""" Context Management for AgentPress Threads. This module handles token counting and thread summarization to prevent reaching the context window limitations of LLM models. """ import json import os import asyncio from typing import List, Dict, Any, Optional, Union from litellm.utils import token_counter from anthropic import Anthropic from core.services.supabase import DBConnection from core.utils.logger import logger from core.ai_models import model_manager from core.agentpress.prompt_caching import apply_anthropic_caching_strategy, supports_prompt_caching DEFAULT_TOKEN_THRESHOLD = 120000 # Module-level singleton clients for memory efficiency # These are lazily initialized once and reused across all ContextManager instances _anthropic_client = None _bedrock_client = None _clients_initialized = False def _get_anthropic_client_singleton(): """Module-level lazy initialization of Anthropic client (singleton).""" global _anthropic_client, _clients_initialized if _anthropic_client is None and not _clients_initialized: api_key = os.environ.get("ANTHROPIC_API_KEY") if api_key: _anthropic_client = Anthropic(api_key=api_key) _clients_initialized = True return _anthropic_client def _get_bedrock_client_singleton(): """Module-level lazy initialization of Bedrock client (singleton).""" global _bedrock_client if _bedrock_client is None: try: import boto3 _bedrock_client = boto3.client('bedrock-runtime', region_name='us-west-2') except Exception as e: logger.debug(f"Could not initialize Bedrock client: {e}") return _bedrock_client class ContextManager: """Manages thread context including token counting and summarization.""" def __init__(self, token_threshold: int = DEFAULT_TOKEN_THRESHOLD, db=None): """Initialize the ContextManager. Args: token_threshold: Token count threshold to trigger summarization db: Optional DBConnection instance to reuse (avoids creating new ones) """ self.db = db if db is not None else DBConnection() self.token_threshold = token_threshold # Tool output management self.keep_recent_tool_outputs = 5 # Number of recent tool outputs to preserve # Compression strategy self.compression_target_ratio = 0.6 # Compress to 60% of max tokens (hysteresis) self.keep_recent_user_messages = 10 # Number of recent user messages to keep uncompressed self.keep_recent_assistant_messages = 10 # Number of recent assistant messages to keep uncompressed def _get_anthropic_client(self): """Get the singleton Anthropic client.""" return _get_anthropic_client_singleton() async def save_compressed_messages(self, compressed_messages: List[Dict[str, Any]]) -> int: """Save compressed message content to database metadata. Updates messages in DB with: - metadata.compressed = True - metadata.compressed_content = the compressed content This allows future reads to use compressed content directly without re-compressing. Always updates compressed_content when provided - allows tiered compression to upgrade already-compressed messages to more aggressive compression. Args: compressed_messages: List of dicts with 'message_id' and 'compressed_content' Returns: Number of messages successfully saved """ if not compressed_messages: return 0 from core.threads import repo as threads_repo saved_count = await threads_repo.save_compressed_messages_batch(compressed_messages) if saved_count > 0: logger.info(f"πŸ’Ύ Compression save: {saved_count} messages saved") return saved_count def _get_bedrock_client(self): """Get the singleton Bedrock client.""" return _get_bedrock_client_singleton() async def count_tokens(self, model: str, messages: List[Dict[str, Any]], system_prompt: Optional[Dict[str, Any]] = None, apply_caching: bool = True) -> int: """Count tokens using the correct tokenizer for the model. For Anthropic/Claude models: Uses Anthropic's official tokenizer For Bedrock models: Uses Bedrock's count_tokens API For other models: Uses LiteLLM's token_counter IMPORTANT: By default, applies caching transformation before counting to match the actual token count that will be sent to the API. Args: model: Model name messages: List of messages system_prompt: Optional system prompt apply_caching: If True, temporarily apply caching transformation before counting Returns: Token count (with caching overhead if apply_caching=True) """ # Apply caching transformation if requested (to match API reality) messages_to_count = messages system_to_count = system_prompt if apply_caching and supports_prompt_caching(model): try: # Temporarily apply caching transformation prepared = await apply_anthropic_caching_strategy( system_prompt, messages, model, thread_id=None, force_recalc=False ) # Separate system from messages system_to_count = None messages_to_count = [] for msg in prepared: if msg.get('role') == 'system': system_to_count = msg else: messages_to_count.append(msg) except Exception as e: logger.debug(f"Failed to apply caching for counting: {e}") # Continue with uncached messages # Check if this is an Anthropic model if 'claude' in model.lower() or 'anthropic' in model.lower(): # Use Anthropic's official tokenizer try: client = self._get_anthropic_client() if client: # Strip provider prefix clean_model = model.split('/')[-1] if '/' in model else model # Clean messages - only role and content clean_messages = [] for msg in messages_to_count: if msg.get('role') == 'system': continue # System passed separately clean_messages.append({ 'role': msg.get('role'), 'content': msg.get('content') }) # Extract system content system_content = None if system_to_count and isinstance(system_to_count, dict): system_content = system_to_count.get('content') # Build parameters count_params = {'model': clean_model, 'messages': clean_messages} if system_content: count_params['system'] = system_content result = client.messages.count_tokens(**count_params) return result.input_tokens except Exception as e: logger.debug(f"Anthropic token counting failed, falling back to LiteLLM: {e}") # Check if this is a Bedrock model elif 'bedrock' in model.lower(): try: bedrock_client = self._get_bedrock_client() if bedrock_client: # Import profile IDs from core.ai_models.registry import KIMI_K2_PROFILE_ID, SONNET_4_5_PROFILE_ID, HAIKU_4_5_PROFILE_ID, MINIMAX_M2_PROFILE_ID model_id_mapping = { KIMI_K2_PROFILE_ID: "moonshot.kimi-k2-thinking", # Kimi K2 (Basic Mode) SONNET_4_5_PROFILE_ID: "us.anthropic.claude-sonnet-4-5-20250929-v1:0", # Sonnet 4.5 (Power Mode) "tyj1ks3nj9qf": "anthropic.claude-sonnet-4-20250514-v1:0", # Sonnet 4 (no constant defined) HAIKU_4_5_PROFILE_ID: "anthropic.claude-haiku-4-5-20251001-v1:0", # HAIKU 4.5 (Legacy - replaced by Kimi K2) MINIMAX_M2_PROFILE_ID: "minimax.minimax-m2", # MiniMax M2 } # Extract profile ID from ARN bedrock_model_id = None if "application-inference-profile" in model: profile_id = model.split("/")[-1] bedrock_model_id = model_id_mapping.get(profile_id) if not bedrock_model_id: # Import Kimi K2 profile ID for default from core.ai_models.registry import KIMI_K2_PROFILE_ID bedrock_model_id = model_id_mapping.get(KIMI_K2_PROFILE_ID, "moonshot.kimi-k2-thinking") # Default to Kimi K2 # Clean content blocks for Bedrock Converse API def clean_content_for_bedrock(content): """ Convert Anthropic format to Bedrock Converse API format. Converts cache_control -> cachePoint to preserve cache overhead in token counts. """ if isinstance(content, str): return [{'text': content}] elif isinstance(content, list): cleaned = [] for block in content: if isinstance(block, dict): # Extract text if 'text' in block: cleaned.append({'text': block['text']}) # Convert cache_control to cachePoint (separate block) if 'cache_control' in block: cleaned.append({'cachePoint': {'type': 'default'}}) return cleaned if cleaned else [{'text': str(content)}] return [{'text': str(content)}] # Format messages for Bedrock bedrock_messages = [] system_content = None for msg in messages_to_count: if msg.get('role') == 'system': system_content = clean_content_for_bedrock(msg.get('content')) continue bedrock_messages.append({ 'role': msg.get('role'), 'content': clean_content_for_bedrock(msg.get('content')) }) # Build input input_to_count = {'messages': bedrock_messages} if system_content: input_to_count['system'] = system_content elif system_to_count: input_to_count['system'] = clean_content_for_bedrock(system_to_count.get('content')) # Call Bedrock count_tokens API response = bedrock_client.count_tokens( modelId=bedrock_model_id, input={'converse': input_to_count} ) return response['inputTokens'] except Exception as e: logger.debug(f"Bedrock token counting failed, falling back to LiteLLM: {e}") # Fallback to LiteLLM token_counter (wrap in thread pool - CPU-heavy tiktoken operation) if system_to_count: return await asyncio.to_thread(token_counter, model=model, messages=[system_to_count] + messages_to_count) else: return await asyncio.to_thread(token_counter, model=model, messages=messages_to_count) async def estimate_token_usage(self, prompt_messages: List[Dict[str, Any]], completion_content: str, model: str) -> Dict[str, Any]: """ Estimate token usage for billing when exact usage is unavailable. This is critical for billing on timeouts, crashes, disconnects, etc. Uses provider-specific APIs (Anthropic/Bedrock) when available for accuracy, with fallbacks to LiteLLM token_counter and word count estimation. Args: prompt_messages: The prompt messages sent to the LLM completion_content: The accumulated completion text model: Model name Returns: Dict with prompt_tokens, completion_tokens, total_tokens, estimated=True """ try: # Count prompt tokens using accurate provider APIs prompt_tokens = await self.count_tokens(model, prompt_messages, apply_caching=False) # Count completion tokens (just the text) - wrap in thread pool (CPU-heavy tiktoken operation) completion_tokens = 0 if completion_content: completion_tokens = await asyncio.to_thread(token_counter, model=model, text=completion_content) total_tokens = prompt_tokens + completion_tokens logger.warning(f"⚠️ ESTIMATED TOKEN USAGE: prompt={prompt_tokens}, completion={completion_tokens}, total={total_tokens}") return { "prompt_tokens": prompt_tokens, "completion_tokens": completion_tokens, "total_tokens": total_tokens, "estimated": True } except Exception as e: logger.error(f"Context manager estimation failed: {e}, falling back to LiteLLM") # Fallback to LiteLLM try: prompt_tokens = token_counter(model=model, messages=prompt_messages) completion_tokens = token_counter(model=model, text=completion_content) if completion_content else 0 logger.warning(f"⚠️ ESTIMATED TOKEN USAGE (LiteLLM): prompt={prompt_tokens}, completion={completion_tokens}") return { "prompt_tokens": prompt_tokens, "completion_tokens": completion_tokens, "total_tokens": prompt_tokens + completion_tokens, "estimated": True } except Exception as e2: logger.error(f"LiteLLM estimation failed: {e2}, using word count fallback") # Final fallback to word count fallback_prompt = len(' '.join(str(m.get('content', '')) for m in prompt_messages).split()) * 1.3 fallback_completion = len(completion_content.split()) * 1.3 if completion_content else 0 logger.warning(f"⚠️ FALLBACK TOKEN ESTIMATION: promptβ‰ˆ{int(fallback_prompt)}, completionβ‰ˆ{int(fallback_completion)}") return { "prompt_tokens": int(fallback_prompt), "completion_tokens": int(fallback_completion), "total_tokens": int(fallback_prompt + fallback_completion), "estimated": True, "fallback": True } def is_tool_result_message(self, msg: Dict[str, Any]) -> bool: """Check if a message is a tool result message. Detects tool results from: 1. Native tool calls: role="tool" 2. Native tool calls: has tool_call_id field 3. XML tool calls: role="user" ONLY if explicitly marked in metadata """ if not isinstance(msg, dict): return False # Native tool calls have role="tool" if msg.get('role') == 'tool': return True # Native tool calls have tool_call_id at top level if 'tool_call_id' in msg: return True # CONSERVATIVE CHECK: For role="user", only consider it a tool result if: # 1. It has explicit tool_call_id in metadata (indicating it's a saved tool result) # This prevents false positives that break message validation if msg.get('role') == 'user': metadata = msg.get('metadata', {}) if isinstance(metadata, dict) and 'tool_call_id' in metadata: return True return False def get_tool_call_ids_from_message(self, msg: Dict[str, Any]) -> List[str]: """Extract tool_call IDs from an assistant message with tool_calls. Returns list of tool_call IDs, or empty list if no tool_calls. """ if not isinstance(msg, dict) or msg.get('role') != 'assistant': return [] tool_calls = msg.get('tool_calls') or [] if not tool_calls or not isinstance(tool_calls, list): return [] ids = [] for tc in tool_calls: if isinstance(tc, dict): tc_id = tc.get('id') if tc_id: ids.append(tc_id) return ids def get_tool_call_id_from_result(self, msg: Dict[str, Any]) -> Optional[str]: """Extract the tool_call_id from a tool result message. Returns the tool_call_id, or None if not a tool result message. """ if not isinstance(msg, dict): return None # Native tool results have tool_call_id directly at top level if 'tool_call_id' in msg: return msg.get('tool_call_id') # role="tool" messages should have tool_call_id if msg.get('role') == 'tool': return msg.get('tool_call_id') # Check metadata for tool_call_id (XML tool results or saved tool results) metadata = msg.get('metadata', {}) if isinstance(metadata, dict) and 'tool_call_id' in metadata: return metadata.get('tool_call_id') return None def group_messages_by_tool_calls(self, messages: List[Dict[str, Any]]) -> List[List[Dict[str, Any]]]: """Group messages into atomic units respecting tool call pairing. CRITICAL: This ensures assistant messages with tool_calls are always grouped with their corresponding tool result messages. These groups must be treated as atomic units that cannot be split during compression or caching. Rules: - An assistant message with tool_calls + all following tool messages with matching tool_call_ids = one atomic group - Regular messages (user, assistant without tool_calls) = standalone groups - Tool results without a preceding assistant (orphans) = standalone groups (with warning) Args: messages: List of conversation messages Returns: List of message groups, where each group is a list of messages """ if not messages: return [] groups: List[List[Dict[str, Any]]] = [] current_group: List[Dict[str, Any]] = [] expected_tool_call_ids: set = set() # IDs we're waiting for results for msg in messages: role = msg.get('role', '') # Check if this is an assistant message with tool_calls tool_call_ids = self.get_tool_call_ids_from_message(msg) if tool_call_ids: # If we have a pending group, save it first if current_group: groups.append(current_group) # Start a new group with this assistant message current_group = [msg] expected_tool_call_ids = set(tool_call_ids) elif self.is_tool_result_message(msg): # This is a tool result message tool_call_id = self.get_tool_call_id_from_result(msg) if tool_call_id and tool_call_id in expected_tool_call_ids: # This tool result belongs to the current group current_group.append(msg) expected_tool_call_ids.discard(tool_call_id) # If we've received all expected tool results, close the group if not expected_tool_call_ids: groups.append(current_group) current_group = [] else: # Orphaned tool result - doesn't match any expected ID # Log warning and treat as standalone if tool_call_id: logger.warning(f"⚠️ Orphaned tool result detected: tool_call_id={tool_call_id} has no matching assistant message") # Close current group if any if current_group: groups.append(current_group) current_group = [] expected_tool_call_ids = set() # Add orphan as standalone group groups.append([msg]) else: # Regular message (user or assistant without tool_calls) # Close current group if we have pending tool calls if current_group: if expected_tool_call_ids: logger.warning(f"⚠️ Closing tool call group with {len(expected_tool_call_ids)} missing tool results") groups.append(current_group) current_group = [] expected_tool_call_ids = set() # Add as standalone group groups.append([msg]) # Don't forget the last group if current_group: if expected_tool_call_ids: logger.warning(f"⚠️ Final group has {len(expected_tool_call_ids)} missing tool results") groups.append(current_group) return groups def flatten_message_groups(self, groups: List[List[Dict[str, Any]]]) -> List[Dict[str, Any]]: """Flatten message groups back into a flat list of messages. Args: groups: List of message groups Returns: Flat list of messages preserving order """ result = [] for group in groups: result.extend(group) return result def validate_tool_call_pairing(self, messages: List[Dict[str, Any]]) -> tuple[bool, List[str], List[str]]: """Validate that tool calls and tool results are properly paired in BOTH directions. Bedrock requires: 1. Every tool result must have a preceding assistant message with matching tool_call_id 2. Every assistant tool_call must have a following tool result with matching tool_call_id This should be called before sending messages to the LLM to catch any issues that would cause Bedrock errors. Args: messages: List of messages to validate Returns: Tuple of (is_valid, orphaned_tool_result_ids, unanswered_tool_call_ids) """ # Track all tool_call IDs from assistant messages all_tool_call_ids: set = set() # Track all tool_call_ids that have been answered with tool results answered_tool_call_ids: set = set() # Track orphaned tool results (results without matching assistant) orphaned_tool_result_ids: List[str] = [] # First pass: collect all tool_call IDs from assistant messages for msg in messages: tool_call_ids = self.get_tool_call_ids_from_message(msg) all_tool_call_ids.update(tool_call_ids) # Second pass: check tool results and track which tool_calls are answered for msg in messages: if self.is_tool_result_message(msg): tool_call_id = self.get_tool_call_id_from_result(msg) if tool_call_id: if tool_call_id not in all_tool_call_ids: # Tool result without matching assistant orphaned_tool_result_ids.append(tool_call_id) else: # This tool_call has been answered answered_tool_call_ids.add(tool_call_id) # Find unanswered tool_calls (assistant tool_calls without matching tool results) unanswered_tool_call_ids = list(all_tool_call_ids - answered_tool_call_ids) is_valid = len(orphaned_tool_result_ids) == 0 and len(unanswered_tool_call_ids) == 0 if orphaned_tool_result_ids: logger.error(f"🚨 VALIDATION FAILED: {len(orphaned_tool_result_ids)} orphaned tool results (no matching assistant): {orphaned_tool_result_ids}") if unanswered_tool_call_ids: logger.error(f"🚨 VALIDATION FAILED: {len(unanswered_tool_call_ids)} unanswered tool calls (no matching tool result): {unanswered_tool_call_ids}") return is_valid, orphaned_tool_result_ids, unanswered_tool_call_ids def needs_tool_ordering_repair(self, messages: List[Dict[str, Any]]) -> bool: """Fast O(n) check if messages have tool ordering issues. Exits early in common case. This is a lightweight pre-check that detects if user messages interrupt the tool_use -> tool_result flow. Only returns True if a problem is detected, allowing callers to skip expensive full validation in the common case. Args: messages: List of messages to check Returns: True if tool ordering issues detected (needs repair), False otherwise """ for i, msg in enumerate(messages): if not isinstance(msg, dict): continue # Check if this is an assistant with tool_calls tool_call_ids = self.get_tool_call_ids_from_message(msg) if not tool_call_ids: continue # Check if immediately following messages are the expected tool results expected_count = len(tool_call_ids) found_tool_call_ids = set() for j in range(i + 1, min(i + 1 + expected_count + 1, len(messages))): next_msg = messages[j] if j < len(messages) else None if not next_msg: # Missing tool results - needs repair return True # If we hit a user message (not tool result) before all results, needs repair if next_msg.get('role') == 'user' and not self.is_tool_result_message(next_msg): return True # Check if it's a tool result for one of our tool_calls tc_id = self.get_tool_call_id_from_result(next_msg) if tc_id and tc_id in tool_call_ids: found_tool_call_ids.add(tc_id) elif not self.is_tool_result_message(next_msg): # Non-tool-result message interrupting the flow return True # If we didn't find all expected tool results, needs repair if len(found_tool_call_ids) < expected_count: return True return False # No issues found def remove_orphaned_tool_results(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """Remove orphaned tool results that have no matching assistant message. This is a repair function to fix invalid message structures before sending to LLM. Args: messages: List of messages Returns: Messages with orphaned tool results removed """ # First pass: collect all valid tool_call IDs from assistant messages valid_tool_call_ids: set = set() for msg in messages: tool_call_ids = self.get_tool_call_ids_from_message(msg) valid_tool_call_ids.update(tool_call_ids) # Second pass: filter out orphaned tool results result = [] removed_count = 0 for msg in messages: if self.is_tool_result_message(msg): tool_call_id = self.get_tool_call_id_from_result(msg) if tool_call_id and tool_call_id not in valid_tool_call_ids: logger.warning(f"πŸ—‘οΈ Removing orphaned tool result: tool_call_id={tool_call_id}") removed_count += 1 continue result.append(msg) if removed_count > 0: logger.info(f"πŸ”§ Removed {removed_count} orphaned tool results to fix message structure") return result def remove_unanswered_tool_calls(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """Remove or fix assistant messages with tool_calls that have no matching tool results. This is a repair function to fix invalid message structures before sending to LLM. If an assistant message has ONLY unanswered tool_calls and no content, it's removed. If it has content, the tool_calls are removed but content is preserved. Args: messages: List of messages Returns: Messages with unanswered tool_calls fixed """ # First pass: collect all tool_call_ids that have results answered_tool_call_ids: set = set() for msg in messages: if self.is_tool_result_message(msg): tool_call_id = self.get_tool_call_id_from_result(msg) if tool_call_id: answered_tool_call_ids.add(tool_call_id) # Second pass: fix assistant messages with unanswered tool_calls result = [] fixed_count = 0 removed_count = 0 for msg in messages: tool_call_ids = self.get_tool_call_ids_from_message(msg) if tool_call_ids: # Check which tool_calls are unanswered unanswered = [tc_id for tc_id in tool_call_ids if tc_id not in answered_tool_call_ids] if unanswered: # Some tool_calls don't have results answered = [tc_id for tc_id in tool_call_ids if tc_id in answered_tool_call_ids] content = msg.get('content', '') has_content = bool(content and str(content).strip()) if not answered and not has_content: # All tool_calls are unanswered and no content - remove the message entirely logger.warning(f"πŸ—‘οΈ Removing assistant message with {len(unanswered)} unanswered tool_calls and no content: {unanswered}") removed_count += 1 continue elif not answered and has_content: # All tool_calls are unanswered but has content - keep content, remove tool_calls fixed_msg = msg.copy() fixed_msg.pop('tool_calls', None) logger.warning(f"πŸ”§ Removing {len(unanswered)} unanswered tool_calls from assistant message (keeping content): {unanswered}") result.append(fixed_msg) fixed_count += 1 continue else: # Some tool_calls are answered - keep only answered ones fixed_msg = msg.copy() original_tool_calls = fixed_msg.get('tool_calls') or [] fixed_msg['tool_calls'] = [ tc for tc in original_tool_calls if isinstance(tc, dict) and tc.get('id') in answered_tool_call_ids ] if isinstance(original_tool_calls, list) else [] logger.warning(f"πŸ”§ Removed {len(unanswered)} unanswered tool_calls from assistant message (kept {len(answered)}): {unanswered}") result.append(fixed_msg) fixed_count += 1 continue result.append(msg) if fixed_count > 0 or removed_count > 0: logger.info(f"πŸ”§ Fixed {fixed_count} assistant messages, removed {removed_count} to fix unanswered tool_calls") return result def repair_tool_call_pairing(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """Repair both directions of tool call pairing issues. This handles: 1. Orphaned tool results (tool results without matching assistant) - removed 2. Unanswered tool calls (assistant tool_calls without matching results) - removed/fixed Args: messages: List of messages Returns: Messages with all tool call pairing issues fixed """ # Fix orphaned tool results first result = self.remove_orphaned_tool_results(messages) # Then fix unanswered tool calls result = self.remove_unanswered_tool_calls(result) # Validate the result is_valid, orphaned, unanswered = self.validate_tool_call_pairing(result) if not is_valid: logger.error(f"🚨 CRITICAL: Could not fully repair message structure. Orphaned: {len(orphaned)}, Unanswered: {len(unanswered)}") else: logger.info(f"βœ… Message structure successfully repaired") return result def validate_tool_call_ordering(self, messages: List[Dict[str, Any]]) -> tuple[bool, List[str], List[str]]: """Validate that tool results immediately follow their assistant tool_calls. Some LLM providers (like Minimax) require tool results to appear right after the assistant message that made the tool_call, not separated by other messages. Args: messages: List of messages to validate Returns: Tuple of (is_valid, out_of_order_tool_call_ids, out_of_order_tool_result_ids) """ out_of_order_tool_call_ids: List[str] = [] out_of_order_tool_result_ids: List[str] = [] for i, msg in enumerate(messages): if not isinstance(msg, dict): continue # Get tool_calls from this message tool_call_ids = self.get_tool_call_ids_from_message(msg) if not tool_call_ids: continue # Check that the next N messages are tool results for these tool_calls expected_count = len(tool_call_ids) found_tool_call_ids = set() for j in range(i + 1, min(i + 1 + expected_count, len(messages))): next_msg = messages[j] if not isinstance(next_msg, dict): break tc_id = self.get_tool_call_id_from_result(next_msg) if tc_id and tc_id in tool_call_ids: found_tool_call_ids.add(tc_id) else: break # Non-matching message found # Any tool_calls not found in the immediate next messages are out-of-order missing = set(tool_call_ids) - found_tool_call_ids if missing: out_of_order_tool_call_ids.extend(missing) # Now find the tool results for these out-of-order tool_calls out_of_order_set = set(out_of_order_tool_call_ids) for msg in messages: if not isinstance(msg, dict): continue tc_id = self.get_tool_call_id_from_result(msg) if tc_id and tc_id in out_of_order_set: out_of_order_tool_result_ids.append(tc_id) is_valid = len(out_of_order_tool_call_ids) == 0 if not is_valid: logger.error(f"🚨 ORDERING VALIDATION FAILED: {len(out_of_order_tool_call_ids)} tool_calls have delayed results: {out_of_order_tool_call_ids}") return is_valid, out_of_order_tool_call_ids, out_of_order_tool_result_ids def remove_out_of_order_tool_pairs(self, messages: List[Dict[str, Any]], out_of_order_ids: List[str]) -> List[Dict[str, Any]]: """Remove assistant tool_calls and their delayed tool results that are out of order. Args: messages: List of messages out_of_order_ids: List of tool_call_ids that are out of order Returns: Messages with out-of-order tool pairs removed/fixed """ if not out_of_order_ids: return messages out_of_order_set = set(out_of_order_ids) result = [] removed_count = 0 for msg in messages: if not isinstance(msg, dict): result.append(msg) continue # Remove tool results for out-of-order tool_calls tc_id = self.get_tool_call_id_from_result(msg) if tc_id and tc_id in out_of_order_set: logger.warning(f"πŸ—‘οΈ Removing out-of-order tool result: tool_call_id={tc_id}") removed_count += 1 continue # Remove tool_calls from assistant messages that are out of order tool_call_ids = self.get_tool_call_ids_from_message(msg) if tool_call_ids: remaining_ids = [tc_id for tc_id in tool_call_ids if tc_id not in out_of_order_set] removed_ids = [tc_id for tc_id in tool_call_ids if tc_id in out_of_order_set] if removed_ids: logger.warning(f"πŸ—‘οΈ Removing out-of-order tool_calls: {removed_ids}") removed_count += len(removed_ids) if not remaining_ids: # All tool_calls were removed, check if message has content content = msg.get('content') if content and content != []: # Keep the message but remove tool_calls new_msg = {k: v for k, v in msg.items() if k != 'tool_calls'} result.append(new_msg) else: # No content, skip the entire message continue else: # Keep only the remaining tool_calls new_msg = msg.copy() new_tool_calls = [tc for tc in msg.get('tool_calls', []) if tc.get('id') not in out_of_order_set] new_msg['tool_calls'] = new_tool_calls result.append(new_msg) continue result.append(msg) if removed_count > 0: logger.info(f"πŸ”§ Removed {removed_count} out-of-order tool call/result pairs") return result def strip_all_tool_content_as_fallback(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """EMERGENCY FALLBACK: Strip all tool-related content if repair completely fails. This is a last-resort method that removes: 1. All tool result messages (role="tool") 2. All tool_calls from assistant messages 3. Any message with tool_call_id This ensures the LLM can at least continue with a clean message history, even if it means losing tool call context. Args: messages: List of messages Returns: Messages with all tool content stripped """ logger.warning("🚨 EMERGENCY FALLBACK: Stripping all tool content from messages") result = [] stripped_count = 0 for msg in messages: if not isinstance(msg, dict): result.append(msg) continue # Skip tool result messages entirely if msg.get('role') == 'tool' or 'tool_call_id' in msg: stripped_count += 1 continue # For assistant messages, remove tool_calls if msg.get('role') == 'assistant' and msg.get('tool_calls'): clean_msg = msg.copy() del clean_msg['tool_calls'] # Keep the message if it has content if clean_msg.get('content'): result.append(clean_msg) else: stripped_count += 1 continue # Keep all other messages result.append(msg) logger.warning(f"🚨 EMERGENCY FALLBACK: Stripped {stripped_count} tool-related messages") return result def remove_old_tool_outputs( self, messages: List[Dict[str, Any]], keep_last_n: int = 8 ) -> List[Dict[str, Any]]: """Compress old tool output messages IN-MEMORY, keeping only the most recent N uncompressed. CRITICAL: This method compresses CONTENT only - it never removes messages. This preserves the tool_call_id field and maintains the assistant+tool_result pairing required by Bedrock. This is a pure in-memory operation with no side effects. Args: messages: List of conversation messages keep_last_n: Number of most recent tool outputs to preserve uncompressed Returns: Messages with old tool outputs' content replaced by summaries (structure intact) """ if not messages: return messages # First, validate input has valid tool call pairing is_valid, orphaned_ids, unanswered_ids = self.validate_tool_call_pairing(messages) if not is_valid: logger.warning(f"⚠️ Input to remove_old_tool_outputs has pairing issues (orphaned: {len(orphaned_ids)}, unanswered: {len(unanswered_ids)}) - repairing first") messages = self.repair_tool_call_pairing(messages) # Group messages to understand which tool results belong to which assistant messages message_groups = self.group_messages_by_tool_calls(messages) # Flatten to get back to list form for processing # But now we understand the structure # First pass: identify tool result messages and their positions tool_result_positions = [] for i, msg in enumerate(messages): if self.is_tool_result_message(msg): tool_result_positions.append(i) total_tool_results = len(tool_result_positions) if total_tool_results <= keep_last_n: # No compression needed logger.debug(f"Only {total_tool_results} tool outputs found, keeping all uncompressed (threshold: {keep_last_n})") return messages # Calculate how many to compress (oldest ones) num_to_compress = total_tool_results - keep_last_n positions_to_compress = set(tool_result_positions[:num_to_compress]) logger.debug(f"Compressing {num_to_compress} old tool outputs in-memory (keeping last {keep_last_n} of {total_tool_results} uncompressed)") # Second pass: compress old tool outputs' CONTENT (keep structure intact) result = [] for i, msg in enumerate(messages): if i in positions_to_compress: # Compress content but PRESERVE all other fields (especially tool_call_id!) message_id = msg.get('message_id', 'unknown') tool_call_id = self.get_tool_call_id_from_result(msg) summary_content = f"[Tool output compressed for token management] message_id: \"{message_id}\". Use expand-message tool to view full output." # Copy message and only replace content - preserve tool_call_id, role, name, etc. compressed_msg = msg.copy() compressed_msg['content'] = summary_content result.append(compressed_msg) logger.debug(f"Compressed tool output at position {i}, tool_call_id={tool_call_id}") else: result.append(msg) # Final validation - structure should still be valid is_valid_after, orphaned_after, unanswered_after = self.validate_tool_call_pairing(result) if not is_valid_after: logger.error(f"🚨 BUG: remove_old_tool_outputs broke tool call pairing! Orphaned: {orphaned_after}, Unanswered: {unanswered_after}") return result def compress_user_messages_in_memory( self, messages: List[Dict[str, Any]], keep_last_n: int = 10 ) -> List[Dict[str, Any]]: """Compress user messages IN-MEMORY, keeping only the most recent N uncompressed. Args: messages: List of conversation messages keep_last_n: Number of most recent user messages to preserve Returns: Messages with old user messages compressed """ if not messages: return messages # Find user message positions user_positions = [] for i, msg in enumerate(messages): if isinstance(msg, dict) and msg.get('role') == 'user': user_positions.append(i) total_user_messages = len(user_positions) if total_user_messages <= keep_last_n: return messages # Positions to compress (all except last N) num_to_compress = total_user_messages - keep_last_n positions_to_compress = user_positions[:num_to_compress] logger.debug(f"Compressing {num_to_compress} user messages in-memory (keeping last {keep_last_n})") # Compress old user messages result = [] for i, msg in enumerate(messages): if i in positions_to_compress: original_content = msg.get('content', '') if isinstance(original_content, str) and len(original_content) > 3000: summary = original_content[:3000] + "... (truncated)" compressed_msg = msg.copy() compressed_msg['content'] = summary result.append(compressed_msg) else: result.append(msg) else: result.append(msg) return result def compress_assistant_messages_in_memory( self, messages: List[Dict[str, Any]], keep_last_n: int = 10 ) -> List[Dict[str, Any]]: """Compress assistant messages IN-MEMORY, keeping only the most recent N uncompressed. Args: messages: List of conversation messages keep_last_n: Number of most recent assistant messages to preserve Returns: Messages with old assistant messages compressed """ if not messages: return messages # Find assistant message positions assistant_positions = [] for i, msg in enumerate(messages): if isinstance(msg, dict) and msg.get('role') == 'assistant': assistant_positions.append(i) total_assistant_messages = len(assistant_positions) if total_assistant_messages <= keep_last_n: return messages # Positions to compress (all except last N) num_to_compress = total_assistant_messages - keep_last_n positions_to_compress = assistant_positions[:num_to_compress] logger.debug(f"Compressing {num_to_compress} assistant messages in-memory (keeping last {keep_last_n})") # Compress old assistant messages result = [] for i, msg in enumerate(messages): if i in positions_to_compress: original_content = msg.get('content', '') if isinstance(original_content, str) and len(original_content) > 3000: summary = original_content[:3000] + "... (truncated)" compressed_msg = msg.copy() compressed_msg['content'] = summary result.append(compressed_msg) else: result.append(msg) else: result.append(msg) return result def compress_message(self, msg_content: Union[str, dict], message_id: Optional[str] = None, max_length: int = 3000) -> Union[str, dict]: """Compress the message content.""" if isinstance(msg_content, str): if len(msg_content) > max_length: return msg_content[:max_length] + "... (truncated)" + f"\n\nmessage_id \"{message_id}\"\nUse expand-message tool to see contents" else: return msg_content def safe_truncate(self, msg_content: Union[str, dict], max_length: int = 100000) -> Union[str, dict]: """Truncate the message content safely by removing the middle portion.""" max_length = min(max_length, 100000) if isinstance(msg_content, str): if len(msg_content) > max_length: # Calculate how much to keep from start and end keep_length = max_length - 150 # Reserve space for truncation message start_length = keep_length // 2 end_length = keep_length - start_length start_part = msg_content[:start_length] end_part = msg_content[-end_length:] if end_length > 0 else "" return start_part + f"\n\n... (middle truncated) ...\n\n" + end_part + f"\n\nThis message is too long, repeat relevant information in your response to remember it" else: return msg_content elif isinstance(msg_content, dict): json_str = json.dumps(msg_content) if len(json_str) > max_length: # Calculate how much to keep from start and end keep_length = max_length - 150 # Reserve space for truncation message start_length = keep_length // 2 end_length = keep_length - start_length start_part = json_str[:start_length] end_part = json_str[-end_length:] if end_length > 0 else "" return start_part + f"\n\n... (middle truncated) ...\n\n" + end_part + f"\n\nThis message is too long, repeat relevant information in your response to remember it" else: return msg_content async def compress_tool_result_messages(self, messages: List[Dict[str, Any]], llm_model: str, max_tokens: Optional[int], token_threshold: int = 1000, uncompressed_total_token_count: Optional[int] = None) -> List[Dict[str, Any]]: """Compress the tool result messages except the most recent N (configured by keep_recent_tool_outputs). Compression is deterministic (simple truncation), ensuring consistent results across requests. This allows prompt caching (applied later) to produce cache hits on identical compressed content. """ if uncompressed_total_token_count is None: uncompressed_total_token_count = await self.count_tokens(llm_model, messages) max_tokens_value = max_tokens or (100 * 1000) if uncompressed_total_token_count > max_tokens_value: _i = 0 # Count the number of ToolResult messages for msg in reversed(messages): # Start from the end and work backwards if not isinstance(msg, dict): continue # Skip non-dict messages if self.is_tool_result_message(msg): # Only compress ToolResult messages _i += 1 # Count the number of ToolResult messages msg_token_count = token_counter(messages=[msg]) # Count the number of tokens in the message if msg_token_count > token_threshold: # If the message is too long if _i > self.keep_recent_tool_outputs: # If this is not one of the most recent N ToolResult messages message_id = msg.get('message_id') # Get the message_id if message_id: msg["content"] = self.compress_message(msg["content"], message_id, token_threshold * 3) else: logger.warning(f"UNEXPECTED: Message has no message_id {str(msg)[:100]}") else: msg["content"] = self.safe_truncate(msg["content"], int(max_tokens_value * 2)) return messages async def compress_user_messages(self, messages: List[Dict[str, Any]], llm_model: str, max_tokens: Optional[int], token_threshold: int = 1000, uncompressed_total_token_count: Optional[int] = None) -> List[Dict[str, Any]]: """Compress the user messages except the most recent one. Compression is deterministic (simple truncation), ensuring consistent results across requests. This allows prompt caching (applied later) to produce cache hits on identical compressed content. """ if uncompressed_total_token_count is None: uncompressed_total_token_count = await self.count_tokens(llm_model, messages) max_tokens_value = max_tokens or (100 * 1000) if uncompressed_total_token_count > max_tokens_value: _i = 0 # Count the number of User messages for msg in reversed(messages): # Start from the end and work backwards if not isinstance(msg, dict): continue # Skip non-dict messages if msg.get('role') == 'user': # Only compress User messages _i += 1 # Count the number of User messages msg_token_count = token_counter(messages=[msg]) # Count the number of tokens in the message if msg_token_count > token_threshold: # If the message is too long if _i > self.keep_recent_user_messages: # If this is not one of the most recent N User messages message_id = msg.get('message_id') # Get the message_id if message_id: # Secondary compression - just placeholder msg["content"] = f"[Content compressed]\n\nmessage_id \"{message_id}\"\nUse expand-message tool to see contents" else: logger.warning(f"UNEXPECTED: Message has no message_id {str(msg)[:100]}") else: msg["content"] = self.safe_truncate(msg["content"], int(max_tokens_value * 2)) return messages async def compress_assistant_messages(self, messages: List[Dict[str, Any]], llm_model: str, max_tokens: Optional[int], token_threshold: int = 1000, uncompressed_total_token_count: Optional[int] = None) -> List[Dict[str, Any]]: """Compress the assistant messages except the most recent one. Compression is deterministic (simple truncation), ensuring consistent results across requests. This allows prompt caching (applied later) to produce cache hits on identical compressed content. """ if uncompressed_total_token_count is None: uncompressed_total_token_count = await self.count_tokens(llm_model, messages) max_tokens_value = max_tokens or (100 * 1000) if uncompressed_total_token_count > max_tokens_value: _i = 0 # Count the number of Assistant messages for msg in reversed(messages): # Start from the end and work backwards if not isinstance(msg, dict): continue # Skip non-dict messages if msg.get('role') == 'assistant': # Only compress Assistant messages _i += 1 # Count the number of Assistant messages msg_token_count = token_counter(messages=[msg]) # Count the number of tokens in the message if msg_token_count > token_threshold: # If the message is too long if _i > self.keep_recent_assistant_messages: # If this is not one of the most recent N Assistant messages message_id = msg.get('message_id') # Get the message_id if message_id: # Secondary compression - just placeholder msg["content"] = f"[Content compressed]\n\nmessage_id \"{message_id}\"\nUse expand-message tool to see contents" else: logger.warning(f"UNEXPECTED: Message has no message_id {str(msg)[:100]}") else: msg["content"] = self.safe_truncate(msg["content"], int(max_tokens_value * 2)) return messages def remove_meta_messages(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """Remove meta messages from the messages.""" result: List[Dict[str, Any]] = [] for msg in messages: msg_content = msg.get('content') # Try to parse msg_content as JSON if it's a string if isinstance(msg_content, str): try: msg_content = json.loads(msg_content) except json.JSONDecodeError: pass if isinstance(msg_content, dict): # Create a copy to avoid modifying the original msg_content_copy = msg_content.copy() if "tool_execution" in msg_content_copy: tool_execution = msg_content_copy["tool_execution"].copy() if "arguments" in tool_execution: del tool_execution["arguments"] msg_content_copy["tool_execution"] = tool_execution # Create a new message dict with the modified content new_msg = msg.copy() new_msg["content"] = json.dumps(msg_content_copy) result.append(new_msg) else: result.append(msg) return result async def compress_messages(self, messages: List[Dict[str, Any]], llm_model: str, max_tokens: Optional[int] = 41000, token_threshold: int = 4096, max_iterations: int = 5, actual_total_tokens: Optional[int] = None, system_prompt: Optional[Dict[str, Any]] = None, thread_id: Optional[str] = None) -> List[Dict[str, Any]]: """Compress the messages WITHOUT applying caching during iterations. Caching should be applied ONCE at the end by the caller, not during compression. Compressed messages are saved to the database for future reads. """ # Capture original content for messages with message_id (for tracking compression) original_content_map: Dict[str, Any] = {} for msg in messages: message_id = msg.get('message_id') if message_id: original_content_map[message_id] = msg.get('content') # Get model-specific token limits from constants context_window = model_manager.get_context_window(llm_model) # Reserve tokens for output generation and safety margin if context_window >= 1_000_000: # Very large context models (Gemini) max_tokens = context_window - 300_000 # Large safety margin for huge contexts elif context_window >= 400_000: # Large context models (GPT-5) max_tokens = context_window - 64_000 # Reserve for output + margin elif context_window >= 200_000: # Medium context models (Claude Sonnet) max_tokens = context_window - 32_000 # Reserve for output + margin elif context_window >= 100_000: # Standard large context models max_tokens = context_window - 16_000 # Reserve for output + margin else: # Smaller context models max_tokens = context_window - 8_000 # Reserve for output + margin # logger.debug(f"Model {llm_model}: context_window={context_window}, effective_limit={max_tokens}") result = messages result = self.remove_meta_messages(result) # Calculate initial token count with caching to match API reality if actual_total_tokens is not None: uncompressed_total_token_count = actual_total_tokens else: # Count conversation + system prompt WITH caching (to match API reality) uncompressed_total_token_count = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) logger.info(f"Initial token count (with caching): {uncompressed_total_token_count}") # Calculate target tokens (hysteresis: compress to 60% of max to avoid repeated compressions) target_tokens = int(max_tokens * self.compression_target_ratio) logger.info(f"Compression threshold: {max_tokens}, target: {target_tokens} (ratio: {self.compression_target_ratio})") # Check if we're already under threshold - no compression needed! if uncompressed_total_token_count <= max_tokens: logger.info(f"βœ… Token count ({uncompressed_total_token_count}) under threshold ({max_tokens}), skipping compression") return await self.middle_out_messages(result) # PRIMARY STRATEGY: Remove old tool outputs if over threshold if uncompressed_total_token_count > max_tokens: logger.info(f"Context over limit ({uncompressed_total_token_count} > {max_tokens}), starting tiered compression...") # Tier 1: Compress old tool outputs in-memory result = self.remove_old_tool_outputs(result, keep_last_n=self.keep_recent_tool_outputs) # Recalculate WITH caching current_token_count = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) logger.info(f"After tool compression: {uncompressed_total_token_count} -> {current_token_count} tokens") # Tier 2: Compress user messages if still above target if current_token_count > target_tokens: logger.info(f"Still above target ({current_token_count} > {target_tokens}), compressing user messages...") # Compress in-memory for this request result = self.compress_user_messages_in_memory(result, keep_last_n=self.keep_recent_user_messages) # Recalculate with in-memory compressed messages WITH caching current_token_count = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) logger.info(f"After user compression: {current_token_count} tokens") # Tier 3: Compress assistant messages if still above target if current_token_count > target_tokens: logger.info(f"Still above target ({current_token_count} > {target_tokens}), compressing assistant messages...") # Compress in-memory for this request result = self.compress_assistant_messages_in_memory(result, keep_last_n=self.keep_recent_assistant_messages) # Recalculate with in-memory compressed messages WITH caching current_token_count = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) logger.info(f"After assistant compression: {current_token_count} tokens") logger.info(f"Tiered compression complete: {uncompressed_total_token_count} -> {current_token_count} tokens (target: {target_tokens})") uncompressed_total_token_count = current_token_count # SECONDARY STRATEGY: Apply compression to remaining messages if still above target # Use target_tokens as threshold to ensure we reach hysteresis goal if uncompressed_total_token_count > target_tokens: logger.info(f"Applying secondary compression to reach target ({uncompressed_total_token_count} > {target_tokens})") # Use lower token_threshold (500) to compress more messages, including smaller ones aggressive_threshold = 500 result = await self.compress_tool_result_messages(result, llm_model, target_tokens, aggressive_threshold, uncompressed_total_token_count) result = await self.compress_user_messages(result, llm_model, target_tokens, aggressive_threshold, uncompressed_total_token_count) result = await self.compress_assistant_messages(result, llm_model, target_tokens, aggressive_threshold, uncompressed_total_token_count) else: # Still run with original max_tokens in case there's any remaining content to compress result = await self.compress_tool_result_messages(result, llm_model, max_tokens, token_threshold, uncompressed_total_token_count) result = await self.compress_user_messages(result, llm_model, max_tokens, token_threshold, uncompressed_total_token_count) result = await self.compress_assistant_messages(result, llm_model, max_tokens, token_threshold, uncompressed_total_token_count) # Recalculate WITH caching (to match API reality) compressed_total = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) if compressed_total != uncompressed_total_token_count: logger.info(f"Context compression: {uncompressed_total_token_count} -> {compressed_total} tokens (saved {uncompressed_total_token_count - compressed_total})") else: logger.info(f"Context compression: {compressed_total} tokens (no compression needed, under threshold)") # Recurse if still too large if max_iterations <= 0: logger.warning(f"Max iterations reached, omitting messages to target ({target_tokens})") result = await self.compress_messages_by_omitting_messages(result, llm_model, target_tokens, system_prompt=system_prompt) compressed_total = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) # Fall through to last_usage update elif compressed_total > max_tokens: logger.warning(f"Further compression needed: {compressed_total} > {max_tokens}") # Recursive call - will handle its own last_usage update return await self.compress_messages( result, llm_model, max_tokens, token_threshold // 2, max_iterations - 1, compressed_total, system_prompt, thread_id=thread_id ) elif compressed_total > target_tokens: # Still over target but under max_tokens - use omit_messages to reach target logger.info(f"Secondary compression didn't reach target ({compressed_total} > {target_tokens}). Using message omission to reach target.") result = await self.compress_messages_by_omitting_messages(result, llm_model, target_tokens, system_prompt=system_prompt) compressed_total = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) logger.info(f"After message omission to target: {compressed_total} tokens") logger.info(f"✨ Final compression complete: {compressed_total} tokens (target: {target_tokens}, max: {max_tokens})") # Identify messages that were compressed and save to database compressed_to_save: List[Dict[str, Any]] = [] for msg in result: message_id = msg.get('message_id') if message_id and message_id in original_content_map: original_content = original_content_map[message_id] current_content = msg.get('content') # If content changed, it was compressed if current_content != original_content: # Save the FULL message structure (matching DB content format) # Remove message_id from the saved content as it's stored separately msg_to_save = {k: v for k, v in msg.items() if k != 'message_id'} compressed_to_save.append({ 'message_id': message_id, 'compressed_content': json.dumps(msg_to_save) }) # Save compressed messages to database (fire and forget for performance) if compressed_to_save: try: await self.save_compressed_messages(compressed_to_save) except Exception as e: logger.warning(f"Failed to save compressed messages: {e}") return await self.middle_out_messages(result) async def compress_messages_by_omitting_messages( self, messages: List[Dict[str, Any]], llm_model: str, max_tokens: Optional[int] = 41000, removal_batch_size: int = 3, # Now operates on groups, not individual messages min_groups_to_keep: int = 5, # Minimum number of groups to preserve system_prompt: Optional[Dict[str, Any]] = None ) -> List[Dict[str, Any]]: """Compress the messages by omitting message GROUPS from the middle. CRITICAL: This method operates on atomic message groups to preserve the assistant+tool_calls / tool_result pairing required by Bedrock. Args: messages: List of messages to compress llm_model: Model name for token counting max_tokens: Maximum allowed tokens removal_batch_size: Number of groups to remove per iteration min_groups_to_keep: Minimum number of groups to preserve system_prompt: Optional system prompt for token counting """ if not messages: return messages result = messages result = self.remove_meta_messages(result) # Early exit if no compression needed - WITH caching initial_token_count = await self.count_tokens(llm_model, result, system_prompt, apply_caching=True) max_allowed_tokens = max_tokens or (100 * 1000) if initial_token_count <= max_allowed_tokens: return result # Group messages into atomic units (assistant+tool_calls grouped with their tool results) message_groups = self.group_messages_by_tool_calls(result) logger.info(f"πŸ“¦ Grouped {len(result)} messages into {len(message_groups)} atomic groups for compression") system_message = system_prompt safety_limit = 500 current_token_count = initial_token_count # Track all omitted messages to save them with placeholder content all_omitted_messages: List[Dict[str, Any]] = [] while current_token_count > max_allowed_tokens and safety_limit > 0: safety_limit -= 1 if len(message_groups) <= min_groups_to_keep: logger.warning(f"Cannot compress further: only {len(message_groups)} groups remain (min: {min_groups_to_keep})") break # Calculate removal strategy based on current group count if len(message_groups) > (removal_batch_size * 2): # Remove from middle, keeping recent and early context middle_start = len(message_groups) // 2 - (removal_batch_size // 2) middle_end = middle_start + removal_batch_size # Log what we're removing removed_groups = message_groups[middle_start:middle_end] removed_msg_count = sum(len(g) for g in removed_groups) logger.debug(f"Removing {len(removed_groups)} groups ({removed_msg_count} messages) from middle") # Track omitted messages for group in removed_groups: all_omitted_messages.extend(group) message_groups = message_groups[:middle_start] + message_groups[middle_end:] else: # Remove from earlier groups, preserving recent context groups_to_remove = min(removal_batch_size, len(message_groups) // 2) if groups_to_remove > 0: removed_groups = message_groups[:groups_to_remove] removed_msg_count = sum(len(g) for g in removed_groups) logger.debug(f"Removing {groups_to_remove} early groups ({removed_msg_count} messages)") # Track omitted messages for group in removed_groups: all_omitted_messages.extend(group) message_groups = message_groups[groups_to_remove:] else: # Can't remove any more groups break # Flatten groups back to messages for token counting conversation_messages = self.flatten_message_groups(message_groups) # Recalculate token count WITH caching current_token_count = await self.count_tokens(llm_model, conversation_messages, system_message, apply_caching=True) # Save omitted messages with placeholder content so they're read as compressed next time if all_omitted_messages: omitted_to_save: List[Dict[str, Any]] = [] for msg in all_omitted_messages: message_id = msg.get('message_id') if message_id: # Create placeholder content preserving message structure placeholder_msg = { 'role': msg.get('role', 'user'), 'content': f"[Message omitted for context management. message_id: \"{message_id}\". Use expand-message tool to view full content.]" } # Preserve tool_call_id for tool messages if msg.get('tool_call_id'): placeholder_msg['tool_call_id'] = msg['tool_call_id'] # Preserve tool_calls for assistant messages (critical for pairing validation) if msg.get('tool_calls'): placeholder_msg['tool_calls'] = msg['tool_calls'] omitted_to_save.append({ 'message_id': message_id, 'compressed_content': json.dumps(placeholder_msg), 'is_omission': True # Mark as omission so it can override compression }) if omitted_to_save: try: await self.save_compressed_messages(omitted_to_save) # Also mark any tool results belonging to omitted assistant messages # This handles the case where tool results were compressed separately omitted_tool_call_ids = [] for msg in all_omitted_messages: if msg.get('tool_calls'): for tc in msg['tool_calls']: tc_id = tc.get('id') if tc_id: omitted_tool_call_ids.append(tc_id) if omitted_tool_call_ids and self.thread_id: from core.threads import repo as threads_repo marked_count = await threads_repo.mark_tool_results_as_omitted( self.thread_id, omitted_tool_call_ids ) if marked_count > 0: logger.info(f"πŸ“ Also marked {marked_count} orphaned tool results as omitted") except Exception as e: logger.warning(f"Failed to save omitted messages: {e}") # Flatten final groups to messages final_messages = self.flatten_message_groups(message_groups) # Validate tool call pairing is intact is_valid, orphaned_ids, unanswered_ids = self.validate_tool_call_pairing(final_messages) if not is_valid: logger.warning(f"⚠️ Post-compression validation found pairing issues (orphaned: {len(orphaned_ids)}, unanswered: {len(unanswered_ids)}) - repairing") final_messages = self.repair_tool_call_pairing(final_messages) # Log with system prompt included for accurate token reporting WITH caching final_token_count = await self.count_tokens(llm_model, final_messages, system_message, apply_caching=True) logger.info(f"Context compression (omit): {initial_token_count} -> {final_token_count} tokens ({len(messages)} -> {len(final_messages)} messages, {len(message_groups)} groups)") return final_messages async def middle_out_messages(self, messages: List[Dict[str, Any]], max_messages: int = 320) -> List[Dict[str, Any]]: """Remove message GROUPS from the middle of the list, keeping approximately max_messages total. CRITICAL: This method operates on atomic message groups to preserve the assistant+tool_calls / tool_result pairing required by Bedrock. Args: messages: List of messages max_messages: Approximate maximum messages to keep (actual may vary due to group sizes) Returns: Messages with middle groups removed, preserving tool call pairing """ if len(messages) <= max_messages: return messages # Group messages into atomic units message_groups = self.group_messages_by_tool_calls(messages) # If already few enough groups, return as-is total_messages = sum(len(g) for g in message_groups) if total_messages <= max_messages: return messages # Estimate how many groups we need to keep # Use average group size to estimate avg_group_size = total_messages / len(message_groups) if message_groups else 1 target_groups = int(max_messages / avg_group_size) # Ensure we keep at least 4 groups (2 from start, 2 from end) target_groups = max(4, target_groups) if len(message_groups) <= target_groups: return messages # Keep half from the beginning and half from the end (by groups) keep_start_groups = target_groups // 2 keep_end_groups = target_groups - keep_start_groups # Ensure we keep at least 1 group from each end keep_start_groups = max(1, keep_start_groups) keep_end_groups = max(1, keep_end_groups) # Build the result by keeping start and end groups kept_groups = message_groups[:keep_start_groups] + message_groups[-keep_end_groups:] # Track removed groups for saving as compressed removed_groups = message_groups[keep_start_groups:-keep_end_groups] if keep_end_groups > 0 else message_groups[keep_start_groups:] removed_count = len(message_groups) - len(kept_groups) if removed_count > 0: logger.info(f"πŸ“¦ Middle-out: removed {removed_count} groups from middle ({len(message_groups)} -> {len(kept_groups)} groups)") # Save removed messages with placeholder content omitted_to_save: List[Dict[str, Any]] = [] for group in removed_groups: for msg in group: message_id = msg.get('message_id') if message_id: placeholder_msg = { 'role': msg.get('role', 'user'), 'content': f"[Message omitted for context management. message_id: \"{message_id}\". Use expand-message tool to view full content.]" } if msg.get('tool_call_id'): placeholder_msg['tool_call_id'] = msg['tool_call_id'] # Preserve tool_calls for assistant messages (critical for pairing validation) if msg.get('tool_calls'): placeholder_msg['tool_calls'] = msg['tool_calls'] omitted_to_save.append({ 'message_id': message_id, 'compressed_content': json.dumps(placeholder_msg), 'is_omission': True # Mark as omission so it can override compression }) if omitted_to_save: try: await self.save_compressed_messages(omitted_to_save) # Also mark any tool results belonging to omitted assistant messages omitted_tool_call_ids = [] for group in removed_groups: for msg in group: if msg.get('tool_calls'): for tc in msg['tool_calls']: tc_id = tc.get('id') if tc_id: omitted_tool_call_ids.append(tc_id) if omitted_tool_call_ids and self.thread_id: from core.threads import repo as threads_repo marked_count = await threads_repo.mark_tool_results_as_omitted( self.thread_id, omitted_tool_call_ids ) if marked_count > 0: logger.info(f"πŸ“ Also marked {marked_count} orphaned tool results as omitted") except Exception as e: logger.warning(f"Failed to save middle-out omitted messages: {e}") # Flatten groups back to messages result = self.flatten_message_groups(kept_groups) # Validate tool call pairing is intact is_valid, orphaned_ids, unanswered_ids = self.validate_tool_call_pairing(result) if not is_valid: logger.warning(f"⚠️ Middle-out validation found pairing issues (orphaned: {len(orphaned_ids)}, unanswered: {len(unanswered_ids)}) - repairing") result = self.repair_tool_call_pairing(result) return result