feat: add persistent chat agent and indexed batch workflows

backend/wisegrade/llm.py

@@ -101,6 +101,36 @@ class DeepSeekClient:
                     time.sleep(min(2 ** attempt, 5))
         raise ModelError(f"模型请求或JSON校验失败（{last_error}），已重试{self.retries}次")
 
+    def tool_call(self, messages: list[dict], tools: list[dict], *, tool_choice="auto"):
+        """OpenAI-compatible native tool calls. Credentials and provider bodies never enter logs."""
+        body = {"model": self.model, "messages": messages, "tools": tools,
+                "tool_choice": tool_choice, "temperature": 0}
+        try:
+            with httpx.Client(timeout=min(self.timeout, 180), transport=self.transport) as client:
+                response = client.post(self.base_url + "/chat/completions", json=body,
+                                       headers={"Authorization": f"Bearer {self.api_key}"})
+            if response.status_code in (400, 404, 422):
+                raise ModelError("当前端点或模型不支持原生 tool_calls；请在模型设置中配置支持工具调用的 OpenAI 兼容端点和模型")
+            response.raise_for_status()
+            message = response.json()["choices"][0]["message"]
+            if not isinstance(message, dict) or not isinstance(message.get("tool_calls", []), list):
+                raise ModelError("模型返回格式不包含有效的 tool_calls 消息")
+            return {"role": "assistant", "content": message.get("content"), "tool_calls": message.get("tool_calls", [])}
+        except ModelError:
+            raise
+        except (httpx.HTTPError, KeyError, IndexError, ValueError, TypeError) as exc:
+            raise ModelError(f"对话模型请求失败（{type(exc).__name__}）；请检查端点、模型和网络") from exc
+
+    def verify_tool_support(self):
+        """Small capability probe without student or course content."""
+        probe = [{"type": "function", "function": {"name": "wisegrade_ping",
+                  "description": "Return a ping for compatibility verification.",
+                  "parameters": {"type": "object", "properties": {}, "additionalProperties": False}}}]
+        # Some compatible endpoints implement auto tool selection but reject "required".
+        response = self.tool_call([{"role": "user", "content": "For compatibility verification, call wisegrade_ping now and do not answer in text."}], probe)
+        if not any(call.get("function", {}).get("name") == "wisegrade_ping" for call in response["tool_calls"]):
+            raise ModelError("当前端点或模型未返回原生 tool_calls；请配置支持工具调用的端点和模型")
+
 
 def validate_score(result: ScoreResult, criterion: dict, chunks: list[dict]):
     if result.criterion_id != criterion["criterion_id"] or result.max_score != criterion["max_score"]: