fix(adapters): 修复 GovernedLLMClient 流式消费和规格偏差

- _stream_request + _consume_stream 合并为 _call_streaming，在 async with self._http.stream() 内直接消费流，确保看门狗作用于真实 HTTP 流而非内存列表（Critical 1） - breaker 检查移到 call_id 生成之前（Critical 2） - cache/ttft_timeout_s/inter_token_timeout_s 支持 None（Important 1） - 重试失败路径遥测使用统一 call_id（Important 2） Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-07-06 23:08:52 -04:00
parent cfd3277f80
commit ee90fb8e88
2 changed files with 96 additions and 111 deletions
@@ -23,9 +23,7 @@ class _FakeRedisCache:
    def __init__(self) -> None:
        self._store: dict[str, LLMResponse] = {}

-    async def get(
-        self, model: str, messages: list[dict[str, str]]
-    ) -> LLMResponse | None:
+    async def get(self, model: str, messages: list[dict[str, str]]) -> LLMResponse | None:
        key = f"{model}:{json.dumps(messages, sort_keys=True)}"
        return self._store.get(key)

@@ -49,52 +47,6 @@ class _FakeTelemetry:
        self.calls.append(kwargs)


-# ── 工具函数 ─────────────────────────────────────────────────────────
-
-
-async def _async_line_iter(lines: list[str]):
-    """将字符串列表转为异步行迭代器。"""
-    for line in lines:
-        yield line
-
-
-def _make_sse_lines(
-    content_chunks: list[str],
-    *,
-    reasoning_chunks: list[str] | None = None,
-    model: str = "test-model",
-    usage: dict[str, int] | None = None,
-) -> list[str]:
-    """构造 SSE 文本行序列，模拟 OpenAI 兼容流式响应。"""
-    lines: list[str] = []
-
-    # 先产出 reasoning 帧
-    for chunk in reasoning_chunks or []:
-        frame = {
-            "choices": [{"delta": {"reasoning_content": chunk}}],
-            "model": model,
-        }
-        lines.append(f"data: {json.dumps(frame)}\n")
-
-    # 再产出 content 帧
-    for chunk in content_chunks:
-        frame = {
-            "choices": [{"delta": {"content": chunk}}],
-            "model": model,
-        }
-        lines.append(f"data: {json.dumps(frame)}\n")
-
-    # usage 帧
-    if usage is None:
-        usage = {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}
-    frame_usage = {"choices": [], "usage": usage, "model": model}
-    lines.append(f"data: {json.dumps(frame_usage)}\n")
-
-    # 终止标记
-    lines.append("data: [DONE]\n")
-    return lines
-
-
 def _build_client(
    *,
    breaker: CircuitBreaker | None = None,
@@ -180,16 +132,18 @@ async def test_successful_streaming_call():
    client = _build_client(telemetry=telemetry)
    messages = [{"role": "user", "content": "hello"}]

-    sse_lines = _make_sse_lines(
-        ["Hello", " world"],
-        reasoning_chunks=["Let me", " think"],
-        usage={"prompt_tokens": 20, "completion_tokens": 10, "total_tokens": 30},
-    )
+    async def fake_call_streaming(
+        _messages: Any,
+    ) -> tuple[str, str, float | None, float | None, dict]:
+        return (
+            "Hello world",
+            "Let me think",
+            1.5,  # ttft_ms
+            0.8,  # max_inter_token_ms
+            {"prompt_tokens": 20, "completion_tokens": 10, "total_tokens": 30},
+        )

-    async def fake_stream_request(_messages: Any) -> list[str]:
-        return sse_lines
-
-    with patch.object(client, "_stream_request", side_effect=fake_stream_request):
+    with patch.object(client, "_call_streaming", side_effect=fake_call_streaming):
        result = await client.chat(messages)

    assert result.content == "Hello world"
@@ -213,17 +167,24 @@ async def test_transient_error_retries_and_records_telemetry():
    client = _build_client(telemetry=telemetry)
    messages = [{"role": "user", "content": "hello"}]

-    sse_lines = _make_sse_lines(["OK"])
    call_count = 0

-    async def flaky_stream(_messages: Any) -> list[str]:
+    async def flaky_call_streaming(
+        _messages: Any,
+    ) -> tuple[str, str, float | None, float | None, dict]:
        nonlocal call_count
        call_count += 1
        if call_count <= 2:
            raise httpx.ConnectError("connection refused")
-        return sse_lines
+        return (
+            "OK",
+            "",
+            1.0,
+            None,
+            {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+        )

-    with patch.object(client, "_stream_request", side_effect=flaky_stream):
+    with patch.object(client, "_call_streaming", side_effect=flaky_call_streaming):
        result = await client.chat(messages)

    assert result.content == "OK"
@@ -236,6 +197,10 @@ async def test_transient_error_retries_and_records_telemetry():
    success_calls = [c for c in telemetry.calls if c.get("error") is None]
    assert len(success_calls) == 1

+    # 所有遥测记录应使用同一个 call_id（Important 2 修复验证）
+    call_ids = {c["call_id"] for c in telemetry.calls}
+    assert len(call_ids) == 1
+

@pytest.mark.asyncio
 async def test_fatal_error_force_opens_breaker():
@@ -250,13 +215,17 @@ async def test_fatal_error_force_opens_breaker():
        request=httpx.Request("POST", "http://localhost:8000/v1/chat/completions"),
    )

-    async def auth_fail(_messages: Any) -> list[str]:
+    async def auth_fail(
+        _messages: Any,
+    ) -> tuple[str, str, float | None, float | None, dict]:
        raise httpx.HTTPStatusError(
            "Unauthorized", request=mock_response.request, response=mock_response
        )

-    with patch.object(client, "_stream_request", side_effect=auth_fail), \
-         pytest.raises(httpx.HTTPStatusError):
+    with (
+        patch.object(client, "_call_streaming", side_effect=auth_fail),
+        pytest.raises(httpx.HTTPStatusError),
+    ):
        await client.chat(messages)

    # 熔断器已被 force_open
@@ -291,16 +260,20 @@ async def test_parent_call_id_forwarded_to_telemetry():
    client = _build_client(telemetry=telemetry)
    messages = [{"role": "user", "content": "hello"}]

-    sse_lines = _make_sse_lines(["result"])
-
-    async def fake_stream(_messages: Any) -> list[str]:
-        return sse_lines
-
-    with patch.object(client, "_stream_request", side_effect=fake_stream):
-        await client.chat(
-            messages, session_id="sess-42", parent_call_id="parent-99"
+    async def fake_call_streaming(
+        _messages: Any,
+    ) -> tuple[str, str, float | None, float | None, dict]:
+        return (
+            "result",
+            "",
+            1.0,
+            None,
+            {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
        )

+    with patch.object(client, "_call_streaming", side_effect=fake_call_streaming):
+        await client.chat(messages, session_id="sess-42", parent_call_id="parent-99")
+
    assert len(telemetry.calls) == 1
    assert telemetry.calls[0]["session_id"] == "sess-42"
    assert telemetry.calls[0]["parent_call_id"] == "parent-99"