The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
"""Model-metadata probes have a total wall-clock bound, not just per-op httpx timeouts.
Probes run inside requests (a new session resolves its context window through them), so a
``/models`` body that trickles in under the read timeout must not hold the caller.
"""
import threading
import time
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
import pytest
import agent.model_metadata as mm
import agent.model_metadata_http as http
class _Trickle(BaseHTTPRequestHandler):
def log_message(self, *args):
pass
def do_GET(self):
try:
self.send_response(200)
self.send_header("Content-Type", "application/json")
self.send_header("Content-Length", "1000000")
self.end_headers()
for _ in range(80): # ~8 s of one byte per 0.1 s: far under any per-read timeout
self.wfile.write(b" ")
self.wfile.flush()
if self.server.stop.wait(0.1):
return
except (BrokenPipeError, ConnectionResetError):
pass
@pytest.fixture
def trickling_endpoint(monkeypatch):
server = ThreadingHTTPServer(("127.0.0.1", 0), _Trickle)
server.stop = threading.Event()
threading.Thread(target=server.serve_forever, daemon=True).start()
monkeypatch.setattr(http, "TOTAL_DEADLINE_S", 0.5, raising=False)
yield f"http://127.0.0.1:{server.server_port}"
server.stop.set()
server.shutdown()
server.server_close()
def test_endpoint_models_probe_on_a_trickling_server_returns_within_the_deadline(trickling_endpoint, monkeypatch):
monkeypatch.setattr(mm, "is_local_endpoint", lambda _url: False) # remote shape: /models probes only
started = time.monotonic()
assert mm.fetch_endpoint_model_metadata(trickling_endpoint + "/v1", force_refresh=True) == {}
# Two candidates (/v1 and bare), each capped: well under the ~8 s the body takes to arrive.
assert time.monotonic() - started < 3.0