beswarm 0.1.34__tar.gz → 0.1.35__tar.gz

{beswarm-0.1.34 → beswarm-0.1.35}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: beswarm
-Version: 0.1.34
+Version: 0.1.35
 Summary: MAS
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
@@ -19,8 +19,11 @@ Requires-Dist: networkx>=3.4.2
 Requires-Dist: numpy>=2.2.4
 Requires-Dist: pdfminer-six==20240706
 Requires-Dist: pillow>=11.2.1
+Requires-Dist: pip>=25.1.1
+Requires-Dist: pyautogui>=0.9.54
 Requires-Dist: pyexecjs>=1.5.1
 Requires-Dist: pygments>=2.19.1
+Requires-Dist: pyperclip>=1.9.0
 Requires-Dist: pytz>=2025.2
 Requires-Dist: requests>=2.32.3
 Requires-Dist: scipy>=1.15.2
@@ -67,6 +70,22 @@ cd ~/Downloads/GitHub/beswarm && docker run --rm \
 --goal "分析这个仓库 https://github.com/cloneofsimo/minRF"
 ```
 
+测试 docker 是否可以用 GPU：
+
+```
+docker run --gpus all -it --rm --entrypoint nvidia-smi yym68686/beswarm
+```
+
+服务器安装
+
+```
+pip install pipx
+pipx ensurepath
+source ~/.bashrc
+pipx install nvitop
+pip install beswarm -i https://pypi.tuna.tsinghua.edu.cn/simple
+```
+
 main.py
 
 ```python

{beswarm-0.1.34 → beswarm-0.1.35}/README.md

@@ -38,6 +38,22 @@ cd ~/Downloads/GitHub/beswarm && docker run --rm \
 --goal "分析这个仓库 https://github.com/cloneofsimo/minRF"
 ```
 
+测试 docker 是否可以用 GPU：
+
+```
+docker run --gpus all -it --rm --entrypoint nvidia-smi yym68686/beswarm
+```
+
+服务器安装
+
+```
+pip install pipx
+pipx ensurepath
+source ~/.bashrc
+pipx install nvitop
+pip install beswarm -i https://pypi.tuna.tsinghua.edu.cn/simple
+```
+
 main.py
 
 ```python

{beswarm-0.1.34 → beswarm-0.1.35}/beswarm/aient/src/aient/core/request.py

@@ -31,7 +31,10 @@ async def get_gemini_payload(request, engine, provider, api_key=None):
     model_dict = get_model_dict(provider)
     original_model = model_dict[request.model]
 
-    gemini_stream = "streamGenerateContent"
+    if request.stream:
+        gemini_stream = "streamGenerateContent"
+    else:
+        gemini_stream = "generateContent"
     url = provider['base_url']
     parsed_url = urllib.parse.urlparse(url)
     if "/v1beta" in parsed_url.path:
@@ -295,7 +298,10 @@ async def get_vertex_gemini_payload(request, engine, provider, api_key=None):
     if provider.get("project_id"):
         project_id = provider.get("project_id")
 
-    gemini_stream = "streamGenerateContent"
+    if request.stream:
+        gemini_stream = "streamGenerateContent"
+    else:
+        gemini_stream = "generateContent"
     model_dict = get_model_dict(provider)
     original_model = model_dict[request.model]
     search_tool = None
@@ -941,10 +947,6 @@ async def get_gpt_payload(request, engine, provider, api_key=None):
         payload.pop("tools", None)
         payload.pop("tool_choice", None)
 
-    # if "models.inference.ai.azure.com" in url:
-    #     payload["stream"] = False
-    #     payload.pop("stream_options", None)
-
     if "api.x.ai" in url:
         payload.pop("stream_options", None)
         payload.pop("presence_penalty", None)

{beswarm-0.1.34 → beswarm-0.1.35}/beswarm/aient/src/aient/core/response.py

@@ -35,6 +35,7 @@ async def fetch_gemini_response_stream(client, url, headers, payload, model):
         promptTokenCount = 0
         candidatesTokenCount = 0
         totalTokenCount = 0
+        parts_json = ""
         # line_index = 0
         # last_text_line = 0
         # if "thinking" in model:
@@ -63,24 +64,25 @@ async def fetch_gemini_response_stream(client, url, headers, payload, model):
                     json_data = parse_json_safely( "{" + line + "}")
                     totalTokenCount = json_data.get('totalTokenCount', 0)
 
-                # print(line)
-                if line and '\"text\": \"' in line and is_finish == False:
+                if (line and '"parts": [' in line or parts_json != "") and is_finish == False:
+                    parts_json += line
+                if parts_json != "" and line and '],' == line.strip():
+                    parts_json =  "{" + parts_json.strip().rstrip(",} ]}") + "}]}"
                     try:
-                        json_data = json.loads( "{" + line + "}")
-                        content = json_data.get('text', '')
-                        # content = content.replace("\n", "\n\n")
-                        # if last_text_line == 0 and is_thinking:
-                        #     content = "> " + content.lstrip()
-                        # if is_thinking:
-                        #     content = content.replace("\n", "\n> ")
-                        # if last_text_line == line_index - 3:
-                        #     is_thinking = False
-                        #     content = "\n\n\n" + content.lstrip()
-                        sse_string = await generate_sse_response(timestamp, model, content=content)
-                        yield sse_string
+                        json_data = json.loads(parts_json)
+
+                        content = safe_get(json_data, "parts", 0, "text", default="")
+
+                        is_thinking = safe_get(json_data, "parts", 0, "thought", default=False)
+                        if is_thinking:
+                            sse_string = await generate_sse_response(timestamp, model, reasoning_content=content)
+                            yield sse_string
+                        else:
+                            sse_string = await generate_sse_response(timestamp, model, content=content)
+                            yield sse_string
                     except json.JSONDecodeError:
-                        logger.error(f"无法解析JSON: {line}")
-                    # last_text_line = line_index
+                        logger.error(f"无法解析JSON: {parts_json}")
+                    parts_json = ""
 
                 if line and ('\"functionCall\": {' in line or revicing_function_call):
                     revicing_function_call = True
@@ -142,7 +144,7 @@ async def fetch_vertex_claude_response_stream(client, url, headers, payload, mod
 
                 if line and '\"text\": \"' in line and is_finish == False:
                     try:
-                        json_data = json.loads( "{" + line + "}")
+                        json_data = json.loads( "{" + line.strip().rstrip(",") + "}")
                         content = json_data.get('text', '')
                         sse_string = await generate_sse_response(timestamp, model, content=content)
                         yield sse_string
@@ -525,21 +527,28 @@ async def fetch_response(client, url, headers, payload, engine, model):
             parsed_data = ast.literal_eval(str(response_json))
         elif isinstance(response_json, list):
             parsed_data = response_json
+        elif isinstance(response_json, dict):
+            parsed_data = [response_json]
         else:
             logger.error(f"error fetch_response: Unknown response_json type: {type(response_json)}")
             parsed_data = response_json
         # print("parsed_data", json.dumps(parsed_data, indent=4, ensure_ascii=False))
         content = ""
+        reasoning_content = ""
         for item in parsed_data:
             chunk = safe_get(item, "candidates", 0, "content", "parts", 0, "text")
+            is_think = safe_get(item, "candidates", 0, "content", "parts", 0, "thought", default=False)
             # logger.info(f"chunk: {repr(chunk)}")
             if chunk:
-                content += chunk
+                if is_think:
+                    reasoning_content += chunk
+                else:
+                    content += chunk
 
         usage_metadata = safe_get(parsed_data, -1, "usageMetadata")
-        prompt_tokens = usage_metadata.get("promptTokenCount", 0)
-        candidates_tokens = usage_metadata.get("candidatesTokenCount", 0)
-        total_tokens = usage_metadata.get("totalTokenCount", 0)
+        prompt_tokens = safe_get(usage_metadata, "promptTokenCount", default=0)
+        candidates_tokens = safe_get(usage_metadata, "candidatesTokenCount", default=0)
+        total_tokens = safe_get(usage_metadata, "totalTokenCount", default=0)
 
         role = safe_get(parsed_data, -1, "candidates", 0, "content", "role")
         if role == "model":
@@ -552,7 +561,7 @@ async def fetch_response(client, url, headers, payload, engine, model):
         function_call_content = safe_get(parsed_data, -1, "candidates", 0, "content", "parts", 0, "functionCall", "args", default=None)
 
         timestamp = int(datetime.timestamp(datetime.now()))
-        yield await generate_no_stream_response(timestamp, model, content=content, tools_id=None, function_call_name=function_call_name, function_call_content=function_call_content, role=role, total_tokens=total_tokens, prompt_tokens=prompt_tokens, completion_tokens=candidates_tokens)
+        yield await generate_no_stream_response(timestamp, model, content=content, tools_id=None, function_call_name=function_call_name, function_call_content=function_call_content, role=role, total_tokens=total_tokens, prompt_tokens=prompt_tokens, completion_tokens=candidates_tokens, reasoning_content=reasoning_content)
 
     elif engine == "claude":
         response_json = response.json()

beswarm-0.1.35/beswarm/aient/src/aient/core/test/test_geminimask.py

@@ -0,0 +1,330 @@
+import httpx
+import base64
+import os # 用于处理 API 密钥
+import asyncio
+import json
+import re
+# python -m core.test.test_geminimask
+from ..utils import get_image_message, safe_get
+
+# --- 请替换为您的实际值 ---
+MODEL_NAME = "gemini-2.5-flash-preview-04-17"
+API_KEY = os.environ.get("GEMINI_API_KEY") # 从环境变量读取密钥
+IMAGE_PATH = os.environ.get("IMAGE_PATH")
+# --------------------------
+
+# 检查 API 密钥是否存在
+if not API_KEY:
+    raise ValueError("请设置 GOOGLE_API_KEY 环境变量或在代码中提供 API 密钥。")
+
+# 确定图片的 MIME 类型
+# 您可以根据文件扩展名进行猜测，或者使用更可靠的库如 python-magic
+if IMAGE_PATH.lower().endswith(".png"):
+    IMAGE_MIME_TYPE = "image/png"
+elif IMAGE_PATH.lower().endswith(".jpg") or IMAGE_PATH.lower().endswith(".jpeg"):
+    IMAGE_MIME_TYPE = "image/jpeg"
+# 添加其他您需要支持的图片类型
+else:
+    raise ValueError(f"不支持的图片格式: {IMAGE_PATH}")
+
+# 读取图片文件并进行 Base64 编码
+try:
+    with open(IMAGE_PATH, "rb") as image_file:
+        image_data = image_file.read()
+        base64_encoded_image = base64.b64encode(image_data).decode("utf-8")
+        # print(base64_encoded_image)
+except FileNotFoundError:
+    print(f"错误：找不到图片文件 '{IMAGE_PATH}'")
+    exit()
+
+# image_message = get_image_message(base64_encoded_image, "gemini")
+image_message = asyncio.run(get_image_message(f"data:{IMAGE_MIME_TYPE};base64," + base64_encoded_image, "gemini"))
+
+# 构建请求 URL
+url = f"https://generativelanguage.googleapis.com/v1beta/models/{MODEL_NAME}:generateContent"
+
+prompt = "Give the segmentation masks for the Search box. Output a JSON list of segmentation masks where each entry contains the 2D bounding box in \"box_2d\" and the mask in \"mask\"."
+# 定义查询参数 (API Key)
+params = {"key": API_KEY}
+
+# 定义请求头
+headers = {"Content-Type": "application/json"}
+
+# 定义请求体 (JSON payload)
+payload = {
+    "contents": [
+        {
+            "parts": [
+                {
+                    "text": prompt
+                },
+                image_message
+            ]
+        }
+    ],
+    "generationConfig": {"thinkingConfig": {"thinkingBudget": 0}},
+}
+
+# 发送 POST 请求
+try:
+    with httpx.Client() as client:
+        response = client.post(url, params=params, headers=headers, json=payload, timeout=60.0) # 增加超时时间
+        response.raise_for_status() # 如果状态码不是 2xx，则抛出异常
+
+    # 您可以在这里添加代码来解析 response.json() 并提取分割掩码
+    text = safe_get(response.json(), "candidates", 0, "content", "parts", 0, "text")
+    # print(text)
+    # 例如: segmentation_masks = response.json().get('candidates', [{}])[0].get('content', {}).get('parts', [{}])[0].get('text')
+
+except httpx.HTTPStatusError as exc:
+    print(f"HTTP 错误发生: {exc.response.status_code} - {exc.response.text}")
+except httpx.RequestError as exc:
+    print(f"发送请求时出错: {exc}")
+except Exception as e:
+    print(f"发生意外错误: {e}")
+
+
+regex_pattern = r'(\[\s*\{.*?\}\s*\])' # 匹配包含至少一个对象的数组
+
+# 使用 re.search 查找第一个匹配项，re.DOTALL 使点号能匹配换行符
+match = re.search(regex_pattern, text, re.DOTALL)
+
+if match:
+    # 提取匹配到的整个 JSON 数组字符串 (group 1 因为模式中有括号)
+    json_string = match.group(1)
+
+    try:
+        # 使用 json.loads() 解析字符串
+        parsed_data = json.loads(json_string)
+        # 使用 json.dumps 美化打印输出
+        print(json.dumps(parsed_data, indent=2, ensure_ascii=False))
+
+        # 例如，获取第一个元素的 label
+        if isinstance(parsed_data, list) and len(parsed_data) > 0:
+            first_item = parsed_data[0]
+            if isinstance(first_item, dict):
+                label = first_item.get('label')
+                print(f"\n第一个元素的 label 是: {label}")
+
+    except json.JSONDecodeError as e:
+        print(f"JSON 解析错误: {e}")
+        print(f"出错的字符串是: {json_string}")
+else:
+    print("在文本中未找到匹配的 JSON 数组。")
+
+
+import io
+from PIL import Image, ImageDraw, ImageFont # pip install Pillow
+
+def extract_box_and_mask_py(parsed_data):
+    """
+    从已解析的 JSON 数据中提取边界框 (box_2d) 和 Base64 编码的掩码 (mask) 数据。
+
+    Args:
+        parsed_data (list): 一个包含字典的列表，每个字典至少包含 'box_2d' 和 'mask' 键。
+                            例如: [{'box_2d': [y1, x1, y2, x2], 'mask': 'data:image/png;base64,...'}, ...]
+
+    Returns:
+        list: 一个包含字典的列表，每个字典包含 'box' (坐标列表)
+              和 'mask_base64' (Base64 字符串)。
+              例如: [{'box': [y1, x1, y2, x2], 'mask_base64': '...'}, ...]
+              坐标系假定为 0-1000 范围。
+    """
+    # 不再需要正则表达式
+    results = []
+    # 检查 parsed_data 是否为列表
+    if not isinstance(parsed_data, list):
+        print(f"Error: Input data is not a list. Received type: {type(parsed_data)}")
+        return results
+
+    for item in parsed_data:
+        if not isinstance(item, dict):
+            print(f"Skipping non-dictionary item in list: {item}")
+            continue
+
+        try:
+            box = item.get('box_2d')
+            mask_data_uri = item.get('mask')
+
+            # 检查 'box_2d' 和 'mask' 是否存在且不为 None
+            if box is None or mask_data_uri is None:
+                print(f"Skipping item due to missing 'box_2d' or 'mask': {item}")
+                continue
+
+            # 从 mask 数据 URI 中提取 Base64 部分
+            # 格式: "data:image/[^;]+;base64,..."
+            if isinstance(mask_data_uri, str) and mask_data_uri.startswith('data:image/') and ';base64,' in mask_data_uri:
+                mask_b64 = mask_data_uri.split(';base64,', 1)[1]
+            else:
+                print(f"Skipping item due to invalid mask format: {mask_data_uri}")
+                continue
+
+            # 验证 box 数据
+            if isinstance(box, list) and len(box) == 4 and all(isinstance(n, int) for n in box):
+                 results.append({"box": box, "mask_base64": mask_b64})
+            else:
+                 print(f"Skipping invalid box format: {box}")
+
+        # 捕捉可能的 KeyError 或其他在字典访问/处理中发生的错误
+        except Exception as e:
+            print(f"Error processing item: {item}. Error: {e}")
+
+    return results
+
+def display_image_with_bounding_boxes_and_masks_py(
+    original_image_path,
+    box_and_mask_data,
+    output_overlay_path="overlay_image.png",
+    output_compare_dir="comparison_outputs"
+):
+    """
+    在原始图像上绘制边界框和掩码，并生成裁剪区域与掩码的对比图。
+
+    Args:
+        original_image_path (str): 原始图像的文件路径。
+        box_and_mask_data (list): extract_box_and_mask_py 的输出列表。
+        output_overlay_path (str): 保存带有叠加效果的图像的路径。
+        output_compare_dir (str): 保存对比图像的目录路径。
+    """
+    try:
+        img_original = Image.open(original_image_path).convert("RGBA")
+        img_width, img_height = img_original.size
+    except FileNotFoundError:
+        print(f"Error: Original image not found at {original_image_path}")
+        return
+    except Exception as e:
+        print(f"Error opening original image: {e}")
+        return
+
+    # 创建一个副本用于绘制叠加效果
+    img_overlay = img_original.copy()
+    draw = ImageDraw.Draw(img_overlay, "RGBA") # 使用 RGBA 模式以支持透明度
+
+    # 定义颜色列表
+    colors_hex = ['#FF0000', '#00FF00', '#0000FF', '#FFFF00', '#FF00FF', '#00FFFF']
+    # 将十六进制颜色转换为 RGBA 元组 (用于绘制)
+    colors_rgba = []
+    for hex_color in colors_hex:
+        h = hex_color.lstrip('#')
+        rgb = tuple(int(h[i:i+2], 16) for i in (0, 2, 4))
+        colors_rgba.append(rgb + (255,)) # (R, G, B, Alpha) - 边框完全不透明
+
+    # 创建输出目录（如果不存在）
+    import os
+    os.makedirs(output_compare_dir, exist_ok=True)
+
+    print(f"Found {len(box_and_mask_data)} box/mask pairs to process.")
+
+    for i, data in enumerate(box_and_mask_data):
+        box_0_1000 = data['box'] # [ymin, xmin, ymax, xmax] in 0-1000 range
+        mask_b64 = data['mask_base64']
+        color_index = i % len(colors_rgba)
+        outline_color = colors_rgba[color_index]
+        # 叠加掩码时使用半透明颜色
+        mask_fill_color = outline_color[:3] + (int(255 * 0.7),) # 70% Alpha
+
+        # --- 1. 坐标转换与验证 ---
+        # 将 0-1000 坐标转换为图像像素坐标 (left, top, right, bottom)
+        # 假设 box 是 [ymin, xmin, ymax, xmax]
+        try:
+            ymin_norm, xmin_norm, ymax_norm, xmax_norm = [c / 1000.0 for c in box_0_1000]
+
+            left   = int(xmin_norm * img_width)
+            top    = int(ymin_norm * img_height)
+            right  = int(xmax_norm * img_width)
+            bottom = int(ymax_norm * img_height)
+
+            # 确保坐标在图像范围内且有效
+            left = max(0, left)
+            top = max(0, top)
+            right = min(img_width, right)
+            bottom = min(img_height, bottom)
+
+            box_width_px = right - left
+            box_height_px = bottom - top
+
+            if box_width_px <= 0 or box_height_px <= 0:
+                print(f"Skipping box {i+1} due to zero or negative dimensions after conversion.")
+                continue
+
+        except Exception as e:
+            print(f"Error processing coordinates for box {i+1}: {box_0_1000}. Error: {e}")
+            continue
+
+        print(f"Processing Box {i+1}: Pixels(L,T,R,B)=({left},{top},{right},{bottom}) Color={colors_hex[color_index]}")
+
+        # --- 2. 在叠加图像上绘制边界框 ---
+        try:
+            draw.rectangle([left, top, right, bottom], outline=outline_color, width=5)
+        except Exception as e:
+             print(f"Error drawing rectangle for box {i+1}: {e}")
+             continue
+
+        # --- 3. 处理并绘制掩码 ---
+        try:
+            # 解码 Base64 掩码数据
+            mask_bytes = base64.b64decode(mask_b64)
+            mask_img_raw = Image.open(io.BytesIO(mask_bytes)).convert("RGBA")
+
+            # 将掩码图像缩放到边界框的像素尺寸
+            mask_img_resized = mask_img_raw.resize((box_width_px, box_height_px), Image.Resampling.NEAREST)
+
+            # 创建一个纯色块，应用掩码的 alpha 通道
+            color_block = Image.new('RGBA', mask_img_resized.size, mask_fill_color)
+
+            # 将带有透明度的颜色块粘贴到叠加图像上，使用掩码的 alpha 通道作为粘贴蒙版
+            # mask_img_resized.split()[-1] 提取 alpha 通道
+            img_overlay.paste(color_block, (left, top), mask=mask_img_resized.split()[-1])
+
+        except base64.binascii.Error:
+             print(f"Error: Invalid Base64 data for mask {i+1}.")
+             continue
+        except Exception as e:
+             print(f"Error processing or drawing mask for box {i+1}: {e}")
+             continue
+
+        # --- 4. 生成对比图 ---
+        try:
+            # 从原始图像中裁剪出边界框区域
+            img_crop = img_original.crop((left, top, right, bottom))
+
+            # 准备掩码预览图（使用原始解码后的掩码，调整大小以匹配裁剪区域）
+            # 这里直接使用缩放后的 mask_img_resized 的 RGB 部分可能更直观
+            mask_preview = mask_img_resized.convert("RGB") # 转换为 RGB 以便保存为常见格式
+
+            # 保存裁剪图和掩码预览图
+            crop_filename = os.path.join(output_compare_dir, f"compare_{i+1}_crop.png")
+            mask_filename = os.path.join(output_compare_dir, f"compare_{i+1}_mask.png")
+            img_crop.save(crop_filename)
+            mask_preview.save(mask_filename)
+            print(f" - Saved comparison: {crop_filename}, {mask_filename}")
+
+        except Exception as e:
+            print(f"Error creating or saving comparison images for box {i+1}: {e}")
+
+    # --- 5. 保存最终的叠加图像 ---
+    try:
+        img_overlay.save(output_overlay_path)
+        print(f"\nOverlay image saved to: {output_overlay_path}")
+        print(f"Comparison images saved in: {output_compare_dir}")
+    except Exception as e:
+        print(f"Error saving the final overlay image: {e}")
+
+
+extracted_data = extract_box_and_mask_py(parsed_data)
+
+if extracted_data:
+    # 确保原始图像存在
+    import os
+    if os.path.exists(IMAGE_PATH):
+            display_image_with_bounding_boxes_and_masks_py(
+                IMAGE_PATH,
+                extracted_data,
+                output_overlay_path="python_overlay_output.png", # 输出带叠加效果的图片名
+                output_compare_dir="python_comparison_outputs" # 输出对比图的文件夹名
+            )
+    else:
+            print(f"Error: Cannot proceed with visualization, image file not found: {IMAGE_PATH}")
+            print("Please update the 'IMAGE_PATH' variable in the script.")
+else:
+    print("No valid box and mask data found in the response text.")

{beswarm-0.1.34 → beswarm-0.1.35}/beswarm/aient/src/aient/core/utils.py

@@ -70,7 +70,7 @@ def get_engine(provider, endpoint=None, original_model=""):
         engine = "gemini"
     elif parsed_url.netloc.rstrip('/').endswith('aiplatform.googleapis.com') or (parsed_url.netloc.rstrip('/').endswith('gateway.ai.cloudflare.com') and "google-vertex-ai" in parsed_url.path):
         engine = "vertex"
-    elif parsed_url.netloc.rstrip('/').endswith('openai.azure.com') or parsed_url.netloc.rstrip('/').endswith('services.ai.azure.com'):
+    elif parsed_url.netloc.rstrip('/').endswith('azure.com'):
         engine = "azure"
     elif parsed_url.netloc == 'api.cloudflare.com':
         engine = "cloudflare"
@@ -127,6 +127,9 @@ def get_engine(provider, endpoint=None, original_model=""):
         engine = "tts"
         stream = False
 
+    if "stream" in safe_get(provider, "preferences", "post_body_parameter_overrides", default={}):
+        stream = safe_get(provider, "preferences", "post_body_parameter_overrides", "stream")
+
     return engine, stream
 
 from httpx_socks import AsyncProxyTransport
@@ -484,9 +487,17 @@ async def generate_sse_response(timestamp, model, content=None, tools_id=None, f
 
     return sse_response
 
-async def generate_no_stream_response(timestamp, model, content=None, tools_id=None, function_call_name=None, function_call_content=None, role=None, total_tokens=0, prompt_tokens=0, completion_tokens=0):
+async def generate_no_stream_response(timestamp, model, content=None, tools_id=None, function_call_name=None, function_call_content=None, role=None, total_tokens=0, prompt_tokens=0, completion_tokens=0, reasoning_content=None):
     random.seed(timestamp)
     random_str = ''.join(random.choices(string.ascii_letters + string.digits, k=29))
+    message = {
+        "role": role,
+        "content": content,
+        "refusal": None
+    }
+    if reasoning_content:
+        message["reasoning_content"] = reasoning_content
+
     sample_data = {
         "id": f"chatcmpl-{random_str}",
         "object": "chat.completion",
@@ -495,11 +506,7 @@ async def generate_no_stream_response(timestamp, model, content=None, tools_id=N
         "choices": [
             {
                 "index": 0,
-                "message": {
-                    "role": role,
-                    "content": content,
-                    "refusal": None
-                },
+                "message": message,
                 "logprobs": None,
                 "finish_reason": "stop"
             }