """Check if a rule covers the given tool call."""
-        try:
-            # Check if rule references this tool
-            if hasattr(rule, 'references') and rule.references:
-                if hasattr(rule.references, 'tools') and rule.references.tools:
-                    if tool_name not in rule.references.tools:
-                        return False
-            
-            # Check trigger patterns if they exist
-            if hasattr(rule, 'trigger_pattern') and rule.trigger_pattern:
-                # For now, assume all tool_start rules cover their referenced tools
-                # This can be made more sophisticated later
-                pass
-            
-            return True
-        except Exception as e:
-            self._vprint(f"[DASEIN][COVERAGE] Error checking rule coverage: {e}")
-            return False
-    
-    def _fire_micro_turn_for_tool_rules(self, rules, tool_name: str, tool_input: str) -> str:
-        """Fire a micro-turn LLM call to apply tool rules."""
-        try:
-            # Use the first rule for now (can be extended to handle multiple rules)
-            rule = rules[0]
-            rule_id = getattr(rule, 'id', 'unknown')
-            
-            self._vprint(f"[DASEIN][MICROTURN] rule_id={rule_id} tool={tool_name}")
-            
-            # Create micro-turn prompt
-            micro_turn_prompt = self._create_micro_turn_prompt(rule, tool_name, tool_input)
-            
-            # Fire actual micro-turn LLM call
-            modified_input = self._execute_micro_turn_llm_call(micro_turn_prompt, tool_input)
-            
-            # Store the modified input for retrieval during tool execution
-            input_key = f"{tool_name}:{hash(tool_input)}"
-            _MODIFIED_TOOL_INPUTS[input_key] = modified_input
-            
-            self._vprint(f"[DASEIN][MICROTURN] Applied rule {rule_id}: {str(tool_input)[:50]}... -> {str(modified_input)[:50]}...")
-            
-            return modified_input
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][MICROTURN] Error in micro-turn: {e}")
-            return tool_input
-    
-    def _create_micro_turn_prompt(self, rule, tool_name: str, tool_input: str) -> str:
-        """Create the micro-turn prompt for rule application."""
-        advice = getattr(rule, 'advice', '')
-        return f"""Apply this rule to the tool input:
-
-Rule: {advice}
-Tool: {tool_name}
-Current Input: {tool_input}
-
-Output only the corrected tool input:"""
-    
-    def _execute_micro_turn_llm_call(self, prompt: str, original_input: str) -> str:
-        """Execute the actual micro-turn LLM call."""
-        try:
-            if not self._llm:
-                self._vprint(f"[DASEIN][MICROTURN] No LLM available for micro-turn call")
-                return original_input
-            
-            self._vprint(f"[DASEIN][MICROTURN] Executing micro-turn LLM call")
-            self._vprint(f"[DASEIN][MICROTURN] Prompt: {prompt[:200]}...")
-            
-            # Make the micro-turn LLM call
-            # Create a simple message list for the LLM
-            messages = [{"role": "user", "content": prompt}]
-            
-            # Call the LLM
-            response = self._llm.invoke(messages)
-            
-            # Extract the response content
-            if hasattr(response, 'content'):
-                modified_input = response.content.strip()
-            elif isinstance(response, str):
-                modified_input = response.strip()
-            else:
-                modified_input = str(response).strip()
-            
-            self._vprint(f"[DASEIN][MICROTURN] LLM response: {modified_input[:100]}...")
-            
-            # 🚨 CRITICAL: Parse JSON responses with markdown fences
-            if modified_input.startswith('```json') or modified_input.startswith('```'):
-                try:
-                    # Extract JSON from markdown fences
-                    import re
-                    import json
-                    json_match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', modified_input, re.DOTALL)
-                    if json_match:
-                        json_str = json_match.group(1)
-                        parsed_json = json.loads(json_str)
-                        # Convert back to the expected format
-                        if isinstance(parsed_json, dict) and 'name' in parsed_json and 'args' in parsed_json:
-                            modified_input = parsed_json
-                            self._vprint(f"[DASEIN][MICROTURN] Parsed JSON from markdown fences: {parsed_json}")
-                        else:
-                            self._vprint(f"[DASEIN][MICROTURN] JSON doesn't have expected structure, using as-is")
-                    else:
-                        self._vprint(f"[DASEIN][MICROTURN] Could not extract JSON from markdown fences")
-                except Exception as e:
-                    self._vprint(f"[DASEIN][MICROTURN] Error parsing JSON: {e}")
-            
-            # Validate the response - only fallback if completely empty
-            if not modified_input:
-                self._vprint(f"[DASEIN][MICROTURN] LLM response empty, using original input")
-                return original_input
-            
-            return modified_input
-            
-        except Exception as e:
-            self._vprint(f"[DASEIN][MICROTURN] Error executing micro-turn LLM call: {e}")
-            return original_input
-
-
-def get_trace() -> List[Dict[str, Any]]:
-    """
-    DEPRECATED: Legacy function for backward compatibility.
-    Get the current trace from active CognateProxy instances.
-    
-    Returns:
-        List of trace step dictionaries (empty if no active traces)
-    """
-    # Try to get trace from active CognateProxy instances
-    try:
-        import gc
-        for obj in gc.get_objects():
-            if hasattr(obj, '_last_run_trace') and obj._last_run_trace:
-                return obj._last_run_trace.copy()
-            if hasattr(obj, '_callback_handler') and hasattr(obj._callback_handler, '_trace'):
-                return obj._callback_handler._trace.copy()
-    except Exception:
-        pass
-    
-    return []  # Return empty list if no trace found
-
-
-def get_modified_tool_input(tool_name: str, original_input: str) -> str:
-    """
-    Get the modified tool input if it exists.
-    
-    Args:
-        tool_name: Name of the tool
-        original_input: Original tool input
-        
-    Returns:
-        Modified tool input if available, otherwise original input
-    """
-    input_key = f"{tool_name}:{hash(original_input)}"
-    return _MODIFIED_TOOL_INPUTS.get(input_key, original_input)
-
-
-def clear_modified_tool_inputs():
-    """Clear all modified tool inputs."""
-    global _MODIFIED_TOOL_INPUTS
-    _MODIFIED_TOOL_INPUTS.clear()
-
-
-def clear_trace() -> None:
-    """
-    DEPRECATED: Legacy function for backward compatibility.
-    Clear traces in active CognateProxy instances.
-    """
-    # Try to clear traces in active CognateProxy instances
-    try:
-        import gc
-        for obj in gc.get_objects():
-            if hasattr(obj, '_callback_handler') and hasattr(obj._callback_handler, 'reset_run_state'):
-                obj._callback_handler.reset_run_state()
-    except Exception:
-        pass  # Ignore if not available
-
-
-def print_trace(max_chars: int = 240, only: tuple[str, ...] | None = None, suppress: tuple[str, ...] = ("chain_end",), show_tree: bool = True, show_summary: bool = True) -> None:
-    """
-    Print a compact fixed-width table of the trace with tree-like view and filtering.
-    
-    Args:
-        max_chars: Maximum characters per line (default 240)
-        only: Filter by step_type if provided (e.g., ("llm_start", "llm_end"))
-        suppress: Suppress any step_type in this tuple (default: ("chain_end",))
-        show_tree: If True, left-pad args_excerpt by 2*depth spaces for tree-like view
-        show_summary: If True, show step_type counts and deduped rows summary
-    """
-    # Try to get trace from active CognateProxy instances
-    trace = None
-    try:
-        # Import here to avoid circular imports
-        from dasein.api import _global_cognate_proxy
-        if _global_cognate_proxy and hasattr(_global_cognate_proxy, '_wrapped_llm') and _global_cognate_proxy._wrapped_llm:
-            trace = _global_cognate_proxy._wrapped_llm.get_trace()
-    except:
-        pass
-    
-    if not trace:
-        trace = get_trace()  # Use the updated get_trace() function
-    
-    # If global trace is empty, try to get it from the last completed run
-    if not trace:
-        # Try to get trace from any active CognateProxy instances
-        try:
-            import gc
-            for obj in gc.get_objects():
-                # Look for CognateProxy instances with captured traces
-                if hasattr(obj, '_last_run_trace') and obj._last_run_trace:
-                    trace = obj._last_run_trace
-                    print(f"[DASEIN][TRACE] Retrieved trace from CognateProxy: {len(trace)} steps")
-                    break
-                # Fallback: try callback handler
-                elif hasattr(obj, '_callback_handler') and hasattr(obj._callback_handler, 'get_trace'):
-                    potential_trace = obj._callback_handler.get_trace()
-                    if potential_trace:
-                        trace = potential_trace
-                        print(f"[DASEIN][TRACE] Retrieved trace from callback handler: {len(trace)} steps")
-                        break
-        except Exception as e:
-            pass
-    
-    if not trace:
-        print("No trace data available.")
-        return
-    
-    # Print execution state if available
-    try:
-        from dasein.api import _global_cognate_proxy
-        if _global_cognate_proxy and hasattr(_global_cognate_proxy, '_callback_handler'):
-            handler = _global_cognate_proxy._callback_handler
-            if hasattr(handler, '_function_calls_made') and handler._function_calls_made:
-                print("\n" + "=" * 80)
-                print("EXECUTION STATE (Functions Called During Run):")
-                print("=" * 80)
-                for func_name in sorted(handler._function_calls_made.keys()):
-                    calls = handler._function_calls_made[func_name]
-                    count = len(calls)
-                    print(f"  • {func_name}: called {count}x")
-                    # Hybrid window: show all if ≤5, else show most recent 3 (matches injection logic)
-                    if count <= 5:
-                        # Show all calls
-                        for call in calls:
-                            preview = call.get('preview', '(no preview)')
-                            if len(preview) > 80:
-                                preview = preview[:80] + '...'
-                            print(f"      [step {call['step']}] {preview}")
-                    else:
-                        # Show recent 3
-                        print(f"      ... (showing most recent 3 of {count}):")
-                        for call in calls[-3:]:
-                            preview = call.get('preview', '(no preview)')
-                            if len(preview) > 80:
-                                preview = preview[:80] + '...'
-                            print(f"      [step {call['step']}] {preview}")
-                print("=" * 80 + "\n")
-    except Exception as e:
-        pass  # Silently skip if state not available
-    
-    # Filter by step_type if only is provided
-    filtered_trace = trace
-    if only:
-        filtered_trace = [step for step in trace if step.get("step_type") in only]
-    
-    # Suppress any step_type in suppress tuple
-    if suppress:
-        filtered_trace = [step for step in filtered_trace if step.get("step_type") not in suppress]
-    
-    if not filtered_trace:
-        print("No trace data matching filter criteria.")
-        return
-    
-    # Build depth map from parent_run_id
-    depth_map = {}
-    for step in filtered_trace:
-        run_id = step.get("run_id")
-        parent_run_id = step.get("parent_run_id")
-        
-        if run_id is None or parent_run_id is None or parent_run_id not in depth_map:
-            depth_map[run_id] = 0
-        else:
-            depth_map[run_id] = depth_map[parent_run_id] + 1
-    
-    # Calculate column widths based on max_chars
-    # Reserve space for: # (3), step_type (15), tool_name (25), separators (6)
-    available_width = max_chars - 3 - 15 - 25 - 6
-    excerpt_width = available_width // 2
-    outcome_width = available_width - excerpt_width
-    
-    # Print header
-    print(f"{'#':<3} {'step_type':<15} {'tool_name':<25} {'args_excerpt':<{excerpt_width}} {'outcome':<{outcome_width}}")
-    print("-" * max_chars)
-    
-    # Print each step
-    for i, step in enumerate(filtered_trace, 1):
-        step_type = step.get("step_type", "")[:15]
-        tool_name = str(step.get("tool_name", ""))[:25]
-        args_excerpt = step.get("args_excerpt", "")
-        outcome = step.get("outcome", "")
-        
-        # Apply tree indentation if show_tree is True
-        if show_tree:
-            run_id = step.get("run_id")
-            depth = depth_map.get(run_id, 0)
-            args_excerpt = "  " * depth + args_excerpt
-        
-        # Truncate to fit column widths
-        args_excerpt = args_excerpt[:excerpt_width]
-        outcome = outcome[:outcome_width]
-        
-        print(f"{i:<3} {step_type:<15} {tool_name:<25} {args_excerpt:<{excerpt_width}} {outcome:<{outcome_width}}")
-    
-    # Show summary if requested
-    if show_summary:
-        print("\n" + "=" * max_chars)
-        
-        # Count steps by step_type
-        step_counts = {}
-        for step in filtered_trace:
-            step_type = step.get("step_type", "unknown")
-            step_counts[step_type] = step_counts.get(step_type, 0) + 1
-        
-        print("Step counts:")
-        for step_type, count in sorted(step_counts.items()):
-            print(f"  {step_type}: {count}")
-        
-        # Add compact function call summary
-        try:
-            from dasein.api import _global_cognate_proxy
-            if _global_cognate_proxy and hasattr(_global_cognate_proxy, '_callback_handler'):
-                handler = _global_cognate_proxy._callback_handler
-                if hasattr(handler, '_function_calls_made') and handler._function_calls_made:
-                    print("\nFunction calls:")
-                    for func_name in sorted(handler._function_calls_made.keys()):
-                        count = len(handler._function_calls_made[func_name])
-                        print(f"  {func_name}: {count}")
-        except Exception:
-            pass
-        
-        # Count deduped rows skipped (steps that were filtered out)
-        total_steps = len(trace)
-        shown_steps = len(filtered_trace)
-        skipped_steps = total_steps - shown_steps
-        
-        if skipped_steps > 0:
-            print(f"Deduped rows skipped: {skipped_steps}")
+"""
+Trace capture functionality for Dasein.
+"""
+
+# Suppress third-party warnings triggered by pipecleaner dependencies
+import warnings
+warnings.filterwarnings('ignore', category=FutureWarning, message='.*torch.distributed.reduce_op.*')
+warnings.filterwarnings('ignore', category=DeprecationWarning, message='.*Importing chat models from langchain.*')
+
+import hashlib
+from typing import Any, Dict, List, Optional, Union
+from datetime import datetime
+from langchain_core.callbacks.base import BaseCallbackHandler
+from langchain_core.callbacks.manager import CallbackManagerForToolRun
+from langchain_core.tools import BaseTool
+
+
+# ============================================================================
+# VERBOSE LOGGING HELPER
+# ============================================================================
+
+def _vprint(message: str, verbose: bool = False, force: bool = False):
+    """
+    Helper function for verbose printing.
+    
+    Args:
+        message: Message to print
+        verbose: Whether verbose mode is enabled
+        force: If True, always print regardless of verbose setting
+    """
+    if force or verbose:
+        print(message)
+
+
+# DEPRECATED: Global trace store removed for thread-safety
+# Traces are now stored instance-level in DaseinCallbackHandler._trace
+# _TRACE: List[Dict[str, Any]] = []
+
+# Hook cache for agent fingerprinting
+_HOOK_CACHE: Dict[str, Any] = {}
+
+# Store for modified tool inputs
+_MODIFIED_TOOL_INPUTS: Dict[str, str] = {}
+
+
+class DaseinToolWrapper(BaseTool):
+    """Wrapper for tools that applies micro-turn modifications."""
+    
+    name: str = ""
+    description: str = ""
+    original_tool: Any = None
+    callback_handler: Any = None
+    
+    def __init__(self, original_tool, callback_handler=None, verbose: bool = False):
+        super().__init__(
+            name=original_tool.name,
+            description=original_tool.description
+        )
+        self.original_tool = original_tool
+        self.callback_handler = callback_handler
+        self._verbose = verbose
+    
+    def _vprint(self, message: str, force: bool = False):
+        """Helper for verbose printing."""
+        _vprint(message, self._verbose, force)
+    
+    def _run(self, *args, **kwargs):
+        """Run the tool with micro-turn injection at execution level."""
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] _run called for {self.name} - VERSION 2.0")
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] Args: {args}")
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] Kwargs: {kwargs}")
+        
+        try:
+            # Get the original input
+            original_input = args[0] if args else ""
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Original input: {original_input[:100]}...")
+            
+            # Apply micro-turn injection if we have rules
+            modified_input = self._apply_micro_turn_injection(str(original_input))
+            
+            if modified_input != original_input:
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
+                # Use modified input
+                result = self.original_tool._run(modified_input, *args[1:], **kwargs)
+            else:
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] No micro-turn injection applied for {self.name}")
+                # Use original input
+                result = self.original_tool._run(*args, **kwargs)
+            
+            # Capture the tool output in the trace
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] About to capture tool output for {self.name}")
+            self._capture_tool_output(self.name, args, kwargs, result)
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Finished capturing tool output for {self.name}")
+            
+            return result
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Exception in _run: {e}")
+            import traceback
+            traceback.print_exc()
+            # Still try to call the original tool
+            result = self.original_tool._run(*args, **kwargs)
+            return result
+    
+    def invoke(self, input_data, config=None, **kwargs):
+        """Invoke the tool with micro-turn injection."""
+        # Get the original input
+        original_input = str(input_data)
+        
+        # Apply micro-turn injection if we have rules
+        modified_input = self._apply_micro_turn_injection(original_input)
+        
+        if modified_input != original_input:
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
+            # Use modified input
+            return self.original_tool.invoke(modified_input, config, **kwargs)
+        else:
+            # Use original input
+            return self.original_tool.invoke(input_data, config, **kwargs)
+    
+    async def _arun(self, *args, **kwargs):
+        """Async run the tool with micro-turn injection at execution level."""
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] _arun called for {self.name} - ASYNC VERSION")
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] Args: {args}")
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] Kwargs: {kwargs}")
+        
+        try:
+            # Get the original input
+            original_input = args[0] if args else ""
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Original input: {original_input[:100]}...")
+            
+            # Apply micro-turn injection if we have rules
+            modified_input = self._apply_micro_turn_injection(str(original_input))
+            
+            if modified_input != original_input:
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
+                # Use modified input
+                result = await self.original_tool._arun(modified_input, *args[1:], **kwargs)
+            else:
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] No micro-turn injection applied for {self.name}")
+                # Use original input
+                result = await self.original_tool._arun(*args, **kwargs)
+            
+            # Capture the tool output in the trace
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] About to capture tool output for {self.name}")
+            self._capture_tool_output(self.name, args, kwargs, result)
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Finished capturing tool output for {self.name}")
+            
+            return result
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Exception in _arun: {e}")
+            import traceback
+            traceback.print_exc()
+            # Still try to call the original tool
+            result = await self.original_tool._arun(*args, **kwargs)
+            return result
+    
+    async def ainvoke(self, input_data, config=None, **kwargs):
+        """Async invoke the tool with micro-turn injection."""
+        self._vprint(f"[DASEIN][TOOL_WRAPPER] ainvoke called for {self.name} - ASYNC VERSION")
+        
+        # Get the original input
+        original_input = str(input_data)
+        
+        # Apply micro-turn injection if we have rules
+        modified_input = self._apply_micro_turn_injection(original_input)
+        
+        if modified_input != original_input:
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Applied micro-turn injection for {self.name}: {original_input[:50]}... -> {modified_input[:50]}...")
+            # Use modified input
+            result = await self.original_tool.ainvoke(modified_input, config, **kwargs)
+        else:
+            # Use original input
+            result = await self.original_tool.ainvoke(input_data, config, **kwargs)
+        
+        return result
+    
+    def _apply_micro_turn_injection(self, original_input: str) -> str:
+        """Apply micro-turn injection to the tool input."""
+        try:
+            # Check if we have a callback handler with rules and LLM
+            if not self.callback_handler:
+                return original_input
+                
+            # Normalize selected rules into Rule objects (handle (rule, metadata) tuples)
+            normalized_rules = []
+            for rule_meta in getattr(self.callback_handler, "_selected_rules", []) or []:
+                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
+                    rule_obj, _metadata = rule_meta
+                else:
+                    rule_obj = rule_meta
+                normalized_rules.append(rule_obj)
+                
+            # Filter tool_start rules
+            tool_rules = [r for r in normalized_rules if getattr(r, 'target_step_type', '') == "tool_start"]
+            
+            if not tool_rules:
+                self._vprint(f"[DASEIN][MICROTURN] No tool rules selected - skipping micro-turn for {self.name}")
+                return original_input
+                
+            # Check if any rule covers this tool
+            covered_rules = [rule for rule in tool_rules 
+                             if self._rule_covers_tool(rule, self.name, original_input)]
+            
+            if not covered_rules:
+                return original_input
+                
+            # Fire micro-turn LLM call (use first matching rule)
+            rule = covered_rules[0]
+            self._vprint(f"[DASEIN][MICROTURN] rule_id={rule.id} tool={self.name}")
+            
+            # Create micro-turn prompt
+            micro_turn_prompt = self._create_micro_turn_prompt(rule, self.name, original_input)
+            
+            # Execute micro-turn LLM call
+            modified_input = self._execute_micro_turn_llm_call(micro_turn_prompt, original_input)
+            
+            self._vprint(f"[DASEIN][MICROTURN] Applied rule {rule.id}: {str(original_input)[:50]}... -> {str(modified_input)[:50]}...")
+            return modified_input
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][MICROTURN] Error in micro-turn injection: {e}")
+            return original_input
+    
+    def _rule_covers_tool(self, rule, tool_name: str, tool_input: str) -> bool:
+        """Check if a rule covers this tool call."""
+        if not hasattr(rule, 'references') or not rule.references:
+            return False
+            
+        # Check if the rule references this tool
+        tools = rule.references.get('tools', [])
+        return tool_name in tools
+    
+    def _create_micro_turn_prompt(self, rule, tool_name: str, tool_input: str) -> str:
+        """Create the prompt for the micro-turn LLM call."""
+        return f"""You are applying a rule to fix a tool input.
+
+Rule: {rule.advice_text}
+
+Tool: {tool_name}
+Current Input: {tool_input}
+
+Apply the rule to fix the input. Return only the corrected input, nothing else."""
+
+    def _execute_micro_turn_llm_call(self, prompt: str, original_input: str) -> str:
+        """Execute the actual micro-turn LLM call."""
+        try:
+            if not self.callback_handler or not self.callback_handler._llm:
+                self._vprint(f"[DASEIN][MICROTURN] No LLM available for micro-turn call")
+                return original_input
+
+            self._vprint(f"[DASEIN][MICROTURN] Executing micro-turn LLM call")
+            self._vprint(f"[DASEIN][MICROTURN] Prompt: {prompt[:200]}...")
+
+            # Make the micro-turn LLM call
+            messages = [{"role": "user", "content": prompt}]
+            response = self.callback_handler._llm.invoke(messages)
+
+            # Extract the response content
+            if hasattr(response, 'content'):
+                modified_input = response.content.strip()
+            elif isinstance(response, str):
+                modified_input = response.strip()
+            else:
+                modified_input = str(response).strip()
+
+            self._vprint(f"[DASEIN][MICROTURN] LLM response: {modified_input[:100]}...")
+
+            # 🚨 CRITICAL: Parse JSON responses with markdown fences
+            if modified_input.startswith('```json') or modified_input.startswith('```'):
+                try:
+                    # Extract JSON from markdown fences
+                    import re
+                    import json
+                    json_match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', modified_input, re.DOTALL)
+                    if json_match:
+                        json_str = json_match.group(1)
+                        parsed_json = json.loads(json_str)
+                        # Convert back to the expected format
+                        if isinstance(parsed_json, dict) and 'name' in parsed_json and 'args' in parsed_json:
+                            modified_input = parsed_json
+                            self._vprint(f"[DASEIN][MICROTURN] Parsed JSON from markdown fences: {parsed_json}")
+                        else:
+                            self._vprint(f"[DASEIN][MICROTURN] JSON doesn't have expected structure, using as-is")
+                    else:
+                        self._vprint(f"[DASEIN][MICROTURN] Could not extract JSON from markdown fences")
+                except Exception as e:
+                    self._vprint(f"[DASEIN][MICROTURN] Error parsing JSON: {e}")
+
+            # Validate the response - only fallback if completely empty
+            if not modified_input:
+                self._vprint(f"[DASEIN][MICROTURN] LLM response empty, using original input")
+                return original_input
+
+            return modified_input
+
+        except Exception as e:
+            self._vprint(f"[DASEIN][MICROTURN] Error executing micro-turn LLM call: {e}")
+            return original_input
+    
+    def _capture_tool_output(self, tool_name, args, kwargs, result):
+        """Capture tool output in the trace."""
+        try:
+            # Create args excerpt
+            args_str = str(args) if args else ""
+            if len(args_str) > 1000:
+                args_str = args_str[:1000] + "..."
+            
+            # Create result excerpt (with 10k limit)
+            result_str = str(result) if result else ""
+            if len(result_str) > 10000:
+                result_str = result_str[:10000] + "..."
+            
+            # Add tool_end step to trace
+            step = {
+                "step_type": "tool_end",
+                "tool_name": tool_name,
+                "args_excerpt": args_str,
+                "outcome": result_str,
+                "ts": datetime.now().isoformat(),
+                "run_id": f"tool_{id(self)}_{datetime.now().timestamp()}",
+                "parent_run_id": None,
+            }
+            
+            # Add to LLM wrapper's trace if available
+            if self.callback_handler and hasattr(self.callback_handler, '_llm') and self.callback_handler._llm:
+                if hasattr(self.callback_handler._llm, '_trace'):
+                    self.callback_handler._llm._trace.append(step)
+                    self._vprint(f"[DASEIN][TOOL_WRAPPER] Added to LLM wrapper trace")
+                else:
+                    self._vprint(f"[DASEIN][TOOL_WRAPPER] LLM wrapper has no _trace attribute")
+            else:
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] No LLM wrapper available")
+            
+            # Also add to callback handler's trace if it has one
+            if self.callback_handler and hasattr(self.callback_handler, '_trace'):
+                self.callback_handler._trace.append(step)
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] Added to callback handler trace")
+            
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Captured tool output for {tool_name}")
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Output length: {len(result_str)} chars")
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] First 200 chars: {result_str[:200]}")
+            if self.callback_handler and hasattr(self.callback_handler, '_trace'):
+                self._vprint(f"[DASEIN][TOOL_WRAPPER] Callback handler trace length after capture: {len(self.callback_handler._trace)}")
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][TOOL_WRAPPER] Error capturing tool output: {e}")
+
+
+class DaseinCallbackHandler(BaseCallbackHandler):
+    """
+    Callback handler that captures step-by-step traces and implements rule injection.
+    """
+    
+    def __init__(self, weights=None, llm=None, is_langgraph=False, coordinator_node=None, planning_nodes=None, verbose: bool = False, agent=None, extract_tools_fn=None):
+        super().__init__()
+        self._weights = weights
+        self._selected_rules = []  # Rules selected for this run
+        self._injection_guard = set()  # Prevent duplicate injections
+        self._last_modified_prompts = []  # Store modified prompts for LLM wrapper
+        self._llm = llm  # Store reference to LLM for micro-turn calls
+        self._tool_name_by_run_id = {}  # Track tool names by run_id
+        self._discovered_tools = set()  # Track tools discovered during execution
+        self._wrapped_dynamic_tools = {}  # Cache of wrapped dynamic tools
+        self._is_langgraph = is_langgraph  # Flag to skip planning rule injection for LangGraph
+        self._run_number = 1  # Track which run this is (for microturn testing)
+        self._coordinator_node = coordinator_node  # Coordinator node (for future targeted injection)
+        self._planning_nodes = planning_nodes if planning_nodes else set()  # Planning-capable nodes (including subgraph children)
+        self._current_chain_node = None  # Track current LangGraph node
+        self._agent_was_recreated = False  # Track if agent was successfully recreated
+        self._function_calls_made = {}  # Track function calls: {function_name: [{'step': N, 'ts': timestamp}]}
+        self._trace = []  # Instance-level trace storage (not global) for thread-safety
+        self._verbose = verbose
+        self._start_times = {}  # Track start times for duration calculation: {step_index: datetime}
+        self._agent = agent  # CRITICAL: Reference to agent for runtime tool extraction
+        self._extract_tools_fn = extract_tools_fn  # Function to extract tools
+        self._runtime_tools_extracted = False  # Flag to extract tools only once during execution
+        self._compiled_tools_metadata = []  # Store extracted tools
+        self._pipecleaner_embedding_model = None  # Cache embedding model for this run
+        self._current_tool_name = None  # Track currently executing tool for hotpath deduplication
+        
+        # Generate stable run_id for corpus deduplication
+        import uuid
+        self.run_id = str(uuid.uuid4())
+        
+        self._vprint(f"[DASEIN][CALLBACK] Initialized callback handler (LangGraph: {is_langgraph}, run_id: {self.run_id[:8]})")
+        if coordinator_node:
+            self._vprint(f"[DASEIN][CALLBACK] Coordinator: {coordinator_node}")
+        if planning_nodes:
+            self._vprint(f"[DASEIN][CALLBACK] Planning nodes: {planning_nodes}")
+        self._vprint(f"[DASEIN][CALLBACK] Dynamic tool detection enabled (tools discovered at runtime)")
+    
+    def _vprint(self, message: str, force: bool = False):
+        """Helper for verbose printing."""
+        _vprint(message, self._verbose, force)
+    
+    def reset_run_state(self):
+        """Reset state that should be cleared between runs."""
+        self._function_calls_made = {}
+        self._injection_guard = set()
+        self._trace = []  # Clear instance trace
+        self._start_times = {}  # Clear start times
+        self._run_number = getattr(self, '_run_number', 1) + 1  # Increment run number
+        self._vprint(f"[DASEIN][CALLBACK] Reset run state (trace, function calls, injection guard, and start times cleared) - now on RUN {self._run_number}")
+    
+    def get_compiled_tools_summary(self):
+        """Return 1-line summary of extracted tools."""
+        if not self._compiled_tools_metadata:
+            return None
+        # Group by node
+        by_node = {}
+        for tool in self._compiled_tools_metadata:
+            node = tool.get('node', 'unknown')
+            if node not in by_node:
+                by_node[node] = []
+            by_node[node].append(tool['name'])
+        # Format as: node1:[tool1,tool2] node2:[tool3]
+        parts = [f"{node}:[{','.join(tools)}]" for node, tools in sorted(by_node.items())]
+        return f"{len(self._compiled_tools_metadata)} tools extracted: {' '.join(parts)}"
+    
+    def _patch_tools_for_node(self, node_name: str):
+        """
+        Patch tool objects for a specific node when they're discovered at runtime.
+        
+        Called from on_llm_start when tools are detected for a node.
+        """
+        try:
+            print(f"\n{'='*70}")
+            print(f"[DASEIN][TOOL_PATCH] 🔧 Patching tools for node: {node_name}")
+            print(f"{'='*70}")
+            
+            from .wrappers import patch_tool_instance
+            
+            # Track patched tools to avoid double-patching
+            if not hasattr(self, '_patched_tools'):
+                self._patched_tools = set()
+                print(f"[DASEIN][TOOL_PATCH] Initialized patched tools tracker")
+            
+            # Find the actual tool objects for this node in the agent graph
+            print(f"[DASEIN][TOOL_PATCH] Searching for tool objects in node '{node_name}'...")
+            tool_objects = self._find_tool_objects_for_node(node_name)
+            
+            if not tool_objects:
+                print(f"[DASEIN][TOOL_PATCH] ⚠️  No tool objects found for node '{node_name}'")
+                print(f"{'='*70}\n")
+                return
+            
+            print(f"[DASEIN][TOOL_PATCH] ✓ Found {len(tool_objects)} tool object(s)")
+            
+            # Patch each tool
+            patched_count = 0
+            for i, tool_obj in enumerate(tool_objects, 1):
+                tool_name = getattr(tool_obj, 'name', 'unknown')
+                tool_type = type(tool_obj).__name__
+                tool_id = f"{node_name}:{tool_name}"
+                
+                print(f"[DASEIN][TOOL_PATCH] [{i}/{len(tool_objects)}] Tool: '{tool_name}' (type: {tool_type})")
+                
+                if tool_id in self._patched_tools:
+                    print(f"[DASEIN][TOOL_PATCH]   ⏭️  Already patched, skipping")
+                else:
+                    print(f"[DASEIN][TOOL_PATCH]   🔨 Patching...")
+                    if patch_tool_instance(tool_obj, self):
+                        self._patched_tools.add(tool_id)
+                        patched_count += 1
+                        print(f"[DASEIN][TOOL_PATCH]   ✅ Successfully patched '{tool_name}'")
+                    else:
+                        print(f"[DASEIN][TOOL_PATCH]   ❌ Failed to patch '{tool_name}'")
+            
+            print(f"[DASEIN][TOOL_PATCH] Summary: Patched {patched_count}/{len(tool_objects)} tools")
+            print(f"[DASEIN][TOOL_PATCH] Total tools patched so far: {len(self._patched_tools)}")
+            print(f"{'='*70}\n")
+        
+        except Exception as e:
+            print(f"[DASEIN][TOOL_PATCH] ❌ ERROR patching tools for node {node_name}: {e}")
+            import traceback
+            traceback.print_exc()
+            print(f"{'='*70}\n")
+    
+    def _search_node_recursively(self, node_name: str, nodes: dict, depth: int = 0) -> list:
+        """Recursively search for a node by name in graphs and subgraphs."""
+        indent = "  " * depth
+        tool_objects = []
+        
+        for parent_name, parent_node in nodes.items():
+            if parent_name.startswith('__'):
+                continue
+            
+            print(f"[DASEIN][TOOL_PATCH]{indent}   Checking node: {parent_name}")
+            print(f"[DASEIN][TOOL_PATCH]{indent}     Node type: {type(parent_node).__name__}")
+            print(f"[DASEIN][TOOL_PATCH]{indent}     Has .data: {hasattr(parent_node, 'data')}")
+            if hasattr(parent_node, 'data'):
+                print(f"[DASEIN][TOOL_PATCH]{indent}     .data type: {type(parent_node.data).__name__}")
+                print(f"[DASEIN][TOOL_PATCH]{indent}     .data has .nodes: {hasattr(parent_node.data, 'nodes')}")
+            
+            # Check if this parent has a subgraph
+            if hasattr(parent_node, 'data') and hasattr(parent_node.data, 'nodes'):
+                print(f"[DASEIN][TOOL_PATCH]{indent}     Has subgraph!")
+                try:
+                    subgraph = parent_node.data.get_graph()
+                    sub_nodes = subgraph.nodes
+                    print(f"[DASEIN][TOOL_PATCH]{indent}     Subgraph nodes: {list(sub_nodes.keys())}")
+                    
+                    # Check if target node is in this subgraph
+                    if node_name in sub_nodes:
+                        print(f"[DASEIN][TOOL_PATCH]{indent}     ✓ Found '{node_name}' in subgraph!")
+                        target_node = sub_nodes[node_name]
+                        if hasattr(target_node, 'node'):
+                            actual_node = target_node.node
+                            tool_objects = self._extract_tools_from_node_object(actual_node)
+                            if tool_objects:
+                                return tool_objects
+                    
+                    # Not found here, recurse deeper into this subgraph
+                    print(f"[DASEIN][TOOL_PATCH]{indent}     Recursing into subgraph nodes...")
+                    tool_objects = self._search_node_recursively(node_name, sub_nodes, depth + 1)
+                    if tool_objects:
+                        return tool_objects
+                        
+                except Exception as e:
+                    print(f"[DASEIN][TOOL_PATCH]{indent}     Error: {e}")
+                    import traceback
+                    traceback.print_exc()
+            else:
+                print(f"[DASEIN][TOOL_PATCH]{indent}     No subgraph")
+        
+        return tool_objects
+    
+    def _find_tool_objects_for_node(self, node_name: str):
+        """Find actual Python tool objects for a given node."""
+        tool_objects = []
+        
+        try:
+            if not hasattr(self._agent, 'get_graph'):
+                print(f"[DASEIN][TOOL_PATCH]   Agent has no get_graph method")
+                return tool_objects
+            
+            graph = self._agent.get_graph()
+            nodes = graph.nodes
+            node_names = list(nodes.keys())
+            print(f"[DASEIN][TOOL_PATCH]   Graph has {len(nodes)} nodes: {node_names}")
+            
+            # Check if node_name contains a dot (subgraph notation like "research_supervisor.ConductResearch")
+            if '.' in node_name:
+                print(f"[DASEIN][TOOL_PATCH]   Node is subgraph: {node_name}")
+                parent_name, sub_name = node_name.split('.', 1)
+                parent_node = nodes.get(parent_name)
+                
+                if parent_node and hasattr(parent_node, 'data'):
+                    print(f"[DASEIN][TOOL_PATCH]   Found parent node, getting subgraph...")
+                    subgraph = parent_node.data.get_graph()
+                    sub_nodes = subgraph.nodes
+                    print(f"[DASEIN][TOOL_PATCH]   Subgraph has {len(sub_nodes)} nodes")
+                    target_node = sub_nodes.get(sub_name)
+                    
+                    if target_node and hasattr(target_node, 'node'):
+                        print(f"[DASEIN][TOOL_PATCH]   Found target subnode, extracting tools...")
+                        actual_node = target_node.node
+                        tool_objects = self._extract_tools_from_node_object(actual_node)
+                    else:
+                        print(f"[DASEIN][TOOL_PATCH]   ⚠️  Subnode not found or has no .node attribute")
+                else:
+                    print(f"[DASEIN][TOOL_PATCH]   ⚠️  Parent node not found or has no .data attribute")
+            else:
+                # Top-level node
+                print(f"[DASEIN][TOOL_PATCH]   Node is top-level: {node_name}")
+                target_node = nodes.get(node_name)
+                
+                if target_node:
+                    print(f"[DASEIN][TOOL_PATCH]   Found node, checking for .node attribute...")
+                    if hasattr(target_node, 'node'):
+                        print(f"[DASEIN][TOOL_PATCH]   Has .node attribute, extracting tools...")
+                        actual_node = target_node.node
+                        tool_objects = self._extract_tools_from_node_object(actual_node)
+                    else:
+                        print(f"[DASEIN][TOOL_PATCH]   ⚠️  Node has no .node attribute")
+                else:
+                    # Not found as top-level, search in subgraphs
+                    print(f"[DASEIN][TOOL_PATCH]   ⚠️  Node '{node_name}' not found in top-level graph")
+                    print(f"[DASEIN][TOOL_PATCH]   Searching in subgraphs...")
+                    
+                    # Recursively search all subgraphs
+                    tool_objects = self._search_node_recursively(node_name, nodes)
+                    
+                    if not tool_objects:
+                        print(f"[DASEIN][TOOL_PATCH]   ⚠️  Node '{node_name}' not found in any subgraph")
+        
+        except Exception as e:
+            print(f"[DASEIN][TOOL_PATCH]   ❌ Exception while finding tools: {e}")
+            import traceback
+            traceback.print_exc()
+        
+        return tool_objects
+    
+    def _extract_tools_from_node_object(self, node_obj):
+        """Extract tool objects from a node object."""
+        tools = []
+        
+        print(f"[DASEIN][TOOL_PATCH]     Checking node_obj type: {type(node_obj).__name__}")
+        
+        # Check tools_by_name
+        if hasattr(node_obj, 'tools_by_name'):
+            print(f"[DASEIN][TOOL_PATCH]     ✓ Has tools_by_name with {len(node_obj.tools_by_name)} tools")
+            tools.extend(node_obj.tools_by_name.values())
+        else:
+            print(f"[DASEIN][TOOL_PATCH]     ✗ No tools_by_name")
+        
+        # Check runnable.tools
+        if hasattr(node_obj, 'runnable'):
+            print(f"[DASEIN][TOOL_PATCH]     ✓ Has runnable")
+            if hasattr(node_obj.runnable, 'tools'):
+                print(f"[DASEIN][TOOL_PATCH]       ✓ runnable.tools exists")
+                runnable_tools = node_obj.runnable.tools
+                if callable(runnable_tools):
+                    print(f"[DASEIN][TOOL_PATCH]       runnable.tools is callable, calling...")
+                    try:
+                        runnable_tools = runnable_tools()
+                        print(f"[DASEIN][TOOL_PATCH]       Got {len(runnable_tools) if isinstance(runnable_tools, list) else 1} tool(s)")
+                    except Exception as e:
+                        print(f"[DASEIN][TOOL_PATCH]       ❌ Failed to call: {e}")
+                if isinstance(runnable_tools, list):
+                    tools.extend(runnable_tools)
+                elif runnable_tools:
+                    tools.append(runnable_tools)
+            else:
+                print(f"[DASEIN][TOOL_PATCH]       ✗ No runnable.tools")
+        else:
+            print(f"[DASEIN][TOOL_PATCH]     ✗ No runnable")
+        
+        # Check bound.tools
+        if hasattr(node_obj, 'bound'):
+            print(f"[DASEIN][TOOL_PATCH]     ✓ Has bound")
+            if hasattr(node_obj.bound, 'tools'):
+                print(f"[DASEIN][TOOL_PATCH]       ✓ bound.tools exists")
+                bound_tools = node_obj.bound.tools
+                if isinstance(bound_tools, list):
+                    print(f"[DASEIN][TOOL_PATCH]       Got {len(bound_tools)} tool(s)")
+                    tools.extend(bound_tools)
+                elif bound_tools:
+                    print(f"[DASEIN][TOOL_PATCH]       Got 1 tool")
+                    tools.append(bound_tools)
+            else:
+                print(f"[DASEIN][TOOL_PATCH]       ✗ No bound.tools")
+        else:
+            print(f"[DASEIN][TOOL_PATCH]     ✗ No bound")
+        
+        # Check steps
+        if hasattr(node_obj, 'steps'):
+            print(f"[DASEIN][TOOL_PATCH]     ✓ Has steps ({len(node_obj.steps)})")
+            for i, step in enumerate(node_obj.steps):
+                if hasattr(step, 'tools_by_name'):
+                    print(f"[DASEIN][TOOL_PATCH]       ✓ Step {i} has tools_by_name with {len(step.tools_by_name)} tools")
+                    tools.extend(step.tools_by_name.values())
+                    break
+        else:
+            print(f"[DASEIN][TOOL_PATCH]     ✗ No steps")
+        
+        print(f"[DASEIN][TOOL_PATCH]     Total tools extracted: {len(tools)}")
+        
+        return tools
+    
+    def on_llm_start(
+        self,
+        serialized: Dict[str, Any],
+        prompts: List[str],
+        *,
+        run_id: str = None,
+        parent_run_id: Optional[str] = None,
+        tags: Optional[List[str]] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
+    ) -> None:
+        """Called when an LLM starts running."""
+        model_name = serialized.get("name", "unknown") if serialized else "unknown"
+        
+        # PIPECLEANER: Intercept Summary LLM calls
+        tools_in_call = None
+        if 'invocation_params' in kwargs:
+            tools_in_call = kwargs['invocation_params'].get('tools') or kwargs['invocation_params'].get('functions')
+        
+        if tools_in_call:
+            tool_names = [t.get('name') or t.get('function', {}).get('name', 'unknown') for t in tools_in_call]
+            
+            if 'Summary' in tool_names:
+                # NOTE: Deduplication now happens in the HOT PATH (monkey-patched LLM methods)
+                # This callback is just for tracking, not deduplication
+                pass
+                
+            if False and 'Summary' in tool_names:  # DISABLED: Deduplication moved to hotpath
+                # Check if run-scoped corpus is enabled (has filter search rules)
+                has_filter_rules = False
+                if hasattr(self, '_selected_rules'):
+                    from .pipecleaner import _find_filter_search_rules
+                    filter_rules = _find_filter_search_rules('summary', self._selected_rules)
+                    has_filter_rules = len(filter_rules) > 0
+                
+                if not has_filter_rules:
+                    # Silent fail - no corpus deduplication if no rules
+                    pass
+                else:
+                    # Only print when we actually have rules and will deduplicate
+                    print(f"[CORPUS] 📥 Summary LLM detected with {len(prompts)} prompts")
+                    # Re-entrancy guard: prevent nested calls from corrupting state
+                    from contextvars import ContextVar
+                    if not hasattr(DaseinCallbackHandler, '_in_corpus_processing'):
+                        DaseinCallbackHandler._in_corpus_processing = ContextVar('in_corpus', default=False)
+                        DaseinCallbackHandler._reentrancy_count = 0
+                    
+                    if DaseinCallbackHandler._in_corpus_processing.get():
+                        # Already processing corpus in this call stack, fail-open
+                        DaseinCallbackHandler._reentrancy_count += 1
+                        print(f"[CORPUS] ⚠️  Re-entrancy detected #{DaseinCallbackHandler._reentrancy_count}, skipping nested call")
+                        return
+                    
+                    # Set re-entrancy guard
+                    token = DaseinCallbackHandler._in_corpus_processing.set(True)
+                    
+                    try:
+                        # Get or create run-scoped corpus
+                        from .pipecleaner import get_or_create_corpus
+                        import threading
+                        corpus = get_or_create_corpus(self.run_id, verbose=self._verbose)
+                        
+                        # Module-level lock for atomic snapshot/swap (shared across all instances)
+                        if not hasattr(DaseinCallbackHandler, '_prompts_lock'):
+                            DaseinCallbackHandler._prompts_lock = threading.Lock()
+                        
+                        # STEP 1: Snapshot under lock (atomic read, NEVER iterate live dict)
+                        with DaseinCallbackHandler._prompts_lock:
+                            try:
+                                snapshot = tuple(prompts)  # Immutable snapshot, safe to iterate
+                            except RuntimeError:
+                                print(f"[CORPUS] ⚠️  Skipping (prompts being iterated)")
+                                return
+                    
+                        # STEP 2: Process outside lock (no contention)
+                        cleaned_prompts = []
+                        total_original_chars = 0
+                        total_cleaned_chars = 0
+                        total_original_tokens_est = 0
+                        total_cleaned_tokens_est = 0
+                        
+                        for i, prompt in enumerate(snapshot):
+                            prompt_str = str(prompt)
+                            
+                            # Skip if too short
+                            if len(prompt_str) < 2500:
+                                cleaned_prompts.append(prompt_str)
+                                continue
+                            
+                            # Track original
+                            original_chars = len(prompt_str)
+                            original_tokens_est = original_chars // 4  # Rough estimate: 4 chars/token
+                            total_original_chars += original_chars
+                            total_original_tokens_est += original_tokens_est
+                            
+                            # Split: first 2000 chars (system prompt) + rest (content to dedupe)
+                            system_part = prompt_str[:2000]
+                            content_part = prompt_str[2000:]
+                            
+                            # Generate unique prompt_id
+                            import hashlib
+                            prompt_id = f"p{i}_{hashlib.md5(content_part[:100].encode()).hexdigest()[:8]}"
+                            
+                            # Enqueue into corpus (barrier will handle batching, blocks until ready)
+                            # Call synchronous enqueue (will block until batch is processed, then released sequentially)
+                            deduplicated_content = corpus.enqueue_prompt(prompt_id, content_part)
+                            
+                            # Reassemble
+                            cleaned_prompt = system_part + deduplicated_content
+                            
+                            # Track cleaned
+                            cleaned_chars = len(cleaned_prompt)
+                            cleaned_tokens_est = cleaned_chars // 4
+                            total_cleaned_chars += cleaned_chars
+                            total_cleaned_tokens_est += cleaned_tokens_est
+                            
+                            reduction_pct = 100*(original_chars-cleaned_chars)//original_chars if original_chars > 0 else 0
+                            # Always show reduction results (key metric)
+                            print(f"[🧹 CORPUS] Prompt {prompt_id}: {original_chars} → {cleaned_chars} chars ({reduction_pct}% saved)")
+                            cleaned_prompts.append(cleaned_prompt)
+                        
+                        # Store token delta for later adjustment in on_llm_end
+                        if total_original_tokens_est > 0:
+                            tokens_saved = total_original_tokens_est - total_cleaned_tokens_est
+                            if not hasattr(self, '_corpus_token_savings'):
+                                self._corpus_token_savings = {}
+                            self._corpus_token_savings[run_id] = tokens_saved
+                            print(f"[🔬 TOKEN TRACKING] Pre-prune: {total_original_chars} chars (~{total_original_tokens_est} tokens)")
+                            print(f"[🔬 TOKEN TRACKING] Post-prune: {total_cleaned_chars} chars (~{total_cleaned_tokens_est} tokens)")
+                            print(f"[🔬 TOKEN TRACKING] Estimated savings: ~{tokens_saved} tokens ({100*tokens_saved//total_original_tokens_est if total_original_tokens_est > 0 else 0}%)")
+                            print(f"[🔬 TOKEN TRACKING] Stored savings for run_id={str(run_id)[:8]} to adjust on_llm_end")
+                        
+                        # STEP 3: Atomic swap under lock (copy-on-write, no in-place mutation)
+                        print(f"[🔬 CORPUS DEBUG] About to swap prompts - have {len(cleaned_prompts)} cleaned prompts")
+                        with DaseinCallbackHandler._prompts_lock:
+                            try:
+                                print(f"[🔬 CORPUS DEBUG] Inside lock, swapping...")
+                                # Atomic slice assignment (replaces entire contents in one operation)
+                                prompts[:] = cleaned_prompts
+                                # CRITICAL: Update _last_modified_prompts so DaseinLLMWrapper sees deduplicated prompts
+                                self._last_modified_prompts = cleaned_prompts
+                                print(f"[🔬 CORPUS] ✅ Updated _last_modified_prompts with {len(cleaned_prompts)} deduplicated prompts")
+                            except RuntimeError as e:
+                                print(f"[CORPUS] ⚠️  Could not swap prompts (framework collision): {e}")
+                            except Exception as e:
+                                print(f"[CORPUS] ⚠️  Unexpected error swapping: {e}")
+                                import traceback
+                                traceback.print_exc()
+                    finally:
+                        # Always reset re-entrancy guard
+                        DaseinCallbackHandler._in_corpus_processing.reset(token)
+        
+        # DEBUG: Print run context
+        # print(f"🔧 [LLM_START DEBUG] run_id: {run_id}, parent: {parent_run_id}")
+        
+        # 🎯 CRITICAL: Track current node from kwargs metadata FIRST (needed for tool extraction)
+        if self._is_langgraph and 'metadata' in kwargs and isinstance(kwargs['metadata'], dict):
+            if 'langgraph_node' in kwargs['metadata']:
+                node_name = kwargs['metadata']['langgraph_node']
+                self._current_chain_node = node_name
+        
+        # CRITICAL: Extract tools incrementally from each tool-bearing call
+        # Tools are bound node-by-node as they're invoked
+        if self._is_langgraph and self._agent:
+            # Check if THIS call has tools (signal that THIS node's tools are now bound)
+            tools_in_call = None
+            if 'invocation_params' in kwargs:
+                tools_in_call = kwargs['invocation_params'].get('tools') or kwargs['invocation_params'].get('functions')
+            elif 'tools' in kwargs:
+                tools_in_call = kwargs['tools']
+            elif 'functions' in kwargs:
+                tools_in_call = kwargs['functions']
+            
+            if tools_in_call:
+                node_name = self._current_chain_node or 'unknown'
+                
+                # Extract tool names from the schemas
+                tool_names = []
+                for tool in tools_in_call:
+                    name = tool.get('name') or tool.get('function', {}).get('name', 'unknown')
+                    tool_names.append(name)
+                
+                # print(f"🔧 [TOOLS DETECTED] Node '{node_name}' has {len(tool_names)} tools: {tool_names}")  # Commented out - too noisy
+                
+                # Check if we've already extracted tools for this node
+                existing_nodes = {t.get('node') for t in self._compiled_tools_metadata}
+                if node_name not in existing_nodes:
+                    try:
+                        # Extract tools from this specific call (provider-resolved schemas)
+                        for tool in tools_in_call:
+                            tool_meta = {
+                                'name': tool.get('name') or tool.get('function', {}).get('name', 'unknown'),
+                                'description': tool.get('description') or tool.get('function', {}).get('description', ''),
+                                'node': node_name
+                            }
+                            
+                            # Get args schema
+                            if 'parameters' in tool:
+                                tool_meta['args_schema'] = tool['parameters']
+                            elif 'function' in tool and 'parameters' in tool['function']:
+                                tool_meta['args_schema'] = tool['function']['parameters']
+                            else:
+                                tool_meta['args_schema'] = {}
+                            
+                            self._compiled_tools_metadata.append(tool_meta)
+                        
+                        # print(f"🔧 [TOOLS METADATA] Extracted metadata for {len(tool_names)} tools from node '{node_name}'")  # Commented out - too noisy
+                    except Exception as e:
+                        print(f"🔧 [TOOLS ERROR] Failed to extract metadata: {e}")
+                        pass  # Silently fail
+                # else:
+                    # print(f"🔧 [TOOLS SKIP] Already extracted tools for node '{node_name}'")  # Commented out - too noisy
+        
+        # Inject rules if applicable
+        modified_prompts = self._inject_rule_if_applicable("llm_start", model_name, prompts)
+        
+        # Store the modified prompts for the LLM wrapper to use
+        self._last_modified_prompts = modified_prompts
+        
+        # Note: Pipecleaner deduplication now happens at ToolExecutor level (see wrappers.py)
+        
+        # 🚨 OPTIMIZED: For LangGraph, check if kwargs contains 'invocation_params' with messages
+        # Extract the most recent message instead of full history
+        # Use from_end=True to capture the END of system prompts (where user's actual query is)
+        if 'invocation_params' in kwargs and 'messages' in kwargs['invocation_params']:
+            args_excerpt = self._extract_recent_message({'messages': kwargs['invocation_params']['messages']})
+        else:
+            args_excerpt = self._excerpt(" | ".join(modified_prompts), from_end=True)
+        
+        # GNN-related fields
+        step_index = len(self._trace)
+        
+        # Track which rules triggered at this step (llm_start rules)
+        rule_triggered_here = []
+        if hasattr(self, '_selected_rules') and self._selected_rules:
+            for rule_meta in self._selected_rules:
+                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
+                    rule_obj, _metadata = rule_meta
+                else:
+                    rule_obj = rule_meta
+                target_step_type = getattr(rule_obj, 'target_step_type', '')
+                if target_step_type in ['llm_start', 'chain_start']:
+                    rule_triggered_here.append(getattr(rule_obj, 'id', 'unknown'))
+        
+        # Record start time for duration calculation
+        start_time = datetime.now()
+        self._start_times[step_index] = start_time
+        
+        step = {
+            "step_type": "llm_start",
+            "tool_name": model_name,
+            "args_excerpt": args_excerpt,
+            "outcome": "",
+            "ts": start_time.isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+            "node": self._current_chain_node,  # LangGraph node name (if available)
+            # GNN step-level fields
+            "step_index": step_index,
+            "rule_triggered_here": rule_triggered_here,
+        }
+        self._trace.append(step)
+        # self._vprint(f"[DASEIN][CALLBACK] Captured llm_start: {len(_TRACE)} total steps")  # Commented out - too noisy
+    
+    def on_llm_end(
+        self,
+        response: Any,
+        **kwargs: Any,
+    ) -> None:
+        """Called when an LLM ends running."""
+        outcome = ""
+        try:
+            # Debug: Print ALL available data to see what we're getting
+            # print(f"[DEBUG] on_llm_end called")
+            # print(f"  response type: {type(response)}")
+            # print(f"  kwargs keys: {kwargs.keys()}")
+            
+            # Try multiple extraction strategies
+            # Strategy 1: Standard LangChain LLMResult structure
+            if hasattr(response, 'generations') and response.generations:
+                if len(response.generations) > 0:
+                    first_gen = response.generations[0]
+                    if isinstance(first_gen, list) and len(first_gen) > 0:
+                        generation = first_gen[0]
+                    else:
+                        generation = first_gen
+                    
+                    # Try multiple content fields
+                    if hasattr(generation, 'text') and generation.text:
+                        outcome = self._excerpt(generation.text)
+                    elif hasattr(generation, 'message'):
+                        if hasattr(generation.message, 'content'):
+                            outcome = self._excerpt(generation.message.content)
+                        elif hasattr(generation.message, 'text'):
+                            outcome = self._excerpt(generation.message.text)
+                    elif hasattr(generation, 'content'):
+                        outcome = self._excerpt(generation.content)
+                    else:
+                        outcome = self._excerpt(str(generation))
+            
+            # Strategy 2: Check if response itself has content
+            elif hasattr(response, 'content'):
+                outcome = self._excerpt(response.content)
+            
+            # Strategy 3: Check kwargs for output/response
+            elif 'output' in kwargs:
+                outcome = self._excerpt(str(kwargs['output']))
+            elif 'result' in kwargs:
+                outcome = self._excerpt(str(kwargs['result']))
+            
+            # Fallback
+            if not outcome:
+                outcome = self._excerpt(str(response))
+            
+            # Debug: Warn if still empty
+            if not outcome or len(outcome) == 0:
+                self._vprint(f"[DASEIN][CALLBACK] WARNING: on_llm_end got empty outcome!")
+                print(f"  Response: {str(response)[:1000]}")
+                print(f"  kwargs keys: {list(kwargs.keys())}")
+                
+        except (AttributeError, IndexError, TypeError) as e:
+            self._vprint(f"[DASEIN][CALLBACK] Error in on_llm_end: {e}")
+            outcome = self._excerpt(str(response))
+        
+        # # 🎯 PRINT FULL LLM OUTPUT (RAW, UNTRUNCATED) - COMMENTED OUT FOR TESTING
+        # node_name = getattr(self, '_current_chain_node', 'agent')
+        # run_number = getattr(self, '_run_number', 1)
+        # print(f"\n{'='*80}")
+        # print(f"[DASEIN][LLM_END] RUN {run_number} | Node: {node_name}")
+        # print(f"{'='*80}")
+        # print(f"FULL OUTPUT:\n{str(response)}")
+        # print(f"{'='*80}\n")
+        
+        # 🎯 CRITICAL: Extract function calls for state tracking (agent-agnostic)
+        try:
+            if hasattr(response, 'generations') and response.generations:
+                first_gen = response.generations[0]
+                if isinstance(first_gen, list) and len(first_gen) > 0:
+                    generation = first_gen[0]
+                else:
+                    generation = first_gen
+                
+                # Check for function_call in message additional_kwargs
+                if hasattr(generation, 'message') and hasattr(generation.message, 'additional_kwargs'):
+                    func_call = generation.message.additional_kwargs.get('function_call')
+                    if func_call and isinstance(func_call, dict) and 'name' in func_call:
+                        func_name = func_call['name']
+                        step_num = len(self._trace)
+                        
+                        # Extract arguments and create preview
+                        args_str = func_call.get('arguments', '')
+                        preview = ''
+                        if args_str and len(args_str) > 0:
+                            # Take first 100 chars as preview
+                            preview = args_str[:100].replace('\n', ' ').replace('\r', '')
+                            if len(args_str) > 100:
+                                preview += '...'
+                        
+                        call_info = {
+                            'step': step_num,
+                            'ts': datetime.now().isoformat(),
+                            'preview': preview
+                        }
+                        
+                        if func_name not in self._function_calls_made:
+                            self._function_calls_made[func_name] = []
+                        self._function_calls_made[func_name].append(call_info)
+                        
+                        # 🔥 HOTPATH: Set current tool name for next LLM call (which will be inside the tool)
+                        self._current_tool_name = func_name
+                        
+                        self._vprint(f"[DASEIN][STATE] Tracked function call: {func_name} (count: {len(self._function_calls_made[func_name])})")
+        except Exception as e:
+            pass  # Silently skip if function call extraction fails
+        
+        # Extract token usage from response metadata
+        input_tokens = 0
+        output_tokens = 0
+        try:
+            # Try LangChain's standard llm_output field
+            if hasattr(response, 'llm_output') and response.llm_output:
+                llm_output = response.llm_output
+                # Different providers use different field names
+                if 'token_usage' in llm_output:
+                    usage = llm_output['token_usage']
+                    input_tokens = usage.get('prompt_tokens', 0) or usage.get('input_tokens', 0)
+                    output_tokens = usage.get('completion_tokens', 0) or usage.get('output_tokens', 0)
+                elif 'usage_metadata' in llm_output:
+                    usage = llm_output['usage_metadata']
+                    input_tokens = usage.get('input_tokens', 0) or usage.get('prompt_tokens', 0)
+                    output_tokens = usage.get('output_tokens', 0) or usage.get('completion_tokens', 0)
+            
+            if (input_tokens == 0 and output_tokens == 0) and hasattr(response, 'generations') and response.generations:
+                first_gen = response.generations[0]
+                if isinstance(first_gen, list) and len(first_gen) > 0:
+                    gen = first_gen[0]
+                else:
+                    gen = first_gen
+                
+                # Check message.usage_metadata (Google GenAI stores it here!)
+                if hasattr(gen, 'message') and hasattr(gen.message, 'usage_metadata'):
+                    usage = gen.message.usage_metadata
+                    input_tokens = usage.get('input_tokens', 0)
+                    output_tokens = usage.get('output_tokens', 0)
+                
+                # Fallback: Check generation_info
+                elif hasattr(gen, 'generation_info') and gen.generation_info:
+                    gen_info = gen.generation_info
+                    if 'usage_metadata' in gen_info:
+                        usage = gen_info['usage_metadata']
+                        input_tokens = usage.get('prompt_token_count', 0) or usage.get('input_tokens', 0)
+                        output_tokens = usage.get('candidates_token_count', 0) or usage.get('output_tokens', 0)
+            
+            # Check if we have stored savings from corpus deduplication and adjust tokens
+            current_run_id = kwargs.get('run_id', None)
+            if current_run_id and hasattr(self, '_corpus_token_savings') and current_run_id in self._corpus_token_savings:
+                tokens_saved = self._corpus_token_savings[current_run_id]
+                # Adjust input tokens to reflect deduplication savings
+                if input_tokens > 0:
+                    # If provider count is much larger than saved estimate, LLM saw original prompts
+                    if abs(input_tokens - tokens_saved) >= input_tokens * 0.3:
+                        input_tokens = max(0, input_tokens - tokens_saved)
+                # Clean up
+                del self._corpus_token_savings[current_run_id]
+            
+            # Log if we got tokens
+            # if input_tokens > 0 or output_tokens > 0:
+            #     self._vprint(f"[DASEIN][TOKENS] Captured: {input_tokens} in, {output_tokens} out")
+                    
+        except Exception as e:
+            # Print error for debugging
+            self._vprint(f"[DASEIN][CALLBACK] Error extracting tokens: {e}")
+            import traceback
+            traceback.print_exc()
+        
+        # GNN-related fields: compute tokens_delta
+        step_index = len(self._trace)
+        tokens_delta = 0
+        # Find previous step with tokens_output to compute delta
+        for prev_step in reversed(self._trace):
+            if 'tokens_output' in prev_step and prev_step['tokens_output'] > 0:
+                tokens_delta = output_tokens - prev_step['tokens_output']
+                break
+        
+        # Calculate duration_ms by matching with corresponding llm_start
+        duration_ms = 0
+        for i in range(len(self._trace) - 1, -1, -1):
+            if self._trace[i].get('step_type') == 'llm_start':
+                # Found the matching llm_start
+                if i in self._start_times:
+                    start_time = self._start_times[i]
+                    end_time = datetime.now()
+                    duration_ms = int((end_time - start_time).total_seconds() * 1000)
+                    # Update the llm_start step with duration_ms
+                    self._trace[i]['duration_ms'] = duration_ms
+                break
+        
+        step = {
+            "step_type": "llm_end",
+            "tool_name": "",
+            "args_excerpt": "",
+            "outcome": self._excerpt(outcome, max_len=1000),  # Truncate to 1000 chars
+            "ts": datetime.now().isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+            "tokens_input": input_tokens,
+            "tokens_output": output_tokens,
+            "node": self._current_chain_node,  # LangGraph node name (if available)
+            # GNN step-level fields
+            "step_index": step_index,
+            "tokens_delta": tokens_delta,
+            "duration_ms": duration_ms,
+        }
+        self._trace.append(step)
+    
+    def on_agent_action(
+        self,
+        action: Any,
+        **kwargs: Any,
+    ) -> None:
+        """Called when an agent takes an action."""
+        tool_name = getattr(action, 'tool', 'unknown')
+        args_excerpt = self._excerpt(str(getattr(action, 'tool_input', '')))
+        outcome = self._excerpt(str(getattr(action, 'log', '')))
+        
+        step = {
+            "step_type": "agent_action",
+            "tool_name": tool_name,
+            "args_excerpt": args_excerpt,
+            "outcome": outcome,
+            "ts": datetime.now().isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+        }
+        self._trace.append(step)
+    
+    def on_agent_finish(
+        self,
+        finish: Any,
+        **kwargs: Any,
+    ) -> None:
+        """Called when an agent finishes."""
+        outcome = self._excerpt(str(getattr(finish, 'return_values', '')))
+        
+        step = {
+            "step_type": "agent_finish",
+            "tool_name": None,
+            "args_excerpt": "",
+            "outcome": outcome,
+            "ts": datetime.now().isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+        }
+        self._trace.append(step)
+    
+    def on_tool_start(
+        self,
+        serialized: Dict[str, Any],
+        input_str: str,
+        *,
+        run_id: str,
+        parent_run_id: Optional[str] = None,
+        tags: Optional[List[str]] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        inputs: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
+    ) -> None:
+        """Called when a tool starts running.
+        
+        This is where we detect and track dynamic tools that weren't
+        statically attached to the agent at init time.
+        """
+        import time
+        tool_name = serialized.get("name", "unknown") if serialized else "unknown"
+        
+        # Track discovered tools for reporting
+        if tool_name != "unknown" and tool_name not in self._discovered_tools:
+            self._discovered_tools.add(tool_name)
+            # Tool discovered and tracked (silently)
+        
+        # Store tool name for later use in on_tool_end
+        self._tool_name_by_run_id[run_id] = tool_name
+        
+        # 🔥 HOTPATH: Track current tool for pipecleaner deduplication
+        self._current_tool_name = tool_name
+        
+        # Apply tool-level rule injection
+        # self._vprint(f"[DASEIN][CALLBACK] on_tool_start called!")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][CALLBACK] Tool: {tool_name}")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][CALLBACK] Input: {input_str[:100]}...")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][APPLY] on_tool_start: selected_rules={len(self._selected_rules)}")  # Commented out - too noisy
+        modified_input = self._inject_tool_rule_if_applicable("tool_start", tool_name, input_str)
+        
+        args_excerpt = self._excerpt(modified_input)
+        
+        # GNN-related fields: capture step-level metrics
+        step_index = len(self._trace)
+        tool_input_chars = len(str(input_str))
+        
+        # Track which rules triggered at this step
+        rule_triggered_here = []
+        if hasattr(self, '_selected_rules') and self._selected_rules:
+            for rule_meta in self._selected_rules:
+                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
+                    rule_obj, _metadata = rule_meta
+                else:
+                    rule_obj = rule_meta
+                if getattr(rule_obj, 'target_step_type', '') == "tool_start":
+                    rule_triggered_here.append(getattr(rule_obj, 'id', 'unknown'))
+        
+        # Record start time for duration calculation (keyed by run_id for tools)
+        start_time = datetime.now()
+        self._start_times[run_id] = start_time
+        
+        step = {
+            "step_type": "tool_start",
+            "tool_name": tool_name,
+            "args_excerpt": args_excerpt,
+            "outcome": "",
+            "ts": start_time.isoformat(),
+            "run_id": run_id,
+            "parent_run_id": parent_run_id,
+            "node": self._current_chain_node,  # LangGraph node name (if available)
+            # GNN step-level fields
+            "step_index": step_index,
+            "tool_input_chars": tool_input_chars,
+            "rule_triggered_here": rule_triggered_here,
+        }
+        self._trace.append(step)
+    
+    def on_tool_end(
+        self,
+        output: str,
+        *,
+        run_id: str,
+        parent_run_id: Optional[str] = None,
+        tags: Optional[List[str]] = None,
+        **kwargs: Any,
+    ) -> Any:
+        """Called when a tool ends running."""
+        import time
+        # Get the tool name from the corresponding tool_start
+        tool_name = self._tool_name_by_run_id.get(run_id, "unknown")
+        
+        # Handle different output types (LangGraph may pass ToolMessage objects)
+        output_str = str(output)
+        
+        # Note: Pipecleaner deduplication happens at ToolExecutor level (see wrappers.py)
+        
+        outcome = self._excerpt(output_str)
+        
+        # self._vprint(f"[DASEIN][CALLBACK] on_tool_end called!")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][CALLBACK] Tool: {tool_name}")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][CALLBACK] Output length: {len(output_str)} chars")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][CALLBACK] Outcome length: {len(outcome)} chars")  # Commented out - too noisy
+        
+        # GNN-related fields: capture tool output metrics
+        step_index = len(self._trace)
+        tool_output_chars = len(output_str)
+        
+        # Estimate tool_output_items (heuristic: count lines, or rows if SQL-like)
+        tool_output_items = 0
+        try:
+            # Try to count lines as a proxy for items
+            if output_str:
+                tool_output_items = output_str.count('\n') + 1
+        except:
+            tool_output_items = 0
+        
+        # Calculate duration_ms using run_id to match with tool_start
+        duration_ms = 0
+        if run_id in self._start_times:
+            start_time = self._start_times[run_id]
+            end_time = datetime.now()
+            duration_ms = int((end_time - start_time).total_seconds() * 1000)
+            # Update the corresponding tool_start step with duration_ms
+            for i in range(len(self._trace) - 1, -1, -1):
+                if self._trace[i].get('step_type') == 'tool_start' and self._trace[i].get('run_id') == run_id:
+                    self._trace[i]['duration_ms'] = duration_ms
+                    break
+            # Clean up start time
+            del self._start_times[run_id]
+        
+        # Extract available selectors from DOM-like output (web browse agents)
+        available_selectors = None
+        if tool_name in ['extract_text', 'get_elements', 'extract_hyperlinks', 'extract_content']:
+            available_selectors = self._extract_semantic_selectors(output_str)
+        
+        step = {
+            "step_type": "tool_end",
+            "tool_name": tool_name,
+            "args_excerpt": "",
+            "outcome": self._excerpt(outcome, max_len=1000),  # Truncate to 1000 chars
+            "ts": datetime.now().isoformat(),
+            "run_id": run_id,
+            "parent_run_id": parent_run_id,
+            "node": self._current_chain_node,  # LangGraph node name (if available)
+            # GNN step-level fields
+            "step_index": step_index,
+            "tool_output_chars": tool_output_chars,
+            "tool_output_items": tool_output_items,
+            "duration_ms": duration_ms,
+        }
+        
+        # Add available_selectors only if found (keep trace light)
+        if available_selectors:
+            step["available_selectors"] = available_selectors
+        self._trace.append(step)
+        
+        # Clean up the stored tool name
+        if run_id in self._tool_name_by_run_id:
+            del self._tool_name_by_run_id[run_id]
+        
+        # 🔥 HOTPATH: Clear current tool
+        self._current_tool_name = None
+    
+    def on_tool_error(
+        self,
+        error: BaseException,
+        *,
+        run_id: str,
+        parent_run_id: Optional[str] = None,
+        tags: Optional[List[str]] = None,
+        **kwargs: Any,
+    ) -> None:
+        """Called when a tool encounters an error."""
+        error_msg = self._excerpt(str(error))
+        
+        step = {
+            "step_type": "tool_error",
+            "tool_name": "",
+            "args_excerpt": "",
+            "outcome": f"ERROR: {error_msg}",
+            "ts": datetime.now().isoformat(),
+            "run_id": run_id,
+            "parent_run_id": parent_run_id,
+        }
+        self._trace.append(step)
+    
+    def on_chain_start(
+        self,
+        serialized: Dict[str, Any],
+        inputs: Dict[str, Any],
+        **kwargs: Any,
+    ) -> None:
+        """Called when a chain starts running."""
+        chain_name = serialized.get("name", "unknown") if serialized else "unknown"
+        # self._vprint(f"[DASEIN][CALLBACK] on_chain_start called!")  # Commented out - too noisy
+        # self._vprint(f"[DASEIN][CALLBACK] Chain: {chain_name}")  # Commented out - too noisy
+        
+        # 🚨 OPTIMIZED: For LangGraph agents, suppress redundant chain_start events
+        # LangGraph fires on_chain_start for every internal node, creating noise
+        # We already capture llm_start, llm_end, tool_start, tool_end which are more meaningful
+        if self._is_langgraph:
+            # Track current chain node for future targeted injection
+            # 🎯 CRITICAL: Extract actual node name from metadata (same as on_llm_start)
+            if 'metadata' in kwargs and isinstance(kwargs['metadata'], dict):
+                if 'langgraph_node' in kwargs['metadata']:
+                    self._current_chain_node = kwargs['metadata']['langgraph_node']
+                    # print(f"🔵 [NODE EXEC] {self._current_chain_node}")  # Commented out - too noisy
+                else:
+                    self._current_chain_node = chain_name
+                    # print(f"🔵 [NODE EXEC] {chain_name}")  # Commented out - too noisy
+            else:
+                self._current_chain_node = chain_name
+                # print(f"🔵 [NODE EXEC] {chain_name}")  # Commented out - too noisy
+            
+            # self._vprint(f"[DASEIN][CALLBACK] Suppressing redundant chain_start for LangGraph agent")  # Commented out - too noisy
+            # Still handle tool executors
+            if chain_name in {"tools", "ToolNode", "ToolExecutor"}:
+                # self._vprint(f"[DASEIN][CALLBACK] Bridging chain_start to tool_start for {chain_name}")  # Commented out - too noisy
+                pass
+                self._handle_tool_executor_start(serialized, inputs, **kwargs)
+            return
+        
+        # For standard LangChain agents, keep chain_start events
+        # Bridge to tool_start for tool executors
+        if chain_name in {"tools", "ToolNode", "ToolExecutor"}:
+            # self._vprint(f"[DASEIN][CALLBACK] Bridging chain_start to tool_start for {chain_name}")  # Commented out - too noisy
+            self._handle_tool_executor_start(serialized, inputs, **kwargs)
+        
+        args_excerpt = self._excerpt(str(inputs))
+        
+        # Record start time for duration calculation
+        step_index = len(self._trace)
+        start_time = datetime.now()
+        self._start_times[f"chain_{step_index}"] = start_time
+        
+        step = {
+            "step_type": "chain_start",
+            "tool_name": chain_name,
+            "args_excerpt": args_excerpt,
+            "outcome": "",
+            "ts": start_time.isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+            "step_index": step_index,
+        }
+        self._trace.append(step)
+    
+    def on_chain_end(
+        self,
+        outputs: Dict[str, Any],
+        **kwargs: Any,
+    ) -> None:
+        """Called when a chain ends running."""
+        # 🚨 OPTIMIZED: Suppress redundant chain_end for LangGraph agents
+        if self._is_langgraph:
+            return
+        
+        outcome = self._excerpt(str(outputs))
+        
+        # Calculate duration_ms by matching with corresponding chain_start
+        duration_ms = 0
+        for i in range(len(self._trace) - 1, -1, -1):
+            if self._trace[i].get('step_type') == 'chain_start':
+                # Found the matching chain_start
+                chain_key = f"chain_{i}"
+                if chain_key in self._start_times:
+                    start_time = self._start_times[chain_key]
+                    end_time = datetime.now()
+                    duration_ms = int((end_time - start_time).total_seconds() * 1000)
+                    # Update the chain_start step with duration_ms
+                    self._trace[i]['duration_ms'] = duration_ms
+                    # Clean up start time
+                    del self._start_times[chain_key]
+                break
+        
+        step = {
+            "step_type": "chain_end",
+            "tool_name": "",
+            "args_excerpt": "",
+            "outcome": outcome,
+            "ts": datetime.now().isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+            "duration_ms": duration_ms,
+        }
+        self._trace.append(step)
+    
+    def on_chain_error(
+        self,
+        error: BaseException,
+        **kwargs: Any,
+    ) -> None:
+        """Called when a chain encounters an error."""
+        error_msg = self._excerpt(str(error))
+        
+        step = {
+            "step_type": "chain_error",
+            "tool_name": "",
+            "args_excerpt": "",
+            "outcome": f"ERROR: {error_msg}",
+            "ts": datetime.now().isoformat(),
+            "run_id": None,
+            "parent_run_id": None,
+        }
+        self._trace.append(step)
+    
+    def _extract_recent_message(self, inputs: Dict[str, Any]) -> str:
+        """
+        Extract the most recent message from LangGraph inputs to show thought progression.
+        
+        For LangGraph agents, inputs contain {'messages': [msg1, msg2, ...]}.
+        Instead of showing the entire history, we extract just the last message.
+        """
+        try:
+            # Check if this is a LangGraph message format
+            if isinstance(inputs, dict) and 'messages' in inputs:
+                messages = inputs['messages']
+                if isinstance(messages, list) and len(messages) > 0:
+                    # Get the most recent message
+                    last_msg = messages[-1]
+                    
+                    # Extract content based on message type
+                    if hasattr(last_msg, 'content'):
+                        # LangChain message object
+                        content = last_msg.content
+                        msg_type = getattr(last_msg, 'type', 'unknown')
+                        return self._excerpt(f"[{msg_type}] {content}")
+                    elif isinstance(last_msg, tuple) and len(last_msg) >= 2:
+                        # Tuple format: (role, content)
+                        return self._excerpt(f"[{last_msg[0]}] {last_msg[1]}")
+                    else:
+                        # Unknown format, convert to string
+                        return self._excerpt(str(last_msg))
+            
+            # For non-message inputs, check if it's a list of actions/tool calls
+            if isinstance(inputs, list) and len(inputs) > 0:
+                # This might be tool call info
+                return self._excerpt(str(inputs[0]))
+            
+            # Fall back to original behavior for non-LangGraph agents
+            return self._excerpt(str(inputs))
+            
+        except Exception as e:
+            # On any error, fall back to original behavior
+            return self._excerpt(str(inputs))
+    
+    def _excerpt(self, obj: Any, max_len: int = 250, from_end: bool = False) -> str:
+        """
+        Truncate text to max_length with ellipsis.
+        
+        Args:
+            obj: Object to convert to string and truncate
+            max_len: Maximum length of excerpt
+            from_end: If True, take LAST max_len chars (better for system prompts).
+                     If False, take FIRST max_len chars (better for tool args).
+        """
+        text = str(obj)
+        if len(text) <= max_len:
+            return text
+        
+        if from_end:
+            # Take last X chars - better for system prompts where the end contains user's actual query
+            return "..." + text[-(max_len-3):]
+        else:
+            # Take first X chars - better for tool inputs
+            return text[:max_len-3] + "..."
+    
+    def _extract_semantic_selectors(self, html_text: str) -> List[Dict[str, int]]:
+        """
+        Extract semantic HTML tags from output for grounding web browse rules.
+        Only extracts semantic tags (nav, header, h1, etc.) to keep trace lightweight.
+        
+        Args:
+            html_text: Output text that may contain HTML
+            
+        Returns:
+            List of {"tag": str, "count": int} sorted by count descending, or None if no HTML
+        """
+        import re
+        
+        # Quick check: does this look like HTML?
+        if '<' not in html_text or '>' not in html_text:
+            return None
+        
+        # Semantic tags we care about (prioritized for web browse agents)
+        semantic_tags = [
+            # Navigation/Structure (highest priority)
+            'nav', 'header', 'footer', 'main', 'article', 'section', 'aside',
+            
+            # Headers (critical for "find headers" queries!)
+            'h1', 'h2', 'h3', 'h4', 'h5', 'h6',
+            
+            # Interactive
+            'a', 'button', 'form', 'input', 'textarea', 'select', 'label',
+            
+            # Lists (often used for navigation)
+            'ul', 'ol', 'li',
+            
+            # Tables (data extraction)
+            'table', 'thead', 'tbody', 'tr', 'th', 'td',
+            
+            # Media
+            'img', 'video', 'audio'
+        ]
+        
+        # Count occurrences of each semantic tag
+        found_tags = {}
+        for tag in semantic_tags:
+            # Pattern: <tag ...> or <tag> (opening tags only)
+            pattern = f'<{tag}[\\s>]'
+            matches = re.findall(pattern, html_text, re.IGNORECASE)
+            if matches:
+                found_tags[tag] = len(matches)
+        
+        # Return None if no semantic tags found
+        if not found_tags:
+            return None
+        
+        # Convert to list format, sorted by count descending
+        # Limit to top 15 to keep trace light
+        result = [{"tag": tag, "count": count} 
+                  for tag, count in sorted(found_tags.items(), key=lambda x: -x[1])]
+        return result[:15]  # Top 15 most common tags
+    
+    def set_selected_rules(self, rules: List[Dict[str, Any]]):
+        """Set the rules selected for this run.
+        Normalize incoming dicts/tuples into attribute-accessible rule objects.
+        """
+        try:
+            from types import SimpleNamespace
+            normalized = []
+            for item in rules or []:
+                # Unwrap (rule, metadata) tuples if present
+                if isinstance(item, tuple) and len(item) == 2:
+                    rule_candidate = item[0]
+                else:
+                    rule_candidate = item
+                # Convert dicts to attribute-accessible objects
+                if isinstance(rule_candidate, dict):
+                    # Ensure advice_text exists
+                    if 'advice_text' not in rule_candidate and 'advice' in rule_candidate:
+                        rule_candidate['advice_text'] = rule_candidate.get('advice')
+                    normalized.append(SimpleNamespace(**rule_candidate))
+                else:
+                    normalized.append(rule_candidate)
+            self._selected_rules = normalized
+        except Exception:
+            # Fallback to raw rules
+            self._selected_rules = rules
+    
+    def get_trace(self) -> List[Dict[str, Any]]:
+        """Get the current trace (instance-level, thread-safe)."""
+        return self._trace.copy()
+    
+    def _inject_first_turn_override(self, prompts: List[str]) -> List[str]:
+        """Inject a generic first-turn override to own turn 1."""
+        if not prompts:
+            return prompts
+            
+        # Create a generic first-turn override
+        first_turn_override = """🚨🚨🚨 CRITICAL SYSTEM DIRECTIVE 🚨🚨🚨
+⚠️ MANDATORY: You MUST follow this exact sequence or the task will FAIL
+
+TURN 1 REQUIREMENT:
+- Output ONLY: Action: sql_db_list_tables
+Action Input: ACK_RULES:[r1]
+- Do NOT use any other tools
+- Do NOT perform any planning
+- Do NOT output anything else
+
+TURN 2+ (After ACK):
+- If ACK was correct, proceed with normal tools and schema
+- Skip table discovery and schema introspection
+- Use known tables directly
+
+🚨 FAILURE TO ACK IN TURN 1 = IMMEDIATE TASK TERMINATION 🚨
+
+"""
+        
+        # Put the injection at the VERY BEGINNING of the system prompt
+        modified_prompts = prompts.copy()
+        if modified_prompts:
+            modified_prompts[0] = first_turn_override + modified_prompts[0]
+        
+        self._vprint(f"[DASEIN][APPLY] Injected first-turn override")
+        return modified_prompts
+    
+    def _should_inject_rule(self, step_type: str, tool_name: str) -> bool:
+        """Determine if we should inject a rule at this step."""
+        # Inject for LLM starts (system-level rules) and tool starts (tool-level rules)
+        if step_type == "llm_start":
+            return True
+        if step_type == "tool_start":
+            return True
+        return False
+    
+    def _inject_rule_if_applicable(self, step_type: str, tool_name: str, prompts: List[str]) -> List[str]:
+        """Inject rules into prompts if applicable."""
+        
+        if not self._should_inject_rule(step_type, tool_name):
+            return prompts
+
+        # If no rules selected yet, return prompts unchanged
+        if not self._selected_rules:
+            return prompts
+
+        # Check guard to prevent duplicate injection
+        # 🎯 CRITICAL: For LangGraph planning nodes, SKIP the guard - we need to inject on EVERY call
+        # because the same node (e.g., supervisor) can be called multiple times dynamically
+        use_guard = True
+        if hasattr(self, '_is_langgraph') and self._is_langgraph:
+            if step_type == 'llm_start' and hasattr(self, '_current_chain_node'):
+                # For planning nodes, skip guard to allow re-injection on subsequent calls
+                if hasattr(self, '_planning_nodes') and self._current_chain_node in self._planning_nodes:
+                    use_guard = False
+        
+        if use_guard:
+            guard_key = (step_type, tool_name)
+            if guard_key in self._injection_guard:
+                return prompts
+        
+        try:
+            # Inject rules that target llm_start and tool_start (both go to system prompt)
+            system_rules = []
+            for rule_meta in self._selected_rules:
+                # Handle tuple format from select_rules: (rule, metadata)
+                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
+                    rule, metadata = rule_meta
+                elif isinstance(rule_meta, dict):
+                    if 'rule' in rule_meta:
+                        rule = rule_meta.get('rule', {})
+                    else:
+                        rule = rule_meta
+                else:
+                    rule = rule_meta
+                
+                # Check if this rule targets system-level injection (llm_start only)
+                target_step_type = getattr(rule, 'target_step_type', '')
+                
+                # 🚨 CRITICAL: For LangGraph agents, only skip planning rules if agent was successfully recreated
+                # If recreation failed, we need to inject via callback as fallback
+                if step_type == 'llm_start' and hasattr(self, '_is_langgraph') and self._is_langgraph:
+                    # Only skip if agent was actually recreated with planning rules embedded
+                    if hasattr(self, '_agent_was_recreated') and self._agent_was_recreated:
+                        if target_step_type in ['llm_start', 'chain_start']:
+                            self._vprint(f"[DASEIN][CALLBACK] Skipping planning rule {getattr(rule, 'id', 'unknown')} for LangGraph agent (already injected at creation)")
+                            continue
+                
+                # 🎯 NODE-SCOPED INJECTION: Check target_node if specified (for node-specific rules)
+                if target_step_type in ['llm_start', 'chain_start']:
+                    current_node = getattr(self, '_current_chain_node', None)
+                    
+                    # Check if this rule targets a specific node
+                    target_node = getattr(rule, 'target_node', None)
+                    if target_node:
+                        # Rule has explicit target_node - ONLY inject if we're in that node
+                        if current_node != target_node:
+                            # Silently skip - not the target node
+                            continue
+                    else:
+                        # No target_node specified - use existing planning_nodes logic (backward compatibility)
+                        if hasattr(self, '_planning_nodes') and self._planning_nodes:
+                            # Check if current node is in the planning nodes set
+                            if current_node not in self._planning_nodes:
+                                # Silently skip non-planning nodes
+                                continue
+                        # Injecting into planning node (logged in detailed injection log below)
+                    
+                    advice = getattr(rule, 'advice_text', getattr(rule, 'advice', ''))
+                    if advice:
+                        system_rules.append(advice)
+            
+            # Apply system-level rules if any
+            if system_rules and prompts:
+                modified_prompts = prompts.copy()
+                system_prompt = modified_prompts[0]
+                
+                # Combine all system rules with much stronger language
+                rule_injections = []
+                for advice in system_rules:
+                    if "TOOL RULE:" in advice:
+                        # Make tool rules even more explicit
+                        rule_injections.append(f"🚨 CRITICAL TOOL OVERRIDE: {advice}")
+                    else:
+                        rule_injections.append(f"🚨 CRITICAL SYSTEM OVERRIDE: {advice}")
+                
+                # Build execution state context (agent-agnostic, with argument previews)
+                # Strategy: Show all if ≤5 calls, else show most recent 3
+                # Rationale: Small counts get full context; larger counts show recent to prevent duplicates
+                state_context = ""
+                if hasattr(self, '_function_calls_made') and self._function_calls_made:
+                    state_lines = []
+                    for func_name in sorted(self._function_calls_made.keys()):
+                        calls = self._function_calls_made[func_name]
+                        count = len(calls)
+                        
+                        # Hybrid window: show all if ≤5 calls, else show recent 3
+                        if count <= 5:
+                            # Show all calls with previews
+                            state_lines.append(f"  • {func_name}: called {count}x:")
+                            for call in calls:
+                                preview = call.get('preview', '')
+                                if preview:
+                                    state_lines.append(f"      [step {call['step']}] {preview}")
+                                else:
+                                    state_lines.append(f"      [step {call['step']}] (no args)")
+                        else:
+                            # Show summary + recent 3 with previews
+                            state_lines.append(f"  • {func_name}: called {count}x (most recent 3):")
+                            for call in calls[-3:]:
+                                preview = call.get('preview', '')
+                                if preview:
+                                    state_lines.append(f"      [step {call['step']}] {preview}")
+                                else:
+                                    state_lines.append(f"      [step {call['step']}] (no args)")
+                    
+                    if state_lines:
+                        state_context = f"""
+EXECUTION STATE (functions called so far in this run):
+{chr(10).join(state_lines)}
+
+"""
+                
+                combined_injection = f""" SYSTEM OVERRIDE — PLANNING TURN ONLY 
+These rules OVERRIDE all defaults. You MUST enforce them exactly or the task FAILS.
+
+Tags: AVOID (absolute ban), SKIP (force bypass), FIX (mandatory params), PREFER (ranked choice), HINT (optional).
+Precedence: AVOID/SKIP > FIX > PREFER > HINT. On conflict, the higher rule ALWAYS wins.
+
+{state_context}Checklist (non-negotiable):
+- AVOID: no banned targets under ANY condition.
+- SKIP: bypass skipped steps/tools; NEVER retry them.
+- FIX: all required params/settings MUST be included.
+- PREFER: when multiple compliant options exist, choose the preferred—NO exceptions.
+- Recovery: if a banned/skipped item already failed, IMMEDIATELY switch to a compliant alternative.
+
+Output Contract: Produce ONE compliant tool/function call (or direct answer if none is needed).  
+NO reasoning, NO justification, NO markdown.
+
+Rules to Enforce:
+
+
+{chr(10).join(rule_injections)}
+
+
+"""
+                # Put the injection at the VERY BEGINNING of the system prompt
+                modified_prompts[0] = combined_injection + system_prompt
+                
+                # Add to guard (only if we're using the guard)
+                if use_guard:
+                    self._injection_guard.add(guard_key)
+                
+                # Log the complete injection for debugging
+                # Compact injection summary
+                if hasattr(self, '_is_langgraph') and self._is_langgraph:
+                    # LangGraph: show node name
+                    func_count = len(self._function_calls_made) if hasattr(self, '_function_calls_made') and state_context else 0
+                    node_name = getattr(self, '_current_chain_node', 'unknown')
+                    print(f"[DASEIN] 🎯 Injecting {len(system_rules)} rule(s) into {node_name} | State: {func_count} functions tracked")
+                else:
+                    # LangChain: simpler logging without node name
+                    print(f"[DASEIN] 🎯 Injecting {len(system_rules)} rule(s) into agent")
+                
+                return modified_prompts
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][APPLY] Injection failed: {e}")
+        
+        return prompts
+    
+    def _inject_tool_rule_if_applicable(self, step_type: str, tool_name: str, input_str: str) -> str:
+        """Inject rules into tool input if applicable."""
+        if not self._should_inject_rule(step_type, tool_name):
+            return input_str
+
+        # If no rules selected yet, return input unchanged
+        if not self._selected_rules:
+            return input_str
+
+        # Check guard to prevent duplicate injection
+        guard_key = (step_type, tool_name)
+        if guard_key in self._injection_guard:
+            return input_str
+        
+        try:
+            # Inject rules that target tool_start
+            tool_rules = []
+            current_node = getattr(self, '_current_chain_node', None)
+            
+            for rule_meta in self._selected_rules:
+                # Handle tuple format from select_rules: (rule, metadata)
+                if isinstance(rule_meta, tuple) and len(rule_meta) == 2:
+                    rule, metadata = rule_meta
+                else:
+                    rule = rule_meta
+                    metadata = {}
+
+                # Only apply rules that target tool_start
+                if rule.target_step_type == "tool_start":
+                    # 🎯 NODE-SCOPED INJECTION: Check target_node if specified
+                    target_node = getattr(rule, 'target_node', None)
+                    if target_node:
+                        # Rule has explicit target_node - ONLY inject if we're in that node
+                        if current_node != target_node:
+                            # Silently skip - not the target node
+                            continue
+                    # No target_node specified - inject into any node using this tool (backward compat)
+                    
+                    tool_rules.append(rule)
+                    self._vprint(f"[DASEIN][APPLY] Tool rule: {rule.advice_text[:100]}...")
+
+            if tool_rules:
+                # Apply tool-level rule injection
+                modified_input = self._apply_tool_rules(input_str, tool_rules)
+                self._injection_guard.add(guard_key)
+                return modified_input
+            else:
+                return input_str
+
+        except Exception as e:
+            self._vprint(f"[DASEIN][APPLY] Error injecting tool rules: {e}")
+            return input_str
+    
+    def _apply_tool_rules(self, input_str: str, rules: List) -> str:
+        """Apply tool-level rules to modify the input string."""
+        modified_input = input_str
+        
+        for rule in rules:
+            try:
+                # Apply the rule's advice to modify the tool input
+                if "strip" in rule.advice_text.lower() and "fence" in rule.advice_text.lower():
+                    # Strip markdown code fences
+                    import re
+                    # Remove ```sql...``` or ```...``` patterns
+                    modified_input = re.sub(r'```(?:sql)?\s*(.*?)\s*```', r'\1', modified_input, flags=re.DOTALL)
+                    self._vprint(f"[DASEIN][APPLY] Stripped code fences from tool input")
+                elif "strip" in rule.advice_text.lower() and "whitespace" in rule.advice_text.lower():
+                    # Strip leading/trailing whitespace
+                    modified_input = modified_input.strip()
+                    self._vprint(f"[DASEIN][APPLY] Stripped whitespace from tool input")
+                # Add more rule types as needed
+                
+            except Exception as e:
+                self._vprint(f"[DASEIN][APPLY] Error applying tool rule: {e}")
+                continue
+        
+        return modified_input
+    
+    def _handle_tool_executor_start(
+        self,
+        serialized: Dict[str, Any],
+        inputs: Dict[str, Any],
+        **kwargs: Any,
+    ) -> None:
+        """Handle tool executor start - bridge from chain_start to tool_start."""
+        self._vprint(f"[DASEIN][CALLBACK] tool_start (from chain_start)")
+        
+        # Extract tool information from inputs
+        tool_name = "unknown"
+        tool_input = ""
+        
+        if isinstance(inputs, dict):
+            if "tool" in inputs:
+                tool_name = inputs["tool"]
+            elif "tool_name" in inputs:
+                tool_name = inputs["tool_name"]
+            
+            if "tool_input" in inputs:
+                tool_input = str(inputs["tool_input"])
+            elif "input" in inputs:
+                tool_input = str(inputs["input"])
+            else:
+                tool_input = str(inputs)
+        else:
+            tool_input = str(inputs)
+        
+        self._vprint(f"[DASEIN][CALLBACK] Tool: {tool_name}")
+        self._vprint(f"[DASEIN][CALLBACK] Input: {tool_input[:100]}...")
+        
+        # Check if we have tool_start rules that cover this tool
+        tool_rules = [rule for rule in self._selected_rules if rule.target_step_type == "tool_start"]
+        covered_rules = [rule for rule in tool_rules if self._rule_covers_tool(rule, tool_name, tool_input)]
+        
+        if covered_rules:
+            self._vprint(f"[DASEIN][APPLY] tool_start: {len(covered_rules)} rules cover this tool call")
+            # Fire micro-turn for rule application
+            modified_input = self._fire_micro_turn_for_tool_rules(covered_rules, tool_name, tool_input)
+        else:
+            self._vprint(f"[DASEIN][APPLY] tool_start: no rules cover this tool call")
+            modified_input = tool_input
+        
+        args_excerpt = self._excerpt(modified_input)
+        
+        step = {
+            "step_type": "tool_start",
+            "tool_name": tool_name,
+            "args_excerpt": args_excerpt,
+            "outcome": "",
+            "ts": datetime.now().isoformat(),
+            "run_id": kwargs.get("run_id"),
+            "parent_run_id": kwargs.get("parent_run_id"),
+        }
+        self._trace.append(step)
+    
+    def _rule_covers_tool(self, rule, tool_name: str, tool_input: str) -> bool:
+        """Check if a rule covers the given tool call."""
+        try:
+            # Check if rule references this tool
+            if hasattr(rule, 'references') and rule.references:
+                if hasattr(rule.references, 'tools') and rule.references.tools:
+                    if tool_name not in rule.references.tools:
+                        return False
+            
+            # Check trigger patterns if they exist
+            if hasattr(rule, 'trigger_pattern') and rule.trigger_pattern:
+                # For now, assume all tool_start rules cover their referenced tools
+                # This can be made more sophisticated later
+                pass
+            
+            return True
+        except Exception as e:
+            self._vprint(f"[DASEIN][COVERAGE] Error checking rule coverage: {e}")
+            return False
+    
+    def _fire_micro_turn_for_tool_rules(self, rules, tool_name: str, tool_input: str) -> str:
+        """Fire a micro-turn LLM call to apply tool rules."""
+        try:
+            # Use the first rule for now (can be extended to handle multiple rules)
+            rule = rules[0]
+            rule_id = getattr(rule, 'id', 'unknown')
+            
+            self._vprint(f"[DASEIN][MICROTURN] rule_id={rule_id} tool={tool_name}")
+            
+            # Create micro-turn prompt
+            micro_turn_prompt = self._create_micro_turn_prompt(rule, tool_name, tool_input)
+            
+            # Fire actual micro-turn LLM call
+            modified_input = self._execute_micro_turn_llm_call(micro_turn_prompt, tool_input)
+            
+            # Store the modified input for retrieval during tool execution
+            input_key = f"{tool_name}:{hash(tool_input)}"
+            _MODIFIED_TOOL_INPUTS[input_key] = modified_input
+            
+            self._vprint(f"[DASEIN][MICROTURN] Applied rule {rule_id}: {str(tool_input)[:50]}... -> {str(modified_input)[:50]}...")
+            
+            return modified_input
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][MICROTURN] Error in micro-turn: {e}")
+            return tool_input
+    
+    def _create_micro_turn_prompt(self, rule, tool_name: str, tool_input: str) -> str:
+        """Create the micro-turn prompt for rule application."""
+        advice = getattr(rule, 'advice', '')
+        return f"""Apply this rule to the tool input:
+
+Rule: {advice}
+Tool: {tool_name}
+Current Input: {tool_input}
+
+Output only the corrected tool input:"""
+    
+    def _execute_micro_turn_llm_call(self, prompt: str, original_input: str) -> str:
+        """Execute the actual micro-turn LLM call."""
+        try:
+            if not self._llm:
+                self._vprint(f"[DASEIN][MICROTURN] No LLM available for micro-turn call")
+                return original_input
+            
+            self._vprint(f"[DASEIN][MICROTURN] Executing micro-turn LLM call")
+            self._vprint(f"[DASEIN][MICROTURN] Prompt: {prompt[:200]}...")
+            
+            # Make the micro-turn LLM call
+            # Create a simple message list for the LLM
+            messages = [{"role": "user", "content": prompt}]
+            
+            # Call the LLM
+            response = self._llm.invoke(messages)
+            
+            # Extract the response content
+            if hasattr(response, 'content'):
+                modified_input = response.content.strip()
+            elif isinstance(response, str):
+                modified_input = response.strip()
+            else:
+                modified_input = str(response).strip()
+            
+            self._vprint(f"[DASEIN][MICROTURN] LLM response: {modified_input[:100]}...")
+            
+            # 🚨 CRITICAL: Parse JSON responses with markdown fences
+            if modified_input.startswith('```json') or modified_input.startswith('```'):
+                try:
+                    # Extract JSON from markdown fences
+                    import re
+                    import json
+                    json_match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', modified_input, re.DOTALL)
+                    if json_match:
+                        json_str = json_match.group(1)
+                        parsed_json = json.loads(json_str)
+                        # Convert back to the expected format
+                        if isinstance(parsed_json, dict) and 'name' in parsed_json and 'args' in parsed_json:
+                            modified_input = parsed_json
+                            self._vprint(f"[DASEIN][MICROTURN] Parsed JSON from markdown fences: {parsed_json}")
+                        else:
+                            self._vprint(f"[DASEIN][MICROTURN] JSON doesn't have expected structure, using as-is")
+                    else:
+                        self._vprint(f"[DASEIN][MICROTURN] Could not extract JSON from markdown fences")
+                except Exception as e:
+                    self._vprint(f"[DASEIN][MICROTURN] Error parsing JSON: {e}")
+            
+            # Validate the response - only fallback if completely empty
+            if not modified_input:
+                self._vprint(f"[DASEIN][MICROTURN] LLM response empty, using original input")
+                return original_input
+            
+            return modified_input
+            
+        except Exception as e:
+            self._vprint(f"[DASEIN][MICROTURN] Error executing micro-turn LLM call: {e}")
+            return original_input
+
+
+def get_trace() -> List[Dict[str, Any]]:
+    """
+    DEPRECATED: Legacy function for backward compatibility.
+    Get the current trace from active CognateProxy instances.
+    
+    Returns:
+        List of trace step dictionaries (empty if no active traces)
+    """
+    # Try to get trace from active CognateProxy instances
+    try:
+        import gc
+        for obj in gc.get_objects():
+            if hasattr(obj, '_last_run_trace') and obj._last_run_trace:
+                return obj._last_run_trace.copy()
+            if hasattr(obj, '_callback_handler') and hasattr(obj._callback_handler, '_trace'):
+                return obj._callback_handler._trace.copy()
+    except Exception:
+        pass
+    
+    return []  # Return empty list if no trace found
+
+
+def get_modified_tool_input(tool_name: str, original_input: str) -> str:
+    """
+    Get the modified tool input if it exists.
+    
+    Args:
+        tool_name: Name of the tool
+        original_input: Original tool input
+        
+    Returns:
+        Modified tool input if available, otherwise original input
+    """
+    input_key = f"{tool_name}:{hash(original_input)}"
+    return _MODIFIED_TOOL_INPUTS.get(input_key, original_input)
+
+
+def clear_modified_tool_inputs():
+    """Clear all modified tool inputs."""
+    global _MODIFIED_TOOL_INPUTS
+    _MODIFIED_TOOL_INPUTS.clear()
+
+
+def clear_trace() -> None:
+    """
+    DEPRECATED: Legacy function for backward compatibility.
+    Clear traces in active CognateProxy instances.
+    """
+    # Try to clear traces in active CognateProxy instances
+    try:
+        import gc
+        for obj in gc.get_objects():
+            if hasattr(obj, '_callback_handler') and hasattr(obj._callback_handler, 'reset_run_state'):
+                obj._callback_handler.reset_run_state()
+    except Exception:
+        pass  # Ignore if not available
+
+
+def print_trace(max_chars: int = 240, only: tuple[str, ...] | None = None, suppress: tuple[str, ...] = ("chain_end",), show_tree: bool = True, show_summary: bool = True) -> None:
+    """
+    Print a compact fixed-width table of the trace with tree-like view and filtering.
+    
+    Args:
+        max_chars: Maximum characters per line (default 240)
+        only: Filter by step_type if provided (e.g., ("llm_start", "llm_end"))
+        suppress: Suppress any step_type in this tuple (default: ("chain_end",))
+        show_tree: If True, left-pad args_excerpt by 2*depth spaces for tree-like view
+        show_summary: If True, show step_type counts and deduped rows summary
+    """
+    # Try to get trace from active CognateProxy instances
+    trace = None
+    try:
+        # Import here to avoid circular imports
+        from dasein.api import _global_cognate_proxy
+        if _global_cognate_proxy and hasattr(_global_cognate_proxy, '_wrapped_llm') and _global_cognate_proxy._wrapped_llm:
+            trace = _global_cognate_proxy._wrapped_llm.get_trace()
+    except:
+        pass
+    
+    if not trace:
+        trace = get_trace()  # Use the updated get_trace() function
+    
+    # If global trace is empty, try to get it from the last completed run
+    if not trace:
+        # Try to get trace from any active CognateProxy instances
+        try:
+            import gc
+            for obj in gc.get_objects():
+                # Look for CognateProxy instances with captured traces
+                if hasattr(obj, '_last_run_trace') and obj._last_run_trace:
+                    trace = obj._last_run_trace
+                    print(f"[DASEIN][TRACE] Retrieved trace from CognateProxy: {len(trace)} steps")
+                    break
+                # Fallback: try callback handler
+                elif hasattr(obj, '_callback_handler') and hasattr(obj._callback_handler, 'get_trace'):
+                    potential_trace = obj._callback_handler.get_trace()
+                    if potential_trace:
+                        trace = potential_trace
+                        print(f"[DASEIN][TRACE] Retrieved trace from callback handler: {len(trace)} steps")
+                        break
+        except Exception as e:
+            pass
+    
+    if not trace:
+        print("No trace data available.")
+        return
+    
+    # Print execution state if available
+    try:
+        from dasein.api import _global_cognate_proxy
+        if _global_cognate_proxy and hasattr(_global_cognate_proxy, '_callback_handler'):
+            handler = _global_cognate_proxy._callback_handler
+            if hasattr(handler, '_function_calls_made') and handler._function_calls_made:
+                print("\n" + "=" * 80)
+                print("EXECUTION STATE (Functions Called During Run):")
+                print("=" * 80)
+                for func_name in sorted(handler._function_calls_made.keys()):
+                    calls = handler._function_calls_made[func_name]
+                    count = len(calls)
+                    print(f"  • {func_name}: called {count}x")
+                    # Hybrid window: show all if ≤5, else show most recent 3 (matches injection logic)
+                    if count <= 5:
+                        # Show all calls
+                        for call in calls:
+                            preview = call.get('preview', '(no preview)')
+                            if len(preview) > 80:
+                                preview = preview[:80] + '...'
+                            print(f"      [step {call['step']}] {preview}")
+                    else:
+                        # Show recent 3
+                        print(f"      ... (showing most recent 3 of {count}):")
+                        for call in calls[-3:]:
+                            preview = call.get('preview', '(no preview)')
+                            if len(preview) > 80:
+                                preview = preview[:80] + '...'
+                            print(f"      [step {call['step']}] {preview}")
+                print("=" * 80 + "\n")
+    except Exception as e:
+        pass  # Silently skip if state not available
+    
+    # Filter by step_type if only is provided
+    filtered_trace = trace
+    if only:
+        filtered_trace = [step for step in trace if step.get("step_type") in only]
+    
+    # Suppress any step_type in suppress tuple
+    if suppress:
+        filtered_trace = [step for step in filtered_trace if step.get("step_type") not in suppress]
+    
+    if not filtered_trace:
+        print("No trace data matching filter criteria.")
+        return
+    
+    # Build depth map from parent_run_id
+    depth_map = {}
+    for step in filtered_trace:
+        run_id = step.get("run_id")
+        parent_run_id = step.get("parent_run_id")
+        
+        if run_id is None or parent_run_id is None or parent_run_id not in depth_map:
+            depth_map[run_id] = 0
+        else:
+            depth_map[run_id] = depth_map[parent_run_id] + 1
+    
+    # Calculate column widths based on max_chars
+    # Reserve space for: # (3), step_type (15), tool_name (25), separators (6)
+    available_width = max_chars - 3 - 15 - 25 - 6
+    excerpt_width = available_width // 2
+    outcome_width = available_width - excerpt_width
+    
+    # Print header
+    print(f"{'#':<3} {'step_type':<15} {'tool_name':<25} {'args_excerpt':<{excerpt_width}} {'outcome':<{outcome_width}}")
+    print("-" * max_chars)
+    
+    # Print each step
+    for i, step in enumerate(filtered_trace, 1):
+        step_type = step.get("step_type", "")[:15]
+        tool_name = str(step.get("tool_name", ""))[:25]
+        args_excerpt = step.get("args_excerpt", "")
+        outcome = step.get("outcome", "")
+        
+        # Apply tree indentation if show_tree is True
+        if show_tree:
+            run_id = step.get("run_id")
+            depth = depth_map.get(run_id, 0)
+            args_excerpt = "  " * depth + args_excerpt
+        
+        # Truncate to fit column widths
+        args_excerpt = args_excerpt[:excerpt_width]
+        outcome = outcome[:outcome_width]
+        
+        print(f"{i:<3} {step_type:<15} {tool_name:<25} {args_excerpt:<{excerpt_width}} {outcome:<{outcome_width}}")
+    
+    # Show summary if requested
+    if show_summary:
+        print("\n" + "=" * max_chars)
+        
+        # Count steps by step_type
+        step_counts = {}
+        for step in filtered_trace:
+            step_type = step.get("step_type", "unknown")
+            step_counts[step_type] = step_counts.get(step_type, 0) + 1
+        
+        print("Step counts:")
+        for step_type, count in sorted(step_counts.items()):
+            print(f"  {step_type}: {count}")
+        
+        # Add compact function call summary
+        try:
+            from dasein.api import _global_cognate_proxy
+            if _global_cognate_proxy and hasattr(_global_cognate_proxy, '_callback_handler'):
+                handler = _global_cognate_proxy._callback_handler
+                if hasattr(handler, '_function_calls_made') and handler._function_calls_made:
+                    print("\nFunction calls:")
+                    for func_name in sorted(handler._function_calls_made.keys()):
+                        count = len(handler._function_calls_made[func_name])
+                        print(f"  {func_name}: {count}")
+        except Exception:
+            pass
+        
+        # Count deduped rows skipped (steps that were filtered out)
+        total_steps = len(trace)
+        shown_steps = len(filtered_trace)
+        skipped_steps = total_steps - shown_steps
+        
+        if skipped_steps > 0:
+            print(f"Deduped rows skipped: {skipped_steps}")