Enterprise agent platform for building, deploying, and operating AI teammates.
"""Timing-only launcher for benchmark-owned hosts, not a production entry point.
Records task ID and first OpenAI-compatible model-adapter entry per process.
This is NOT socket-send time, token latency, or the earlier llm_call_start trace.
Non-OpenAI-compatible adapters are deliberately unsupported: missing observations
fail validation instead of being substituted with a different clock.
"""
from __future__ import annotations
import atexit
import json
import os
import runpy
import sys
import time
from functools import wraps
from pathlib import Path
def install(output):
from xagent.core.model.chat.basic.openai import OpenAICompatibleLLM
from xagent.web.services.task_lease_service import current_task_lease
entries = {}
def record():
lease = current_task_lease()
if lease is not None:
entries.setdefault(lease.task_id, time.time())
original_chat = OpenAICompatibleLLM.chat
original_stream = OpenAICompatibleLLM.stream_chat
@wraps(original_chat)
async def chat(self, *args, **kwargs):
record()
return await original_chat(self, *args, **kwargs)
@wraps(original_stream)
async def stream(self, *args, **kwargs):
record()
iterator = original_stream(self, *args, **kwargs)
try:
async for chunk in iterator:
yield chunk
finally:
await iterator.aclose()
OpenAICompatibleLLM.chat = chat
OpenAICompatibleLLM.stream_chat = stream
@atexit.register
def dump():
with (Path(output) / f"model-entries-{os.getpid()}.json").open("x") as handle:
json.dump(
[{"task_id": key, "time": value} for key, value in entries.items()],
handle,
)
def main():
module, output, *arguments = sys.argv[1:]
if module not in {"xagent.web", "xagent.web.worker"}:
raise SystemExit("Unsupported benchmark host")
install(output)
sys.argv = [module, *arguments]
runpy.run_module(module, run_name="__main__")
if __name__ == "__main__":
main()