mcpmark / src /agents /mcpmark_agent.py
haochengsama's picture
Add files using upload-large-folder tool
a2ec7b6 verified
Raw
History Blame Contribute Delete
54.3 kB
"""
MCPMark Agent Implementation
============================
Unified agent using LiteLLM for all model interactions with minimal MCP support.
"""
import asyncio
import json
import os
import time
from typing import Any, Dict, List, Optional, Callable
from pydantic import AnyUrl
import httpx
import litellm
import nest_asyncio
from src.logger import get_logger
from .base_agent import BaseMCPAgent
from .mcp import MCPStdioServer, MCPHttpServer
# Apply nested asyncio support
nest_asyncio.apply()
# Configure LiteLLM
litellm.suppress_debug_info = True
logger = get_logger(__name__)
# To fix the "Object of type AnyUrl is not JSON serializable" error in the find_file_contents function.
class CustomJSONEncoder(json.JSONEncoder):
def default(self, obj):
if isinstance(obj, AnyUrl):
return str(obj)
return super().default(obj)
class MCPMarkAgent(BaseMCPAgent):
"""
Unified agent for LLM and MCP server management using LiteLLM.
- Anthropic models: Native MCP support via extra_body
- Other models: Manual MCP server management with function calling
"""
MAX_TURNS = 100
SYSTEM_PROMPT = (
"You are a helpful agent that uses tools iteratively to complete the user's task, "
'and when finished, provides the final answer or simply states "Task completed" without further tool calls.'
)
COMPACTION_PROMPT = (
"You are performing a CONTEXT CHECKPOINT COMPACTION.\n"
"Summarize the conversation so far for another model to continue.\n\n"
"Include:\n"
"- Current progress and key decisions made\n"
"- Important context, constraints, or user preferences\n"
"- What remains to be done (clear next steps)\n"
"- Any critical data, examples, or references needed to continue\n\n"
"Be concise and structured. Do NOT call tools."
)
# Per-tool-response summarizer (mirrors mcp-universe's
# TOOL_RESPONSE_SUMMARIZER_PROMPT). Applied to browser tool results BEFORE
# they enter the history, so the model conditions on — and the trajectory
# records — the SAME compressed text. "Extract ALL relevant info" + "keep
# structure" is what preserves the exact values a webarena task must read
# off the page (counts, votes, prices) through the compression.
TOOL_RESPONSE_SUMMARIZER_PROMPT = (
"Extract all information from the tool response that is relevant to the context.\n\n"
"Tool Call Context:\n{context}\n\n"
"Tool Response:\n{tool_response}\n\n"
"Directly output the extracted information. Try to maintain the original "
"response structure. Use fewer than 500 words."
)
DEFAULT_TIMEOUT = BaseMCPAgent.DEFAULT_TIMEOUT
def __init__(
self,
litellm_input_model_name: str,
api_key: str,
base_url: str,
mcp_service: str,
timeout: int = DEFAULT_TIMEOUT,
service_config: Optional[Dict[str, Any]] = None,
service_config_provider: Optional[Callable[[], Dict[str, Any]]] = None,
reasoning_effort: Optional[str] = "default",
compaction_token: int = BaseMCPAgent.COMPACTION_DISABLED_TOKEN,
extra_body: Optional[Dict[str, Any]] = None,
summarize_tool_response: bool = False,
):
super().__init__(
litellm_input_model_name=litellm_input_model_name,
api_key=api_key,
base_url=base_url,
mcp_service=mcp_service,
timeout=timeout,
service_config=service_config,
service_config_provider=service_config_provider,
reasoning_effort=reasoning_effort,
compaction_token=compaction_token,
extra_body=extra_body,
summarize_tool_response=summarize_tool_response,
)
logger.debug(
"Initialized MCPMarkAgent for '%s' with model '%s' (Claude: %s, Thinking: %s, Reasoning: %s)",
mcp_service,
litellm_input_model_name,
self.is_claude,
self.use_claude_thinking,
reasoning_effort,
)
# ==================== Public Interface Methods ====================
async def execute(
self, instruction: str, tool_call_log_file: Optional[str] = None
) -> Dict[str, Any]:
"""
Execute instruction with the agent.
Args:
instruction: The instruction/prompt to execute
tool_call_log_file: Optional path to log tool calls
Returns:
Dictionary containing execution results
"""
start_time = time.time()
try:
# Reset partial progress for this run
self._reset_progress()
# Refresh service configuration
self._refresh_service_config()
# Execute with timeout control
async def _execute_with_strategy():
if self.use_claude_thinking:
# Claude with thinking -> native Anthropic API with tools
return await self._execute_claude_native_with_tools(
instruction, tool_call_log_file
)
else:
# All other cases -> LiteLLM with tools
return await self._execute_litellm_with_tools(
instruction, tool_call_log_file
)
# Apply timeout to the entire execution
result = await asyncio.wait_for(
_execute_with_strategy(), timeout=self.timeout
)
execution_time = time.time() - start_time
# Update usage statistics
self.usage_tracker.update(
success=result["success"],
token_usage=result.get("token_usage", {}),
turn_count=result.get("turn_count", 0),
execution_time=execution_time,
)
result["execution_time"] = execution_time
return result
except Exception as e:
execution_time = time.time() - start_time
if isinstance(e, asyncio.TimeoutError):
error_msg = f"Execution timed out after {self.timeout} seconds"
logger.error(error_msg)
else:
error_msg = f"Agent execution failed: {e}"
logger.error(error_msg, exc_info=True)
self.usage_tracker.update(
success=False,
token_usage=self._partial_token_usage or {},
turn_count=self._partial_turn_count or 0,
execution_time=execution_time,
)
if self._partial_messages:
if not self.is_claude:
final_msg = self._convert_to_sdk_format(self._partial_messages)
else:
final_msg = self._partial_messages
else:
final_msg = []
return {
"success": False,
"output": final_msg,
"token_usage": self._partial_token_usage or {},
"turn_count": self._partial_turn_count or 0,
"execution_time": execution_time,
"error": error_msg,
"litellm_run_model_name": self.litellm_run_model_name,
}
def execute_sync(
self, instruction: str, tool_call_log_file: Optional[str] = None
) -> Dict[str, Any]:
"""
Synchronous wrapper for execute method.
"""
return asyncio.run(self.execute(instruction, tool_call_log_file))
def get_usage_stats(self) -> Dict[str, Any]:
"""Get usage statistics."""
return self.usage_tracker.get_stats()
def reset_usage_stats(self):
"""Reset usage statistics."""
self.usage_tracker.reset()
# ==================== Claude Native API Execution Path ====================
async def _execute_claude_native_with_tools(
self, instruction: str, tool_call_log_file: Optional[str] = None
) -> Dict[str, Any]:
"""
Execute Claude with thinking using native Anthropic API.
Creates MCP server, gets tools, and executes with thinking.
"""
logger.debug("Using Claude native API with thinking")
thinking_budget = self._get_claude_thinking_budget()
# Create and start MCP server
mcp_server = await self._create_mcp_server()
async with mcp_server:
# Get available tools
tools = await mcp_server.list_tools()
# Convert MCP tools to Anthropic format
anthropic_tools = self._convert_to_anthropic_format(tools)
# Execute with function calling loop
return await self._execute_anthropic_native_tool_loop(
instruction,
anthropic_tools,
mcp_server,
thinking_budget,
tool_call_log_file,
)
async def _call_claude_native_api(
self,
messages: List[Dict],
thinking_budget: int,
tools: Optional[List[Dict]] = None,
mcp_servers: Optional[List[Dict]] = None,
system: Optional[str] = None,
) -> Dict[str, Any]:
"""
Call Claude's native API directly using httpx.
Args:
messages: Conversation messages
thinking_budget: Token budget for thinking
tools: Tool definitions for function calling
mcp_servers: MCP server configurations
system: System prompt
Returns:
API response as dictionary
"""
# Get API base and headers
import os
api_base = os.getenv("ANTHROPIC_API_BASE", "https://api.anthropic.com")
headers = {
"x-api-key": self.api_key,
"anthropic-version": "2023-06-01",
"content-type": "application/json",
"anthropic-beta": "context-1m-2025-08-07", # by default
}
# Build payload
max_tokens = max(thinking_budget + 4096, 4096)
payload = {
"model": self.litellm_input_model_name.replace("anthropic/", ""),
"max_tokens": max_tokens,
"messages": messages,
}
# Add thinking configuration
if thinking_budget:
payload["thinking"] = {"type": "enabled", "budget_tokens": thinking_budget}
# Add tools if provided
if tools:
payload["tools"] = tools
payload["tool_choice"] = {"type": "auto"}
# Add MCP servers if provided
if mcp_servers:
headers["anthropic-beta"] = "mcp-client-2025-04-04"
payload["mcp_servers"] = mcp_servers
# Add system prompt if provided
if system:
payload["system"] = system
# Make the API call
async with httpx.AsyncClient() as client:
try:
response = await client.post(
f"{api_base}/v1/messages",
headers=headers,
json=payload,
timeout=self.timeout,
)
response.raise_for_status()
return response.json(), None
except httpx.HTTPStatusError as e:
return None, e.response.text
except Exception as e:
return None, e
async def _count_claude_input_tokens(
self,
messages: List[Dict[str, Any]],
tools: Optional[List[Dict]] = None,
system: Optional[str] = None,
) -> int:
import os
api_base = os.getenv("ANTHROPIC_API_BASE", "https://api.anthropic.com")
headers = {
"x-api-key": self.api_key,
"anthropic-version": "2023-06-01",
"content-type": "application/json",
}
payload: Dict[str, Any] = {
"model": self.litellm_input_model_name.replace("anthropic/", ""),
"messages": messages,
}
if tools:
payload["tools"] = tools
if system:
payload["system"] = system
async with httpx.AsyncClient() as client:
response = await client.post(
f"{api_base}/v1/messages/count_tokens",
headers=headers,
json=payload,
timeout=self.timeout,
)
response.raise_for_status()
data = response.json() or {}
return int(data.get("input_tokens", 0) or 0)
def _extract_litellm_text(self, response: Any) -> str:
try:
choices = getattr(response, "choices", None) or []
if not choices:
return ""
msg = getattr(choices[0], "message", None)
if msg is not None:
return str(getattr(msg, "content", "") or "")
return str(getattr(choices[0], "text", "") or "")
except Exception: # pragma: no cover - best effort
return ""
def _extract_anthropic_text(self, response_json: Dict[str, Any]) -> str:
pieces: List[str] = []
for block in response_json.get("content", []) or []:
if isinstance(block, dict) and block.get("type") == "text":
text = block.get("text")
if text:
pieces.append(str(text))
return "\n".join(pieces).strip()
def _merge_usage(self, total_tokens: Dict[str, int], usage: Dict[str, Any]) -> None:
try:
input_tokens = int(usage.get("input_tokens", 0) or 0)
output_tokens = int(usage.get("output_tokens", 0) or 0)
total_tokens_count = int(
usage.get("total_tokens", 0) or (input_tokens + output_tokens)
)
total_tokens["input_tokens"] += input_tokens
total_tokens["output_tokens"] += output_tokens
total_tokens["total_tokens"] += total_tokens_count
except Exception: # pragma: no cover - best effort
return
async def _maybe_compact_litellm_messages(
self,
messages: List[Dict[str, Any]],
total_tokens: Dict[str, int],
tool_call_log_file: Optional[str],
current_prompt_tokens: int,
) -> List[Dict[str, Any]]:
if not self._compaction_enabled():
return messages
if current_prompt_tokens < self.compaction_token:
return messages
logger.info(
f"| [compaction] Triggered at prompt tokens: {current_prompt_tokens:,}"
)
if tool_call_log_file:
try:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(
f"| [compaction] Triggered at prompt tokens: {current_prompt_tokens:,}\n"
)
except Exception:
pass
compact_messages = [
{"role": "system", "content": self.COMPACTION_PROMPT},
{"role": "user", "content": json.dumps(messages, ensure_ascii=False)},
]
completion_kwargs = {
"model": self.litellm_input_model_name,
"messages": compact_messages,
"api_key": self.api_key,
}
if self.base_url:
completion_kwargs["base_url"] = self.base_url
if self.extra_body:
completion_kwargs["extra_body"] = self.extra_body
response = await litellm.acompletion(**completion_kwargs)
usage = getattr(response, "usage", None)
if usage:
input_tokens = (
getattr(usage, "prompt_tokens", None)
or getattr(usage, "input_tokens", None)
or 0
)
output_tokens = (
getattr(usage, "completion_tokens", None)
or getattr(usage, "output_tokens", None)
or 0
)
total_tokens_count = getattr(usage, "total_tokens", None)
if total_tokens_count is None:
total_tokens_count = input_tokens + output_tokens
total_tokens["input_tokens"] += int(input_tokens or 0)
total_tokens["output_tokens"] += int(output_tokens or 0)
total_tokens["total_tokens"] += int(total_tokens_count or 0)
summary = self._extract_litellm_text(response).strip() or "(no summary)"
system_msg = (
messages[0]
if messages
else {"role": "system", "content": self.SYSTEM_PROMPT}
)
first_user = (
messages[1] if len(messages) > 1 else {"role": "user", "content": ""}
)
return [
system_msg,
first_user,
{
"role": "user",
"content": f"Context summary (auto-compacted due to token limit):\n{summary}",
},
]
def _should_summarize_tool(self) -> bool:
"""Summarize tool responses only for playwright (browser) runs.
Browser tool results are giant accessibility trees that are navigation
context, not the final answer — safe to compress. Data services
(postgres/github/notion/filesystem) often return the exact value the
evaluator checks, so they are intentionally NOT summarized.
"""
return bool(self.summarize_tool_response) and self.mcp_service in (
"playwright",
"playwright_webarena",
)
async def _summarize_tool_response_litellm(
self,
tool_response: str,
context: str,
total_tokens: Dict[str, int],
) -> str:
"""Compress a tool response via the LLM BEFORE it enters the message
history (mirrors mcp-universe's summarize_tool_response).
The returned text is appended to ``messages`` — which is exactly what
gets saved as the trajectory — so the model conditions on, and the
record stores, the SAME content (no post-hoc divergence). On any
failure the raw response is returned unchanged so a rollout never
breaks. The summarizer's own token usage is folded into total_tokens.
"""
prompt = self.TOOL_RESPONSE_SUMMARIZER_PROMPT.format(
context=context, tool_response=tool_response
)
completion_kwargs = {
"model": self.litellm_input_model_name,
"messages": [{"role": "user", "content": prompt}],
"api_key": self.api_key,
}
if self.base_url:
completion_kwargs["base_url"] = self.base_url
if self.extra_body:
completion_kwargs["extra_body"] = self.extra_body
try:
response = await asyncio.wait_for(
litellm.acompletion(**completion_kwargs),
timeout=self.timeout / 2,
)
except Exception as e: # noqa: BLE001 - keep raw on any failure
logger.warning(f"| [summarize] failed, keeping raw tool response: {e}")
return tool_response
usage = getattr(response, "usage", None)
if usage:
input_tokens = getattr(usage, "prompt_tokens", 0) or 0
total_tokens_count = getattr(usage, "total_tokens", 0) or 0
output_tokens = (
total_tokens_count - input_tokens
if total_tokens_count > 0
else (getattr(usage, "completion_tokens", 0) or 0)
)
total_tokens["input_tokens"] += int(input_tokens)
total_tokens["output_tokens"] += int(output_tokens)
total_tokens["total_tokens"] += int(
total_tokens_count or (input_tokens + output_tokens)
)
summary = self._extract_litellm_text(response).strip()
return summary or tool_response
async def _maybe_compact_anthropic_messages(
self,
messages: List[Dict[str, Any]],
total_tokens: Dict[str, int],
thinking_budget: int,
tool_call_log_file: Optional[str],
current_input_tokens: int,
) -> List[Dict[str, Any]]:
if not self._compaction_enabled():
return messages
if current_input_tokens < self.compaction_token:
return messages
logger.info(
f"| [compaction] Triggered at input tokens: {current_input_tokens:,}"
)
if tool_call_log_file:
try:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(
f"| [compaction] Triggered at input tokens: {current_input_tokens:,}\n"
)
except Exception:
pass
compact_messages = [
{"role": "user", "content": self.COMPACTION_PROMPT},
{"role": "user", "content": json.dumps(messages, ensure_ascii=False)},
]
response, error_msg = await self._call_claude_native_api(
messages=compact_messages,
thinking_budget=thinking_budget,
tools=None,
system=None,
)
if error_msg or not response:
logger.warning(f"| [compaction] Failed: {error_msg}")
return messages
usage = response.get("usage", {}) or {}
input_tokens = usage.get("input_tokens", 0) or 0
output_tokens = usage.get("output_tokens", 0) or 0
total_tokens["input_tokens"] += int(input_tokens)
total_tokens["output_tokens"] += int(output_tokens)
total_tokens["total_tokens"] += int(input_tokens + output_tokens)
summary = self._extract_anthropic_text(response) or "(no summary)"
first_user = messages[0] if messages else {"role": "user", "content": ""}
return [
first_user,
{
"role": "user",
"content": f"Context summary (auto-compacted due to token limit):\n{summary}",
},
]
async def _execute_anthropic_native_tool_loop(
self,
instruction: str,
tools: List[Dict],
mcp_server: Any,
thinking_budget: int,
tool_call_log_file: Optional[str] = None,
) -> Dict[str, Any]:
"""
Execute Claude thinking loop with function calling.
Handles thinking blocks, tool calls, and message formatting.
"""
messages = [{"role": "user", "content": instruction}]
total_tokens = {
"input_tokens": 0,
"output_tokens": 0,
"total_tokens": 0,
"reasoning_tokens": 0,
}
turn_count = 0
max_turns = self.MAX_TURNS
hit_turn_limit = False
ended_normally = False
system_text = self.SYSTEM_PROMPT
# Record initial state
self._update_progress(messages, total_tokens, turn_count)
for _ in range(max_turns):
turn_count += 1
current_input_tokens = 0
if self._compaction_enabled():
try:
current_input_tokens = await self._count_claude_input_tokens(
messages=messages,
tools=tools,
system=system_text,
)
except Exception as exc: # noqa: BLE001
logger.debug("Claude token counting failed: %s", exc)
messages = await self._maybe_compact_anthropic_messages(
messages=messages,
total_tokens=total_tokens,
thinking_budget=thinking_budget,
tool_call_log_file=tool_call_log_file,
current_input_tokens=current_input_tokens,
)
self._update_progress(messages, total_tokens, turn_count)
# Call Claude native API
response, error_msg = await self._call_claude_native_api(
messages=messages,
thinking_budget=thinking_budget,
tools=tools,
system=system_text,
)
if turn_count == 1:
self.litellm_run_model_name = response["model"].split("/")[-1]
if error_msg:
break
# Update token usage
if "usage" in response:
usage = response["usage"]
input_tokens = usage.get("input_tokens", 0)
output_tokens = usage.get("output_tokens", 0)
# Calculate output tokens as total - input for consistency
total_tokens_count = output_tokens + input_tokens
total_tokens["input_tokens"] += input_tokens
total_tokens["output_tokens"] += output_tokens
total_tokens["total_tokens"] += total_tokens_count
## TODO: add reasoning tokens for claude
# Extract blocks from response
blocks = response.get("content", [])
tool_uses = [b for b in blocks if b.get("type") == "tool_use"]
thinking_blocks = [b for b in blocks if b.get("type") == "thinking"]
text_blocks = [b for b in blocks if b.get("type") == "text"]
# Log text output
for tb in text_blocks:
if tb.get("text") and tool_call_log_file:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"{tb['text']}\n")
if tb.get("text"):
for line in tb["text"].splitlines():
logger.info(f"| {line}")
# Build assistant message with all blocks
assistant_content = []
# Add thinking blocks
for tb in thinking_blocks:
assistant_content.append(
{
"type": "thinking",
"thinking": tb.get("thinking", ""),
"signature": tb.get("signature", ""),
}
)
# Add text blocks
for tb in text_blocks:
if tb.get("text"):
assistant_content.append({"type": "text", "text": tb["text"]})
# Add tool_use blocks
for tu in tool_uses:
assistant_content.append(
{
"type": "tool_use",
"id": tu.get("id"),
"name": tu.get("name"),
"input": tu.get("input", {}),
}
)
messages.append({"role": "assistant", "content": assistant_content})
# Update partial progress after assistant response
self._update_progress(messages, total_tokens, turn_count)
# If no tool calls, we're done
if not tool_uses:
ended_normally = True
break
# Execute tools and add results
tool_results = []
for tu in tool_uses:
name = tu.get("name")
inputs = tu.get("input", {})
# Log tool call
args_str = json.dumps(inputs, separators=(",", ": "))
display_args = (
args_str[:140] + "..." if len(args_str) > 140 else args_str
)
logger.info(f"| \033[1m{name}\033[0m \033[2;37m{display_args}\033[0m")
if tool_call_log_file:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"| {name} {args_str}\n")
# Execute tool
try:
result = await asyncio.wait_for(
mcp_server.call_tool(name, inputs), timeout=60
)
tool_results.append(
{
"type": "tool_result",
"tool_use_id": tu["id"],
"content": [
{
"type": "text",
"text": json.dumps(result, cls=CustomJSONEncoder),
}
],
}
)
except Exception as e:
logger.error(f"Tool call failed: {e}")
tool_results.append(
{
"type": "tool_result",
"tool_use_id": tu["id"],
"content": [{"type": "text", "text": f"Error: {str(e)}"}],
}
)
messages.append({"role": "user", "content": tool_results})
# Update partial progress after tool results
self._update_progress(messages, total_tokens, turn_count)
# Detect if we exited due to hitting the turn limit
if not ended_normally:
if turn_count >= max_turns:
hit_turn_limit = True
logger.warning(
f"| Max turns ({max_turns}) exceeded; returning failure with partial output."
)
if tool_call_log_file:
try:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"| Max turns ({max_turns}) exceeded\n")
except Exception:
pass
elif error_msg:
logger.warning(f"| {error_msg}\n")
if tool_call_log_file:
try:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"| {error_msg}\n")
except Exception:
pass
# Display final token usage
if total_tokens["total_tokens"] > 0:
log_msg = (
f"|\n| Token usage: Total: {total_tokens['total_tokens']:,} | "
f"Input: {total_tokens['input_tokens']:,} | "
f"Output: {total_tokens['output_tokens']:,}"
)
if total_tokens.get("reasoning_tokens", 0) > 0:
log_msg += f" | Reasoning: {total_tokens['reasoning_tokens']:,}"
logger.info(log_msg)
logger.info(f"| Turns: {turn_count}")
# Convert messages to SDK format
sdk_format_messages = self._convert_to_sdk_format(messages)
if hit_turn_limit:
return {
"success": False,
"output": sdk_format_messages,
"token_usage": total_tokens,
"turn_count": turn_count,
"error": f"Max turns ({max_turns}) exceeded",
"litellm_run_model_name": self.litellm_run_model_name,
}
if error_msg:
return {
"success": False,
"output": sdk_format_messages,
"token_usage": total_tokens,
"turn_count": turn_count,
"error": error_msg,
"litellm_run_model_name": self.litellm_run_model_name,
}
return {
"success": True,
"output": sdk_format_messages,
"token_usage": total_tokens,
"turn_count": turn_count,
"error": None,
"litellm_run_model_name": self.litellm_run_model_name,
}
# ==================== LiteLLM Execution Path ====================
async def _execute_litellm_with_tools(
self, instruction: str, tool_call_log_file: Optional[str] = None
) -> Dict[str, Any]:
"""
Execute with manual MCP server management.
Used for all non-Anthropic models and Anthropic models with STDIO services.
"""
logger.debug("Using manual MCP execution with function calling loop")
# Create and start MCP server
mcp_server = await self._create_mcp_server()
try:
async with mcp_server:
# Get available tools
tools = await mcp_server.list_tools()
# Convert MCP tools to OpenAI function format
functions = self._convert_to_openai_format(tools)
# Execute with function calling loop
return await self._execute_litellm_tool_loop(
instruction, functions, mcp_server, tool_call_log_file
)
except Exception as e:
logger.error(f"Manual MCP execution failed: {e}")
raise
async def _execute_litellm_tool_loop(
self,
instruction: str,
functions: List[Dict],
mcp_server: Any,
tool_call_log_file: Optional[str] = None,
) -> Dict[str, Any]:
"""Execute function calling loop with LiteLLM."""
messages = [
{"role": "system", "content": self.SYSTEM_PROMPT},
{"role": "user", "content": instruction},
]
total_tokens = {
"input_tokens": 0,
"output_tokens": 0,
"total_tokens": 0,
"reasoning_tokens": 0,
}
turn_count = 0
max_turns = self.MAX_TURNS # Limit turns to prevent infinite loops
consecutive_failures = 0
max_consecutive_failures = 3
hit_turn_limit = False
ended_normally = False
# Convert functions to tools format for newer models
tools = (
[{"type": "function", "function": func} for func in functions]
if functions
else None
)
if tool_call_log_file and tools:
max_name_length = (
max(len(tool.get("function", {}).get("name", "")) for tool in tools)
if tools
else 15
)
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write("===== Available Tools =====\n")
for tool in tools:
function_info = tool.get("function", {})
tool_name = function_info.get("name", "N/A")
description = function_info.get("description", "N/A")
f.write(
f"- ToolName: {tool_name:<{max_name_length}} Description: {description}\n"
)
f.write("\n===== Execution Logs =====\n")
# Record initial state
self._update_progress(messages, total_tokens, turn_count)
try:
while turn_count < max_turns:
current_prompt_tokens = 0
if self._compaction_enabled():
current_prompt_tokens = self._count_prompt_tokens_litellm(messages)
messages = await self._maybe_compact_litellm_messages(
messages=messages,
total_tokens=total_tokens,
tool_call_log_file=tool_call_log_file,
current_prompt_tokens=current_prompt_tokens,
)
self._update_progress(messages, total_tokens, turn_count)
# Build completion kwargs
completion_kwargs = {
"model": self.litellm_input_model_name,
"messages": messages,
"api_key": self.api_key,
}
# Always use tools format if available - LiteLLM will handle conversion
if tools:
completion_kwargs["tools"] = tools
completion_kwargs["tool_choice"] = "auto"
# Add reasoning_effort and base_url if specified
if self.reasoning_effort != "default":
completion_kwargs["reasoning_effort"] = self.reasoning_effort
if self.base_url:
completion_kwargs["base_url"] = self.base_url
if self.extra_body:
completion_kwargs["extra_body"] = self.extra_body
try:
# Call LiteLLM with timeout for individual call
response = await asyncio.wait_for(
litellm.acompletion(**completion_kwargs),
timeout=self.timeout / 2, # Use half of total timeout
)
consecutive_failures = 0 # Reset failure counter on success
except asyncio.TimeoutError:
logger.warning(f"| ✗ LLM call timed out on turn {turn_count + 1}")
consecutive_failures += 1
if consecutive_failures >= max_consecutive_failures:
raise Exception(
f"Too many consecutive failures ({consecutive_failures})"
)
await asyncio.sleep(8**consecutive_failures) # Exponential backoff
continue
except Exception as e:
logger.error(f"| ✗ LLM call failed on turn {turn_count + 1}: {e}")
consecutive_failures += 1
if consecutive_failures >= max_consecutive_failures:
raise
if "ContextWindowExceededError" in str(e):
# Best-effort fallback: compact and retry once.
messages = await self._maybe_compact_litellm_messages(
messages=messages,
total_tokens=total_tokens,
tool_call_log_file=tool_call_log_file,
current_prompt_tokens=self.compaction_token,
)
self._update_progress(messages, total_tokens, turn_count)
continue
elif "RateLimitError" in str(e):
await asyncio.sleep(12**consecutive_failures)
else:
await asyncio.sleep(2**consecutive_failures)
continue
# Extract actual model name from response (first turn only)
if turn_count == 0 and hasattr(response, "model") and response.model:
self.litellm_run_model_name = response.model.split("/")[-1]
# Update token usage including reasoning tokens
if hasattr(response, "usage") and response.usage:
input_tokens = response.usage.prompt_tokens or 0
total_tokens_count = response.usage.total_tokens or 0
# Calculate output tokens as total - input for consistency
output_tokens = (
total_tokens_count - input_tokens
if total_tokens_count > 0
else (response.usage.completion_tokens or 0)
)
total_tokens["input_tokens"] += input_tokens
total_tokens["output_tokens"] += output_tokens
total_tokens["total_tokens"] += total_tokens_count
# Extract reasoning tokens if available
if hasattr(response.usage, "completion_tokens_details"):
details = response.usage.completion_tokens_details
if hasattr(details, "reasoning_tokens"):
total_tokens["reasoning_tokens"] += (
details.reasoning_tokens or 0
)
# Get response message
choices = response.choices
if len(choices):
message = choices[0].message
# deeply dump the message to ensure we capture all fields
message_dict = (
message.model_dump()
if hasattr(message, "model_dump")
else dict(message)
)
# Explicitly preserve function_call if present (even if tool_calls exists),
# as it may contain provider-specific metadata (e.g. Gemini thought_signature)
if hasattr(message, "function_call") and message.function_call:
# Ensure it's in the dict if model_dump missed it or it was excluded
if (
"function_call" not in message_dict
or not message_dict["function_call"]
):
fc = message.function_call
message_dict["function_call"] = (
fc.model_dump() if hasattr(fc, "model_dump") else fc
)
# Log assistant's text content if present
if hasattr(message, "content") and message.content:
# Display the content with line prefix
for line in message.content.splitlines():
logger.info(f"| {line}")
# Also log to file if specified
if tool_call_log_file:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"{message.content}\n")
# Check for tool calls (newer format)
if hasattr(message, "tool_calls") and message.tool_calls:
messages.append(message_dict)
turn_count += 1
# Update progress after assistant with tool calls
self._update_progress(messages, total_tokens, turn_count)
# Process tool calls
for tool_call in message.tool_calls:
func_name = tool_call.function.name
func_args = json.loads(tool_call.function.arguments)
try:
result = await asyncio.wait_for(
mcp_server.call_tool(func_name, func_args), timeout=60
)
_content = json.dumps(result, cls=CustomJSONEncoder)
# Playwright only: compress the browser result before
# it enters history, so the model input == the saved
# trajectory (fidelity-preserving, keeps tokens down).
if self._should_summarize_tool():
_context = json.dumps(
{"tool": func_name, "arguments": func_args},
ensure_ascii=False,
)
_content = await self._summarize_tool_response_litellm(
_content,
context=_context,
total_tokens=total_tokens,
)
messages.append(
{
"role": "tool",
"tool_call_id": tool_call.id,
"content": _content,
}
)
except asyncio.TimeoutError:
error_msg = (
f"Tool call '{func_name}' timed out after 60 seconds"
)
logger.error(error_msg)
messages.append(
{
"role": "tool",
"tool_call_id": tool_call.id,
"content": f"Error: {error_msg}",
}
)
except Exception as e:
logger.error(f"Tool call failed: {e}")
messages.append(
{
"role": "tool",
"tool_call_id": tool_call.id,
"content": f"Error: {str(e)}",
}
)
# Format arguments for display (truncate if too long)
args_str = json.dumps(func_args, separators=(",", ": "))
display_arguments = (
args_str[:140] + "..." if len(args_str) > 140 else args_str
)
# Log with ANSI color codes (bold tool name, dim gray arguments)
logger.info(
f"| \033[1m{func_name}\033[0m \033[2;37m{display_arguments}\033[0m"
)
if tool_call_log_file:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"| {func_name} {args_str}\n")
# Update progress after tool results appended
self._update_progress(messages, total_tokens, turn_count)
continue
else:
# Log end reason
if not choices:
logger.info(
"|\n|\n| Task ended with no messages generated by the model."
)
elif choices[0].finish_reason == "stop":
logger.info(
"|\n|\n| Task ended with the finish reason from messages being 'stop'."
)
# No tool/function call, add message and we're done
messages.append(message_dict)
turn_count += 1
# Update progress before exiting
self._update_progress(messages, total_tokens, turn_count)
ended_normally = True
break
except Exception as loop_error:
# On any error, return partial conversation, token usage, and turn count
logger.error(f"Manual MCP loop failed: {loop_error}", exc_info=True)
sdk_format_messages = self._convert_to_sdk_format(messages)
return {
"success": False,
"output": sdk_format_messages,
"token_usage": total_tokens,
"turn_count": turn_count,
"error": str(loop_error),
"litellm_run_model_name": self.litellm_run_model_name,
}
# Detect if we exited due to hitting the turn limit
if (not ended_normally) and (turn_count >= max_turns):
hit_turn_limit = True
logger.warning(
f"| Max turns ({max_turns}) exceeded); returning failure with partial output."
)
if tool_call_log_file:
try:
with open(tool_call_log_file, "a", encoding="utf-8") as f:
f.write(f"| Max turns ({max_turns}) exceeded\n")
except Exception:
pass
# Display final token usage
if total_tokens["total_tokens"] > 0:
log_msg = (
f"| Token usage: Total: {total_tokens['total_tokens']:,} | "
f"Input: {total_tokens['input_tokens']:,} | "
f"Output: {total_tokens['output_tokens']:,}"
)
if total_tokens.get("reasoning_tokens", 0) > 0:
log_msg += f" | Reasoning: {total_tokens['reasoning_tokens']:,}"
logger.info(log_msg)
logger.info(f"| Turns: {turn_count}")
# Convert messages to SDK format for backward compatibility
sdk_format_messages = self._convert_to_sdk_format(messages)
return {
"success": not hit_turn_limit,
"output": sdk_format_messages,
"token_usage": total_tokens,
"turn_count": turn_count,
"error": (f"Max turns ({max_turns}) exceeded" if hit_turn_limit else None),
"litellm_run_model_name": self.litellm_run_model_name,
}
# ==================== MCP Server Management ====================
async def _create_mcp_server(self) -> Any:
"""Create and return an MCP server instance."""
if self.mcp_service in self.STDIO_SERVICES:
return self._create_stdio_server()
elif self.mcp_service in self.HTTP_SERVICES:
return self._create_http_server()
else:
raise ValueError(f"Unsupported MCP service: {self.mcp_service}")
def _create_stdio_server(self) -> MCPStdioServer:
"""Create stdio-based MCP server."""
if self.mcp_service == "notion":
notion_key = self.service_config.get("notion_key")
if not notion_key:
raise ValueError("Notion API key required")
return MCPStdioServer(
command="npx",
args=["-y", "@notionhq/notion-mcp-server@1.9.1"],
env={
"OPENAPI_MCP_HEADERS": (
'{"Authorization": "Bearer ' + notion_key + '", '
'"Notion-Version": "2022-06-28"}'
)
},
)
elif self.mcp_service == "filesystem":
test_directory = self.service_config.get("test_directory")
if not test_directory:
raise ValueError("Test directory required for filesystem service")
return MCPStdioServer(
command="npx",
args=[
"-y",
"@modelcontextprotocol/server-filesystem",
str(test_directory),
],
)
elif self.mcp_service in ["playwright", "playwright_webarena"]:
browser = self.service_config.get("browser", "chromium")
headless = self.service_config.get("headless", True)
viewport_width = self.service_config.get("viewport_width", 1280)
viewport_height = self.service_config.get("viewport_height", 720)
args = ["-y", "@playwright/mcp@latest"]
if headless:
args.append("--headless")
args.extend(
[
"--isolated",
"--no-sandbox",
"--browser",
browser,
"--viewport-size",
f"{viewport_width},{viewport_height}",
]
)
# Real-web playwright only: route the browser through an egress
# proxy (e.g. US socks5) when MCPMARK_BROWSER_PROXY is set. WebArena
# is localhost docker and must NOT be proxied, so this is gated to
# the real-internet "playwright" service.
if self.mcp_service == "playwright":
_browser_proxy = os.environ.get("MCPMARK_BROWSER_PROXY", "").strip()
if _browser_proxy:
args.extend(["--proxy-server", _browser_proxy])
return MCPStdioServer(command="npx", args=args)
elif self.mcp_service == "postgres":
host = self.service_config.get("host", "localhost")
port = self.service_config.get("port", 5432)
username = self.service_config.get("username")
password = self.service_config.get("password")
database = self.service_config.get(
"current_database"
) or self.service_config.get("database")
if not all([username, password, database]):
raise ValueError("PostgreSQL requires username, password, and database")
database_url = (
f"postgresql://{username}:{password}@{host}:{port}/{database}"
)
return MCPStdioServer(
command="pipx",
args=["run", "postgres-mcp", "--access-mode=unrestricted"],
env={"DATABASE_URI": database_url},
)
elif self.mcp_service == "insforge":
api_key = self.service_config.get("api_key")
backend_url = self.service_config.get("backend_url")
if not all([api_key, backend_url]):
raise ValueError("Insforge requires api_key and backend_url")
return MCPStdioServer(
command="npx",
args=["-y", "@insforge/mcp@dev"],
env={
"INSFORGE_API_KEY": api_key,
"INSFORGE_BACKEND_URL": backend_url,
},
)
elif self.mcp_service == "github":
github_token = self.service_config.get("github_token")
if not github_token:
raise ValueError("GitHub token required")
return MCPStdioServer(
command="docker",
args=[
"run", "-i", "--rm",
"-e", "GITHUB_PERSONAL_ACCESS_TOKEN",
"ghcr.io/github/github-mcp-server:v0.15.0",
],
env={"GITHUB_PERSONAL_ACCESS_TOKEN": github_token},
)
else:
raise ValueError(f"Unsupported stdio service: {self.mcp_service}")
def _create_http_server(self) -> MCPHttpServer:
"""Create HTTP-based MCP server."""
if self.mcp_service == "supabase":
# Use built-in MCP server from Supabase CLI
api_url = self.service_config.get("api_url", "http://localhost:54321")
api_key = self.service_config.get("api_key", "")
if not api_key:
raise ValueError(
"Supabase requires api_key (use secret key from 'supabase status')"
)
# Supabase CLI exposes MCP at /mcp endpoint
mcp_url = f"{api_url}/mcp"
return MCPHttpServer(
url=mcp_url,
headers={
"apikey": api_key,
"Authorization": f"Bearer {api_key}",
},
)
else:
raise ValueError(f"Unsupported HTTP service: {self.mcp_service}")