feat(api): DeepSeek V4 Flash 默认、HTTP 错讯与多供应商分层

- 主链路默认 deepseek-v4-flash，DEEPSEEK_THINKING_ENABLED 对齐旧非思考 chat
- 评测台评审装配迁入 adapters/llm（deepseek_eval_judge、zhipu_eval_judge）与 eval_judge_spec
- 拆分 llm_http_openai_chat_errors 与 llm_errors（DeepSeek/智谱品牌与文档链），llm_call 支持 http_error_vendor
- EvalJudgeService 按 spec.provider 传入 allm_json_call；评测台前端文案改为 V4 Flash
- 更新 .env 示例与 staging/production 的 DEEPSEEK_MODEL；补充 openai/供应商错讯测试

Made-with: Cursor

api/tests/test_eval_judge_llm_spec.py

@@ -51,7 +51,8 @@ def test_build_eval_judge_deepseek_context_budget(
     spec = build_eval_judge_llm_spec("deepseek", None)
     assert spec is not None
     assert spec.provider == "deepseek"
-    assert spec.resolved_model == "deepseek-reasoner"
+    # 旧名 deepseek-reasoner 规范为 v4-flash 思考模式
+    assert spec.resolved_model == "deepseek-v4-flash"
     assert spec.context_window_tokens == 64_000
     n = eval_judge_conversation_transcript_max_chars_for_context(64_000)
     glm_n = eval_judge_conversation_transcript_max_chars_for_context(200_000)

api/tests/test_llm_json_call.py

@@ -2,8 +2,10 @@
 
 from __future__ import annotations
 
+import httpx
 import pytest
 from langchain_core.messages import AIMessage
+from openai import APIStatusError
 from pydantic import BaseModel, Field
 
 from app.core.langchain_llm import ensure_json_object_prompt_has_json_keyword
@@ -128,6 +130,41 @@ def test_llm_json_call_no_fallback_raises() -> None:
     assert ei.value.kind == "validation"
 
 
+def _api_status_402() -> APIStatusError:
+    req = httpx.Request("POST", "https://api.deepseek.com/v1/chat/completions")
+    resp = httpx.Response(
+        402, request=req, json={"error": {"message": "Insufficient balance"}}
+    )
+    return APIStatusError("Payment required", response=resp, body=resp.json())
+
+
+class _LlmInvokeRaises:
+    def bind(self, **_kwargs: object):
+        return self
+
+    def invoke(self, _prompt: str) -> object:
+        raise _api_status_402()
+
+    async def ainvoke(self, _prompt: str) -> object:
+        return self.invoke(_prompt)
+
+
+def test_llm_json_call_openai_status_maps_to_friendly_chinese() -> None:
+    with pytest.raises(LLMCallError) as ei:
+        llm_json_call(
+            _LlmInvokeRaises(),
+            "p",
+            _SmallOut,
+            max_tokens=8,
+            agent="t",
+        )
+    assert ei.value.kind == "invoke"
+    s = str(ei.value)
+    assert "402" in s
+    assert "余额" in s
+    assert "DeepSeek" in s
+
+
 @pytest.mark.asyncio
 async def test_allm_json_call_parity_with_sync() -> None:
     llm = _SyncFakeLlm(['{"answer": "async", "score": 7}'])

api/tests/test_llm_vendor_errors.py

@@ -0,0 +1,36 @@
+"""供应商层 HTTP 错讯：DeepSeek / 智谱 品牌与文档链。"""
+
+import httpx
+from openai import APIStatusError
+
+from app.core.llm_errors import format_llm_http_error_message
+
+
+def _status_402() -> APIStatusError:
+    req = httpx.Request("POST", "https://x/v1/chat/completions")
+    resp = httpx.Response(402, request=req, json={"error": {"message": "x"}})
+    return APIStatusError("u", response=resp, body=resp.json())
+
+
+def test_vendor_deepseek_includes_brand_and_doc() -> None:
+    e = _status_402()
+    m = format_llm_http_error_message(e, "deepseek")
+    assert m is not None
+    assert "DeepSeek" in m
+    assert "api-docs.deepseek.com" in m
+
+
+def test_vendor_zhipu_includes_brand_and_doc() -> None:
+    e = _status_402()
+    m = format_llm_http_error_message(e, "zhipu")
+    assert m is not None
+    assert "智谱" in m
+    assert "bigmodel" in m
+
+
+def test_vendor_none_is_transport_only() -> None:
+    e = _status_402()
+    m = format_llm_http_error_message(e, None)
+    assert m is not None
+    assert "【" not in m
+    assert "402" in m

api/tests/test_openai_compatible_errors.py

@@ -0,0 +1,68 @@
+"""传输层 `llm_http_openai_chat_errors` 的中性错讯；兼容 re-export 仍经 openai_compatible_errors。"""
+
+import httpx
+import pytest
+from openai import APIStatusError
+
+from app.core.openai_compatible_errors import (
+    extract_openai_http_status,
+    format_openai_compatible_http_error_message,
+    should_log_openai_error_as_warning,
+)
+
+
+def _status_error(status: int, *, body: object | None = None) -> APIStatusError:
+    req = httpx.Request("POST", "https://api.deepseek.com/v1/chat/completions")
+    resp = httpx.Response(status, request=req, json=body if body is not None else {})
+    return APIStatusError("upstream", response=resp, body=body)
+
+
+def test_extract_status_from_api_status_error() -> None:
+    e = _status_error(429)
+    assert extract_openai_http_status(e) == 429
+
+
+def test_format_402_balance_chinese_message() -> None:
+    e = _status_error(
+        402,
+        body={"error": {"message": "Insufficient balance", "type": "insufficient_quota"}},
+    )
+    msg = format_openai_compatible_http_error_message(e)
+    assert msg is not None
+    assert "402" in msg
+    assert "余额" in msg
+
+
+def test_format_401_and_warning_flag() -> None:
+    e = _status_error(401, body={"error": {"message": "invalid api key"}})
+    assert should_log_openai_error_as_warning(e) is True
+    m = format_openai_compatible_http_error_message(e)
+    assert m is not None
+    assert "401" in m
+    assert "密钥" in m
+
+
+def test_format_503_server_busy() -> None:
+    e = _status_error(503)
+    m = format_openai_compatible_http_error_message(e)
+    assert m is not None
+    assert "503" in m
+    assert should_log_openai_error_as_warning(e) is False
+
+
+def test_format_httpx_http_status_error() -> None:
+    req = httpx.Request("GET", "https://api.deepseek.com/v1/models")
+    resp = httpx.Response(429, request=req)
+    try:
+        resp.raise_for_status()
+    except httpx.HTTPStatusError as e:
+        m = format_openai_compatible_http_error_message(e)
+        assert m is not None
+        assert "429" in m
+
+
+def test_unknown_status_418() -> None:
+    e = _status_error(418)
+    m = format_openai_compatible_http_error_message(e)
+    assert m is not None
+    assert "418" in m