Skip to content

JiuwenSwarm is an intelligent AI Agent built on openJiuwen. It extends the powerful capabilities of large language models directly to your fingertips through various communication apps you use daily.

Python
5.6 KB
Raw Blame
# Copyright (c) Huawei Technologies Co., Ltd. 2026. All rights reserved.

"""并发会话换模后懒重载:重建模型缓存不能丢掉老会话仍在用的模型。"""

from __future__ import annotations

from types import SimpleNamespace

from openjiuwen.core.foundation.llm import Model, ModelClientConfig, ModelRequestConfig

from jiuwenswarm.common.model_identity import build_model_identity_reference
from jiuwenswarm.server.runtime.agent_adapter.deep.model_factory import (
    _DEFAULT_STREAM_IDLE_TIMEOUT_SECONDS,
    _build_model_from_entry,
)
from jiuwenswarm.server.runtime.agent_adapter.interface_deep import (
    JiuWenSwarmDeepAdapter,
)


def _entry(name: str, *, api_base: str = "https://maas.example/v2", api_key: str = "k1") -> dict:
    return {
        "model_client_config": {
            "model_name": name,
            "api_base": api_base,
            "api_key": api_key,
            "client_provider": "OpenAI",
        },
        "model_config_obj": {},
    }


def test_entry_without_idle_timeout_uses_host_default():
    entry = _entry("glm-5.3")
    model = _build_model_from_entry(entry["model_client_config"], entry["model_config_obj"])
    assert model.model_client_config.stream_idle_timeout == _DEFAULT_STREAM_IDLE_TIMEOUT_SECONDS
    assert "stream_idle_timeout" not in entry["model_client_config"]


def test_entry_idle_timeout_overrides_host_default():
    entry = _entry("glm-5.3")
    entry["model_client_config"]["stream_idle_timeout"] = 30
    model = _build_model_from_entry(entry["model_client_config"], entry["model_config_obj"])
    assert model.model_client_config.stream_idle_timeout == 30


def _config(*entries: dict) -> dict:
    return {"models": {"defaults": list(entries)}}


def _ref(entry: dict) -> str:
    mcc = entry["model_client_config"]
    return build_model_identity_reference(mcc["model_name"], mcc)


def _adapter(monkeypatch) -> tuple[JiuWenSwarmDeepAdapter, list[str]]:
    adapter = JiuWenSwarmDeepAdapter()
    adapter._instance = SimpleNamespace(_react_agent=SimpleNamespace())
    built: list[str] = []

    def _build(mcc, _mco, *_args, **_kwargs):
        built.append(mcc["model_name"])
        return Model(
            model_client_config=ModelClientConfig(
                client_provider="OpenAI",
                api_key=mcc["api_key"],
                api_base=mcc["api_base"],
                verify_ssl=False,
            ),
            model_config=ModelRequestConfig(model_name=mcc["model_name"]),
        )

    monkeypatch.setattr(
        "jiuwenswarm.server.runtime.agent_adapter.interface_deep._build_model_from_entry", _build)
    monkeypatch.setattr(adapter, "_register_models_json_cache_entries", lambda _counter: None)
    return adapter, built


def _name(model: Model) -> str:
    return model.model_config.model_name


def _use(adapter: JiuWenSwarmDeepAdapter, name: str) -> Model:
    model = adapter._resolve_model_by_name(name)
    adapter._apply_model_to_react_agent(model)
    return model


def test_reload_keeps_applied_model_dropped_from_new_config(monkeypatch):
    adapter, built = _adapter(monkeypatch)
    glm = _entry("glm-5.1")
    adapter._create_model(_config(glm))
    old_glm = _use(adapter, "glm-5.1")

    adapter._create_model(_config(_entry("deepseek-v4.1-flash")))

    assert _name(adapter._model) == "deepseek-v4.1-flash"
    resolved = adapter._resolve_model_by_name("glm-5.1")
    assert _name(resolved) == "glm-5.1"
    assert resolved is not old_glm
    assert built.count("glm-5.1") == 2
    assert adapter._resolve_model_by_identity(_ref(glm)) is resolved


def test_only_applied_models_are_retained(monkeypatch):
    adapter, built = _adapter(monkeypatch)
    glm = _entry("glm-5.1")
    adapter._create_model(_config(glm))
    _use(adapter, "glm-5.1")
    adapter._create_model(_config(_entry("glm-5.2")))
    adapter._create_model(_config(_entry("deepseek-v4.1-flash")))

    assert _name(adapter._resolve_model_by_name("glm-5.1")) == "glm-5.1"
    assert _name(adapter._resolve_model_by_identity(_ref(glm))) == "glm-5.1"
    assert _name(adapter._resolve_model_by_name("glm-5.2")) == "deepseek-v4.1-flash"
    assert built.count("glm-5.2") == 1


def test_retained_model_reapplied_after_reload_survives_next_reload(monkeypatch):
    adapter, _ = _adapter(monkeypatch)
    adapter._create_model(_config(_entry("glm-5.1")))
    _use(adapter, "glm-5.1")
    adapter._create_model(_config(_entry("glm-5.2")))
    _use(adapter, "glm-5.1")
    adapter._create_model(_config(_entry("deepseek-v4.1-flash")))

    assert _name(adapter._resolve_model_by_name("glm-5.1")) == "glm-5.1"
    assert len(adapter._model_name_to_keys["glm-5.1"]) == 1


def test_new_config_wins_for_same_identity(monkeypatch):
    adapter, _ = _adapter(monkeypatch)
    adapter._create_model(_config(_entry("glm-5.1", api_key="old")))
    _use(adapter, "glm-5.1")
    adapter._create_model(_config(_entry("glm-5.1", api_key="new")))

    resolved = adapter._resolve_model_by_name("glm-5.1")
    assert resolved.model_client_config.api_key == "new"
    assert len(adapter._model_identity_to_keys[_ref(_entry("glm-5.1"))]) == 1


def test_same_name_other_endpoint_stays_reachable_by_ref_only(monkeypatch):
    adapter, _ = _adapter(monkeypatch)
    old = _entry("glm-5.1", api_base="https://old.example/v1")
    adapter._create_model(_config(old))
    _use(adapter, "glm-5.1")
    adapter._create_model(_config(_entry("glm-5.1", api_base="https://new.example/v1")))

    by_name = adapter._resolve_model_by_name("glm-5.1")
    assert by_name.model_client_config.api_base == "https://new.example/v1"
    by_ref = adapter._resolve_model_by_identity(_ref(old))
    assert by_ref.model_client_config.api_base == "https://old.example/v1"