落地并存测服全端：智能体 Graph/路由、后台切 8180、主账与部署脚本齐备，便于线上重发与验收。

Co-authored-by: Cursor <cursoragent@cursor.com>

inquiry-agent/agent/llm/http_client.py

@@ -0,0 +1,156 @@
+"""
+LLM 出站 HTTP 客户端（DeepSeek / 千问兼容 OpenAI Chat）。
+
+本文件职责：带超时的 chat.completions 调用；缺 key / 未确认数据用途时明确失败。
+禁止：在企微回调线程同步长等；正式长调用走 Worker 槽 + 队列。
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import httpx
+
+from agent.config import Settings, get_settings
+from agent.llm.modes import LlmMode
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class LlmHttpResult:
+    ok: bool
+    content: str = ""
+    raw: Optional[dict[str, Any]] = None
+    error: str = ""
+    provider: str = ""
+    tool_calls: list[dict[str, Any]] | None = None
+
+
+@dataclass
+class LlmHttpClient:
+    """
+    兼容 OpenAI 的 Chat Completions 客户端。
+
+    DeepSeek / 千问 DashScope compatible-mode 共用此壳。
+    """
+
+    settings: Settings
+    timeout_seconds: float = 60.0
+
+    @classmethod
+    def from_settings(cls, settings: Optional[Settings] = None) -> "LlmHttpClient":
+        return cls(settings=settings or get_settings())
+
+    def _provider_for_mode(self, mode: LlmMode | str) -> tuple[str, str, str]:
+        """
+        返回 (provider, base_url, api_key)。
+
+        A/B 文本 → DeepSeek；附件/视觉 → 千问。
+        """
+        mode_val = mode.value if isinstance(mode, LlmMode) else str(mode)
+        visionish = mode_val in {
+            LlmMode.ATTACHMENT_TEXT.value,
+            LlmMode.VISION.value,
+            LlmMode.SHEET_MAP.value,
+        }
+        if visionish:
+            return (
+                "qwen",
+                (self.settings.qwen_base_url or "").rstrip("/"),
+                self.settings.qwen_api_key or "",
+            )
+        return (
+            "deepseek",
+            (self.settings.deepseek_base_url or "").rstrip("/"),
+            self.settings.deepseek_api_key or "",
+        )
+
+    def _model_for(self, provider: str, mode: LlmMode | str) -> str:
+        mode_val = mode.value if isinstance(mode, LlmMode) else str(mode)
+        if provider == "qwen":
+            if mode_val == LlmMode.VISION.value:
+                return self.settings.qwen_vl_model or "qwen-vl-plus"
+            return self.settings.qwen_text_model or "qwen-plus"
+        return self.settings.deepseek_model or "deepseek-chat"
+
+    def chat(
+        self,
+        *,
+        mode: LlmMode | str,
+        messages: list[dict[str, Any]],
+        temperature: float = 0.0,
+        tools: Optional[list[dict[str, Any]]] = None,
+        tool_choice: Any = None,
+    ) -> LlmHttpResult:
+        """
+        发起一次 chat.completions（可带 tools）。
+
+        副作用：出站 HTTPS；超时由 timeout_seconds 限制。
+        未确认 LLM_DATA_USAGE_CONFIRMED 时拒绝调用（prompt/04）。
+        """
+        if not self.settings.llm_data_usage_confirmed:
+            return LlmHttpResult(
+                ok=False,
+                error="llm_data_usage_not_confirmed",
+                provider="none",
+            )
+        provider, base, key = self._provider_for_mode(mode)
+        if not key or not base:
+            return LlmHttpResult(
+                ok=False,
+                error=f"{provider}_not_configured",
+                provider=provider,
+            )
+        model = self._model_for(provider, mode)
+        url = f"{base}/chat/completions"
+        headers = {
+            "Authorization": f"Bearer {key}",
+            "Content-Type": "application/json",
+        }
+        body: dict[str, Any] = {
+            "model": model,
+            "messages": messages,
+            "temperature": temperature,
+        }
+        if tools:
+            body["tools"] = tools
+        if tool_choice is not None:
+            body["tool_choice"] = tool_choice
+        try:
+            with httpx.Client(timeout=self.timeout_seconds) as client:
+                resp = client.post(url, headers=headers, json=body)
+            data = resp.json() if resp.content else {}
+            if resp.status_code >= 400:
+                logger.warning(
+                    "llm.http 失败 provider=%s status=%s",
+                    provider,
+                    resp.status_code,
+                )
+                return LlmHttpResult(
+                    ok=False,
+                    error=f"http_{resp.status_code}",
+                    raw=data if isinstance(data, dict) else None,
+                    provider=provider,
+                )
+            choices = (data or {}).get("choices") or []
+            content = ""
+            tool_calls: list[dict[str, Any]] = []
+            if choices:
+                msg = (choices[0].get("message") or {})
+                content = str(msg.get("content") or "")
+                raw_calls = msg.get("tool_calls") or []
+                if isinstance(raw_calls, list):
+                    tool_calls = [c for c in raw_calls if isinstance(c, dict)]
+            return LlmHttpResult(
+                ok=True,
+                content=content,
+                raw=data if isinstance(data, dict) else None,
+                provider=provider,
+                tool_calls=tool_calls or None,
+            )
+        except httpx.HTTPError as exc:
+            logger.warning("llm.http 异常 provider=%s err=%s", provider, exc)
+            return LlmHttpResult(ok=False, error=str(exc), provider=provider)