Merge branch 'main' into dev_tao

agent/core/runner.py

@@ -1261,6 +1261,16 @@ agent(
 
             if self.trace_store:
                 await self.trace_store.add_message(assistant_msg)
+                # 记录模型使用
+                await self.trace_store.record_model_usage(
+                    trace_id=trace_id,
+                    sequence=sequence - 1,  # assistant_msg的sequence
+                    role="assistant",
+                    model=config.model,
+                    prompt_tokens=prompt_tokens,
+                    completion_tokens=completion_tokens,
+                    cache_read_tokens=cache_read_tokens or 0,
+                )
 
             yield assistant_msg
             head_seq = sequence
@@ -1334,31 +1344,34 @@ agent(
                             logger.info(f"[Knowledge Tracking] 记录保存的知识 ID: {knowledge_id}")
 
                     # --- 支持多模态工具反馈 ---
-                    # execute() 返回 dict{"text", "images", "metadata"} 或 str
-                    if isinstance(tool_result, dict):
-                        tool_result_text = tool_result.get("text", str(tool_result))
-
-                        # 处理图片
-                        if tool_result.get("images"):
-                            # 构建多模态消息格式
-                            tool_content_for_llm = [{"type": "text", "text": tool_result_text}]
-                            for img in tool_result["images"]:
-                                if img.get("type") == "base64" and img.get("data"):
-                                    media_type = img.get("media_type", "image/png")
-                                    tool_content_for_llm.append({
-                                        "type": "image_url",
-                                        "image_url": {
-                                            "url": f"data:{media_type};base64,{img['data']}"
-                                        }
-                                    })
-                            img_count = len(tool_content_for_llm) - 1  # 减去 text 块
-                            print(f"[Runner] 多模态工具反馈: tool={tool_name}, images={img_count}, text_len={len(tool_result_text)}")
-                        else:
-                            # 无图片，只有文本
-                            tool_content_for_llm = tool_result_text
+                    # execute() 返回 dict{"text","images","tool_usage"} 或 str
+                    # 统一为dict格式
+                    if isinstance(tool_result, str):
+                        tool_result = {"text": tool_result}
+
+                    tool_text = tool_result.get("text", str(tool_result))
+                    tool_images = tool_result.get("images", [])
+                    tool_usage = tool_result.get("tool_usage")  # 新增：提取tool_usage
+
+                    # 处理多模态消息
+                    if tool_images:
+                        tool_result_text = tool_text
+                        # 构建多模态消息格式
+                        tool_content_for_llm = [{"type": "text", "text": tool_text}]
+                        for img in tool_images:
+                            if img.get("type") == "base64" and img.get("data"):
+                                media_type = img.get("media_type", "image/png")
+                                tool_content_for_llm.append({
+                                    "type": "image_url",
+                                    "image_url": {
+                                        "url": f"data:{media_type};base64,{img['data']}"
+                                    }
+                                })
+                        img_count = len(tool_content_for_llm) - 1  # 减去 text 块
+                        print(f"[Runner] 多模态工具反馈: tool={tool_name}, images={img_count}, text_len={len(tool_result_text)}")
                     else:
-                        tool_result_text = str(tool_result)
-                        tool_content_for_llm = tool_result_text
+                        tool_result_text = tool_text
+                        tool_content_for_llm = tool_text
 
                     tool_msg = Message.create(
                         trace_id=trace_id,
@@ -1373,10 +1386,22 @@ agent(
 
                     if self.trace_store:
                         await self.trace_store.add_message(tool_msg)
+                        # 记录工具的模型使用
+                        if tool_usage:
+                            await self.trace_store.record_model_usage(
+                                trace_id=trace_id,
+                                sequence=sequence,
+                                role="tool",
+                                tool_name=tool_name,
+                                model=tool_usage.get("model"),
+                                prompt_tokens=tool_usage.get("prompt_tokens", 0),
+                                completion_tokens=tool_usage.get("completion_tokens", 0),
+                                cache_read_tokens=tool_usage.get("cache_read_tokens", 0),
+                            )
                         # 截图单独存为同名 PNG 文件
-                        if isinstance(tool_result, dict) and tool_result.get("images"):
+                        if tool_images:
                             import base64 as b64mod
-                            for img in tool_result["images"]:
+                            for img in tool_images:
                                 if img.get("data"):
                                     png_path = self.trace_store._get_messages_dir(trace_id) / f"{tool_msg.message_id}.png"
                                     png_path.write_bytes(b64mod.b64decode(img["data"]))

agent/tools/builtin/browser/baseClass.py

@@ -226,7 +226,11 @@ async def create_container(url: str, account_name: str = "liuwenwu") -> Dict[str
 
 async def init_browser_session(
     browser_type: str = "local",
-    headless: bool = False,
+    # TEMPORARY FIX (2026-03-02): 改为 True 以解决 CDP 连接时序问题
+    # browser-use 在非 headless 模式下有时会在 Chrome 完全启动前尝试连接 CDP，
+    # 导致 "JSONDecodeError: Expecting value" 错误
+    # TODO: 之后改回 headless: bool = False，或在 browser-use 修复此问题后移除此注释
+    headless: bool = True,  # 原值: False
     url: Optional[str] = None,
     profile_name: str = "default",
     user_data_dir: Optional[str] = None,
@@ -290,6 +294,8 @@ async def init_browser_session(
 
     # 创建会话
     _browser_session = BrowserSession(**session_params)
+    # 添加短暂延迟，确保 Chrome CDP 端点完全就绪
+    await asyncio.sleep(1)
     await _browser_session.start()
 
     _browser_tools = Tools()

agent/tools/models.py

@@ -41,6 +41,9 @@ class ToolResult:
 	attachments: List[str] = field(default_factory=list)  # 文件路径列表
 	images: List[Dict[str, Any]] = field(default_factory=list)  # 图片列表
 
+	# Token追踪（用于工具内部LLM调用）
+	tool_usage: Optional[Dict[str, Any]] = None  # 格式：{"model": "...", "prompt_tokens": 100, "completion_tokens": 50, "cost": 0.0}
+
 	def to_llm_message(self, first_time: bool = True) -> str:
 		"""
 		转换为给 LLM 的消息

agent/tools/registry.py

@@ -241,11 +241,20 @@ class ToolRegistry:
 			# 处理 ToolResult 对象
 			from agent.tools.models import ToolResult
 			if isinstance(result, ToolResult):
-				# 有图片时返回 dict 以便 runner 构建多模态消息
+				ret = {"text": result.to_llm_message()}
+
+				# 保留images
 				if result.images:
-					return {"text": result.to_llm_message(), "images": result.images, "metadata": result.metadata}
-				# 无图片时也返回 dict，保留 metadata
-				return {"text": result.to_llm_message(), "metadata": result.metadata}
+					ret["images"] = result.images
+
+				# 保留tool_usage
+				if result.tool_usage:
+					ret["tool_usage"] = result.tool_usage
+
+				# 向后兼容：只有text时返回字符串
+				if len(ret) == 1:
+					return ret["text"]
+				return ret
 
 			return json.dumps(result, ensure_ascii=False, indent=2)
 

agent/trace/store.py

@@ -61,6 +61,10 @@ class FileSystemTraceStore:
         """获取 events.jsonl 文件路径"""
         return self._get_trace_dir(trace_id) / "events.jsonl"
 
+    def _get_model_usage_file(self, trace_id: str) -> Path:
+        """获取 model_usage.json 文件路径"""
+        return self._get_trace_dir(trace_id) / "model_usage.json"
+
     # ===== Trace 操作 =====
 
     async def create_trace(self, trace: Trace) -> str:
@@ -573,6 +577,104 @@ class FileSystemTraceStore:
 
         return abandoned_ids
 
+    # ===== 模型使用追踪 =====
+
+    async def record_model_usage(
+        self,
+        trace_id: str,
+        sequence: int,
+        role: str,
+        model: str,
+        prompt_tokens: int,
+        completion_tokens: int,
+        cache_read_tokens: int = 0,
+        tool_name: Optional[str] = None,
+    ) -> None:
+        """
+        记录模型使用情况到 model_usage.json
+
+        Args:
+            trace_id: Trace ID
+            sequence: 消息序号
+            role: 角色（assistant/tool）
+            model: 模型名称
+            prompt_tokens: 输入tokens
+            completion_tokens: 输出tokens
+            cache_read_tokens: 缓存读取tokens
+            tool_name: 工具名称（role=tool时）
+        """
+        usage_file = self._get_model_usage_file(trace_id)
+
+        # 读取现有数据
+        if usage_file.exists():
+            data = json.loads(usage_file.read_text())
+        else:
+            data = {
+                "summary": {
+                    "total_models": 0,
+                    "total_tokens": 0,
+                    "total_cache_read_tokens": 0,
+                    "agent_tokens": 0,
+                    "tool_tokens": 0,
+                },
+                "models": [],
+                "timeline": [],
+            }
+
+        # 更新summary
+        total_tokens = prompt_tokens + completion_tokens
+        data["summary"]["total_tokens"] += total_tokens
+        data["summary"]["total_cache_read_tokens"] += cache_read_tokens
+
+        if role == "assistant":
+            data["summary"]["agent_tokens"] += total_tokens
+            source = "agent"
+        else:
+            data["summary"]["tool_tokens"] += total_tokens
+            source = f"tool:{tool_name}" if tool_name else "tool"
+
+        # 更新models列表
+        model_entry = None
+        for m in data["models"]:
+            if m["model"] == model and m["source"] == source:
+                model_entry = m
+                break
+
+        if model_entry:
+            model_entry["prompt_tokens"] += prompt_tokens
+            model_entry["completion_tokens"] += completion_tokens
+            model_entry["total_tokens"] += total_tokens
+            model_entry["cache_read_tokens"] += cache_read_tokens
+            model_entry["call_count"] += 1
+        else:
+            data["models"].append({
+                "model": model,
+                "source": source,
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": total_tokens,
+                "cache_read_tokens": cache_read_tokens,
+                "call_count": 1,
+            })
+            data["summary"]["total_models"] = len(data["models"])
+
+        # 添加到timeline
+        timeline_entry = {
+            "sequence": sequence,
+            "role": role,
+            "model": model,
+            "prompt_tokens": prompt_tokens,
+            "completion_tokens": completion_tokens,
+        }
+        if cache_read_tokens > 0:
+            timeline_entry["cache_read_tokens"] = cache_read_tokens
+        if tool_name:
+            timeline_entry["tool_name"] = tool_name
+        data["timeline"].append(timeline_entry)
+
+        # 写回文件
+        usage_file.write_text(json.dumps(data, indent=2, ensure_ascii=False))
+
     # ===== 事件流操作（用于 WebSocket 断线续传）=====
 
     async def get_events(

examples/how/tool/nanobanana.py

@@ -1,9 +1,8 @@
 """
-NanoBanana Tool - 图像特征提取与图像生成
+NanoBanana Tool - 图像生成
 
-该工具可以提取图片中的特征，也可以根据描述生成图片。
-支持通过 OpenRouter 调用多模态模型，提取结构化的图像特征并保存为 JSON，
-或基于输入图像生成新的图像。
+通用图像生成工具，可以接受自然语言描述和/或图像输入，生成新的图像。
+支持通过 OpenRouter 调用 Gemini 2.5 Flash Image 模型。
 """
 
 import base64
@@ -22,23 +21,10 @@ from agent.tools import tool, ToolResult
 OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
 DEFAULT_TIMEOUT = 120.0
 
-DEFAULT_EXTRACTION_PROMPT = (
-    "请从这张图像中提取跨场景相对稳定、可复用的视觉不变特征。"
-    "输出严格 JSON，字段包含：identity_features、pose_features、appearance_features、"
-    "material_features、style_features、uncertainty、notes。"
-    "每个字段给出简洁要点，避免臆测。"
-)
-
-DEFAULT_IMAGE_PROMPT = (
-    "基于输入图像生成一张保留主体身份与关键视觉特征的新图像。"
-    "保持人物核心特征一致，同时提升清晰度与可用性。"
-)
+DEFAULT_IMAGE_PROMPT = "根据输入生成图像。"
 
 DEFAULT_IMAGE_MODEL_CANDIDATES = [
     "google/gemini-2.5-flash-image",
-    # "google/gemini-3-pro-image-preview",
-    # "black-forest-labs/flux.2-flex",
-    # "black-forest-labs/flux.2-pro",
 ]
 
 
@@ -214,84 +200,76 @@ def _normalize_model_id(model_id: str) -> str:
     return m
 
 
-@tool(description="可以提取图片中的特征，也可以根据描述生成图片")
+@tool(description="通用图像生成工具，可以接受自然语言描述和/或图像输入，生成新的图像")
 async def nanobanana(
     image_path: str = "",
     image_paths: Optional[List[str]] = None,
-    output_file: Optional[str] = None,
     prompt: Optional[str] = None,
     model: Optional[str] = None,
     max_tokens: int = 1200,
-    generate_image: bool = False,
     image_output_path: Optional[str] = None,
 ) -> ToolResult:
     """
-    可以提取图片中的特征，也可以根据描述生成图片。
+    通用图像生成工具，可以接受自然语言描述和/或图像输入，生成新的图像。
 
     Args:
         image_path: 输入图片路径（单图模式，可选）
-        image_paths: 输入图片路径列表（多图整体模式，可选）
-        output_file: 输出 JSON 文件路径（可选，用于特征提取模式）
-        prompt: 自定义提取指令或生成描述（可选）
-        model: OpenRouter 模型名（可选，默认读取 NANOBANANA_MODEL 或使用 Gemini 视觉模型）
+        image_paths: 输入图片路径列表（多图模式，可选）
+        prompt: 自定义生成描述（可选，默认使用通用prompt）
+        model: OpenRouter 模型名（可选，默认使用 gemini-2.5-flash-image）
         max_tokens: 最大输出 token
-        generate_image: 是否生成图片（False=提取特征，True=生成图片）
-        image_output_path: 生成图片保存路径（generate_image=True 时可选）
+        image_output_path: 生成图片保存路径（可选）
 
     Returns:
-        ToolResult: 包含结构化特征和输出文件路径，或生成的图片路径
+        ToolResult: 包含生成的图片路径
     """
     raw_paths: List[str] = []
     if image_paths:
         raw_paths.extend(image_paths)
     if image_path:
         raw_paths.append(image_path)
-    if not raw_paths:
-        return ToolResult(
-            title="NanoBanana 提取失败",
-            output="",
-            error="未提供输入图片，请传入 image_path 或 image_paths",
-        )
 
-    # 去重并检查路径
-    unique_raw: List[str] = []
-    seen = set()
-    for p in raw_paths:
-        if p and p not in seen:
-            unique_raw.append(p)
-            seen.add(p)
-
-    input_paths: List[Path] = [Path(p) for p in unique_raw]
-    invalid = [str(p) for p in input_paths if (not p.exists() or not p.is_file())]
-    if invalid:
-        return ToolResult(
-            title="NanoBanana 提取失败",
-            output="",
-            error=f"以下图片不存在或不可读: {invalid}",
-        )
+    # 图像输入是可选的，但如果提供了就需要验证
+    input_paths: List[Path] = []
+    if raw_paths:
+        # 去重并检查路径
+        unique_raw: List[str] = []
+        seen = set()
+        for p in raw_paths:
+            if p and p not in seen:
+                unique_raw.append(p)
+                seen.add(p)
+
+        input_paths = [Path(p) for p in unique_raw]
+        invalid = [str(p) for p in input_paths if (not p.exists() or not p.is_file())]
+        if invalid:
+            return ToolResult(
+                title="NanoBanana 生成失败",
+                output="",
+                error=f"以下图片不存在或不可读: {invalid}",
+            )
 
     api_key = _resolve_api_key()
     if not api_key:
         return ToolResult(
-            title="NanoBanana 提取失败",
+            title="NanoBanana 生成失败",
             output="",
             error="未找到 OpenRouter API Key，请设置 OPENROUTER_API_KEY 或 OPEN_ROUTER_API_KEY",
         )
 
-    if generate_image:
-        user_prompt = prompt or DEFAULT_IMAGE_PROMPT
-    else:
-        chosen_model = model or os.getenv("NANOBANANA_MODEL") or "google/gemini-2.5-flash"
-        user_prompt = prompt or DEFAULT_EXTRACTION_PROMPT
+    user_prompt = prompt or DEFAULT_IMAGE_PROMPT
 
-    try:
-        image_data_urls = [_image_to_data_url(p) for p in input_paths]
-    except Exception as e:
-        return ToolResult(
-            title="NanoBanana 提取失败",
-            output="",
-            error=f"图片编码失败: {e}",
-        )
+    # 编码图像（如果有）
+    image_data_urls = []
+    if input_paths:
+        try:
+            image_data_urls = [_image_to_data_url(p) for p in input_paths]
+        except Exception as e:
+            return ToolResult(
+                title="NanoBanana 生成失败",
+                output="",
+                error=f"图片编码失败: {e}",
+            )
 
     user_content: List[Dict[str, Any]] = [{"type": "text", "text": user_prompt}]
     for u in image_data_urls:
@@ -301,11 +279,7 @@ async def nanobanana(
         "messages": [
             {
                 "role": "system",
-                "content": (
-                    "你是视觉助手。"
-                    "当任务为特征提取时输出 JSON 对象，不要输出 markdown。"
-                    "当任务为图像生成时请返回图像。"
-                ),
+                "content": "你是图像生成助手。请根据用户的描述和/或输入图像生成新的图像。",
             },
             {
                 "role": "user",
@@ -314,9 +288,8 @@ async def nanobanana(
         ],
         "temperature": 0.2,
         "max_tokens": max_tokens,
+        "modalities": ["image", "text"],
     }
-    if generate_image:
-        payload["modalities"] = ["image", "text"]
 
     headers = {
         "Authorization": f"Bearer {api_key}",
@@ -327,33 +300,28 @@ async def nanobanana(
 
     endpoint = f"{OPENROUTER_BASE_URL}/chat/completions"
 
-    # 图像生成模式：自动尝试多个可用模型，减少 404/invalid model 影响
-    if generate_image:
-        candidates: List[str] = []
-        if model:
-            candidates.append(_normalize_model_id(model))
-        if env_model := os.getenv("NANOBANANA_IMAGE_MODEL"):
-            candidates.append(_normalize_model_id(env_model))
-        candidates.extend([_normalize_model_id(x) for x in DEFAULT_IMAGE_MODEL_CANDIDATES])
-        # 去重并保持顺序
-        dedup: List[str] = []
-        seen = set()
-        for m in candidates:
-            if m and m not in seen:
-                dedup.append(m)
-                seen.add(m)
-        candidates = dedup
-    else:
-        candidates = [chosen_model]
+    # 自动尝试多个可用模型，减少 404/invalid model 影响
+    candidates: List[str] = []
+    if model:
+        candidates.append(_normalize_model_id(model))
+    if env_model := os.getenv("NANOBANANA_IMAGE_MODEL"):
+        candidates.append(_normalize_model_id(env_model))
+    candidates.extend([_normalize_model_id(x) for x in DEFAULT_IMAGE_MODEL_CANDIDATES])
+    # 去重并保持顺序
+    dedup: List[str] = []
+    seen = set()
+    for m in candidates:
+        if m and m not in seen:
+            dedup.append(m)
+            seen.add(m)
+    candidates = dedup
 
     data: Optional[Dict[str, Any]] = None
     used_model: Optional[str] = None
     errors: List[Dict[str, Any]] = []
 
     for cand in candidates:
-        modality_attempts: List[Optional[List[str]]] = [None]
-        if generate_image:
-            modality_attempts = [["image", "text"], ["image"], None]
+        modality_attempts: List[Optional[List[str]]] = [["image", "text"], ["image"], None]
 
         for mods in modality_attempts:
             trial_payload = dict(payload)
@@ -392,9 +360,8 @@ async def nanobanana(
             break
 
     if data is None:
-        title = "NanoBanana 生成失败" if generate_image else "NanoBanana 提取失败"
         return ToolResult(
-            title=title,
+            title="NanoBanana 生成失败",
             output=json.dumps({"attempted_models": candidates, "errors": errors}, ensure_ascii=False, indent=2),
             long_term_memory="All candidate models failed for this request",
             metadata={"attempted_models": candidates, "errors": errors},
@@ -405,168 +372,115 @@ async def nanobanana(
     choices = data.get("choices") or []
     message = choices[0].get("message", {}) if choices else {}
 
-    # 图像生成分支
-    if generate_image:
-        refs = _extract_image_refs(choices[0] if choices else {}, message)
-        if not refs:
-            content = message.get("content")
-            preview = ""
-            if isinstance(content, str):
-                preview = content[:500]
-            elif isinstance(content, list):
-                preview = json.dumps(content[:3], ensure_ascii=False)[:500]
+    # 提取生成的图像
+    refs = _extract_image_refs(choices[0] if choices else {}, message)
+    if not refs:
+        content = message.get("content")
+        preview = ""
+        if isinstance(content, str):
+            preview = content[:500]
+        elif isinstance(content, list):
+            preview = json.dumps(content[:3], ensure_ascii=False)[:500]
 
-            return ToolResult(
-                title="NanoBanana 生成失败",
-                output=json.dumps(data, ensure_ascii=False, indent=2),
-                error="模型未返回可解析图片（未在 message.images/choice.images/content 中发现图片）",
-                metadata={
-                    "model": chosen_model,
-                    "choice_keys": list((choices[0] if choices else {}).keys()),
-                    "message_keys": list(message.keys()) if isinstance(message, dict) else [],
-                    "content_preview": preview,
-                },
-            )
+        return ToolResult(
+            title="NanoBanana 生成失败",
+            output=json.dumps(data, ensure_ascii=False, indent=2),
+            error="模型未返回可解析图片（未在 message.images/choice.images/content 中发现图片）",
+            metadata={
+                "model": chosen_model,
+                "choice_keys": list((choices[0] if choices else {}).keys()),
+                "message_keys": list(message.keys()) if isinstance(message, dict) else [],
+                "content_preview": preview,
+            },
+        )
 
-        output_paths: List[str] = []
-        if image_output_path:
-            base_path = Path(image_output_path)
+    output_paths: List[str] = []
+    if image_output_path:
+        base_path = Path(image_output_path)
+    else:
+        if len(input_paths) > 1:
+            base_path = input_paths[0].parent / "set_generated.png"
         else:
-            if len(input_paths) > 1:
-                base_path = input_paths[0].parent / "set_generated.png"
-            else:
-                base_path = input_paths[0].parent / f"{input_paths[0].stem}_generated.png"
-        base_path.parent.mkdir(parents=True, exist_ok=True)
-
-        for idx, ref in enumerate(refs):
-            kind = ref.get("kind", "")
-            mime_type = "image/png"
-            raw_bytes: Optional[bytes] = None
-
-            if kind == "data_url":
-                m = re.match(r"^data:([^;]+);base64,(.+)$", ref.get("value", ""), flags=re.DOTALL)
-                if not m:
-                    continue
-                mime_type = m.group(1)
-                raw_bytes = base64.b64decode(m.group(2))
-            elif kind == "base64":
-                mime_type = ref.get("mime_type", "image/png")
-                raw_bytes = base64.b64decode(ref.get("value", ""))
-            elif kind == "url":
-                url = ref.get("value", "")
-                try:
-                    with httpx.Client(timeout=DEFAULT_TIMEOUT) as client:
-                        r = client.get(url)
-                        r.raise_for_status()
-                        raw_bytes = r.content
-                        mime_type = r.headers.get("content-type", "image/png").split(";")[0]
-                except Exception:
-                    continue
-            else:
-                continue
+            base_path = input_paths[0].parent / f"{input_paths[0].stem}_generated.png"
+    base_path.parent.mkdir(parents=True, exist_ok=True)
 
-            if not raw_bytes:
-                continue
+    for idx, ref in enumerate(refs):
+        kind = ref.get("kind", "")
+        mime_type = "image/png"
+        raw_bytes: Optional[bytes] = None
 
-            ext = _mime_to_ext(mime_type)
-            if len(refs) == 1:
-                target = base_path
-                if target.suffix.lower() not in [".png", ".jpg", ".jpeg", ".webp"]:
-                    target = target.with_suffix(ext)
-            else:
-                stem = base_path.stem
-                target = base_path.with_name(f"{stem}_{idx+1}{ext}")
+        if kind == "data_url":
+            m = re.match(r"^data:([^;]+);base64,(.+)$", ref.get("value", ""), flags=re.DOTALL)
+            if not m:
+                continue
+            mime_type = m.group(1)
+            raw_bytes = base64.b64decode(m.group(2))
+        elif kind == "base64":
+            mime_type = ref.get("mime_type", "image/png")
+            raw_bytes = base64.b64decode(ref.get("value", ""))
+        elif kind == "url":
+            url = ref.get("value", "")
             try:
-                target.write_bytes(raw_bytes)
-                output_paths.append(str(target))
-            except Exception as e:
-                return ToolResult(
-                    title="NanoBanana 生成失败",
-                    output="",
-                    error=f"写入生成图片失败: {e}",
-                    metadata={"model": chosen_model},
-                )
-
-        if not output_paths:
+                with httpx.Client(timeout=DEFAULT_TIMEOUT) as client:
+                    r = client.get(url)
+                    r.raise_for_status()
+                    raw_bytes = r.content
+                    mime_type = r.headers.get("content-type", "image/png").split(";")[0]
+            except Exception:
+                continue
+        else:
+            continue
+
+        if not raw_bytes:
+            continue
+
+        ext = _mime_to_ext(mime_type)
+        if len(refs) == 1:
+            target = base_path
+            if target.suffix.lower() not in [".png", ".jpg", ".jpeg", ".webp"]:
+                target = target.with_suffix(ext)
+        else:
+            stem = base_path.stem
+            target = base_path.with_name(f"{stem}_{idx+1}{ext}")
+        try:
+            target.write_bytes(raw_bytes)
+            output_paths.append(str(target))
+        except Exception as e:
             return ToolResult(
                 title="NanoBanana 生成失败",
-                output=json.dumps(data, ensure_ascii=False, indent=2),
-                error="检测到图片引用但写入失败（可能是无效 base64 或 URL 不可访问）",
-                metadata={"model": chosen_model, "ref_count": len(refs)},
+                output="",
+                error=f"写入生成图片失败: {e}",
+                metadata={"model": chosen_model},
             )
 
-        usage = data.get("usage", {})
-        prompt_tokens = usage.get("prompt_tokens") or usage.get("input_tokens", 0)
-        completion_tokens = usage.get("completion_tokens") or usage.get("output_tokens", 0)
-        summary = {
-            "model": chosen_model,
-            "input_images": [str(p) for p in input_paths],
-            "input_count": len(input_paths),
-            "generated_images": output_paths,
-            "prompt_tokens": prompt_tokens,
-            "completion_tokens": completion_tokens,
-        }
-        return ToolResult(
-            title="NanoBanana 图片生成完成",
-            output=json.dumps({"summary": summary}, ensure_ascii=False, indent=2),
-            long_term_memory=f"Generated {len(output_paths)} image(s) from {len(input_paths)} input image(s) using {chosen_model}",
-            attachments=output_paths,
-            metadata=summary,
-        )
-
-    content = message.get("content") or ""
-    if not content:
+    if not output_paths:
         return ToolResult(
-            title="NanoBanana 提取失败",
+            title="NanoBanana 生成失败",
             output=json.dumps(data, ensure_ascii=False, indent=2),
-            error="模型未返回内容",
+            error="检测到图片引用但写入失败（可能是无效 base64 或 URL 不可访问）",
+            metadata={"model": chosen_model, "ref_count": len(refs)},
         )
 
-    try:
-        parsed = _safe_json_parse(content)
-    except Exception as e:
-        return ToolResult(
-            title="NanoBanana 提取失败",
-            output=content,
-            error=f"模型返回非 JSON 内容，解析失败: {e}",
-            metadata={"model": chosen_model},
-        )
-
-    if output_file:
-        out_path = Path(output_file)
-    else:
-        if len(input_paths) > 1:
-            out_path = input_paths[0].parent / "set_invariant_features.json"
-        else:
-            out_path = input_paths[0].parent / f"{input_paths[0].stem}_invariant_features.json"
-
-    out_path.parent.mkdir(parents=True, exist_ok=True)
-    out_path.write_text(json.dumps(parsed, ensure_ascii=False, indent=2), encoding="utf-8")
-
     usage = data.get("usage", {})
     prompt_tokens = usage.get("prompt_tokens") or usage.get("input_tokens", 0)
     completion_tokens = usage.get("completion_tokens") or usage.get("output_tokens", 0)
-
     summary = {
         "model": chosen_model,
         "input_images": [str(p) for p in input_paths],
         "input_count": len(input_paths),
-        "output_file": str(out_path),
+        "generated_images": output_paths,
         "prompt_tokens": prompt_tokens,
         "completion_tokens": completion_tokens,
     }
-
     return ToolResult(
-        title="NanoBanana 不变特征提取完成",
-        output=json.dumps(
-            {
-                "summary": summary,
-                "features": parsed,
-            },
-            ensure_ascii=False,
-            indent=2,
-        ),
-        long_term_memory=f"Extracted invariant features from {len(input_paths)} input image(s) using {chosen_model}",
-        attachments=[str(out_path)],
+        title="NanoBanana 图片生成完成",
+        output=json.dumps({"summary": summary}, ensure_ascii=False, indent=2),
+        long_term_memory=f"Generated {len(output_paths)} image(s) from {len(input_paths)} input image(s) using {chosen_model}",
+        attachments=output_paths,
         metadata=summary,
+        tool_usage={
+            "model": chosen_model,
+            "prompt_tokens": prompt_tokens,
+            "completion_tokens": completion_tokens,
+        }
     )