mirror of
https://github.com/simonlin1212/TradingAgents-astock.git
synced 2026-08-31 01:23:38 +00:00
383a38dc31
五条全部实测复现，无一误报。 1. 截断告警对默认 provider 根本不会响 只认 Anthropic 的 stop_reason=max_tokens 与 Chat 的 finish_reason=length， 而 openai 是默认 provider 且走 Responses API（status=incomplete + incomplete_details.reason=max_output_tokens），Gemini 是大写 MAX_TOKENS。 v0.5.1 加的这个告警在默认配置下一次都不会触发。现覆盖四种形状。 2. 非 A 股防护漏了三个 vendor get_dragon_tiger_board / get_lockup_expiry / get_industry_comparison 直接调 safe_ticker_component 绕过了 _normalize_ticker。实测拿 00700 调用返回的是 看起来完全正常的报告（"近30日未上龙虎榜"），模型会当成腾讯的事实。 v0.5.3 声称的"一个卡点覆盖 15 个接口"并不成立。 3. 绩效统计对看空评级把对错算反 给 Sell 后股价下跌是判断正确，原实现只看收益是否为正，记成失败。新增 direction_accuracy（按方向判定，Hold 不计入），原指标改名 up_rate / outperform_rate —— 命名即口径，报告写明它们与判断对错无关。 4. 平均持有期永远缺失 记忆日志写 "5d"，int() 抛异常被吞成 None。改为提取数字部分。 5. mootdx 提前收手会漏掉可用服务器 三台远端拒绝推不出本地封了协议，靠后的服务器可能是好的。移除提前退出， 跑完整张表。实测 18.7s → 23.7s（最初 >170s，bestip 全表测速仍单独规避）。 测试：新增 13 例，316 passed / 13 skipped / 0 failed
153 lines
5.6 KiB
Python
153 lines
5.6 KiB
Python
"""输出上限与 Anthropic 兼容端点配置（#91 / #89）。
|
||
|
||
#91「报告输出到一半结束」：走 anthropic 通道跑第三方模型（Kimi 等）时，
|
||
langchain-anthropic 认不出模型名，会落到一个很小的兜底输出上限，报告写不完就
|
||
被截断——而返回值本身完全合法，没有任何报错。
|
||
|
||
#89「调用 kimi 失败 401」：用第三方模型名走 anthropic 通道却没配端点，请求会被
|
||
发到 api.anthropic.com，拿第三方 token 认证，报一句看不懂的 invalid x-api-key。
|
||
"""
|
||
|
||
import logging
|
||
|
||
import pytest
|
||
|
||
from tradingagents.llm_clients.anthropic_client import (
|
||
_THIRD_PARTY_DEFAULT_MAX_TOKENS,
|
||
AnthropicClient,
|
||
)
|
||
from tradingagents.llm_clients.base_client import warn_if_truncated
|
||
from tradingagents.llm_clients.openai_client import OpenAIClient
|
||
|
||
|
||
class FakeResponse:
|
||
"""够用的 AIMessage 替身：只需要 response_metadata。"""
|
||
|
||
def __init__(self, metadata):
|
||
self.response_metadata = metadata
|
||
self.content = "写到一半就断了的报告"
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# #91 输出上限
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_third_party_model_gets_explicit_max_tokens():
|
||
"""Kimi 这类第三方模型必须拿到显式上限，不能听凭 langchain 的小兜底值。"""
|
||
client = AnthropicClient(
|
||
"kimi-k2-0905-preview", base_url="https://api.kimi.com/coding/"
|
||
)
|
||
llm = client.get_llm()
|
||
|
||
assert llm.max_tokens == _THIRD_PARTY_DEFAULT_MAX_TOKENS
|
||
|
||
|
||
def test_real_claude_model_keeps_provider_default():
|
||
"""真 Claude 模型不要动——它的上限比我们的默认值大得多。"""
|
||
llm = AnthropicClient("claude-sonnet-4-6").get_llm()
|
||
|
||
assert llm.max_tokens > _THIRD_PARTY_DEFAULT_MAX_TOKENS
|
||
|
||
|
||
def test_dated_claude_model_id_is_not_treated_as_third_party():
|
||
"""带日期的正规 Claude ID 不在我们的目录里，但绝不能被砍到第三方默认值。"""
|
||
llm = AnthropicClient("claude-sonnet-4-5-20250929").get_llm()
|
||
|
||
assert llm.max_tokens > _THIRD_PARTY_DEFAULT_MAX_TOKENS
|
||
|
||
|
||
def test_explicit_max_tokens_wins():
|
||
"""用户显式配置的上限优先于任何默认值。"""
|
||
client = AnthropicClient(
|
||
"kimi-k2-0905-preview",
|
||
base_url="https://api.kimi.com/coding/",
|
||
max_tokens=32000,
|
||
)
|
||
|
||
assert client.get_llm().max_tokens == 32000
|
||
|
||
|
||
def test_openai_client_forwards_max_tokens():
|
||
"""OpenAI 兼容通道同样要能配上限，否则这个配置项对半数用户是死的。"""
|
||
client = OpenAIClient(
|
||
"deepseek-chat", provider="deepseek", api_key="test-key", max_tokens=12345
|
||
)
|
||
|
||
assert client.get_llm().max_tokens == 12345
|
||
|
||
|
||
@pytest.mark.parametrize(
|
||
"metadata",
|
||
[
|
||
{"stop_reason": "max_tokens"}, # Anthropic
|
||
{"finish_reason": "length"}, # OpenAI 兼容 Chat
|
||
{"finish_reason": "MAX_TOKENS"}, # Gemini（大写）
|
||
# OpenAI Responses API —— `openai` 是**默认 provider** 且走这条路径，
|
||
# 漏掉它等于默认配置下这个告警根本不会响（codex P1）
|
||
{"status": "incomplete",
|
||
"incomplete_details": {"reason": "max_output_tokens"}},
|
||
],
|
||
)
|
||
def test_truncated_response_is_reported(metadata, caplog):
|
||
"""被截断必须喊出来——否则用户只会以为模型没写完（#91 的真正痛点）。"""
|
||
with caplog.at_level(logging.WARNING):
|
||
warn_if_truncated(FakeResponse(metadata), "some-model")
|
||
|
||
assert any("max_tokens" in r.getMessage() for r in caplog.records)
|
||
|
||
|
||
@pytest.mark.parametrize(
|
||
"metadata",
|
||
[
|
||
{"stop_reason": "end_turn"},
|
||
{"finish_reason": "stop"},
|
||
{"finish_reason": "STOP"},
|
||
# 因为别的原因 incomplete（如内容过滤）不是输出上限，不该报 max_tokens
|
||
{"status": "incomplete", "incomplete_details": {"reason": "content_filter"}},
|
||
{"status": "completed"},
|
||
{},
|
||
],
|
||
)
|
||
def test_normal_response_is_not_reported(metadata, caplog):
|
||
"""正常收尾不要报警，否则告警变噪音就没人看了。"""
|
||
with caplog.at_level(logging.WARNING):
|
||
warn_if_truncated(FakeResponse(metadata), "some-model")
|
||
|
||
assert not caplog.records
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# #89 Anthropic 兼容端点
|
||
# ---------------------------------------------------------------------------
|
||
|
||
|
||
def test_third_party_model_without_endpoint_fails_loudly(monkeypatch):
|
||
"""没配端点就用第三方模型名 → 当场说清楚，而不是让 Anthropic 回一句 401。"""
|
||
monkeypatch.delenv("ANTHROPIC_BASE_URL", raising=False)
|
||
|
||
with pytest.raises(RuntimeError) as exc:
|
||
AnthropicClient("kimi-k2-0905-preview").get_llm()
|
||
|
||
message = str(exc.value)
|
||
assert "backend_url" in message
|
||
assert "ANTHROPIC_API_KEY" in message
|
||
# 必须点破这个常见误配：ANTHROPIC_AUTH_TOKEN 是 Claude Code CLI 的写法
|
||
assert "ANTHROPIC_AUTH_TOKEN" in message
|
||
|
||
|
||
def test_anthropic_base_url_env_is_honoured(monkeypatch):
|
||
"""端点也允许走环境变量给，不是只能写 config。"""
|
||
monkeypatch.setenv("ANTHROPIC_BASE_URL", "https://api.kimi.com/coding/")
|
||
|
||
llm = AnthropicClient("kimi-k2-0905-preview").get_llm()
|
||
|
||
assert "kimi" in str(llm.anthropic_api_url)
|
||
|
||
|
||
def test_claude_model_without_endpoint_still_works(monkeypatch):
|
||
"""真 Claude 模型不配端点是完全正常的用法，不能被新校验误伤。"""
|
||
monkeypatch.delenv("ANTHROPIC_BASE_URL", raising=False)
|
||
|
||
assert AnthropicClient("claude-sonnet-4-6").get_llm() is not None
|