fix(inference): chat 每次调用新建 httpx client，修复同步门禁 Event loop is closed

_AI_USAGE_LOG.md

@@ -107,3 +107,4 @@
 | 2026-08-13 23:11 | Agent 实现 | 接通真实 LLM 引擎工厂（P5-T10 门禁接线）：新增 inference/factory.py，orchestrator/qa_loop 接入，脚本校验适配，补工厂测试 | src/genesis/inference/factory.py; src/genesis/writer/orchestrator.py; src/genesis/qa/qa_loop.py; scripts/run_phase5_slice.py; tests/test_inference_factory.py | hy3-free |
 | 2026-08-13 23:20 | 测试验证 | P5-T10 引擎工厂接线：补 engine=None 委托工厂测试（orchestrator/qa_loop），全量 296 passed / 99.04% | tests/test_phase5_writer_orchestrator.py; tests/test_phase5_qa_loop.py | hy3-free |
 | 2026-08-13 23:35 | 测试验证 | P5-T10 门禁诊断：修复 WriterGenerationError 吞掉底层 LLM 错误（如 401 详情），补透传测试 | src/genesis/writer/writer_agent.py; tests/test_phase5_writer_agent.py | hy3-free |
+| 2026-08-13 23:50 | Agent 实现 | 修复 HttpLLMClient 在同步门禁中多次 asyncio.run 复用已关闭事件循环致 Event loop is closed：chat 改为每次调用新建 client（保留 async with 协议）| src/genesis/inference/client.py | hy3-free |

src/genesis/inference/client.py

@@ -50,6 +50,7 @@ class HttpLLMClient:
         self._api_key = api_key
         self._timeout_sec = timeout_sec
         self._retry_backoff = retry_backoff
+        self._transport = transport
         self._client = httpx.AsyncClient(timeout=timeout_sec, transport=transport)
 
     async def __aenter__(self) -> HttpLLMClient:
@@ -79,42 +80,46 @@ class HttpLLMClient:
             "Content-Type": "application/json",
         }
 
-        attempts = 1 + len(self._retry_backoff)
+        # 每次调用新建 httpx.AsyncClient，绑定到当前事件循环。
-        last_error: Exception | None = None
+        # 兼容「同步门禁中多次 asyncio.run 驱动 async chat_structured」场景，
-        for attempt in range(attempts):
+        # 避免复用 __init__ 中创建的 client 因首个循环关闭而报 Event loop is closed。
-            if attempt > 0:
+        async with httpx.AsyncClient(timeout=self._timeout_sec, transport=self._transport) as client:
-                await asyncio.sleep(self._retry_backoff[attempt - 1])
+            attempts = 1 + len(self._retry_backoff)
-            try:
+            last_error: Exception | None = None
-                resp = await self._client.post(url, json=payload, headers=headers)
+            for attempt in range(attempts):
-            except httpx.TimeoutException as exc:
+                if attempt > 0:
-                last_error = exc
+                    await asyncio.sleep(self._retry_backoff[attempt - 1])
-                continue
+                try:
-            except httpx.HTTPError as exc:
+                    resp = await client.post(url, json=payload, headers=headers)
-                last_error = exc
+                except httpx.TimeoutException as exc:
-                continue
+                    last_error = exc
+                    continue
+                except httpx.HTTPError as exc:
+                    last_error = exc
+                    continue
 
-            if resp.status_code >= 500:
+                if resp.status_code >= 500:
-                last_error = LLMNetworkError(f"LLM 5xx: {resp.status_code}")
+                    last_error = LLMNetworkError(f"LLM 5xx: {resp.status_code}")
-                continue
+                    continue
-            if resp.status_code >= 400:
+                if resp.status_code >= 400:
-                raise LLMNetworkError(f"LLM HTTP {resp.status_code}: {resp.text[:200]}")
+                    raise LLMNetworkError(f"LLM HTTP {resp.status_code}: {resp.text[:200]}")
-            if not (200 <= resp.status_code < 300):
+                if not (200 <= resp.status_code < 300):
-                # 3xx 重定向不自动跟随，不得误判为成功
+                    # 3xx 重定向不自动跟随，不得误判为成功
-                raise LLMNetworkError(f"LLM HTTP {resp.status_code}: {resp.text[:200]}")
+                    raise LLMNetworkError(f"LLM HTTP {resp.status_code}: {resp.text[:200]}")
 
-            try:
+                try:
-                data = resp.json()
+                    data = resp.json()
-                content = data["choices"][0]["message"]["content"]
+                    content = data["choices"][0]["message"]["content"]
-            except (json.JSONDecodeError, KeyError, IndexError, TypeError) as exc:
+                except (json.JSONDecodeError, KeyError, IndexError, TypeError) as exc:
-                # 2xx 但响应结构损坏（非 JSON / 缺字段）→ 结构化错误，不重试
+                    # 2xx 但响应结构损坏（非 JSON / 缺字段）→ 结构化错误，不重试
-                raise LLMResponseError(f"LLM 响应结构损坏: {exc}") from exc
+                    raise LLMResponseError(f"LLM 响应结构损坏: {exc}") from exc
-            usage_raw = data.get("usage", {})
+                usage_raw = data.get("usage", {})
-            usage = TokenUsage(
+                usage = TokenUsage(
-                input_tokens=usage_raw.get("prompt_tokens", 0),
+                    input_tokens=usage_raw.get("prompt_tokens", 0),
-                output_tokens=usage_raw.get("completion_tokens", 0),
+                    output_tokens=usage_raw.get("completion_tokens", 0),
-            )
+                )
-            return content, usage
+                return content, usage
 
-        if isinstance(last_error, httpx.TimeoutException):
+            if isinstance(last_error, httpx.TimeoutException):
-            raise LLMTimeoutError(f"LLM 超时（{self._timeout_sec}s）") from last_error
+                raise LLMTimeoutError(f"LLM 超时（{self._timeout_sec}s）") from last_error
-        raise LLMNetworkError(f"LLM 调用失败（重试耗尽）: {last_error}") from last_error
+            raise LLMNetworkError(f"LLM 调用失败（重试耗尽）: {last_error}") from last_error