mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-20 18:50:12 +03:00
edb2b07999
Resolving the main endpoint's probe key loads its credential pool, which can persist auth.json and wait up to AUTH_LOCK_TIMEOUT_SECONDS (15s) for its lock; an OAuth refresh holds that lock across its network call. The lookup ran before the 5s budget started, so /api/model/info could stall past it. Resolve the key inside the bounded worker, which binds the request's profile scope itself (and copies the request context, so an unnamed profile keeps a task-local home override).
104 lines
3.8 KiB
Python
104 lines
3.8 KiB
Python
"""/api/model/info must bound its context-length probe (#63214).
|
|
|
|
The resolver chain (``agent.model_metadata.get_model_context_length``) runs
|
|
several sequential provider probes, each with its own multi-second timeout, so
|
|
an unreachable ``model.base_url`` held the response for tens of seconds — and
|
|
the Desktop Model Settings page, which awaits this endpoint inside one
|
|
``Promise.all``, showed loading skeletons indefinitely.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import time
|
|
|
|
|
|
def _config(model: str, provider: str, base_url: str) -> dict:
|
|
return {"model": {"default": model, "provider": provider, "base_url": base_url}}
|
|
|
|
|
|
def test_model_info_degrades_when_the_context_probe_exceeds_its_budget(monkeypatch):
|
|
from hermes_cli.web_routers import models as router
|
|
|
|
monkeypatch.setattr(
|
|
router, "_load_config_scoped",
|
|
lambda profile: _config("some-model", "custom-proxy", "http://localhost:9/v1"),
|
|
)
|
|
monkeypatch.setattr(router, "_MODEL_INFO_PROBE_BUDGET_S", 0.2)
|
|
|
|
import agent.model_metadata as metadata
|
|
|
|
def _hanging_probe(model, base_url="", api_key="", config_context_length=None, provider="", custom_providers=None):
|
|
time.sleep(1.0)
|
|
raise AssertionError("the abandoned probe should never be awaited")
|
|
|
|
monkeypatch.setattr(metadata, "get_model_context_length", _hanging_probe)
|
|
|
|
started = time.monotonic()
|
|
info = router.get_model_info()
|
|
elapsed = time.monotonic() - started
|
|
|
|
# The response degrades to "auto context unknown" instead of hanging…
|
|
assert info["model"] == "some-model"
|
|
assert info["provider"] == "custom-proxy"
|
|
assert info["auto_context_length"] == 0
|
|
# …and it returns within the (shrunk) budget, not the probe's 1s sleep.
|
|
assert elapsed < 0.9
|
|
|
|
|
|
def test_model_info_surfaces_the_context_value_when_the_probe_is_fast(monkeypatch):
|
|
from hermes_cli.web_routers import models as router
|
|
|
|
monkeypatch.setattr(
|
|
router, "_load_config_scoped",
|
|
lambda profile: _config("some-model", "custom-proxy", "http://localhost:9/v1"),
|
|
)
|
|
|
|
import agent.model_metadata as metadata
|
|
|
|
def _fast_probe(model, base_url="", api_key="", config_context_length=None, provider="", custom_providers=None):
|
|
return 131072
|
|
|
|
monkeypatch.setattr(metadata, "get_model_context_length", _fast_probe)
|
|
|
|
info = router.get_model_info()
|
|
|
|
assert info["model"] == "some-model"
|
|
assert info["auto_context_length"] == 131072
|
|
assert info["effective_context_length"] == 131072
|
|
|
|
|
|
def test_model_info_bounds_the_endpoint_key_lookup_behind_a_held_auth_lock(monkeypatch):
|
|
"""Resolving the probe key loads the endpoint's credential pool, which persists config-seeded rows
|
|
to auth.json and waits for its lock (an OAuth refresh holds it across a network call). That wait
|
|
is inside the route's budget, not in front of it."""
|
|
import threading
|
|
|
|
from hermes_cli.auth import _auth_store_lock
|
|
from hermes_cli.config import save_config
|
|
from hermes_cli.web_routers import models as router
|
|
|
|
base = "http://127.0.0.1:47970/v1"
|
|
save_config({"model": {"default": "qwen-lab", "provider": "custom", "base_url": base},
|
|
"providers": {"lab": {"base_url": base, "api_key": "entry-key-1234"}}})
|
|
monkeypatch.setattr(router, "_MODEL_INFO_PROBE_BUDGET_S", 0.3)
|
|
monkeypatch.setattr("agent.model_metadata.get_model_context_length", lambda **kwargs: 131072)
|
|
held, release = threading.Event(), threading.Event()
|
|
|
|
def hold():
|
|
with _auth_store_lock():
|
|
held.set()
|
|
release.wait(5)
|
|
|
|
holder = threading.Thread(target=hold, daemon=True)
|
|
holder.start()
|
|
assert held.wait(5)
|
|
try:
|
|
started = time.monotonic()
|
|
info = router.get_model_info()
|
|
elapsed = time.monotonic() - started
|
|
finally:
|
|
release.set()
|
|
holder.join(5)
|
|
assert info["auto_context_length"] == 0
|
|
assert elapsed < 1.0
|