The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
"""Status reads never load a model.
llama-server's router (b11370, ``--models-autoload``) loads an unloaded model for any request that
names it, /props, /slots and /metrics included. Clicking Eject on the local models page unloaded the
model, and a capability read 76 ms later loaded it straight back. A per-request ``autoload=false``
makes the router answer 400 "model is not loaded" instead (checked against a router serving a fake
model: zero spawns with it, a spawn per request without it).
"""
import io
import urllib.error
from types import SimpleNamespace
import pytest
from hermes_cli.local_runtime import capabilities, endpoint, load_progress
from hermes_cli.local_runtime.supervisor import LlamaServerSupervisor
def test_child_query_opts_out_of_autoload():
assert endpoint.child_query("org/model q4") == "model=org/model%20q4&autoload=false"
def test_capability_read_does_not_load_the_model(monkeypatch):
seen = []
monkeypatch.setattr(endpoint, "managed_root", lambda: ("http://127.0.0.1:1", "k"))
monkeypatch.setattr(endpoint, "managed_get_json", lambda base, key, path, timeout_s: seen.append(path) or {})
capabilities._props_modalities("qwen")
assert seen == ["/props?model=qwen&autoload=false"]
def test_prefill_read_does_not_load_the_model(monkeypatch):
seen = []
monkeypatch.setattr(load_progress, "_endpoint", lambda: ("http://127.0.0.1:1", "k"))
monkeypatch.setattr(endpoint, "managed_get_json", lambda base, key, path, timeout_s: seen.append(path) or [])
assert load_progress.get_prefill_progress("qwen") is None
assert seen == ["/slots?model=qwen&autoload=false"]
class _Client:
"""httpx-shaped client over a router whose model is unloaded."""
def __init__(self):
self.paths = []
def get(self, url):
path = url.split("127.0.0.1:1", 1)[1]
self.paths.append(path)
if path == "/models":
return SimpleNamespace(status_code=200, json=lambda: {"data": [{"id": "qwen", "status": {
"value": "unloaded", "args": ["llama-server.exe", "--ctx-size", "65536", "--jinja"]}}]})
return SimpleNamespace(status_code=400, json=lambda: {"error": {"message": "model is not loaded"}})
def test_context_probe_reads_launch_args_for_an_unloaded_model():
from agent.model_metadata_llamacpp import llamacpp_context
client = _Client()
assert llamacpp_context(client, "http://127.0.0.1:1", "qwen") == 65536
assert "/props?model=qwen&autoload=false" in client.paths
assert "/props?model=qwen" not in client.paths
@pytest.mark.parametrize("body", [["not", "an", "object"], {"data": "oops"}, {"data": [{"id": "qwen", "status": "loading"}]}])
def test_context_probe_survives_an_odd_models_body(body):
from agent.model_metadata_llamacpp import router_launch_ctx
client = SimpleNamespace(get=lambda url: SimpleNamespace(status_code=200, json=lambda: body))
assert router_launch_ctx(client, "http://127.0.0.1:1", "qwen") is None
def test_an_unloaded_model_counts_as_idle(monkeypatch):
"""Growth and the idle sweeper ask whether a model is busy; one that isn't loaded is not."""
sup = LlamaServerSupervisor.__new__(LlamaServerSupervisor)
def request(route, body=None, timeout_s=30):
raise urllib.error.HTTPError(route, 400, "model is not loaded", {}, io.BytesIO(b""))
monkeypatch.setattr(sup, "_request", request, raising=False)
assert sup._probe_idle("qwen") is True