Ling0925
/
suna
mirror of https://github.com/kortix-ai/suna.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
suna/backend/agentpress/response_processor.py

1684 lines
98 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

"""
Response processing module for AgentPress.
This module handles the processing of LLM responses, including:
- Streaming and non-streaming response handling
- XML and native tool call detection and parsing
- Tool execution orchestration
- Message formatting and persistence
"""
import json
import re
import uuid
import asyncio
from datetime import datetime, timezone
from typing import List, Dict, Any, Optional, AsyncGenerator, Tuple, Union, Callable, Literal
from dataclasses import dataclass
from utils.logger import logger
from agentpress.tool import ToolResult
from agentpress.tool_registry import ToolRegistry
from agentpress.xml_tool_parser import XMLToolParser
from langfuse.client import StatefulTraceClient
from services.langfuse import langfuse
from utils.json_helpers import (
ensure_dict, ensure_list, safe_json_parse,
to_json_string, format_for_yield
)
from litellm.utils import token_counter
# Type alias for XML result adding strategy
XmlAddingStrategy = Literal["user_message", "assistant_message", "inline_edit"]
# Type alias for tool execution strategy
ToolExecutionStrategy = Literal["sequential", "parallel"]
@dataclass
class ToolExecutionContext:
"""Context for a tool execution including call details, result, and display info."""
tool_call: Dict[str, Any]
tool_index: int
result: Optional[ToolResult] = None
function_name: Optional[str] = None
xml_tag_name: Optional[str] = None
error: Optional[Exception] = None
assistant_message_id: Optional[str] = None
parsing_details: Optional[Dict[str, Any]] = None
@dataclass
class ProcessorConfig:
"""
Configuration for response processing and tool execution.
This class controls how the LLM's responses are processed, including how tool calls
are detected, executed, and their results handled.
Attributes:
xml_tool_calling: Enable XML-based tool call detection (<tool>...</tool>)
native_tool_calling: Enable OpenAI-style function calling format
execute_tools: Whether to automatically execute detected tool calls
execute_on_stream: For streaming, execute tools as they appear vs. at the end
tool_execution_strategy: How to execute multiple tools ("sequential" or "parallel")
xml_adding_strategy: How to add XML tool results to the conversation
max_xml_tool_calls: Maximum number of XML tool calls to process (0 = no limit)
"""
xml_tool_calling: bool = True
native_tool_calling: bool = False
execute_tools: bool = True
execute_on_stream: bool = False
tool_execution_strategy: ToolExecutionStrategy = "sequential"
xml_adding_strategy: XmlAddingStrategy = "assistant_message"
max_xml_tool_calls: int = 0 # 0 means no limit
def __post_init__(self):
"""Validate configuration after initialization."""
if self.xml_tool_calling is False and self.native_tool_calling is False and self.execute_tools:
raise ValueError("At least one tool calling format (XML or native) must be enabled if execute_tools is True")
if self.xml_adding_strategy not in ["user_message", "assistant_message", "inline_edit"]:
raise ValueError("xml_adding_strategy must be 'user_message', 'assistant_message', or 'inline_edit'")
if self.max_xml_tool_calls < 0:
raise ValueError("max_xml_tool_calls must be a non-negative integer (0 = no limit)")
class ResponseProcessor:
"""Processes LLM responses, extracting and executing tool calls."""
def __init__(self, tool_registry: ToolRegistry, add_message_callback: Callable, trace: Optional[StatefulTraceClient] = None, is_agent_builder: bool = False, target_agent_id: Optional[str] = None, agent_config: Optional[dict] = None):
"""Initialize the ResponseProcessor.
Args:
tool_registry: Registry of available tools
add_message_callback: Callback function to add messages to the thread.
MUST return the full saved message object (dict) or None.
agent_config: Optional agent configuration with version information
"""
self.tool_registry = tool_registry
self.add_message = add_message_callback
self.trace = trace or langfuse.trace(name="anonymous:response_processor")
# Initialize the XML parser
self.xml_parser = XMLToolParser()
self.is_agent_builder = is_agent_builder
self.target_agent_id = target_agent_id
self.agent_config = agent_config
async def _yield_message(self, message_obj: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
"""Helper to yield a message with proper formatting.
Ensures that content and metadata are JSON strings for client compatibility.
"""
if message_obj:
return format_for_yield(message_obj)
return None
async def _add_message_with_agent_info(
self,
thread_id: str,
type: str,
content: Union[Dict[str, Any], List[Any], str],
is_llm_message: bool = False,
metadata: Optional[Dict[str, Any]] = None
):
"""Helper to add a message with agent version information if available."""
agent_id = None
agent_version_id = None
if self.agent_config:
agent_id = self.agent_config.get('agent_id')
agent_version_id = self.agent_config.get('current_version_id')
return await self.add_message(
thread_id=thread_id,
type=type,
content=content,
is_llm_message=is_llm_message,
metadata=metadata,
agent_id=agent_id,
agent_version_id=agent_version_id
)
async def process_streaming_response(
self,
llm_response: AsyncGenerator,
thread_id: str,
prompt_messages: List[Dict[str, Any]],
llm_model: str,
config: ProcessorConfig = ProcessorConfig(),
can_auto_continue: bool = False,
auto_continue_count: int = 0,
continuous_state: Optional[Dict[str, Any]] = None,
) -> AsyncGenerator[Dict[str, Any], None]:
"""Process a streaming LLM response, handling tool calls and execution.
Args:
llm_response: Streaming response from the LLM
thread_id: ID of the conversation thread
prompt_messages: List of messages sent to the LLM (the prompt)
llm_model: The name of the LLM model used
config: Configuration for parsing and execution
can_auto_continue: Whether auto-continue is enabled
auto_continue_count: Number of auto-continue cycles
continuous_state: Previous state of the conversation
Yields:
Complete message objects matching the DB schema, except for content chunks.
"""
# Initialize from continuous state if provided (for auto-continue)
continuous_state = continuous_state or {}
accumulated_content = continuous_state.get('accumulated_content', "")
tool_calls_buffer = {}
current_xml_content = accumulated_content # equal to accumulated_content if auto-continuing, else blank
xml_chunks_buffer = []
pending_tool_executions = []
yielded_tool_indices = set() # Stores indices of tools whose *status* has been yielded
tool_index = 0
xml_tool_call_count = 0
finish_reason = None
should_auto_continue = False
last_assistant_message_object = None # Store the final saved assistant message object
tool_result_message_objects = {} # tool_index -> full saved message object
has_printed_thinking_prefix = False # Flag for printing thinking prefix only once
agent_should_terminate = False # Flag to track if a terminating tool has been executed
complete_native_tool_calls = [] # Initialize early for use in assistant_response_end
# Collect metadata for reconstructing LiteLLM response object
streaming_metadata = {
"model": llm_model,
"created": None,
"usage": {
"prompt_tokens": 0,
"completion_tokens": 0,
"total_tokens": 0
},
"response_ms": None,
"first_chunk_time": None,
"last_chunk_time": None
}
logger.info(f"Streaming Config: XML={config.xml_tool_calling}, Native={config.native_tool_calling}, "
f"Execute on stream={config.execute_on_stream}, Strategy={config.tool_execution_strategy}")
# Reuse thread_run_id for auto-continue or create new one
thread_run_id = continuous_state.get('thread_run_id') or str(uuid.uuid4())
continuous_state['thread_run_id'] = thread_run_id
try:
# --- Save and Yield Start Events (only if not auto-continuing) ---
if auto_continue_count == 0:
start_content = {"status_type": "thread_run_start", "thread_run_id": thread_run_id}
start_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=start_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if start_msg_obj: yield format_for_yield(start_msg_obj)
assist_start_content = {"status_type": "assistant_response_start"}
assist_start_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=assist_start_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if assist_start_msg_obj: yield format_for_yield(assist_start_msg_obj)
# --- End Start Events ---
__sequence = continuous_state.get('sequence', 0) # get the sequence from the previous auto-continue cycle
async for chunk in llm_response:
# Extract streaming metadata from chunks
current_time = datetime.now(timezone.utc).timestamp()
if streaming_metadata["first_chunk_time"] is None:
streaming_metadata["first_chunk_time"] = current_time
streaming_metadata["last_chunk_time"] = current_time
# Extract metadata from chunk attributes
if hasattr(chunk, 'created') and chunk.created:
streaming_metadata["created"] = chunk.created
if hasattr(chunk, 'model') and chunk.model:
streaming_metadata["model"] = chunk.model
if hasattr(chunk, 'usage') and chunk.usage:
# Update usage information if available (including zero values)
if hasattr(chunk.usage, 'prompt_tokens') and chunk.usage.prompt_tokens is not None:
streaming_metadata["usage"]["prompt_tokens"] = chunk.usage.prompt_tokens
if hasattr(chunk.usage, 'completion_tokens') and chunk.usage.completion_tokens is not None:
streaming_metadata["usage"]["completion_tokens"] = chunk.usage.completion_tokens
if hasattr(chunk.usage, 'total_tokens') and chunk.usage.total_tokens is not None:
streaming_metadata["usage"]["total_tokens"] = chunk.usage.total_tokens
if hasattr(chunk, 'choices') and chunk.choices and hasattr(chunk.choices[0], 'finish_reason') and chunk.choices[0].finish_reason:
finish_reason = chunk.choices[0].finish_reason
logger.debug(f"Detected finish_reason: {finish_reason}")
if hasattr(chunk, 'choices') and chunk.choices:
delta = chunk.choices[0].delta if hasattr(chunk.choices[0], 'delta') else None
# Check for and log Anthropic thinking content
if delta and hasattr(delta, 'reasoning_content') and delta.reasoning_content:
if not has_printed_thinking_prefix:
# print("[THINKING]: ", end='', flush=True)
has_printed_thinking_prefix = True
# print(delta.reasoning_content, end='', flush=True)
# Append reasoning to main content to be saved in the final message
accumulated_content += delta.reasoning_content
# Process content chunk
if delta and hasattr(delta, 'content') and delta.content:
chunk_content = delta.content
# print(chunk_content, end='', flush=True)
accumulated_content += chunk_content
current_xml_content += chunk_content
if not (config.max_xml_tool_calls > 0 and xml_tool_call_count >= config.max_xml_tool_calls):
# Yield ONLY content chunk (don't save)
now_chunk = datetime.now(timezone.utc).isoformat()
yield {
"sequence": __sequence,
"message_id": None, "thread_id": thread_id, "type": "assistant",
"is_llm_message": True,
"content": to_json_string({"role": "assistant", "content": chunk_content}),
"metadata": to_json_string({"stream_status": "chunk", "thread_run_id": thread_run_id}),
"created_at": now_chunk, "updated_at": now_chunk
}
__sequence += 1
else:
logger.info("XML tool call limit reached - not yielding more content chunks")
self.trace.event(name="xml_tool_call_limit_reached", level="DEFAULT", status_message=(f"XML tool call limit reached - not yielding more content chunks"))
# --- Process XML Tool Calls (if enabled and limit not reached) ---
if config.xml_tool_calling and not (config.max_xml_tool_calls > 0 and xml_tool_call_count >= config.max_xml_tool_calls):
xml_chunks = self._extract_xml_chunks(current_xml_content)
for xml_chunk in xml_chunks:
current_xml_content = current_xml_content.replace(xml_chunk, "", 1)
xml_chunks_buffer.append(xml_chunk)
result = self._parse_xml_tool_call(xml_chunk)
if result:
tool_call, parsing_details = result
xml_tool_call_count += 1
current_assistant_id = last_assistant_message_object['message_id'] if last_assistant_message_object else None
context = self._create_tool_context(
tool_call, tool_index, current_assistant_id, parsing_details
)
if config.execute_tools and config.execute_on_stream:
# Save and Yield tool_started status
started_msg_obj = await self._yield_and_save_tool_started(context, thread_id, thread_run_id)
if started_msg_obj: yield format_for_yield(started_msg_obj)
yielded_tool_indices.add(tool_index) # Mark status as yielded
execution_task = asyncio.create_task(self._execute_tool(tool_call))
pending_tool_executions.append({
"task": execution_task, "tool_call": tool_call,
"tool_index": tool_index, "context": context
})
tool_index += 1
if config.max_xml_tool_calls > 0 and xml_tool_call_count >= config.max_xml_tool_calls:
logger.debug(f"Reached XML tool call limit ({config.max_xml_tool_calls})")
finish_reason = "xml_tool_limit_reached"
break # Stop processing more XML chunks in this delta
# --- Process Native Tool Call Chunks ---
if config.native_tool_calling and delta and hasattr(delta, 'tool_calls') and delta.tool_calls:
for tool_call_chunk in delta.tool_calls:
# Yield Native Tool Call Chunk (transient status, not saved)
# ... (safe extraction logic for tool_call_data_chunk) ...
tool_call_data_chunk = {} # Placeholder for extracted data
if hasattr(tool_call_chunk, 'model_dump'): tool_call_data_chunk = tool_call_chunk.model_dump()
else: # Manual extraction...
if hasattr(tool_call_chunk, 'id'): tool_call_data_chunk['id'] = tool_call_chunk.id
if hasattr(tool_call_chunk, 'index'): tool_call_data_chunk['index'] = tool_call_chunk.index
if hasattr(tool_call_chunk, 'type'): tool_call_data_chunk['type'] = tool_call_chunk.type
if hasattr(tool_call_chunk, 'function'):
tool_call_data_chunk['function'] = {}
if hasattr(tool_call_chunk.function, 'name'): tool_call_data_chunk['function']['name'] = tool_call_chunk.function.name
if hasattr(tool_call_chunk.function, 'arguments'): tool_call_data_chunk['function']['arguments'] = tool_call_chunk.function.arguments if isinstance(tool_call_chunk.function.arguments, str) else to_json_string(tool_call_chunk.function.arguments)
now_tool_chunk = datetime.now(timezone.utc).isoformat()
yield {
"message_id": None, "thread_id": thread_id, "type": "status", "is_llm_message": True,
"content": to_json_string({"role": "assistant", "status_type": "tool_call_chunk", "tool_call_chunk": tool_call_data_chunk}),
"metadata": to_json_string({"thread_run_id": thread_run_id}),
"created_at": now_tool_chunk, "updated_at": now_tool_chunk
}
# --- Buffer and Execute Complete Native Tool Calls ---
if not hasattr(tool_call_chunk, 'function'): continue
idx = tool_call_chunk.index if hasattr(tool_call_chunk, 'index') else 0
# ... (buffer update logic remains same) ...
# ... (check complete logic remains same) ...
has_complete_tool_call = False # Placeholder
if (tool_calls_buffer.get(idx) and
tool_calls_buffer[idx]['id'] and
tool_calls_buffer[idx]['function']['name'] and
tool_calls_buffer[idx]['function']['arguments']):
try:
safe_json_parse(tool_calls_buffer[idx]['function']['arguments'])
has_complete_tool_call = True
except json.JSONDecodeError: pass
if has_complete_tool_call and config.execute_tools and config.execute_on_stream:
current_tool = tool_calls_buffer[idx]
tool_call_data = {
"function_name": current_tool['function']['name'],
"arguments": safe_json_parse(current_tool['function']['arguments']),
"id": current_tool['id']
}
current_assistant_id = last_assistant_message_object['message_id'] if last_assistant_message_object else None
context = self._create_tool_context(
tool_call_data, tool_index, current_assistant_id
)
# Save and Yield tool_started status
started_msg_obj = await self._yield_and_save_tool_started(context, thread_id, thread_run_id)
if started_msg_obj: yield format_for_yield(started_msg_obj)
yielded_tool_indices.add(tool_index) # Mark status as yielded
execution_task = asyncio.create_task(self._execute_tool(tool_call_data))
pending_tool_executions.append({
"task": execution_task, "tool_call": tool_call_data,
"tool_index": tool_index, "context": context
})
tool_index += 1
if finish_reason == "xml_tool_limit_reached":
logger.info("Stopping stream processing after loop due to XML tool call limit")
self.trace.event(name="stopping_stream_processing_after_loop_due_to_xml_tool_call_limit", level="DEFAULT", status_message=(f"Stopping stream processing after loop due to XML tool call limit"))
break
# print() # Add a final newline after the streaming loop finishes
# --- After Streaming Loop ---
if (
streaming_metadata["usage"]["total_tokens"] == 0
):
logger.info("🔥 No usage data from provider, counting with litellm.token_counter")
try:
# prompt side
prompt_tokens = token_counter(
model=llm_model,
messages=prompt_messages # chat or plain; token_counter handles both
)
# completion side
completion_tokens = token_counter(
model=llm_model,
text=accumulated_content or "" # empty string safe
)
streaming_metadata["usage"]["prompt_tokens"] = prompt_tokens
streaming_metadata["usage"]["completion_tokens"] = completion_tokens
streaming_metadata["usage"]["total_tokens"] = prompt_tokens + completion_tokens
logger.info(
f"🔥 Estimated tokens – prompt: {prompt_tokens}, "
f"completion: {completion_tokens}, total: {prompt_tokens + completion_tokens}"
)
self.trace.event(name="usage_calculated_with_litellm_token_counter", level="DEFAULT", status_message=(f"Usage calculated with litellm.token_counter"))
except Exception as e:
logger.warning(f"Failed to calculate usage: {str(e)}")
self.trace.event(name="failed_to_calculate_usage", level="WARNING", status_message=(f"Failed to calculate usage: {str(e)}"))
# Wait for pending tool executions from streaming phase
tool_results_buffer = [] # Stores (tool_call, result, tool_index, context)
if pending_tool_executions:
logger.info(f"Waiting for {len(pending_tool_executions)} pending streamed tool executions")
self.trace.event(name="waiting_for_pending_streamed_tool_executions", level="DEFAULT", status_message=(f"Waiting for {len(pending_tool_executions)} pending streamed tool executions"))
# ... (asyncio.wait logic) ...
pending_tasks = [execution["task"] for execution in pending_tool_executions]
done, _ = await asyncio.wait(pending_tasks)
for execution in pending_tool_executions:
tool_idx = execution.get("tool_index", -1)
context = execution["context"]
tool_name = context.function_name
# Check if status was already yielded during stream run
if tool_idx in yielded_tool_indices:
logger.debug(f"Status for tool index {tool_idx} already yielded.")
# Still need to process the result for the buffer
try:
if execution["task"].done():
result = execution["task"].result()
context.result = result
tool_results_buffer.append((execution["tool_call"], result, tool_idx, context))
if tool_name in ['ask', 'complete']:
logger.info(f"Terminating tool '{tool_name}' completed during streaming. Setting termination flag.")
self.trace.event(name="terminating_tool_completed_during_streaming", level="DEFAULT", status_message=(f"Terminating tool '{tool_name}' completed during streaming. Setting termination flag."))
agent_should_terminate = True
else: # Should not happen with asyncio.wait
logger.warning(f"Task for tool index {tool_idx} not done after wait.")
self.trace.event(name="task_for_tool_index_not_done_after_wait", level="WARNING", status_message=(f"Task for tool index {tool_idx} not done after wait."))
except Exception as e:
logger.error(f"Error getting result for pending tool execution {tool_idx}: {str(e)}")
self.trace.event(name="error_getting_result_for_pending_tool_execution", level="ERROR", status_message=(f"Error getting result for pending tool execution {tool_idx}: {str(e)}"))
context.error = e
# Save and Yield tool error status message (even if started was yielded)
error_msg_obj = await self._yield_and_save_tool_error(context, thread_id, thread_run_id)
if error_msg_obj: yield format_for_yield(error_msg_obj)
continue # Skip further status yielding for this tool index
# If status wasn't yielded before (shouldn't happen with current logic), yield it now
try:
if execution["task"].done():
result = execution["task"].result()
context.result = result
tool_results_buffer.append((execution["tool_call"], result, tool_idx, context))
# Check if this is a terminating tool
if tool_name in ['ask', 'complete']:
logger.info(f"Terminating tool '{tool_name}' completed during streaming. Setting termination flag.")
self.trace.event(name="terminating_tool_completed_during_streaming", level="DEFAULT", status_message=(f"Terminating tool '{tool_name}' completed during streaming. Setting termination flag."))
agent_should_terminate = True
# Save and Yield tool completed/failed status
completed_msg_obj = await self._yield_and_save_tool_completed(
context, None, thread_id, thread_run_id
)
if completed_msg_obj: yield format_for_yield(completed_msg_obj)
yielded_tool_indices.add(tool_idx)
except Exception as e:
logger.error(f"Error getting result/yielding status for pending tool execution {tool_idx}: {str(e)}")
self.trace.event(name="error_getting_result_yielding_status_for_pending_tool_execution", level="ERROR", status_message=(f"Error getting result/yielding status for pending tool execution {tool_idx}: {str(e)}"))
context.error = e
# Save and Yield tool error status
error_msg_obj = await self._yield_and_save_tool_error(context, thread_id, thread_run_id)
if error_msg_obj: yield format_for_yield(error_msg_obj)
yielded_tool_indices.add(tool_idx)
# Save and yield finish status if limit was reached
if finish_reason == "xml_tool_limit_reached":
finish_content = {"status_type": "finish", "finish_reason": "xml_tool_limit_reached"}
finish_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=finish_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if finish_msg_obj: yield format_for_yield(finish_msg_obj)
logger.info(f"Stream finished with reason: xml_tool_limit_reached after {xml_tool_call_count} XML tool calls")
self.trace.event(name="stream_finished_with_reason_xml_tool_limit_reached_after_xml_tool_calls", level="DEFAULT", status_message=(f"Stream finished with reason: xml_tool_limit_reached after {xml_tool_call_count} XML tool calls"))
# Calculate if auto-continue is needed if the finish reason is length
should_auto_continue = (can_auto_continue and finish_reason == 'length')
# --- SAVE and YIELD Final Assistant Message ---
# Only save assistant message if NOT auto-continuing due to length to avoid duplicate messages
if accumulated_content and not should_auto_continue:
# ... (Truncate accumulated_content logic) ...
if config.max_xml_tool_calls > 0 and xml_tool_call_count >= config.max_xml_tool_calls and xml_chunks_buffer:
last_xml_chunk = xml_chunks_buffer[-1]
last_chunk_end_pos = accumulated_content.find(last_xml_chunk) + len(last_xml_chunk)
if last_chunk_end_pos > 0:
accumulated_content = accumulated_content[:last_chunk_end_pos]
# ... (Extract complete_native_tool_calls logic) ...
# Update complete_native_tool_calls from buffer (initialized earlier)
if config.native_tool_calling:
for idx, tc_buf in tool_calls_buffer.items():
if tc_buf['id'] and tc_buf['function']['name'] and tc_buf['function']['arguments']:
try:
args = safe_json_parse(tc_buf['function']['arguments'])
complete_native_tool_calls.append({
"id": tc_buf['id'], "type": "function",
"function": {"name": tc_buf['function']['name'],"arguments": args}
})
except json.JSONDecodeError: continue
message_data = { # Dict to be saved in 'content'
"role": "assistant", "content": accumulated_content,
"tool_calls": complete_native_tool_calls or None
}
last_assistant_message_object = await self._add_message_with_agent_info(
thread_id=thread_id, type="assistant", content=message_data,
is_llm_message=True, metadata={"thread_run_id": thread_run_id}
)
if last_assistant_message_object:
# Yield the complete saved object, adding stream_status metadata just for yield
yield_metadata = ensure_dict(last_assistant_message_object.get('metadata'), {})
yield_metadata['stream_status'] = 'complete'
# Format the message for yielding
yield_message = last_assistant_message_object.copy()
yield_message['metadata'] = yield_metadata
yield format_for_yield(yield_message)
else:
logger.error(f"Failed to save final assistant message for thread {thread_id}")
self.trace.event(name="failed_to_save_final_assistant_message_for_thread", level="ERROR", status_message=(f"Failed to save final assistant message for thread {thread_id}"))
# Save and yield an error status
err_content = {"role": "system", "status_type": "error", "message": "Failed to save final assistant message"}
err_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=err_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if err_msg_obj: yield format_for_yield(err_msg_obj)
# --- Process All Tool Results Now ---
if config.execute_tools:
final_tool_calls_to_process = []
# ... (Gather final_tool_calls_to_process from native and XML buffers) ...
# Gather native tool calls from buffer
if config.native_tool_calling and complete_native_tool_calls:
for tc in complete_native_tool_calls:
final_tool_calls_to_process.append({
"function_name": tc["function"]["name"],
"arguments": tc["function"]["arguments"], # Already parsed object
"id": tc["id"]
})
# Gather XML tool calls from buffer (up to limit)
parsed_xml_data = []
if config.xml_tool_calling:
# Reparse remaining content just in case (should be empty if processed correctly)
xml_chunks = self._extract_xml_chunks(current_xml_content)
xml_chunks_buffer.extend(xml_chunks)
# Process only chunks not already handled in the stream loop
remaining_limit = config.max_xml_tool_calls - xml_tool_call_count if config.max_xml_tool_calls > 0 else len(xml_chunks_buffer)
xml_chunks_to_process = xml_chunks_buffer[:remaining_limit] # Ensure limit is respected
for chunk in xml_chunks_to_process:
parsed_result = self._parse_xml_tool_call(chunk)
if parsed_result:
tool_call, parsing_details = parsed_result
# Avoid adding if already processed during streaming
if not any(exec['tool_call'] == tool_call for exec in pending_tool_executions):
final_tool_calls_to_process.append(tool_call)
parsed_xml_data.append({'tool_call': tool_call, 'parsing_details': parsing_details})
all_tool_data_map = {} # tool_index -> {'tool_call': ..., 'parsing_details': ...}
# Add native tool data
native_tool_index = 0
if config.native_tool_calling and complete_native_tool_calls:
for tc in complete_native_tool_calls:
# Find the corresponding entry in final_tool_calls_to_process if needed
# For now, assume order matches if only native used
exec_tool_call = {
"function_name": tc["function"]["name"],
"arguments": tc["function"]["arguments"],
"id": tc["id"]
}
all_tool_data_map[native_tool_index] = {"tool_call": exec_tool_call, "parsing_details": None}
native_tool_index += 1
# Add XML tool data
xml_tool_index_start = native_tool_index
for idx, item in enumerate(parsed_xml_data):
all_tool_data_map[xml_tool_index_start + idx] = item
tool_results_map = {} # tool_index -> (tool_call, result, context)
# Populate from buffer if executed on stream
if config.execute_on_stream and tool_results_buffer:
logger.info(f"Processing {len(tool_results_buffer)} buffered tool results")
self.trace.event(name="processing_buffered_tool_results", level="DEFAULT", status_message=(f"Processing {len(tool_results_buffer)} buffered tool results"))
for tool_call, result, tool_idx, context in tool_results_buffer:
if last_assistant_message_object: context.assistant_message_id = last_assistant_message_object['message_id']
tool_results_map[tool_idx] = (tool_call, result, context)
# Or execute now if not streamed
elif final_tool_calls_to_process and not config.execute_on_stream:
logger.info(f"Executing {len(final_tool_calls_to_process)} tools ({config.tool_execution_strategy}) after stream")
self.trace.event(name="executing_tools_after_stream", level="DEFAULT", status_message=(f"Executing {len(final_tool_calls_to_process)} tools ({config.tool_execution_strategy}) after stream"))
results_list = await self._execute_tools(final_tool_calls_to_process, config.tool_execution_strategy)
current_tool_idx = 0
for tc, res in results_list:
# Map back using all_tool_data_map which has correct indices
if current_tool_idx in all_tool_data_map:
tool_data = all_tool_data_map[current_tool_idx]
context = self._create_tool_context(
tc, current_tool_idx,
last_assistant_message_object['message_id'] if last_assistant_message_object else None,
tool_data.get('parsing_details')
)
context.result = res
tool_results_map[current_tool_idx] = (tc, res, context)
else:
logger.warning(f"Could not map result for tool index {current_tool_idx}")
self.trace.event(name="could_not_map_result_for_tool_index", level="WARNING", status_message=(f"Could not map result for tool index {current_tool_idx}"))
current_tool_idx += 1
# Save and Yield each result message
if tool_results_map:
logger.info(f"Saving and yielding {len(tool_results_map)} final tool result messages")
self.trace.event(name="saving_and_yielding_final_tool_result_messages", level="DEFAULT", status_message=(f"Saving and yielding {len(tool_results_map)} final tool result messages"))
for tool_idx in sorted(tool_results_map.keys()):
tool_call, result, context = tool_results_map[tool_idx]
context.result = result
if not context.assistant_message_id and last_assistant_message_object:
context.assistant_message_id = last_assistant_message_object['message_id']
# Yield start status ONLY IF executing non-streamed (already yielded if streamed)
if not config.execute_on_stream and tool_idx not in yielded_tool_indices:
started_msg_obj = await self._yield_and_save_tool_started(context, thread_id, thread_run_id)
if started_msg_obj: yield format_for_yield(started_msg_obj)
yielded_tool_indices.add(tool_idx) # Mark status yielded
# Save the tool result message to DB
saved_tool_result_object = await self._add_tool_result( # Returns full object or None
thread_id, tool_call, result, config.xml_adding_strategy,
context.assistant_message_id, context.parsing_details
)
# Yield completed/failed status (linked to saved result ID if available)
completed_msg_obj = await self._yield_and_save_tool_completed(
context,
saved_tool_result_object['message_id'] if saved_tool_result_object else None,
thread_id, thread_run_id
)
if completed_msg_obj: yield format_for_yield(completed_msg_obj)
# Don't add to yielded_tool_indices here, completion status is separate yield
# Yield the saved tool result object
if saved_tool_result_object:
tool_result_message_objects[tool_idx] = saved_tool_result_object
yield format_for_yield(saved_tool_result_object)
else:
logger.error(f"Failed to save tool result for index {tool_idx}, not yielding result message.")
self.trace.event(name="failed_to_save_tool_result_for_index", level="ERROR", status_message=(f"Failed to save tool result for index {tool_idx}, not yielding result message."))
# Optionally yield error status for saving failure?
# --- Final Finish Status ---
if finish_reason and finish_reason != "xml_tool_limit_reached":
finish_content = {"status_type": "finish", "finish_reason": finish_reason}
finish_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=finish_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if finish_msg_obj: yield format_for_yield(finish_msg_obj)
# Check if agent should terminate after processing pending tools
if agent_should_terminate:
logger.info("Agent termination requested after executing ask/complete tool. Stopping further processing.")
self.trace.event(name="agent_termination_requested", level="DEFAULT", status_message="Agent termination requested after executing ask/complete tool. Stopping further processing.")
# Set finish reason to indicate termination
finish_reason = "agent_terminated"
# Save and yield termination status
finish_content = {"status_type": "finish", "finish_reason": "agent_terminated"}
finish_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=finish_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if finish_msg_obj: yield format_for_yield(finish_msg_obj)
# Save assistant_response_end BEFORE terminating
if last_assistant_message_object:
try:
# Calculate response time if we have timing data
if streaming_metadata["first_chunk_time"] and streaming_metadata["last_chunk_time"]:
streaming_metadata["response_ms"] = (streaming_metadata["last_chunk_time"] - streaming_metadata["first_chunk_time"]) * 1000
# Create a LiteLLM-like response object for streaming (before termination)
# Check if we have any actual usage data
has_usage_data = (
streaming_metadata["usage"]["prompt_tokens"] > 0 or
streaming_metadata["usage"]["completion_tokens"] > 0 or
streaming_metadata["usage"]["total_tokens"] > 0
)
assistant_end_content = {
"choices": [
{
"finish_reason": finish_reason or "stop",
"index": 0,
"message": {
"role": "assistant",
"content": accumulated_content,
"tool_calls": complete_native_tool_calls or None
}
}
],
"created": streaming_metadata.get("created"),
"model": streaming_metadata.get("model", llm_model),
"usage": streaming_metadata["usage"], # Always include usage like LiteLLM does
"streaming": True, # Add flag to indicate this was reconstructed from streaming
}
# Only include response_ms if we have timing data
if streaming_metadata.get("response_ms"):
assistant_end_content["response_ms"] = streaming_metadata["response_ms"]
await self.add_message(
thread_id=thread_id,
type="assistant_response_end",
content=assistant_end_content,
is_llm_message=False,
metadata={"thread_run_id": thread_run_id}
)
logger.info("Assistant response end saved for stream (before termination)")
except Exception as e:
logger.error(f"Error saving assistant response end for stream (before termination): {str(e)}")
self.trace.event(name="error_saving_assistant_response_end_for_stream_before_termination", level="ERROR", status_message=(f"Error saving assistant response end for stream (before termination): {str(e)}"))
# Skip all remaining processing and go to finally block
return
# --- Save and Yield assistant_response_end ---
# Only save assistant_response_end if not auto-continuing (response is actually complete)
if not should_auto_continue:
if last_assistant_message_object: # Only save if assistant message was saved
try:
# Calculate response time if we have timing data
if streaming_metadata["first_chunk_time"] and streaming_metadata["last_chunk_time"]:
streaming_metadata["response_ms"] = (streaming_metadata["last_chunk_time"] - streaming_metadata["first_chunk_time"]) * 1000
# Create a LiteLLM-like response object for streaming
# Check if we have any actual usage data
has_usage_data = (
streaming_metadata["usage"]["prompt_tokens"] > 0 or
streaming_metadata["usage"]["completion_tokens"] > 0 or
streaming_metadata["usage"]["total_tokens"] > 0
)
assistant_end_content = {
"choices": [
{
"finish_reason": finish_reason or "stop",
"index": 0,
"message": {
"role": "assistant",
"content": accumulated_content,
"tool_calls": complete_native_tool_calls or None
}
}
],
"created": streaming_metadata.get("created"),
"model": streaming_metadata.get("model", llm_model),
"usage": streaming_metadata["usage"], # Always include usage like LiteLLM does
"streaming": True, # Add flag to indicate this was reconstructed from streaming
}
# Only include response_ms if we have timing data
if streaming_metadata.get("response_ms"):
assistant_end_content["response_ms"] = streaming_metadata["response_ms"]
await self.add_message(
thread_id=thread_id,
type="assistant_response_end",
content=assistant_end_content,
is_llm_message=False,
metadata={"thread_run_id": thread_run_id}
)
logger.info("Assistant response end saved for stream")
except Exception as e:
logger.error(f"Error saving assistant response end for stream: {str(e)}")
self.trace.event(name="error_saving_assistant_response_end_for_stream", level="ERROR", status_message=(f"Error saving assistant response end for stream: {str(e)}"))
except Exception as e:
logger.error(f"Error processing stream: {str(e)}", exc_info=True)
self.trace.event(name="error_processing_stream", level="ERROR", status_message=(f"Error processing stream: {str(e)}"))
# Save and yield error status message
err_content = {"role": "system", "status_type": "error", "message": str(e)}
if (not "AnthropicException - Overloaded" in str(e)):
err_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=err_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id if 'thread_run_id' in locals() else None}
)
if err_msg_obj: yield format_for_yield(err_msg_obj) # Yield the saved error message
# Re-raise the same exception (not a new one) to ensure proper error propagation
logger.critical(f"Re-raising error to stop further processing: {str(e)}")
self.trace.event(name="re_raising_error_to_stop_further_processing", level="ERROR", status_message=(f"Re-raising error to stop further processing: {str(e)}"))
else:
logger.error(f"AnthropicException - Overloaded detected - Falling back to OpenRouter: {str(e)}", exc_info=True)
self.trace.event(name="anthropic_exception_overloaded_detected", level="ERROR", status_message=(f"AnthropicException - Overloaded detected - Falling back to OpenRouter: {str(e)}"))
raise # Use bare 'raise' to preserve the original exception with its traceback
finally:
# Update continuous state for potential auto-continue
if should_auto_continue:
continuous_state['accumulated_content'] = accumulated_content
continuous_state['sequence'] = __sequence
logger.info(f"Updated continuous state for auto-continue with {len(accumulated_content)} chars")
else:
# Save and Yield the final thread_run_end status (only if not auto-continuing and finish_reason is not 'length')
try:
end_content = {"status_type": "thread_run_end"}
end_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=end_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id if 'thread_run_id' in locals() else None}
)
if end_msg_obj: yield format_for_yield(end_msg_obj)
except Exception as final_e:
logger.error(f"Error in finally block: {str(final_e)}", exc_info=True)
self.trace.event(name="error_in_finally_block", level="ERROR", status_message=(f"Error in finally block: {str(final_e)}"))
async def process_non_streaming_response(
self,
llm_response: Any,
thread_id: str,
prompt_messages: List[Dict[str, Any]],
llm_model: str,
config: ProcessorConfig = ProcessorConfig(),
) -> AsyncGenerator[Dict[str, Any], None]:
"""Process a non-streaming LLM response, handling tool calls and execution.
Args:
llm_response: Response from the LLM
thread_id: ID of the conversation thread
prompt_messages: List of messages sent to the LLM (the prompt)
llm_model: The name of the LLM model used
config: Configuration for parsing and execution
Yields:
Complete message objects matching the DB schema.
"""
content = ""
thread_run_id = str(uuid.uuid4())
all_tool_data = [] # Stores {'tool_call': ..., 'parsing_details': ...}
tool_index = 0
assistant_message_object = None
tool_result_message_objects = {}
finish_reason = None
native_tool_calls_for_message = []
try:
# Save and Yield thread_run_start status message
start_content = {"status_type": "thread_run_start", "thread_run_id": thread_run_id}
start_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=start_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if start_msg_obj: yield format_for_yield(start_msg_obj)
# Extract finish_reason, content, tool calls
if hasattr(llm_response, 'choices') and llm_response.choices:
if hasattr(llm_response.choices[0], 'finish_reason'):
finish_reason = llm_response.choices[0].finish_reason
logger.info(f"Non-streaming finish_reason: {finish_reason}")
self.trace.event(name="non_streaming_finish_reason", level="DEFAULT", status_message=(f"Non-streaming finish_reason: {finish_reason}"))
response_message = llm_response.choices[0].message if hasattr(llm_response.choices[0], 'message') else None
if response_message:
if hasattr(response_message, 'content') and response_message.content:
content = response_message.content
if config.xml_tool_calling:
parsed_xml_data = self._parse_xml_tool_calls(content)
if config.max_xml_tool_calls > 0 and len(parsed_xml_data) > config.max_xml_tool_calls:
# Truncate content and tool data if limit exceeded
# ... (Truncation logic similar to streaming) ...
if parsed_xml_data:
xml_chunks = self._extract_xml_chunks(content)[:config.max_xml_tool_calls]
if xml_chunks:
last_chunk = xml_chunks[-1]
last_chunk_pos = content.find(last_chunk)
if last_chunk_pos >= 0: content = content[:last_chunk_pos + len(last_chunk)]
parsed_xml_data = parsed_xml_data[:config.max_xml_tool_calls]
finish_reason = "xml_tool_limit_reached"
all_tool_data.extend(parsed_xml_data)
if config.native_tool_calling and hasattr(response_message, 'tool_calls') and response_message.tool_calls:
for tool_call in response_message.tool_calls:
if hasattr(tool_call, 'function'):
exec_tool_call = {
"function_name": tool_call.function.name,
"arguments": safe_json_parse(tool_call.function.arguments) if isinstance(tool_call.function.arguments, str) else tool_call.function.arguments,
"id": tool_call.id if hasattr(tool_call, 'id') else str(uuid.uuid4())
}
all_tool_data.append({"tool_call": exec_tool_call, "parsing_details": None})
native_tool_calls_for_message.append({
"id": exec_tool_call["id"], "type": "function",
"function": {
"name": tool_call.function.name,
"arguments": tool_call.function.arguments if isinstance(tool_call.function.arguments, str) else to_json_string(tool_call.function.arguments)
}
})
# --- SAVE and YIELD Final Assistant Message ---
message_data = {"role": "assistant", "content": content, "tool_calls": native_tool_calls_for_message or None}
assistant_message_object = await self._add_message_with_agent_info(
thread_id=thread_id, type="assistant", content=message_data,
is_llm_message=True, metadata={"thread_run_id": thread_run_id}
)
if assistant_message_object:
yield assistant_message_object
else:
logger.error(f"Failed to save non-streaming assistant message for thread {thread_id}")
self.trace.event(name="failed_to_save_non_streaming_assistant_message_for_thread", level="ERROR", status_message=(f"Failed to save non-streaming assistant message for thread {thread_id}"))
err_content = {"role": "system", "status_type": "error", "message": "Failed to save assistant message"}
err_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=err_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if err_msg_obj: yield format_for_yield(err_msg_obj)
# --- Execute Tools and Yield Results ---
tool_calls_to_execute = [item['tool_call'] for item in all_tool_data]
if config.execute_tools and tool_calls_to_execute:
logger.info(f"Executing {len(tool_calls_to_execute)} tools with strategy: {config.tool_execution_strategy}")
self.trace.event(name="executing_tools_with_strategy", level="DEFAULT", status_message=(f"Executing {len(tool_calls_to_execute)} tools with strategy: {config.tool_execution_strategy}"))
tool_results = await self._execute_tools(tool_calls_to_execute, config.tool_execution_strategy)
for i, (returned_tool_call, result) in enumerate(tool_results):
original_data = all_tool_data[i]
tool_call_from_data = original_data['tool_call']
parsing_details = original_data['parsing_details']
current_assistant_id = assistant_message_object['message_id'] if assistant_message_object else None
context = self._create_tool_context(
tool_call_from_data, tool_index, current_assistant_id, parsing_details
)
context.result = result
# Save and Yield start status
started_msg_obj = await self._yield_and_save_tool_started(context, thread_id, thread_run_id)
if started_msg_obj: yield format_for_yield(started_msg_obj)
# Save tool result
saved_tool_result_object = await self._add_tool_result(
thread_id, tool_call_from_data, result, config.xml_adding_strategy,
current_assistant_id, parsing_details
)
# Save and Yield completed/failed status
completed_msg_obj = await self._yield_and_save_tool_completed(
context,
saved_tool_result_object['message_id'] if saved_tool_result_object else None,
thread_id, thread_run_id
)
if completed_msg_obj: yield format_for_yield(completed_msg_obj)
# Yield the saved tool result object
if saved_tool_result_object:
tool_result_message_objects[tool_index] = saved_tool_result_object
yield format_for_yield(saved_tool_result_object)
else:
logger.error(f"Failed to save tool result for index {tool_index}")
self.trace.event(name="failed_to_save_tool_result_for_index", level="ERROR", status_message=(f"Failed to save tool result for index {tool_index}"))
tool_index += 1
# --- Save and Yield Final Status ---
if finish_reason:
finish_content = {"status_type": "finish", "finish_reason": finish_reason}
finish_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=finish_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id}
)
if finish_msg_obj: yield format_for_yield(finish_msg_obj)
# --- Save and Yield assistant_response_end ---
if assistant_message_object: # Only save if assistant message was saved
try:
# Save the full LiteLLM response object directly in content
await self.add_message(
thread_id=thread_id,
type="assistant_response_end",
content=llm_response,
is_llm_message=False,
metadata={"thread_run_id": thread_run_id}
)
logger.info("Assistant response end saved for non-stream")
except Exception as e:
logger.error(f"Error saving assistant response end for non-stream: {str(e)}")
self.trace.event(name="error_saving_assistant_response_end_for_non_stream", level="ERROR", status_message=(f"Error saving assistant response end for non-stream: {str(e)}"))
except Exception as e:
logger.error(f"Error processing non-streaming response: {str(e)}", exc_info=True)
self.trace.event(name="error_processing_non_streaming_response", level="ERROR", status_message=(f"Error processing non-streaming response: {str(e)}"))
# Save and yield error status
err_content = {"role": "system", "status_type": "error", "message": str(e)}
err_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=err_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id if 'thread_run_id' in locals() else None}
)
if err_msg_obj: yield format_for_yield(err_msg_obj)
# Re-raise the same exception (not a new one) to ensure proper error propagation
logger.critical(f"Re-raising error to stop further processing: {str(e)}")
self.trace.event(name="re_raising_error_to_stop_further_processing", level="CRITICAL", status_message=(f"Re-raising error to stop further processing: {str(e)}"))
raise # Use bare 'raise' to preserve the original exception with its traceback
finally:
# Save and Yield the final thread_run_end status
end_content = {"status_type": "thread_run_end"}
end_msg_obj = await self.add_message(
thread_id=thread_id, type="status", content=end_content,
is_llm_message=False, metadata={"thread_run_id": thread_run_id if 'thread_run_id' in locals() else None}
)
if end_msg_obj: yield format_for_yield(end_msg_obj)
def _extract_xml_chunks(self, content: str) -> List[str]:
"""Extract complete XML chunks using start and end pattern matching."""
chunks = []
pos = 0
try:
# First, look for new format <function_calls> blocks
start_pattern = '<function_calls>'
end_pattern = '</function_calls>'
while pos < len(content):
# Find the next function_calls block
start_pos = content.find(start_pattern, pos)
if start_pos == -1:
break
# Find the matching end tag
end_pos = content.find(end_pattern, start_pos)
if end_pos == -1:
break
# Extract the complete block including tags
chunk_end = end_pos + len(end_pattern)
chunk = content[start_pos:chunk_end]
chunks.append(chunk)
# Move position past this chunk
pos = chunk_end
# If no new format found, fall back to old format for backwards compatibility
if not chunks:
pos = 0
while pos < len(content):
# Find the next tool tag
next_tag_start = -1
current_tag = None
# Find the earliest occurrence of any registered tool function name
# Check for available function names
available_functions = self.tool_registry.get_available_functions()
for func_name in available_functions.keys():
# Convert function name to potential tag name (underscore to dash)
tag_name = func_name.replace('_', '-')
start_pattern = f'<{tag_name}'
tag_pos = content.find(start_pattern, pos)
if tag_pos != -1 and (next_tag_start == -1 or tag_pos < next_tag_start):
next_tag_start = tag_pos
current_tag = tag_name
if next_tag_start == -1 or not current_tag:
break
# Find the matching end tag
end_pattern = f'</{current_tag}>'
tag_stack = []
chunk_start = next_tag_start
current_pos = next_tag_start
while current_pos < len(content):
# Look for next start or end tag of the same type
next_start = content.find(f'<{current_tag}', current_pos + 1)
next_end = content.find(end_pattern, current_pos)
if next_end == -1: # No closing tag found
break
if next_start != -1 and next_start < next_end:
# Found nested start tag
tag_stack.append(next_start)
current_pos = next_start + 1
else:
# Found end tag
if not tag_stack: # This is our matching end tag
chunk_end = next_end + len(end_pattern)
chunk = content[chunk_start:chunk_end]
chunks.append(chunk)
pos = chunk_end
break
else:
# Pop nested tag
tag_stack.pop()
current_pos = next_end + 1
if current_pos >= len(content): # Reached end without finding closing tag
break
pos = max(pos + 1, current_pos)
except Exception as e:
logger.error(f"Error extracting XML chunks: {e}")
logger.error(f"Content was: {content}")
self.trace.event(name="error_extracting_xml_chunks", level="ERROR", status_message=(f"Error extracting XML chunks: {e}"), metadata={"content": content})
return chunks
def _parse_xml_tool_call(self, xml_chunk: str) -> Optional[Tuple[Dict[str, Any], Dict[str, Any]]]:
"""Parse XML chunk into tool call format and return parsing details.
Returns:
Tuple of (tool_call, parsing_details) or None if parsing fails.
- tool_call: Dict with 'function_name', 'xml_tag_name', 'arguments'
- parsing_details: Dict with 'attributes', 'elements', 'text_content', 'root_content'
"""
try:
# Check if this is the new format (contains <function_calls>)
if '<function_calls>' in xml_chunk and '<invoke' in xml_chunk:
# Use the new XML parser
parsed_calls = self.xml_parser.parse_content(xml_chunk)
if not parsed_calls:
logger.error(f"No tool calls found in XML chunk: {xml_chunk}")
return None
# Take the first tool call (should only be one per chunk)
xml_tool_call = parsed_calls[0]
# Convert to the expected format
tool_call = {
"function_name": xml_tool_call.function_name,
"xml_tag_name": xml_tool_call.function_name.replace('_', '-'), # For backwards compatibility
"arguments": xml_tool_call.parameters
}
# Include the parsing details
parsing_details = xml_tool_call.parsing_details
parsing_details["raw_xml"] = xml_tool_call.raw_xml
logger.debug(f"Parsed new format tool call: {tool_call}")
return tool_call, parsing_details
# If not the expected <function_calls><invoke> format, return None
logger.error(f"XML chunk does not contain expected <function_calls><invoke> format: {xml_chunk}")
return None
except Exception as e:
logger.error(f"Error parsing XML chunk: {e}")
logger.error(f"XML chunk was: {xml_chunk}")
self.trace.event(name="error_parsing_xml_chunk", level="ERROR", status_message=(f"Error parsing XML chunk: {e}"), metadata={"xml_chunk": xml_chunk})
return None
def _parse_xml_tool_calls(self, content: str) -> List[Dict[str, Any]]:
"""Parse XML tool calls from content string.
Returns:
List of dictionaries, each containing {'tool_call': ..., 'parsing_details': ...}
"""
parsed_data = []
try:
xml_chunks = self._extract_xml_chunks(content)
for xml_chunk in xml_chunks:
result = self._parse_xml_tool_call(xml_chunk)
if result:
tool_call, parsing_details = result
parsed_data.append({
"tool_call": tool_call,
"parsing_details": parsing_details
})
except Exception as e:
logger.error(f"Error parsing XML tool calls: {e}", exc_info=True)
self.trace.event(name="error_parsing_xml_tool_calls", level="ERROR", status_message=(f"Error parsing XML tool calls: {e}"), metadata={"content": content})
return parsed_data
# Tool execution methods
async def _execute_tool(self, tool_call: Dict[str, Any]) -> ToolResult:
"""Execute a single tool call and return the result."""
span = self.trace.span(name=f"execute_tool.{tool_call['function_name']}", input=tool_call["arguments"])
try:
function_name = tool_call["function_name"]
arguments = tool_call["arguments"]
logger.info(f"Executing tool: {function_name} with arguments: {arguments}")
self.trace.event(name="executing_tool", level="DEFAULT", status_message=(f"Executing tool: {function_name} with arguments: {arguments}"))
if isinstance(arguments, str):
try:
arguments = safe_json_parse(arguments)
except json.JSONDecodeError:
arguments = {"text": arguments}
# Get available functions from tool registry
available_functions = self.tool_registry.get_available_functions()
# Look up the function by name
tool_fn = available_functions.get(function_name)
if not tool_fn:
logger.error(f"Tool function '{function_name}' not found in registry")
span.end(status_message="tool_not_found", level="ERROR")
return ToolResult(success=False, output=f"Tool function '{function_name}' not found")
logger.debug(f"Found tool function for '{function_name}', executing...")
result = await tool_fn(**arguments)
logger.info(f"Tool execution complete: {function_name} -> {result}")
span.end(status_message="tool_executed", output=result)
return result
except Exception as e:
logger.error(f"Error executing tool {tool_call['function_name']}: {str(e)}", exc_info=True)
span.end(status_message="tool_execution_error", output=f"Error executing tool: {str(e)}", level="ERROR")
return ToolResult(success=False, output=f"Error executing tool: {str(e)}")
async def _execute_tools(
self,
tool_calls: List[Dict[str, Any]],
execution_strategy: ToolExecutionStrategy = "sequential"
) -> List[Tuple[Dict[str, Any], ToolResult]]:
"""Execute tool calls with the specified strategy.
This is the main entry point for tool execution. It dispatches to the appropriate
execution method based on the provided strategy.
Args:
tool_calls: List of tool calls to execute
execution_strategy: Strategy for executing tools:
- "sequential": Execute tools one after another, waiting for each to complete
- "parallel": Execute all tools simultaneously for better performance
Returns:
List of tuples containing the original tool call and its result
"""
logger.info(f"Executing {len(tool_calls)} tools with strategy: {execution_strategy}")
self.trace.event(name="executing_tools_with_strategy", level="DEFAULT", status_message=(f"Executing {len(tool_calls)} tools with strategy: {execution_strategy}"))
if execution_strategy == "sequential":
return await self._execute_tools_sequentially(tool_calls)
elif execution_strategy == "parallel":
return await self._execute_tools_in_parallel(tool_calls)
else:
logger.warning(f"Unknown execution strategy: {execution_strategy}, falling back to sequential")
return await self._execute_tools_sequentially(tool_calls)
async def _execute_tools_sequentially(self, tool_calls: List[Dict[str, Any]]) -> List[Tuple[Dict[str, Any], ToolResult]]:
"""Execute tool calls sequentially and return results.
This method executes tool calls one after another, waiting for each tool to complete
before starting the next one. This is useful when tools have dependencies on each other.
Args:
tool_calls: List of tool calls to execute
Returns:
List of tuples containing the original tool call and its result
"""
if not tool_calls:
return []
try:
tool_names = [t.get('function_name', 'unknown') for t in tool_calls]
logger.info(f"Executing {len(tool_calls)} tools sequentially: {tool_names}")
self.trace.event(name="executing_tools_sequentially", level="DEFAULT", status_message=(f"Executing {len(tool_calls)} tools sequentially: {tool_names}"))
results = []
for index, tool_call in enumerate(tool_calls):
tool_name = tool_call.get('function_name', 'unknown')
logger.debug(f"Executing tool {index+1}/{len(tool_calls)}: {tool_name}")
try:
result = await self._execute_tool(tool_call)
results.append((tool_call, result))
logger.debug(f"Completed tool {tool_name} with success={result.success}")
# Check if this is a terminating tool (ask or complete)
if tool_name in ['ask', 'complete']:
logger.info(f"Terminating tool '{tool_name}' executed. Stopping further tool execution.")
self.trace.event(name="terminating_tool_executed", level="DEFAULT", status_message=(f"Terminating tool '{tool_name}' executed. Stopping further tool execution."))
break # Stop executing remaining tools
except Exception as e:
logger.error(f"Error executing tool {tool_name}: {str(e)}")
self.trace.event(name="error_executing_tool", level="ERROR", status_message=(f"Error executing tool {tool_name}: {str(e)}"))
error_result = ToolResult(success=False, output=f"Error executing tool: {str(e)}")
results.append((tool_call, error_result))
logger.info(f"Sequential execution completed for {len(results)} tools (out of {len(tool_calls)} total)")
self.trace.event(name="sequential_execution_completed", level="DEFAULT", status_message=(f"Sequential execution completed for {len(results)} tools (out of {len(tool_calls)} total)"))
return results
except Exception as e:
logger.error(f"Error in sequential tool execution: {str(e)}", exc_info=True)
# Return partial results plus error results for remaining tools
completed_results = results if 'results' in locals() else []
completed_tool_names = [r[0].get('function_name', 'unknown') for r in completed_results]
remaining_tools = [t for t in tool_calls if t.get('function_name', 'unknown') not in completed_tool_names]
# Add error results for remaining tools
error_results = [(tool, ToolResult(success=False, output=f"Execution error: {str(e)}"))
for tool in remaining_tools]
return completed_results + error_results
async def _execute_tools_in_parallel(self, tool_calls: List[Dict[str, Any]]) -> List[Tuple[Dict[str, Any], ToolResult]]:
"""Execute tool calls in parallel and return results.
This method executes all tool calls simultaneously using asyncio.gather, which
can significantly improve performance when executing multiple independent tools.
Args:
tool_calls: List of tool calls to execute
Returns:
List of tuples containing the original tool call and its result
"""
if not tool_calls:
return []
try:
tool_names = [t.get('function_name', 'unknown') for t in tool_calls]
logger.info(f"Executing {len(tool_calls)} tools in parallel: {tool_names}")
self.trace.event(name="executing_tools_in_parallel", level="DEFAULT", status_message=(f"Executing {len(tool_calls)} tools in parallel: {tool_names}"))
# Create tasks for all tool calls
tasks = [self._execute_tool(tool_call) for tool_call in tool_calls]
# Execute all tasks concurrently with error handling
results = await asyncio.gather(*tasks, return_exceptions=True)
# Process results and handle any exceptions
processed_results = []
for i, (tool_call, result) in enumerate(zip(tool_calls, results)):
if isinstance(result, Exception):
logger.error(f"Error executing tool {tool_call.get('function_name', 'unknown')}: {str(result)}")
self.trace.event(name="error_executing_tool", level="ERROR", status_message=(f"Error executing tool {tool_call.get('function_name', 'unknown')}: {str(result)}"))
# Create error result
error_result = ToolResult(success=False, output=f"Error executing tool: {str(result)}")
processed_results.append((tool_call, error_result))
else:
processed_results.append((tool_call, result))
logger.info(f"Parallel execution completed for {len(tool_calls)} tools")
self.trace.event(name="parallel_execution_completed", level="DEFAULT", status_message=(f"Parallel execution completed for {len(tool_calls)} tools"))
return processed_results
except Exception as e:
logger.error(f"Error in parallel tool execution: {str(e)}", exc_info=True)
self.trace.event(name="error_in_parallel_tool_execution", level="ERROR", status_message=(f"Error in parallel tool execution: {str(e)}"))
# Return error results for all tools if the gather itself fails
return [(tool_call, ToolResult(success=False, output=f"Execution error: {str(e)}"))
for tool_call in tool_calls]
async def _add_tool_result(
self,
thread_id: str,
tool_call: Dict[str, Any],
result: ToolResult,
strategy: Union[XmlAddingStrategy, str] = "assistant_message",
assistant_message_id: Optional[str] = None,
parsing_details: Optional[Dict[str, Any]] = None
) -> Optional[Dict[str, Any]]: # Return the full message object
"""Add a tool result to the conversation thread based on the specified format.
This method formats tool results and adds them to the conversation history,
making them visible to the LLM in subsequent interactions. Results can be
added either as native tool messages (OpenAI format) or as XML-wrapped content
with a specified role (user or assistant).
Args:
thread_id: ID of the conversation thread
tool_call: The original tool call that produced this result
result: The result from the tool execution
strategy: How to add XML tool results to the conversation
("user_message", "assistant_message", or "inline_edit")
assistant_message_id: ID of the assistant message that generated this tool call
parsing_details: Detailed parsing info for XML calls (attributes, elements, etc.)
"""
try:
message_obj = None # Initialize message_obj
# Create metadata with assistant_message_id if provided
metadata = {}
if assistant_message_id:
metadata["assistant_message_id"] = assistant_message_id
logger.info(f"Linking tool result to assistant message: {assistant_message_id}")
self.trace.event(name="linking_tool_result_to_assistant_message", level="DEFAULT", status_message=(f"Linking tool result to assistant message: {assistant_message_id}"))
# --- Add parsing details to metadata if available ---
if parsing_details:
metadata["parsing_details"] = parsing_details
logger.info("Adding parsing_details to tool result metadata")
self.trace.event(name="adding_parsing_details_to_tool_result_metadata", level="DEFAULT", status_message=(f"Adding parsing_details to tool result metadata"), metadata={"parsing_details": parsing_details})
# ---
# Check if this is a native function call (has id field)
if "id" in tool_call:
# Format as a proper tool message according to OpenAI spec
function_name = tool_call.get("function_name", "")
# Format the tool result content - tool role needs string content
if isinstance(result, str):
content = result
elif hasattr(result, 'output'):
# If it's a ToolResult object
if isinstance(result.output, dict) or isinstance(result.output, list):
# If output is already a dict or list, convert to JSON string
content = json.dumps(result.output)
else:
# Otherwise just use the string representation
content = str(result.output)
else:
# Fallback to string representation of the whole result
content = str(result)
logger.info(f"Formatted tool result content: {content[:100]}...")
self.trace.event(name="formatted_tool_result_content", level="DEFAULT", status_message=(f"Formatted tool result content: {content[:100]}..."))
# Create the tool response message with proper format
tool_message = {
"role": "tool",
"tool_call_id": tool_call["id"],
"name": function_name,
"content": content
}
logger.info(f"Adding native tool result for tool_call_id={tool_call['id']} with role=tool")
self.trace.event(name="adding_native_tool_result_for_tool_call_id", level="DEFAULT", status_message=(f"Adding native tool result for tool_call_id={tool_call['id']} with role=tool"))
# Add as a tool message to the conversation history
# This makes the result visible to the LLM in the next turn
message_obj = await self.add_message(
thread_id=thread_id,
type="tool", # Special type for tool responses
content=tool_message,
is_llm_message=True,
metadata=metadata
)
return message_obj # Return the full message object
# For XML and other non-native tools, use the new structured format
# Determine message role based on strategy
result_role = "user" if strategy == "user_message" else "assistant"
# Create two versions of the structured result
# 1. Rich version for the frontend
structured_result_for_frontend = self._create_structured_tool_result(tool_call, result, parsing_details, for_llm=False)
# 2. Concise version for the LLM
structured_result_for_llm = self._create_structured_tool_result(tool_call, result, parsing_details, for_llm=True)
# Add the message with the appropriate role to the conversation history
# This allows the LLM to see the tool result in subsequent interactions
result_message_for_llm = {
"role": result_role,
"content": json.dumps(structured_result_for_llm)
}
# Add rich content to metadata for frontend use
if metadata is None:
metadata = {}
metadata['frontend_content'] = structured_result_for_frontend
message_obj = await self._add_message_with_agent_info(
thread_id=thread_id,
type="tool",
content=result_message_for_llm, # Save the LLM-friendly version
is_llm_message=True,
metadata=metadata
)
# If the message was saved, modify it in-memory for the frontend before returning
if message_obj:
# The frontend expects the rich content in the 'content' field.
# The DB has the rich content in metadata.frontend_content.
# Let's reconstruct the message for yielding.
message_for_yield = message_obj.copy()
message_for_yield['content'] = structured_result_for_frontend
return message_for_yield
return message_obj # Return the modified message object
except Exception as e:
logger.error(f"Error adding tool result: {str(e)}", exc_info=True)
self.trace.event(name="error_adding_tool_result", level="ERROR", status_message=(f"Error adding tool result: {str(e)}"), metadata={"tool_call": tool_call, "result": result, "strategy": strategy, "assistant_message_id": assistant_message_id, "parsing_details": parsing_details})
# Fallback to a simple message
try:
fallback_message = {
"role": "user",
"content": str(result)
}
message_obj = await self.add_message(
thread_id=thread_id,
type="tool",
content=fallback_message,
is_llm_message=True,
metadata={"assistant_message_id": assistant_message_id} if assistant_message_id else {}
)
return message_obj # Return the full message object
except Exception as e2:
logger.error(f"Failed even with fallback message: {str(e2)}", exc_info=True)
self.trace.event(name="failed_even_with_fallback_message", level="ERROR", status_message=(f"Failed even with fallback message: {str(e2)}"), metadata={"tool_call": tool_call, "result": result, "strategy": strategy, "assistant_message_id": assistant_message_id, "parsing_details": parsing_details})
return None # Return None on error
def _create_structured_tool_result(self, tool_call: Dict[str, Any], result: ToolResult, parsing_details: Optional[Dict[str, Any]] = None, for_llm: bool = False):
"""Create a structured tool result format that's tool-agnostic and provides rich information.
Args:
tool_call: The original tool call that was executed
result: The result from the tool execution
parsing_details: Optional parsing details for XML calls
for_llm: If True, creates a concise version for the LLM context.
Returns:
Structured dictionary containing tool execution information
"""
# Extract tool information
function_name = tool_call.get("function_name", "unknown")
xml_tag_name = tool_call.get("xml_tag_name")
arguments = tool_call.get("arguments", {})
tool_call_id = tool_call.get("id")
# Process the output - if it's a JSON string, parse it back to an object
output = result.output if hasattr(result, 'output') else str(result)
if isinstance(output, str):
try:
# Try to parse as JSON to provide structured data to frontend
parsed_output = safe_json_parse(output)
# If parsing succeeded and we got a dict/list, use the parsed version
if isinstance(parsed_output, (dict, list)):
output = parsed_output
# Otherwise keep the original string
except Exception:
# If parsing fails, keep the original string
pass
output_to_use = output
# If this is for the LLM and it's an edit_file tool, create a concise output
if for_llm and function_name == 'edit_file' and isinstance(output, dict):
# The frontend needs original_content and updated_content to render diffs.
# The concise version for the LLM was causing issues.
# We will now pass the full output, and rely on the ContextManager to truncate if needed.
output_to_use = output
# Create the structured result
structured_result_v1 = {
"tool_execution": {
"function_name": function_name,
"xml_tag_name": xml_tag_name,
"tool_call_id": tool_call_id,
"arguments": arguments,
"result": {
"success": result.success if hasattr(result, 'success') else True,
"output": output_to_use, # This will be either rich or concise based on `for_llm`
"error": getattr(result, 'error', None) if hasattr(result, 'error') else None
},
}
}
return structured_result_v1
def _create_tool_context(self, tool_call: Dict[str, Any], tool_index: int, assistant_message_id: Optional[str] = None, parsing_details: Optional[Dict[str, Any]] = None) -> ToolExecutionContext:
"""Create a tool execution context with display name and parsing details populated."""
context = ToolExecutionContext(
tool_call=tool_call,
tool_index=tool_index,
assistant_message_id=assistant_message_id,
parsing_details=parsing_details
)
# Set function_name and xml_tag_name fields
if "xml_tag_name" in tool_call:
context.xml_tag_name = tool_call["xml_tag_name"]
context.function_name = tool_call.get("function_name", tool_call["xml_tag_name"])
else:
# For non-XML tools, use function name directly
context.function_name = tool_call.get("function_name", "unknown")
context.xml_tag_name = None
return context
async def _yield_and_save_tool_started(self, context: ToolExecutionContext, thread_id: str, thread_run_id: str) -> Optional[Dict[str, Any]]:
"""Formats, saves, and returns a tool started status message."""
tool_name = context.xml_tag_name or context.function_name
content = {
"role": "assistant", "status_type": "tool_started",
"function_name": context.function_name, "xml_tag_name": context.xml_tag_name,
"message": f"Starting execution of {tool_name}", "tool_index": context.tool_index,
"tool_call_id": context.tool_call.get("id") # Include tool_call ID if native
}
metadata = {"thread_run_id": thread_run_id}
saved_message_obj = await self.add_message(
thread_id=thread_id, type="status", content=content, is_llm_message=False, metadata=metadata
)
return saved_message_obj # Return the full object (or None if saving failed)
async def _yield_and_save_tool_completed(self, context: ToolExecutionContext, tool_message_id: Optional[str], thread_id: str, thread_run_id: str) -> Optional[Dict[str, Any]]:
"""Formats, saves, and returns a tool completed/failed status message."""
if not context.result:
# Delegate to error saving if result is missing (e.g., execution failed)
return await self._yield_and_save_tool_error(context, thread_id, thread_run_id)
tool_name = context.xml_tag_name or context.function_name
status_type = "tool_completed" if context.result.success else "tool_failed"
message_text = f"Tool {tool_name} {'completed successfully' if context.result.success else 'failed'}"
content = {
"role": "assistant", "status_type": status_type,
"function_name": context.function_name, "xml_tag_name": context.xml_tag_name,
"message": message_text, "tool_index": context.tool_index,
"tool_call_id": context.tool_call.get("id")
}
metadata = {"thread_run_id": thread_run_id}
# Add the *actual* tool result message ID to the metadata if available and successful
if context.result.success and tool_message_id:
metadata["linked_tool_result_message_id"] = tool_message_id
# <<< ADDED: Signal if this is a terminating tool >>>
if context.function_name in ['ask', 'complete']:
metadata["agent_should_terminate"] = "true"
logger.info(f"Marking tool status for '{context.function_name}' with termination signal.")
self.trace.event(name="marking_tool_status_for_termination", level="DEFAULT", status_message=(f"Marking tool status for '{context.function_name}' with termination signal."))
# <<< END ADDED >>>
saved_message_obj = await self.add_message(
thread_id=thread_id, type="status", content=content, is_llm_message=False, metadata=metadata
)
return saved_message_obj
async def _yield_and_save_tool_error(self, context: ToolExecutionContext, thread_id: str, thread_run_id: str) -> Optional[Dict[str, Any]]:
"""Formats, saves, and returns a tool error status message."""
error_msg = str(context.error) if context.error else "Unknown error during tool execution"
tool_name = context.xml_tag_name or context.function_name
content = {
"role": "assistant", "status_type": "tool_error",
"function_name": context.function_name, "xml_tag_name": context.xml_tag_name,
"message": f"Error executing tool {tool_name}: {error_msg}",
"tool_index": context.tool_index,
"tool_call_id": context.tool_call.get("id")
}
metadata = {"thread_run_id": thread_run_id}
# Save the status message with is_llm_message=False
saved_message_obj = await self.add_message(
thread_id=thread_id, type="status", content=content, is_llm_message=False, metadata=metadata
)
return saved_message_obj
Reference in New Issue View Git Blame Copy Permalink