mcpmark / src /evaluator.py
haochengsama's picture
Add files using upload-large-folder tool
97cb846 verified
Raw
History Blame Contribute Delete
19.1 kB
import time
import json
import shutil
from datetime import datetime
from pathlib import Path
from typing import List, Optional
from src.logger import get_logger
from src.factory import MCPServiceFactory
from src.model_config import ModelConfig
from src.results_reporter import EvaluationReport, ResultsReporter, TaskResult
from src.errors import is_retryable_error
from src.agents import AGENT_REGISTRY
# Initialize logger
logger = get_logger(__name__)
class MCPEvaluator:
def __init__(
self,
mcp_service: str,
model: str,
timeout: int = 300,
exp_name: str = "test-run",
output_dir: Path = None,
reasoning_effort: str = "default",
agent_name: str = "mcpmark",
task_suite: str = "standard",
compaction_token: int = 0,
summarize_tool_response: bool = False,
):
# Main configuration
self.mcp_service = mcp_service
self.timeout = timeout
self.agent_name = (agent_name or "mcpmark").lower()
self.task_suite = (task_suite or "standard").lower()
if self.agent_name not in AGENT_REGISTRY:
raise ValueError(f"Unsupported agent '{agent_name}'. Available: {sorted(AGENT_REGISTRY)}")
# Initialize model configuration
self.reasoning_effort = reasoning_effort
self.model_name = model
model_config = ModelConfig(self.model_name)
self.api_key = model_config.api_key
self.base_url = model_config.base_url
self.litellm_input_model_name = model_config.litellm_input_model_name
self.extra_body = model_config.extra_body
# Track the actual model name from LiteLLM responses
self.litellm_run_model_name = None
# Initialize managers using the factory pattern (simplified)
self.task_manager = MCPServiceFactory.create_task_manager(
mcp_service, task_suite=self.task_suite
)
self.state_manager = MCPServiceFactory.create_state_manager(mcp_service)
# Obtain static service configuration from state manager (e.g., notion_key)
self.service_config = self.state_manager.get_service_config_for_agent()
# Initialize agent for LLM and MCP server management. The agent will
# automatically refresh its service configuration from the state
# manager before each execution, so per-task manual updates are no
# longer needed.
agent_cls = AGENT_REGISTRY[self.agent_name]
self.agent = agent_cls(
litellm_input_model_name=self.litellm_input_model_name,
api_key=self.api_key,
base_url=self.base_url,
mcp_service=mcp_service,
timeout=timeout,
service_config=self.service_config,
service_config_provider=self.state_manager.get_service_config_for_agent,
reasoning_effort=self.reasoning_effort,
compaction_token=compaction_token,
extra_body=self.extra_body,
summarize_tool_response=summarize_tool_response,
)
# Initialize results reporter
self.results_reporter = ResultsReporter()
# Output directory handling
if self.reasoning_effort != "default":
model_slug = self.model_name.replace(".", "-") + "-" + self.reasoning_effort
else:
model_slug = self.model_name.replace(".", "-")
service_for_dir = "playwright" if mcp_service == "playwright_webarena" else mcp_service
suite_suffix = "" if self.task_suite in ("standard", "", None) else f"-{self.task_suite}"
service_dir_name = f"{service_for_dir}{suite_suffix}"
self.base_experiment_dir = output_dir / f"{model_slug}__{service_dir_name}" / exp_name
self.base_experiment_dir.mkdir(parents=True, exist_ok=True)
def _format_duration(self, seconds: float) -> str:
"""Format duration: <1s as ms, otherwise seconds."""
return f"{(seconds * 1000):.2f}ms" if seconds < 1 else f"{seconds:.2f}s"
def _get_task_output_dir(self, task) -> Path:
"""Return the directory path for storing this task's reports using '__' separator."""
# Use category_id and task_id with '__' separator
category_id = task.category_id if task.category_id else "uncategorized"
task_id = str(task.task_id)
return self.base_experiment_dir / f"{category_id}__{task_id}"
# ------------------------------------------------------------------
# Resuming helpers
# ------------------------------------------------------------------
def _load_latest_task_result(self, task) -> Optional[TaskResult]:
"""Return the most recent TaskResult for *task* if it has been run before."""
task_dir = self._get_task_output_dir(task)
if not task_dir.exists():
return None
meta_path = task_dir / "meta.json"
if not meta_path.exists():
return None
try:
with meta_path.open("r", encoding="utf-8") as f:
meta_data = json.load(f)
return TaskResult(
task_name=meta_data["task_name"],
success=meta_data["execution_result"]["success"],
error_message=meta_data["execution_result"].get("error_message"),
verification_error=meta_data["execution_result"].get("verification_error"),
verification_output=meta_data["execution_result"].get("verification_output"),
category_id=task.category_id,
task_id=task.task_id,
model_output=None,
token_usage=meta_data.get("token_usage", {}),
turn_count=meta_data.get("turn_count"),
agent_execution_time=meta_data.get("agent_execution_time", 0.0),
task_execution_time=meta_data.get("task_execution_time", 0.0),
)
except Exception as exc:
logger.warning("Failed to load existing result for %s: %s", task.name, exc)
return None
def _gather_all_task_results(self) -> List[TaskResult]:
"""Scan *all* task sub-directories and collect the latest TaskResult from each."""
results: list[TaskResult] = []
if not self.base_experiment_dir.exists():
return results
for task_dir in self.base_experiment_dir.iterdir():
if not task_dir.is_dir():
continue
meta_path = task_dir / "meta.json"
if not meta_path.exists():
continue
try:
with meta_path.open("r", encoding="utf-8") as f:
meta_data = json.load(f)
category_id, task_id = task_dir.name.split("__", 1)
result = TaskResult(
task_name=meta_data["task_name"],
success=meta_data["execution_result"]["success"],
error_message=meta_data["execution_result"].get("error_message"),
verification_error=meta_data["execution_result"].get("verification_error"),
verification_output=meta_data["execution_result"].get("verification_output"),
category_id=category_id,
task_id=task_id,
model_output=None,
token_usage=meta_data.get("token_usage", {}),
turn_count=meta_data.get("turn_count"),
agent_execution_time=meta_data.get("agent_execution_time", 0.0),
task_execution_time=meta_data.get("task_execution_time", 0.0),
)
results.append(result)
except Exception as exc:
logger.warning(
"Failed to parse existing report in %s: %s", task_dir, exc
)
return results
def _run_single_task(self, task) -> TaskResult:
"""
Runs a single task, including setup, agent execution, verification, and cleanup.
"""
# Track overall task start time
task_start_time = time.time()
# ------------------------------------------------------------------
# Stage 1: Set up the initial state for the task
# ------------------------------------------------------------------
setup_start_time = time.time()
logger.info(
"\n┌─ Stage 1: Setup ─────────────────────────────────────────────────────"
)
setup_success = self.state_manager.set_up(task)
setup_time = time.time() - setup_start_time
if not setup_success:
logger.error(f"| State setup failed for task: {task.name}")
task_total_time = time.time() - task_start_time
return TaskResult(
task_name=task.name,
success=False,
error_message="State Duplication Error",
verification_error=None,
verification_output=None,
category_id=task.category_id,
task_id=task.task_id,
agent_execution_time=0.0,
task_execution_time=task_total_time,
)
display_time = self._format_duration(setup_time)
logger.info(f"└─ Completed in {display_time}\n")
# ------------------------------------------------------------------
# Stage 2: Execute the task using the agent
# ------------------------------------------------------------------
logger.info(
"┌─ Stage 2: Execute ───────────────────────────────────────────────────"
)
agent_execution_start_time = time.time()
# Get task instruction from task manager
task_instruction = self.task_manager.get_task_instruction(task)
# Prepare task_output_dir and tool call log file
task_output_dir = self._get_task_output_dir(task)
task_output_dir.mkdir(parents=True, exist_ok=True)
execution_log_path = task_output_dir / "execution.log"
# Remove existing execution.log to ensure clean start
if execution_log_path.exists():
execution_log_path.unlink()
# Execute with agent
agent_result = self.agent.execute_sync(
task_instruction, str(execution_log_path)
)
agent_execution_time = time.time() - agent_execution_start_time
# Extract actual model name from LiteLLM response
if agent_result.get("litellm_run_model_name"):
self.litellm_run_model_name = agent_result["litellm_run_model_name"]
# Write messages.json to task_output_dir
messages_path = task_output_dir / "messages.json"
self.results_reporter.save_messages_json(
agent_result.get("output", []), messages_path
)
# Set service-specific environment variables for verification scripts
self.state_manager.set_verification_environment(str(messages_path))
logger.info(f"└─ Completed in {self._format_duration(agent_execution_time)}\n")
# ------------------------------------------------------------------
# Stage 3: Verify
# ------------------------------------------------------------------
logger.info(
"┌─ Stage 3: Verify ────────────────────────────────────────────────────"
)
verify_start_time = time.time()
try:
result = self.task_manager.execute_task(task, agent_result)
finally:
# Clean up environment variables
import os
os.environ.pop("MCP_MESSAGES", None)
os.environ.pop("MCP_GITHUB_TOKEN", None)
verify_time = time.time() - verify_start_time
logger.info(f"└─ Completed in {self._format_duration(verify_time)}\n")
# ------------------------------------------------------------------
# Stage 4: Clean up
# ------------------------------------------------------------------
logger.info(
"┌─ Stage 4: Cleanup ───────────────────────────────────────────────────"
)
cleanup_start_time = time.time()
self.state_manager.clean_up(task)
cleanup_time = time.time() - cleanup_start_time
logger.info(f"└─ Completed in {self._format_duration(cleanup_time)}\n")
# Calculate total task execution time
task_total_time = time.time() - task_start_time
# Add timing information to the result
result.agent_execution_time = agent_execution_time
result.task_execution_time = task_total_time
return result
def run_evaluation(self, task_filter: str) -> EvaluationReport:
"""
Runs the full evaluation for the specified tasks.
"""
tasks = self.task_manager.filter_tasks(task_filter)
results = []
for task in tasks:
# --------------------------------------------------------------
# Resume check
# --------------------------------------------------------------
existing_result = self._load_latest_task_result(task)
# Decide whether to skip or retry this task
retry_due_to_error = (
existing_result is not None
and not existing_result.success
and is_retryable_error(existing_result.error_message)
)
if existing_result and not retry_due_to_error:
# Existing result is either successful or failed with a non-retryable error – skip.
logger.info(
"↩️ Skipping already-completed task (resume): %s", task.name
)
results.append(existing_result)
continue
if retry_due_to_error:
# Clean previous artifacts so that new results fully replace them.
task_output_dir = self._get_task_output_dir(task)
if task_output_dir.exists():
shutil.rmtree(task_output_dir)
logger.info(
"🔄 Retrying task due to pipeline error (%s): %s",
existing_result.error_message,
task.name,
)
# --------------------------------------------------------------
# Execute new task
# --------------------------------------------------------------
task_start = time.time()
task_result = self._run_single_task(task)
task_end = time.time()
results.append(task_result)
# Prepare directory & save
task_output_dir = self._get_task_output_dir(task)
task_output_dir.mkdir(parents=True, exist_ok=True)
# Save messages.json (conversation trajectory)
messages_path = task_output_dir / "messages.json"
if not messages_path.exists(): # 已经写过就跳过
messages = (
task_result.model_output
if getattr(task_result, "model_output", None)
else []
)
self.results_reporter.save_messages_json(messages, messages_path)
# Save meta.json (all other metadata)
meta_path = task_output_dir / "meta.json"
model_config = {
"mcp_service": self.mcp_service,
"model_name": self.model_name,
"litellm_run_model_name": self.litellm_run_model_name,
"reasoning_effort": self.reasoning_effort,
"timeout": self.timeout,
"agent_name": self.agent_name,
}
self.results_reporter.save_meta_json(
task_result,
model_config,
datetime.fromtimestamp(task_start),
datetime.fromtimestamp(task_end),
meta_path,
)
# --------------------------------------------------------------
# Aggregate results – combine current `results` with any previously
# saved TaskResults that ALSO match the current task_filter.
# --------------------------------------------------------------
# Helper: determine if a TaskResult matches the filter string
def _matches_filter(tr: TaskResult, flt: str) -> bool:
if flt.lower() == "all":
return True
if "/" in flt:
# specific task (category_id/task_id)
category_id, task_id = flt.split("/", 1)
return tr.category_id == category_id and str(tr.task_id) == task_id
# category level
return tr.category_id == flt
# Pull existing reports from disk and merge
existing_results = [
r
for r in self._gather_all_task_results()
if _matches_filter(r, task_filter)
]
# Merge, giving preference to fresh `results` (avoids duplicates)
merged: dict[str, TaskResult] = {r.task_name: r for r in existing_results}
merged.update({r.task_name: r for r in results}) # overwrite with latest run
final_results = list(merged.values())
aggregated_report = EvaluationReport(
model_name=self.model_name,
model_config={
"mcp_service": self.mcp_service,
"model_name": self.model_name,
"litellm_run_model_name": self.litellm_run_model_name,
"reasoning_effort": self.reasoning_effort,
"timeout": self.timeout,
"agent_name": self.agent_name,
},
total_tasks=len(final_results),
successful_tasks=sum(1 for r in final_results if r.success),
failed_tasks=sum(1 for r in final_results if not r.success),
task_results=final_results,
tasks_filter=task_filter,
)
# Save model-level summary
summary_path = self.base_experiment_dir / "summary.json"
self.results_reporter.save_model_summary(aggregated_report, summary_path)
logger.info(
"\n============================================================"
"\nResults Summary"
"\n============================================================"
)
logger.info(
f"✓ Tasks passed: {aggregated_report.successful_tasks}/{aggregated_report.total_tasks} ({aggregated_report.success_rate:.1f}%)"
)
logger.info(f"⏱ Total time: {aggregated_report.total_task_execution_time:.1f}s")
return aggregated_report