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>

adapters/llm.py

@@ -234,11 +234,11 @@ class GovernedLLMClient:
         provider: str,
         thinking: bool,
         breaker: CircuitBreaker,
-        cache: Any,
+        cache: Any | None,
         telemetry: Any,
         timeout_s: float,
-        ttft_timeout_s: float,
-        inter_token_timeout_s: float,
+        ttft_timeout_s: float | None,
+        inter_token_timeout_s: float | None,
         max_retries: int,
         retry_base_delay_s: float,
         retry_max_delay_s: float,
@@ -252,8 +252,10 @@ class GovernedLLMClient:
         self._cache = cache
         self._telemetry = telemetry
         self._timeout_s = timeout_s
-        self._ttft_timeout_s = ttft_timeout_s
-        self._inter_token_timeout_s = inter_token_timeout_s
+        self._ttft_timeout_s = ttft_timeout_s if ttft_timeout_s is not None else timeout_s
+        self._inter_token_timeout_s = (
+            inter_token_timeout_s if inter_token_timeout_s is not None else timeout_s
+        )
         self._max_retries = max_retries
         self._retry_base_delay_s = retry_base_delay_s
         self._retry_max_delay_s = retry_max_delay_s
@@ -288,14 +290,16 @@ class GovernedLLMClient:
             httpx.HTTPStatusError: 不可恢复的 HTTP 错误（401/403）。
         """
         started = time.monotonic()
-        call_id = str(uuid4())
 
-        # ① 熔断器检查
+        # ① 熔断器检查（规格要求在 call_id 生成之前）
         if self._breaker.is_open(self._provider, time.monotonic()):
             raise CircuitOpenError(f"熔断器已开启，拒绝调用 provider={self._provider}")
 
-        # ② 缓存查询
-        cached = await self._cache.get(self._model, messages)
+        # ② call_id 生成
+        call_id = str(uuid4())
+
+        # ③ 缓存查询（cache 为 None 时跳过）
+        cached = await self._cache.get(self._model, messages) if self._cache is not None else None
         if cached is not None:
             response = LLMResponse(
                 content=cached.content,
@@ -329,14 +333,13 @@ class GovernedLLMClient:
             )
             return response
 
-        # ③ 重试循环 + 流式消费
+        # ④ 重试循环 + 流式消费
         last_exc: Exception | None = None
         for attempt in range(self._max_retries):
             attempt_start = time.monotonic()
             try:
-                sse_lines = await self._stream_request(messages)
-                content, thinking_text, ttft_ms, max_itoken_ms, usage = await self._consume_stream(
-                    sse_lines
+                content, thinking_text, ttft_ms, max_itoken_ms, usage = await self._call_streaming(
+                    messages
                 )
 
                 # 熔断器记成功
@@ -366,8 +369,9 @@ class GovernedLLMClient:
                     call_id=call_id,
                 )
 
-                # ④ 写缓存
-                await self._cache.set(self._model, messages, response)
+                # ④ 写缓存（cache 为 None 时跳过）
+                if self._cache is not None:
+                    await self._cache.set(self._model, messages, response)
 
                 # ⑤ 遥测
                 await self._telemetry.record_llm_call(
@@ -420,7 +424,7 @@ class GovernedLLMClient:
                 if _is_transient_error(exc):
                     self._breaker.record_failure(self._provider, time.monotonic())
                     await self._telemetry.record_llm_call(
-                        call_id=str(uuid4()),  # 每次重试用新 call_id
+                        call_id=call_id,
                         parent_call_id=parent_call_id,
                         session_id=session_id,
                         model_name=self._model,
@@ -478,20 +482,14 @@ class GovernedLLMClient:
         assert last_exc is not None
         raise last_exc
 
-    async def _stream_request(self, messages: list[dict[str, Any]]) -> list[str]:
-        """发起流式 HTTP 请求并收集全部 SSE 行。
-
-        此方法为主要的 HTTP 交互点，测试通过 patch 替换以注入假 SSE 行。
+    def _build_request_body(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
+        """构造流式 chat 请求体。
 
         参数:
             messages: OpenAI 格式消息列表。
 
         返回:
-            SSE 文本行列表。
-
-        异常:
-            httpx.HTTPStatusError: HTTP 状态码错误。
-            httpx.ConnectError: 连接错误。
+            请求体字典。
         """
         payload: dict[str, Any] = {
             "model": self._model,
@@ -501,34 +499,48 @@ class GovernedLLMClient:
         }
         if self._thinking:
             payload.update(_build_thinking_body(self._provider))
+        return payload
 
-        collected_lines: list[str] = []
+    async def _call_streaming(
+        self, messages: list[dict[str, Any]]
+    ) -> tuple[str, str, float | None, float | None, dict]:
+        """发起流式 HTTP 请求并在 context manager 内直接消费 SSE 流。
+
+        将 HTTP 请求与流消费合并，确保三层活性看门狗作用于真实 HTTP 流而非内存列表。
+        测试通过 patch 此方法注入假结果。
+
+        参数:
+            messages: OpenAI 格式消息列表。
+
+        返回:
+            (content, thinking, ttft_ms, max_inter_token_ms, usage_dict)。
+
+        异常:
+            httpx.HTTPStatusError: HTTP 状态码错误。
+            httpx.ConnectError: 连接错误。
+            StreamLivenessTimeout: 活性超时。
+        """
+        payload = self._build_request_body(messages)
         async with self._http.stream("POST", "/chat/completions", json=payload) as resp:
             if resp.status_code >= 400:
                 await resp.aread()
                 resp.raise_for_status()
-            async for line in resp.aiter_lines():
-                collected_lines.append(line)
-        return collected_lines
+            # 直接在 context manager 内消费流，看门狗作用于真实 HTTP 流
+            return await self._consume_stream(resp.aiter_lines())
 
     async def _consume_stream(
-        self, sse_lines: list[str]
+        self, lines: AsyncIterator[str]
     ) -> tuple[str, str, float | None, float | None, dict]:
-        """消费 SSE 行序列，累积 content/thinking，测量 TTFT 和 max_inter_token_ms。
+        """消费 SSE 行异步迭代器，累积 content/thinking，测量 TTFT 和 max_inter_token_ms。
 
         参数:
-            sse_lines: SSE 文本行列表。
+            lines: SSE 文本行异步迭代器（直接来自 httpx resp.aiter_lines()）。
 
         返回:
             (content, thinking, ttft_ms, max_inter_token_ms, usage_dict)。
         """
         usage_sink: dict = {}
-
-        async def _lines_iter() -> AsyncIterator[str]:
-            for line in sse_lines:
-                yield line
-
-        raw_deltas = _iter_sse_deltas(_lines_iter(), usage_sink)
+        raw_deltas = _iter_sse_deltas(lines, usage_sink)
         guarded = stream_with_liveness_timeouts(
             raw_deltas,
             ttft_s=self._ttft_timeout_s,

tests/unit/test_governed_llm.py

@@ -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"