mirror of
https://github.com/simonlin1212/TradingAgents-astock.git
synced 2026-08-31 01:23:38 +00:00
fb96e0ac9c
前三轮修的东西里，有两处"看着修好了、实际没生效"：
1. 未来函数告警仍然不触发
v0.5.6 给 get_profit_forecast 补了 curr_date，但给了默认空串——LangChain 只把
ticker 标成必填，模型正常调用时 curr_date 为空，_is_historical("") 为 False，
告警照样一次都不触发。提示词里也没提这个参数。改为必填 + 提示词显式说明。
2. 历史资金流放大窗口后没裁回
为够回溯放大了请求窗口，过滤未来行后没裁回承诺的 20 个交易日——复盘 90 天前
返回约 40 行，改变了趋势窗口且返回体翻倍。
3. 未选中的分析师角色也会被建模型
role_llms 配了 policy 但本次只选 market 时，policy 的模型仍被实例化。一个永远
不执行的节点可能因缺 key 或缺可选依赖，把本来正常的分析在启动时打断。
4. 方向正确率的显著性用错分母
direction_accuracy 排除 Hold，却用已结算总数判断样本是否足够。20 条已结算里
只有 1 条有方向时，噪音提示被抑制，报告却显示"方向正确率 100%"。
测试：新增 7 例，329 passed / 13 skipped / 0 failed。版本 → v0.5.8
321 lines
11 KiB
Python
321 lines
11 KiB
Python
"""分角色模型（#39）。
|
||
|
||
同一个模型分饰多角时倾向于互相附和，多空辩论就失去意义。`role_llms` 允许给
|
||
单个角色指定另一家模型。
|
||
|
||
**默认必须完全维持原行为**——大多数用户只有一家模型，不配这一项时不能有任何
|
||
变化。下面第一组用例就是锁这一点的。
|
||
"""
|
||
|
||
import pytest
|
||
|
||
from tradingagents.graph.setup import DEEP_ROLES, ROLE_KEYS, GraphSetup
|
||
|
||
|
||
class FakeLLM:
|
||
def __init__(self, tag):
|
||
self.tag = tag
|
||
|
||
def __repr__(self): # pragma: no cover - 只为断言失败时好读
|
||
return f"FakeLLM({self.tag})"
|
||
|
||
|
||
@pytest.fixture
|
||
def llms():
|
||
return FakeLLM("quick"), FakeLLM("deep")
|
||
|
||
|
||
def make_setup(quick, deep, resolve=None):
|
||
return GraphSetup(quick, deep, tool_nodes={}, conditional_logic=None, resolve_llm=resolve)
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 默认行为不变
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_without_role_llms_quick_roles_use_quick(llms):
|
||
quick, deep = llms
|
||
setup = make_setup(quick, deep)
|
||
|
||
for role in ROLE_KEYS:
|
||
if role not in DEEP_ROLES:
|
||
assert setup.llm_for(role) is quick, role
|
||
|
||
|
||
def test_without_role_llms_deep_roles_use_deep(llms):
|
||
quick, deep = llms
|
||
setup = make_setup(quick, deep)
|
||
|
||
for role in DEEP_ROLES:
|
||
assert setup.llm_for(role) is deep, role
|
||
|
||
|
||
def test_empty_resolver_falls_back(llms):
|
||
"""resolve_llm 存在但对该角色返回 None → 回落，而不是把 None 传给 agent。"""
|
||
quick, deep = llms
|
||
setup = make_setup(quick, deep, resolve={}.get)
|
||
|
||
assert setup.llm_for("bull") is quick
|
||
assert setup.llm_for("portfolio_manager") is deep
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 配置生效
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_configured_role_overrides_default(llms):
|
||
quick, deep = llms
|
||
bear_llm = FakeLLM("bear-other-vendor")
|
||
setup = make_setup(quick, deep, resolve={"bear": bear_llm}.get)
|
||
|
||
assert setup.llm_for("bear") is bear_llm
|
||
assert setup.llm_for("bull") is quick # 没配的角色不受影响
|
||
|
||
|
||
def test_configured_deep_role_overrides_deep(llms):
|
||
quick, deep = llms
|
||
pm_llm = FakeLLM("pm-other-vendor")
|
||
setup = make_setup(quick, deep, resolve={"portfolio_manager": pm_llm}.get)
|
||
|
||
assert setup.llm_for("portfolio_manager") is pm_llm
|
||
assert setup.llm_for("research_manager") is deep
|
||
|
||
|
||
def test_all_roles_are_addressable(llms):
|
||
"""ROLE_KEYS 里的每个名字都必须真的能指到一个角色。"""
|
||
quick, deep = llms
|
||
for role in ROLE_KEYS:
|
||
marker = FakeLLM(role)
|
||
setup = make_setup(quick, deep, resolve={role: marker}.get)
|
||
assert setup.llm_for(role) is marker, role
|
||
|
||
|
||
def test_role_keys_cover_bull_and_bear():
|
||
"""#39 的核心诉求就是多空分开，这两个键必须在。"""
|
||
assert "bull" in ROLE_KEYS
|
||
assert "bear" in ROLE_KEYS
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# 配置解析：写错要当场报错，相同配置要复用实例
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def build(config, monkeypatch, subscription_on=False):
|
||
"""只跑 _build_role_llms，不做整图初始化（那要真 API key）。"""
|
||
from tradingagents.graph import trading_graph as tg
|
||
|
||
created = []
|
||
|
||
class FakeClient:
|
||
def __init__(self, provider, model, base_url):
|
||
self.spec = (provider, model, base_url)
|
||
|
||
def get_llm(self):
|
||
return FakeLLM(self.spec)
|
||
|
||
def fake_create(provider, model, base_url=None, **kwargs):
|
||
created.append((provider, model, base_url))
|
||
return FakeClient(provider, model, base_url)
|
||
|
||
monkeypatch.setattr(tg, "create_llm_client", fake_create)
|
||
graph = tg.TradingAgentsGraph.__new__(tg.TradingAgentsGraph)
|
||
graph.config = config
|
||
return graph._build_role_llms({}, subscription_on), created
|
||
|
||
|
||
def test_no_config_builds_nothing(monkeypatch):
|
||
resolved, created = build({"llm_provider": "openai", "role_llms": {}}, monkeypatch)
|
||
|
||
assert resolved == {}
|
||
assert created == [] # 一个客户端都不该建
|
||
|
||
|
||
def test_unknown_role_name_raises(monkeypatch):
|
||
"""角色名写错必须当场报错——静默忽略会让人以为配置生效了。"""
|
||
with pytest.raises(ValueError, match="无法识别的角色名"):
|
||
build(
|
||
{"llm_provider": "openai", "role_llms": {"bulls": {"model": "m"}}},
|
||
monkeypatch,
|
||
)
|
||
|
||
|
||
def test_spec_without_model_raises(monkeypatch):
|
||
with pytest.raises(ValueError, match="必须是带 model 的字典"):
|
||
build(
|
||
{"llm_provider": "openai", "role_llms": {"bull": {"provider": "qwen"}}},
|
||
monkeypatch,
|
||
)
|
||
|
||
|
||
def test_identical_specs_share_one_instance(monkeypatch):
|
||
"""两个角色配同一个模型，只该建一个实例，不该开两条连接。"""
|
||
cfg = {
|
||
"llm_provider": "openai",
|
||
"role_llms": {
|
||
"bull": {"provider": "deepseek", "model": "deepseek-chat"},
|
||
"bear": {"provider": "deepseek", "model": "deepseek-chat"},
|
||
},
|
||
}
|
||
resolved, created = build(cfg, monkeypatch)
|
||
|
||
assert resolved["bull"] is resolved["bear"]
|
||
assert len(created) == 1
|
||
|
||
|
||
def test_different_vendors_build_separate_instances(monkeypatch):
|
||
cfg = {
|
||
"llm_provider": "openai",
|
||
"role_llms": {
|
||
"bull": {"provider": "deepseek", "model": "deepseek-chat"},
|
||
"bear": {"provider": "qwen", "model": "qwen-plus"},
|
||
},
|
||
}
|
||
resolved, created = build(cfg, monkeypatch)
|
||
|
||
assert resolved["bull"] is not resolved["bear"]
|
||
assert len(created) == 2
|
||
|
||
|
||
def test_backend_url_not_leaked_across_vendors(monkeypatch):
|
||
"""主 provider 的端点不能带给另一家，否则请求发到别人的网关。"""
|
||
cfg = {
|
||
"llm_provider": "openai",
|
||
"backend_url": "https://my-openai-relay.example/v1",
|
||
"role_llms": {"bear": {"provider": "deepseek", "model": "deepseek-chat"}},
|
||
}
|
||
_, created = build(cfg, monkeypatch)
|
||
|
||
assert created[0][2] is None
|
||
|
||
|
||
def test_backend_url_kept_for_same_vendor(monkeypatch):
|
||
"""同一家 provider 换模型，端点应当继续沿用。"""
|
||
cfg = {
|
||
"llm_provider": "openai",
|
||
"backend_url": "https://my-openai-relay.example/v1",
|
||
"role_llms": {"bear": {"model": "gpt-5.4-mini"}},
|
||
}
|
||
_, created = build(cfg, monkeypatch)
|
||
|
||
assert created[0] == ("openai", "gpt-5.4-mini", "https://my-openai-relay.example/v1")
|
||
|
||
|
||
def test_explicit_backend_url_wins(monkeypatch):
|
||
cfg = {
|
||
"llm_provider": "openai",
|
||
"backend_url": "https://main.example/v1",
|
||
"role_llms": {
|
||
"bear": {"provider": "qwen", "model": "qwen-plus",
|
||
"backend_url": "https://my-qwen.example/v1"},
|
||
},
|
||
}
|
||
_, created = build(cfg, monkeypatch)
|
||
|
||
assert created[0][2] == "https://my-qwen.example/v1"
|
||
|
||
|
||
def test_warns_when_bypassing_subscription(monkeypatch, caplog):
|
||
"""订阅覆盖开着时，绕开它去计费的角色必须被点名，不能悄悄花钱。"""
|
||
import logging
|
||
|
||
cfg = {
|
||
"llm_provider": "openai",
|
||
"role_llms": {"bear": {"provider": "deepseek", "model": "deepseek-chat"}},
|
||
}
|
||
with caplog.at_level(logging.WARNING):
|
||
build(cfg, monkeypatch, subscription_on=True)
|
||
|
||
assert any("bear" in r.getMessage() and "计费" in r.getMessage() for r in caplog.records)
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# codex 复审补的两条
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_provider_specific_kwargs_not_leaked_to_other_vendors(monkeypatch):
|
||
"""openai 的 reasoning_effort 不能带给 qwen —— 别家可能直接拒收这个参数。"""
|
||
from tradingagents.graph import trading_graph as tg
|
||
|
||
created = []
|
||
|
||
class FakeClient:
|
||
def __init__(self, **kw): pass
|
||
def get_llm(self): return FakeLLM("x")
|
||
|
||
def fake_create(provider, model, base_url=None, **kwargs):
|
||
created.append((provider, kwargs))
|
||
return FakeClient()
|
||
|
||
monkeypatch.setattr(tg, "create_llm_client", fake_create)
|
||
graph = tg.TradingAgentsGraph.__new__(tg.TradingAgentsGraph)
|
||
graph.config = {
|
||
"llm_provider": "openai",
|
||
"role_llms": {
|
||
"bull": {"provider": "qwen", "model": "qwen-plus"},
|
||
"bear": {"model": "gpt-5.4-mini"}, # 同一家，应保留
|
||
},
|
||
}
|
||
graph._build_role_llms({"reasoning_effort": "high", "max_tokens": 8000}, False)
|
||
|
||
by_provider = {p: kw for p, kw in created}
|
||
assert "reasoning_effort" not in by_provider["qwen"], "别家 provider 收到了 openai 专属参数"
|
||
assert by_provider["qwen"]["max_tokens"] == 8000, "通用参数不该被一起过滤掉"
|
||
assert by_provider["openai"]["reasoning_effort"] == "high", "同一家应保留专属参数"
|
||
|
||
|
||
def test_unselected_analyst_roles_are_not_instantiated(monkeypatch):
|
||
"""没选中的分析师不会进图，就不该为它建模型。
|
||
|
||
否则一个**永远不执行**的节点会因为缺 API key 或缺可选依赖，把一次本来完全
|
||
正常的分析在启动时就打断（codex 终轮指出）。
|
||
"""
|
||
from tradingagents.graph import trading_graph as tg
|
||
|
||
created = []
|
||
|
||
class FakeClient:
|
||
def __init__(self, **kw): pass
|
||
def get_llm(self): return FakeLLM("x")
|
||
|
||
monkeypatch.setattr(tg, "create_llm_client",
|
||
lambda provider, model, base_url=None, **kw: (
|
||
created.append(provider) or FakeClient()))
|
||
graph = tg.TradingAgentsGraph.__new__(tg.TradingAgentsGraph)
|
||
graph.config = {
|
||
"llm_provider": "openai",
|
||
"role_llms": {
|
||
"market": {"provider": "qwen", "model": "qwen-plus"},
|
||
"policy": {"provider": "glm", "model": "glm-4.6"}, # 未选中
|
||
},
|
||
}
|
||
|
||
resolved = graph._build_role_llms({}, False, selected_analysts=["market"])
|
||
|
||
assert "market" in resolved
|
||
assert "policy" not in resolved
|
||
assert "glm" not in created, "未选中的分析师角色不该建模型"
|
||
|
||
|
||
def test_non_analyst_roles_are_always_built(monkeypatch):
|
||
"""多空/风险/Manager 这些角色不受 selected_analysts 控制，必须照常建。"""
|
||
from tradingagents.graph import trading_graph as tg
|
||
|
||
class FakeClient:
|
||
def __init__(self, **kw): pass
|
||
def get_llm(self): return FakeLLM("x")
|
||
|
||
monkeypatch.setattr(tg, "create_llm_client",
|
||
lambda provider, model, base_url=None, **kw: FakeClient())
|
||
graph = tg.TradingAgentsGraph.__new__(tg.TradingAgentsGraph)
|
||
graph.config = {
|
||
"llm_provider": "openai",
|
||
"role_llms": {"bull": {"provider": "qwen", "model": "qwen-plus"}},
|
||
}
|
||
|
||
resolved = graph._build_role_llms({}, False, selected_analysts=["market"])
|
||
assert "bull" in resolved
|