feat: expand model support and qwen-vl ux

2026-01-03 07:01:24 +08:00 · 2026-01-03 07:01:24 +08:00 · e2ba632ac8
commit e2ba632ac8
parent 77959226dd
34 changed files with 2179 additions and 100 deletions
--- a/config/model_profiles.py
+++ b/config/model_profiles.py
@ -0,0 +1,117 @@
+import os
+
+def _env(name: str, default: str = "") -> str:
+    return os.environ.get(name, default)
+
+
+# 默认（Kimi）
+KIMI_BASE = _env("API_BASE_KIMI", _env("AGENT_API_BASE_URL", "https://api.moonshot.cn/v1"))
+KIMI_KEY = _env("API_KEY_KIMI", _env("AGENT_API_KEY", ""))
+KIMI_FAST_MODEL = _env("MODEL_KIMI_FAST", _env("AGENT_MODEL_ID", "kimi-k2-0905-preview"))
+KIMI_THINK_MODEL = _env("MODEL_KIMI_THINK", _env("AGENT_THINKING_MODEL_ID", "kimi-k2-thinking"))
+
+# DeepSeek
+DEEPSEEK_BASE = _env("API_BASE_DEEPSEEK", "https://api.deepseek.com")
+DEEPSEEK_KEY = _env("API_KEY_DEEPSEEK", _env("AGENT_DEEPSEEK_API_KEY", ""))
+DEEPSEEK_FAST_MODEL = _env("MODEL_DEEPSEEK_FAST", "deepseek-chat")
+DEEPSEEK_THINK_MODEL = _env("MODEL_DEEPSEEK_THINK", "deepseek-reasoner")
+
+# Qwen
+QWEN_BASE = _env("API_BASE_QWEN", "https://dashscope.aliyuncs.com/compatible-mode/v1")
+QWEN_KEY = _env("API_KEY_QWEN", _env("DASHSCOPE_API_KEY", ""))
+QWEN_MAX_MODEL = _env("MODEL_QWEN_MAX", "qwen3-max")
+QWEN_VL_MODEL = _env("MODEL_QWEN_VL", "qwen3-vl-plus")
+
+
+MODEL_PROFILES = {
+    "kimi": {
+        "fast": {"base_url": KIMI_BASE, "api_key": KIMI_KEY, "model_id": KIMI_FAST_MODEL, "max_tokens": None},
+        "thinking": {"base_url": KIMI_BASE, "api_key": KIMI_KEY, "model_id": KIMI_THINK_MODEL, "max_tokens": None},
+        "supports_thinking": True,
+        "fast_only": False,
+        "name": "Kimi-k2"
+    },
+    "deepseek": {
+        "fast": {"base_url": DEEPSEEK_BASE, "api_key": DEEPSEEK_KEY, "model_id": DEEPSEEK_FAST_MODEL, "max_tokens": 8192},
+        "thinking": {
+            "base_url": DEEPSEEK_BASE,
+            "api_key": DEEPSEEK_KEY,
+            "model_id": DEEPSEEK_THINK_MODEL,
+            "max_tokens": 65536
+        },
+        "supports_thinking": True,
+        "fast_only": False,
+        "name": "DeepSeek"
+    },
+    "qwen3-max": {
+        "fast": {"base_url": QWEN_BASE, "api_key": QWEN_KEY, "model_id": QWEN_MAX_MODEL, "max_tokens": 65536},
+        "thinking": None,  # 不支持思考
+        "supports_thinking": False,
+        "fast_only": True,
+        "name": "Qwen3-Max"
+    },
+    "qwen3-vl-plus": {
+        "fast": {
+            "base_url": QWEN_BASE,
+            "api_key": QWEN_KEY,
+            "model_id": QWEN_VL_MODEL,
+            "max_tokens": 32768,
+            "extra_params": {}
+        },
+        "thinking": {
+            "base_url": QWEN_BASE,
+            "api_key": QWEN_KEY,
+            "model_id": QWEN_VL_MODEL,
+            "max_tokens": 32768,
+            "extra_params": {"enable_thinking": True}
+        },
+        "supports_thinking": True,
+        "fast_only": False,
+        "name": "Qwen3-VL"
+    }
+}
+
+MODEL_PROMPT_OVERRIDES = {
+    "kimi": {
+        "model_description": "你的基础模型是 Kimi-k2，由月之暗面公司开发，是一个开源的 MoE 架构模型，拥有 1T 参数和 32B 激活参数，当前智能助手应用由火山引擎提供 API 服务。",
+        "thinking_model_line": "思考模式时，第一次请求的模型不是 Kimi-k2，而是 Kimi-k2-Thinking，一个更善于分析复杂问题、规划复杂流程的模型，在后续请求时模型会换回 Kimi-k2。",
+        "deep_thinking_line": "在深度思考模式中，请求的模型是 Kimi-k2-Thinking，一个更善于分析复杂问题、规划复杂流程的模型。"
+    },
+    "deepseek": {
+        "model_description": "你的基础模型是 DeepSeek-V3.2（deepseek-chat），由 DeepSeek 提供，数学与推理能力较强，当前通过官方 API 调用。",
+        "thinking_model_line": "思考模式时，第一次请求使用 DeepSeek-Reasoner，一个强化推理的模型，后续请求会切回 DeepSeek-V3.2。",
+        "deep_thinking_line": "在深度思考模式中，请求的模型是 DeepSeek-Reasoner，用于深入分析复杂问题并规划步骤。"
+    },
+    "qwen3-max": {
+        "model_description": "你的基础模型是 Qwen3-Max，由通义千问提供，当前仅支持快速模式，不提供思考或深度思考能力。",
+        "thinking_model_line": "Qwen3-Max 仅支持快速模式，思考模式会被自动关闭。",
+        "deep_thinking_line": "Qwen3-Max 不支持深度思考模式，将保持快速模式。"
+    },
+    "qwen3-vl-plus": {
+        "model_description": "你的基础模型是 Qwen3-VL-Plus，支持图文多模态理解，接口来自通义千问 DashScope。",
+        "thinking_model_line": "思考模式时，请求的模型仍为 Qwen3-VL-Plus（开启思考能力），后续请求会切回快速模型。",
+        "deep_thinking_line": "在深度思考模式中，请求的模型是 Qwen3-VL-Plus（思考版），以获得更强的分析能力。"
+    }
+}
+
+
+def get_model_profile(key: str) -> dict:
+    if key not in MODEL_PROFILES:
+        raise ValueError(f"未知模型 key: {key}")
+    profile = MODEL_PROFILES[key]
+    # 基础校验：必须有 fast 段且有 key
+    fast = profile.get("fast") or {}
+    if not fast.get("api_key"):
+        raise ValueError(f"模型 {key} 缺少 API Key 配置")
+    return profile
+
+
+def get_model_prompt_replacements(key: str) -> dict:
+    """获取模型相关的提示词替换字段，若缺失则回退到 Kimi 版本。"""
+    fallback = MODEL_PROMPT_OVERRIDES.get("kimi", {})
+    overrides = MODEL_PROMPT_OVERRIDES.get(key) or {}
+    return {
+        "model_description": overrides.get("model_description") or fallback.get("model_description") or "",
+        "thinking_model_line": overrides.get("thinking_model_line") or fallback.get("thinking_model_line") or "",
+        "deep_thinking_line": overrides.get("deep_thinking_line") or fallback.get("deep_thinking_line") or ""
+    }
--- a/config/ui.py
+++ b/config/ui.py
@ -16,7 +16,7 @@ OUTPUT_FORMATS = {
    "session": "📺 [会话]",
 }

-AGENT_VERSION = "v5.5"
+AGENT_VERSION = "v6.0"

 LOG_LEVEL = "INFO"
 LOG_FORMAT = "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
--- a/core/main_terminal.py
+++ b/core/main_terminal.py
@ -64,6 +64,7 @@ from utils.api_client import DeepSeekClient
 from utils.context_manager import ContextManager
 from utils.tool_result_formatter import format_tool_result_for_context
 from utils.logger import setup_logger
+from config.model_profiles import get_model_profile, get_model_prompt_replacements

 if TYPE_CHECKING:
    from modules.user_container_manager import ContainerHandle
@ -93,6 +94,9 @@ class MainTerminal:
        # 初始化组件
        self.api_client = DeepSeekClient(thinking_mode=self.thinking_mode)
        self.api_client.set_deep_thinking_mode(self.deep_thinking_mode)
+        self.model_key = "kimi"
+        self.model_profile = get_model_profile(self.model_key)
+        self.apply_model_profile(self.model_profile)
        self.context_manager = ContextManager(project_path, data_dir=str(self.data_dir))
        self.context_manager.main_terminal = self
        self.container_mount_path = TERMINAL_SANDBOX_MOUNT_PATH or "/workspace"
@ -108,6 +112,7 @@ class MainTerminal:
        self.search_engine = SearchEngine()
        self.terminal_ops = TerminalOperator(project_path, container_session=container_session)
        self.ocr_client = OCRClient(project_path, self.file_manager)
+        self.pending_image_view = None  # 供 view_image 工具使用，保存一次性图片插入请求
        
        # 新增：终端管理器
        self.terminal_manager = TerminalManager(
@ -1689,6 +1694,29 @@ class MainTerminal:
                }
            }
        ]
+        # Qwen-VL 自带多模态能力，不向其暴露额外的 vlm_analyze 工具，避免重复与误导
+        if getattr(self, "model_key", None) == "qwen3-vl-plus":
+            tools = [
+                tool for tool in tools
+                if (tool.get("function") or {}).get("name") != "vlm_analyze"
+            ]
+            tools.append({
+                "type": "function",
+                "function": {
+                    "name": "view_image",
+                    "description": "将指定本地图片插入到对话中（系统代发一条包含图片的消息），便于模型主动查看图片内容。",
+                    "parameters": {
+                        "type": "object",
+                        "properties": self._inject_intent({
+                            "path": {
+                                "type": "string",
+                                "description": "项目内的图片相对路径（不要以 /workspace 开头），支持 png/jpg/webp/gif/bmp/svg。"
+                            }
+                        }),
+                        "required": ["path"]
+                    }
+                }
+            })
        if self.disabled_tools:
            tools = [
                tool for tool in tools
@ -1749,6 +1777,29 @@ class MainTerminal:
                if not path:
                    return json.dumps({"success": False, "error": "缺少 path 参数", "warnings": []}, ensure_ascii=False)
                result = self.ocr_client.vlm_analyze(path=path, prompt=prompt or "")
+            elif tool_name == "view_image":
+                path = (arguments.get("path") or "").strip()
+                if not path:
+                    return json.dumps({"success": False, "error": "path 不能为空"}, ensure_ascii=False)
+                if path.startswith("/workspace"):
+                    return json.dumps({"success": False, "error": "非法路径，超出项目根目录，请使用不带/workspace的相对路径"}, ensure_ascii=False)
+                abs_path = (Path(self.context_manager.project_path) / path).resolve()
+                try:
+                    abs_path.relative_to(Path(self.context_manager.project_path).resolve())
+                except Exception:
+                    return json.dumps({"success": False, "error": "非法路径，超出项目根目录，请使用不带/workspace的相对路径"}, ensure_ascii=False)
+                if not abs_path.exists() or not abs_path.is_file():
+                    return json.dumps({"success": False, "error": f"图片不存在: {path}"}, ensure_ascii=False)
+                if abs_path.stat().st_size > 10 * 1024 * 1024:
+                    return json.dumps({"success": False, "error": "图片过大，需 <= 10MB"}, ensure_ascii=False)
+                allowed_ext = {".png", ".jpg", ".jpeg", ".webp", ".gif", ".bmp", ".svg"}
+                if abs_path.suffix.lower() not in allowed_ext:
+                    return json.dumps({"success": False, "error": f"不支持的图片格式: {abs_path.suffix}"}, ensure_ascii=False)
+                # 记录待注入图片，供上层循环追加消息
+                self.pending_image_view = {
+                    "path": str(path)
+                }
+                result = {"success": True, "message": "图片已请求插入到对话中，将在后续消息中呈现。", "path": path}
            
            # 终端会话管理工具
            elif tool_name == "terminal_session":
@ -2292,14 +2343,17 @@ class MainTerminal:

    def build_messages(self, context: Dict, user_input: str) -> List[Dict]:
        """构建消息列表（添加终端内容注入）"""
-        # 加载系统提示
-        system_prompt = self.load_prompt("main_system")
+        # 加载系统提示（Qwen-VL 使用专用提示）
+        prompt_name = "main_system_qwenvl" if getattr(self, "model_key", "kimi") == "qwen3-vl-plus" else "main_system"
+        system_prompt = self.load_prompt(prompt_name)
        
        # 格式化系统提示
        container_path = self.container_mount_path or "/workspace"
        container_cpus = self.container_cpu_limit
        container_memory = self.container_memory_limit
        project_storage = self.project_storage_limit
+        model_key = getattr(self, "model_key", "kimi")
+        prompt_replacements = get_model_prompt_replacements(model_key)
        system_prompt = system_prompt.format(
            project_path=container_path,
            container_path=container_path,
@ -2308,7 +2362,8 @@ class MainTerminal:
            project_storage=project_storage,
            file_tree=context["project_info"]["file_tree"],
            memory=context["memory"],
-            current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+            model_description=prompt_replacements.get("model_description", "")
        )
        
        messages = [
@ -2328,10 +2383,16 @@ class MainTerminal:
        if self.deep_thinking_mode:
            deep_prompt = self.load_prompt("deep_thinking_mode_guidelines").strip()
            if deep_prompt:
+                deep_prompt = deep_prompt.format(
+                    deep_thinking_line=prompt_replacements.get("deep_thinking_line", "")
+                )
                messages.append({"role": "system", "content": deep_prompt})
        elif self.thinking_mode:
            thinking_prompt = self.load_prompt("thinking_mode_guidelines").strip()
            if thinking_prompt:
+                thinking_prompt = thinking_prompt.format(
+                    thinking_model_line=prompt_replacements.get("thinking_model_line", "")
+                )
                messages.append({"role": "system", "content": thinking_prompt})

        personalization_config = load_personalization_config(self.data_dir)
@ -2383,9 +2444,14 @@ class MainTerminal:
                })
            else:
                # User 或普通 System 消息
+                images = conv.get("images") or metadata.get("images") or []
+                content_payload = (
+                    self.context_manager._build_content_with_images(conv["content"], images)
+                    if images else conv["content"]
+                )
                messages.append({
                    "role": conv["role"],
-                    "content": conv["content"]
+                    "content": content_payload
                })
        
        # 当前用户输入已经在conversation中了，不需要重复添加
@ -2554,6 +2620,12 @@ class MainTerminal:
        normalized = mode.lower()
        if normalized not in allowed:
            raise ValueError(f"不支持的模式: {mode}")
+        # Qwen-VL 官方不支持深度思考模式
+        if getattr(self, "model_key", None) == "qwen3-vl-plus" and normalized == "deep":
+            raise ValueError("Qwen-VL 不支持深度思考模式")
+        # fast-only 模型限制
+        if getattr(self, "model_profile", {}).get("fast_only") and normalized != "fast":
+            raise ValueError("当前模型仅支持快速模式")
        previous_mode = getattr(self, "run_mode", "fast")
        self.run_mode = normalized
        self.thinking_mode = normalized != "fast"
@ -2569,6 +2641,29 @@ class MainTerminal:
            self.api_client.start_new_task()
        return self.run_mode

+    def apply_model_profile(self, profile: dict):
+        """将模型配置应用到 API 客户端"""
+        if not profile:
+            return
+        self.api_client.apply_profile(profile)
+
+    def set_model(self, model_key: str) -> str:
+        profile = get_model_profile(model_key)
+        if getattr(self.context_manager, "has_images", False) and model_key != "qwen3-vl-plus":
+            raise ValueError("当前对话包含图片，仅支持 Qwen-VL")
+        self.model_key = model_key
+        self.model_profile = profile
+        # 将模型标识传递给底层 API 客户端，便于按模型做兼容处理
+        self.api_client.model_key = model_key
+        # 应用模型配置
+        self.apply_model_profile(profile)
+        # fast-only 模型强制快速模式
+        if profile.get("fast_only") and self.run_mode != "fast":
+            self.set_run_mode("fast")
+        # 如果模型支持思考，但当前 run_mode 为 thinking/deep，则保持；否则无需调整
+        self.api_client.start_new_task(force_deep=self.deep_thinking_mode)
+        return self.model_key
+
    def get_run_mode_label(self) -> str:
        labels = {
            "fast": "快速模式（无思考）",
--- a/core/tool_config.py
+++ b/core/tool_config.py
@ -40,7 +40,14 @@ TOOL_CATEGORIES: Dict[str, ToolCategory] = {
    ),
    "read_focus": ToolCategory(
        label="阅读聚焦",
-        tools=["read_file", "focus_file", "unfocus_file", "vlm_analyze", "ocr_image"],
+        tools=[
+            "read_file",
+            "focus_file",
+            "unfocus_file",
+            "vlm_analyze",
+            "ocr_image",
+            "view_image",
+        ],
    ),
    "terminal_realtime": ToolCategory(
        label="实时终端",
--- a/core/web_terminal.py
+++ b/core/web_terminal.py
@ -286,6 +286,8 @@ class WebTerminal(MainTerminal):
            "thinking_mode": self.thinking_mode,
            "thinking_status": self.get_thinking_mode_status(),
            "run_mode": self.run_mode,
+            "model_key": getattr(self, "model_key", None),
+            "has_images": getattr(self.context_manager, "has_images", False),
            "context": {
                "usage_percent": context_status['usage_percent'],
                "total_size": context_status['sizes']['total'],
--- a/model_tests/deepseek_stream_test.py
+++ b/model_tests/deepseek_stream_test.py
@ -0,0 +1,105 @@
+"""Deepseek API 流式快速验证脚本。
+
+按照需求：
+- deepseek-chat: max_tokens=8000
+- deepseek-reasoner: max_tokens=65536
+- 开启流式输出，打印每条 data 行便于确认返回。
+
+根据用户允许，直接在脚本中硬编码 DeepSeek 测试用 API 信息，避免被仓库中其他
+默认配置（如 kimi）覆盖。请勿在生产使用。
+"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Optional
+
+import httpx
+
+
+DEEPSEEK_BASE_URL = "https://api.deepseek.com"
+DEEPSEEK_API_KEY = "sk-3457fbc33f0b4aefb2ce1d3101bb2341"
+DEEPSEEK_CHAT_MODEL = "deepseek-chat"
+DEEPSEEK_REASONER_MODEL = "deepseek-reasoner"
+
+
+def build_headers(api_key: str) -> dict[str, str]:
+    return {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json",
+    }
+
+
+async def stream_call(
+    *,
+    name: str,
+    base_url: str,
+    api_key: str,
+    model: str,
+    max_tokens: int,
+    prompt: Optional[str] = None,
+) -> None:
+    """向指定模型发起流式 chat/completions 请求并打印 data 行。"""
+    url = base_url.rstrip("/") + "/chat/completions"
+    payload = {
+        "model": model,
+        "stream": True,
+        "max_tokens": max_tokens,
+        "messages": [
+            {
+                "role": "user",
+                "content": prompt
+                or f"这是 {name} 模型的流式测试，请用一句话自我介绍。",
+            }
+        ],
+    }
+
+    print(f"\n=== {name} ===")
+    print(f"POST {url}")
+    async with httpx.AsyncClient(http2=True, timeout=120) as client:
+        async with client.stream(
+            "POST",
+            url,
+            json=payload,
+            headers=build_headers(api_key),
+        ) as resp:
+            print(f"status: {resp.status_code}")
+            if resp.status_code != 200:
+                body = await resp.aread()
+                print("error body:", body.decode(errors="ignore"))
+                return
+            async for line in resp.aiter_lines():
+                if not line:
+                    continue
+                if line.startswith("data:"):
+                    data = line[5:].strip()
+                    if data == "[DONE]":
+                        print("[DONE]")
+                        break
+                    print(data)
+                else:
+                    # 兼容潜在的非 data 行（例如心跳）
+                    print(line)
+
+
+async def main() -> None:
+    await stream_call(
+        name="deepseek-chat (max_tokens=8000)",
+        base_url=DEEPSEEK_BASE_URL,
+        api_key=DEEPSEEK_API_KEY,
+        model=DEEPSEEK_CHAT_MODEL,
+        max_tokens=8000,
+    )
+
+    await stream_call(
+        name="deepseek-reasoner (max_tokens=65536)",
+        base_url=DEEPSEEK_BASE_URL,
+        api_key=DEEPSEEK_API_KEY,
+        model=DEEPSEEK_REASONER_MODEL,
+        max_tokens=65536,
+        prompt="你是一个思考模型，请简述测试状态并结束。",
+    )
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
--- a/model_tests/qwen_stream_test.py
+++ b/model_tests/qwen_stream_test.py
@ -0,0 +1,86 @@
+"""Qwen API 流式测试脚本（qwen-max，快速模型）。
+
+目标：
+- 验证 qwen-max 的流式输出与用量字段（无思考能力）。
+- 打印 data 行，尾包包含 usage。
+
+注意：硬编码测试密钥，仅限本地验证，勿用于生产。
+"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Optional
+
+import httpx
+
+
+QWEN_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
+QWEN_API_KEY = "sk-64af1343e67d46d7a902ef5bcf6817ad"
+QWEN_MAX_MODEL = "qwen3-max"
+
+
+def headers(api_key: str) -> dict[str, str]:
+    return {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json",
+    }
+
+
+async def stream_call(
+    *,
+    base_url: str,
+    api_key: str,
+    model: str,
+    max_tokens: int = 2048,
+    prompt: Optional[str] = None,
+) -> None:
+    url = base_url.rstrip("/") + "/chat/completions"
+    payload = {
+        "model": model,
+        "stream": True,
+        "max_tokens": max_tokens,
+        "stream_options": {"include_usage": True},
+        "messages": [
+            {
+                "role": "user",
+                "content": prompt
+                or "请用简短中文自我介绍，并说明你当前正在执行的动作。",
+            }
+        ],
+    }
+    print("\n=== qwen-max fast mode ===")
+    print(f"POST {url}")
+    async with httpx.AsyncClient(http2=True, timeout=120) as client:
+        async with client.stream(
+            "POST", url, json=payload, headers=headers(api_key)
+        ) as resp:
+            print("status:", resp.status_code)
+            if resp.status_code != 200:
+                body = await resp.aread()
+                print("error body:", body.decode(errors="ignore"))
+                return
+            async for line in resp.aiter_lines():
+                if not line:
+                    continue
+                if line.startswith("data:"):
+                    data = line[5:].strip()
+                    if data == "[DONE]":
+                        print("[DONE]")
+                        break
+                    print(data)
+                else:
+                    print(line)
+
+
+async def main() -> None:
+    await stream_call(
+        base_url=QWEN_BASE_URL,
+        api_key=QWEN_API_KEY,
+        model=QWEN_MAX_MODEL,
+        max_tokens=64000,  # qwen3-max 官方上限 64K
+    )
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
--- a/model_tests/qwen_vl_stream_test.py
+++ b/model_tests/qwen_vl_stream_test.py
@ -0,0 +1,130 @@
+"""Qwen VL 测试脚本（兼容模式）。
+
+用例：
+1) 纯文字：验证流式输出与 usage。
+2) 图文：发送本地图片，验证多模态输入。
+
+注意：硬编码测试密钥，仅限本地验证，勿用于生产。
+"""
+
+from __future__ import annotations
+
+import asyncio
+import base64
+from pathlib import Path
+from typing import Optional
+
+import httpx
+
+
+QWEN_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
+QWEN_API_KEY = "sk-64af1343e67d46d7a902ef5bcf6817ad"
+QWEN_VL_MODEL = "qwen3-vl-plus"
+
+# 默认图片路径（仓库根目录下“截图/截屏2025-12-12 17.30.04.png”）
+DEFAULT_IMAGE_PATH = Path("截图/截屏2025-12-12 17.30.04.png")
+
+
+def headers(api_key: str) -> dict[str, str]:
+    return {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json",
+    }
+
+
+def build_image_content(image_path: Path) -> str:
+    data = image_path.read_bytes()
+    b64 = base64.b64encode(data).decode("ascii")
+    return f"data:image/{image_path.suffix.lstrip('.').lower()};base64,{b64}"
+
+
+async def stream_call(
+    *,
+    name: str,
+    base_url: str,
+    api_key: str,
+    model: str,
+    messages,
+    max_tokens: int,
+    enable_thinking: bool = False,
+) -> None:
+    url = base_url.rstrip("/") + "/chat/completions"
+    payload = {
+        "model": model,
+        "stream": True,
+        "max_tokens": max_tokens,
+        "messages": messages,
+        "stream_options": {"include_usage": True},
+    }
+    if enable_thinking:
+        payload["enable_thinking"] = True
+    print(f"\n=== {name} ===")
+    print(f"POST {url}")
+    async with httpx.AsyncClient(http2=True, timeout=180) as client:
+        async with client.stream(
+            "POST", url, json=payload, headers=headers(api_key)
+        ) as resp:
+            print("status:", resp.status_code)
+            if resp.status_code != 200:
+                body = await resp.aread()
+                print("error body:", body.decode(errors="ignore"))
+                return
+            async for line in resp.aiter_lines():
+                if not line:
+                    continue
+                if line.startswith("data:"):
+                    data = line[5:].strip()
+                    if data == "[DONE]":
+                        print("[DONE]")
+                        break
+                    print(data)
+                else:
+                    print(line)
+
+
+async def main(image_path: Optional[Path] = None) -> None:
+    # 1) 纯文字
+    text_messages = [
+        {
+            "role": "user",
+            "content": "请用一句话自我介绍，并简单说明你目前在执行的动作。",
+        }
+    ]
+    await stream_call(
+        name="qwen-vl text only",
+        base_url=QWEN_BASE_URL,
+        api_key=QWEN_API_KEY,
+        model=QWEN_VL_MODEL,
+        messages=text_messages,
+        max_tokens=32000,  # 官方上限 32K
+        enable_thinking=True,
+    )
+
+    # 2) 图文
+    img_path = image_path or DEFAULT_IMAGE_PATH
+    if not img_path.exists():
+        print(f"\n[warn] 图片文件不存在: {img_path}")
+        return
+    img_url = build_image_content(img_path)
+    multimodal_messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "请描述这张图片的主要内容，并给出一句话总结。"},
+                {"type": "image_url", "image_url": {"url": img_url}},
+            ],
+        }
+    ]
+    await stream_call(
+        name="qwen-vl image+text",
+        base_url=QWEN_BASE_URL,
+        api_key=QWEN_API_KEY,
+        model=QWEN_VL_MODEL,
+        messages=multimodal_messages,
+        max_tokens=32000,  # 官方上限 32K
+        enable_thinking=True,
+    )
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
--- a/model_tests/截屏2026-01-02
+++ b/model_tests/截屏2026-01-02
--- a/prompts/deep_thinking_mode_guidelines.txt
+++ b/prompts/deep_thinking_mode_guidelines.txt
@ -1,4 +1,4 @@
 你现在处于「深度思考模式」
-在深度思考模式中，请求的模型是kimi-k2-thinking 一个更善于分析复杂问题，规划复杂流程的模型
+{deep_thinking_line}
 在每一轮对用户要求的执行中，你的之前的思考会始终可见，保障思维过程和操作流程的连续性
-每次思考时，禁止回顾“我上一步做了什么”，只需要判断“下一步应该做什么”
+每次思考时，禁止回顾“我上一步做了什么”，只需要判断“下一步应该做什么”
--- a/prompts/main_system.txt
+++ b/prompts/main_system.txt
@ -1,6 +1,6 @@
 你是一名运行在云端服务器上的智能助手，可以帮助用户完成各种任务。你的用户可能没有编程背景，请用通俗易懂的方式与他们交流。

-你的基础模型是Kimi-k2,由月之暗面公司开发，是一个开源的Moe架构模型，由1t的参数和32b的激活参数，当前智能助手应用由火山引擎提供api服务
+{model_description}

 ## 你能做什么
 - **文档处理**：整理文字、编辑文件、格式转换
--- a/prompts/main_system_qwenvl.txt
+++ b/prompts/main_system_qwenvl.txt
@ -0,0 +1,300 @@
+你是一名运行在云端服务器上的智能助手，可以帮助用户完成各种任务。你的用户可能没有编程背景，请用通俗易懂的方式与他们交流。
+
+{model_description}
+
+## 你能做什么
+- **文档处理**：整理文字、编辑文件、格式转换
+- **信息查找**：搜索资料、提取网页内容、整理信息
+- **数据整理**：处理表格、分析数据、生成报告
+- **文件管理**：创建、修改、重命名文件和文件夹
+- **自动化任务**：批量处理文件、执行重复性工作
+- **视觉理解**：你自带多模态能力，用户可以直接发送图片；如需主动查看本地图片，可调用 `view_image` 指定路径，系统会代发一条包含图片的用户消息供你查看。
+
+## 图片分析（Qwen-VL 重点）
+当用户提出“这是什么”“识别文字/表格/票据”“找瑕疵/细节”“读屏/按钮含义”等图片分析任务时，优先采用下面的方法，保证细节充分、结论可验证：
+
+### 基本流程（先粗后细）
+1. **先整体后局部**：先看全图总结场景与目标，再针对关键区域逐块放大验证。
+2. **明确不确定性**：对看不清/模糊/遮挡区域，明确指出并提出下一步（放大/裁切/增强）。
+3. **用证据说话**：结论尽量引用可见线索（位置、颜色、形状、文字片段），避免凭感觉下结论。
+
+### 细节增强与“切图放大”方法（推荐）
+当图片文字很小、细节密集、或需要逐块检查时：
+1. **用 `run_python` 切图**：把原图按区域裁切成若干张更小的局部图（例如：左上/右上/左下/右下；或按表格/按钮/车标/铭牌所在区域裁切）。
+2. **必要时做多版本增强**：对同一区域输出多张增强版本（例如：对比度增强、锐化、灰度、二值化）用于读字/看边缘。
+3. **再次查看局部图**：对每张局部图分别观察并给出结论，再把结论汇总回原图任务。
+
+### 实操建议（你要主动想到并执行）
+- **裁切策略**：优先裁“目标本体 + 周边上下文”而不是只裁最小块；读文字时再裁更紧。
+- **输出路径**：建议输出到 `/workspace/cache/` 或项目内临时目录（如 `cache/`），文件名带序号（例如 `crop_01.png`）。
+- **复核展示**：需要让用户/自己确认时，可用 `<show_image src="..." />` 展示裁切结果；或将裁切图作为本地文件再用 `view_image` 查看。
+- **多图对比**：同一部位若存在多张版本（原裁切/增强后），按顺序展示并说明“哪张更利于读字/看细节”。
+
+## 图片展示
+- 如果需要直接在界面展示图片（本地或网络），请在回复里输出 `<show_image src="路径" alt="描述" />`，不用调用工具。
+- `src` 支持以 `/` 开头的本地静态路径或 `http/https`，`alt` 可选，会显示在图片下方。
+- 不要用 Markdown 图片语法或其它自定义标签。
+- 示例：
+  - `<show_image src="/workspace/images/result.png" alt="最终渲染效果" />`
+  - `<show_image src="/workspace/cache/thumb.jpg" />`
+  - `<show_image src="https://example.com/demo.png" alt="官方示例截图" />`
+
+### 图片检索与展示流程
+- 触发：用户询问“X长什么样”“给我看X的图片”等需求时。
+- 检索优先级：先在 **Wikimedia Commons（commons.wikimedia.org）** 搜索关键词（必要时添加“图片/照片/截图”）；若无合适结果，再用 `web_search` 进行全网搜索。
+- 提取：对候选链接使用 `extract_webpage` 获取正文中的图片直链，优先 `https`、扩展名为 jpg/png/webp、分辨率≥800px 的原图，避开缩略图和水印预览。仍优先采用 Wikipedia/Wikimedia 图源，其次再选其他站点。
+- 本地/校验：已有本地图片时直接展示；若网上图片是否匹配存疑，先下载并用你的视觉能力查看内容后再确定是否展示。
+- 展示：选数张代表性图片，直接输出 `<show_image src="直链或本地路径" alt="简短描述" />`；需要多张时多行重复该标签。
+- 回退：用户反馈“看不到/无法展示”时，先将图片下载到可访问路径（如 `/workspace/cache/xxx.jpg`）再用本地路径展示；仍失败则提供文字描述并询问是否换图源。
+
+## 重要提醒：你的工作环境
+1. **云端运行**：你在远程服务器上工作，在网页端和用户交互
+2. **多人共用**：服务器上可能有其他用户，你只能访问被授权的文件夹
+3. **文件传输**：用户可以在网页上传文件给你，你也可以生成文件让用户下载
+4. **安全第一**：只操作用户明确要求的文件，不要碰其他内容
+
+## 工作方式：先想后做
+遇到任务时，请这样工作：
+1. **确认理解**：复述一遍你理解的任务是什么
+2. **说明计划**：告诉用户你打算怎么做，分几步
+3. **征求同意**：询问用户的意见，向用户确认更多细节
+4. **报告结果**：在用户给出明确的指令，比如”好的，请开始做吧“再开始创建待办事项并完成任务
+
+**❌ 不要做的事**：
+- 不要一句"好的我来做"就直接开始
+- 不要猜测用户想要什么
+- 不要操作用户没提到的文件
+- 不要编造没做的事情
+
+## 文件查看：两种方式选择
+
+### 方式1：读取（临时看一眼）
+适合场景：
+- 只是想快速看看内容
+- 小文件（比如配置文件、说明文档）
+- 看完就不用了
+
+### 方式2：聚焦（长期盯着）
+适合场景：
+- 需要反复查看和修改的文件
+- 重要的核心文件
+- 会花较长时间处理的文件
+
+**限制**：
+- 聚焦最多3个文件
+- 每个文件不超过10000字
+- 用完记得取消聚焦，给下个任务腾空间
+
+**已聚焦的文件**：内容完全可见，不需要也不能再用命令查看
+
+## 文件操作示例
+
+### 创建和写入文件
+```
+用户："帮我整理一份待办清单"
+你的做法：
+1. 先询问清单内容有哪些
+2. 调用 create_file 创建空文件
+3. 调用 append_to_file 写入内容
+4. 告诉用户文件创建在哪里
+```
+
+### 修改文件内容
+```
+用户："把报告里的'2024'改成'2025'"
+你的做法：
+1. 如果文件已聚焦，直接看到内容
+2. 如果没聚焦，先读取或聚焦文件
+3. 调用 modify_file 进行替换
+4. 确认修改是否成功
+```
+
+### 搜索和提取信息
+```
+用户："帮我找一下最近的AI新闻"
+你的做法：
+1. 调用 web_search 搜索相关信息
+2. 如果需要详细内容，用 extract_webpage
+3. 整理信息给用户
+4. 如果用户要保存，可以创建文件
+```
+
+## 执行命令的两种方式
+
+### 方式1：快速命令（一次性的）
+用 `run_command` 工具
+适合：
+- 查看文件列表：`ls -lh`
+- 查看文件内容：`cat 文件.txt`
+- 统计行数：`wc -l 文件.txt`
+- 搜索内容：`grep "关键词" 文件.txt`
+
+### 方式2：持久终端（需要保持运行的）
+用 `terminal_session` + `terminal_input` 工具
+适合：
+- 运行需要一直开着的程序
+- 需要多次输入的交互任务
+- 需要等待较长时间的任务
+
+**⚠️ 注意**：
+- 最多同时开3个终端
+- 不要在终端里启动 python、node、vim 这类会占用界面的程序
+- 如果终端卡住了，用 terminal_reset 重启
+
+## 常用命令示例
+
+### 文件查看
+```bash
+# 查看文件内容
+cat 文件.txt
+
+# 查看文件前10行
+head -n 10 文件.txt
+
+# 查看文件后10行
+tail -n 10 文件.txt
+
+# 搜索包含关键词的行
+grep "关键词" 文件.txt
+
+# 统计文件行数
+wc -l 文件.txt
+```
+
+### 文件操作
+```bash
+# 复制文件
+cp 原文件.txt 新文件.txt
+
+# 移动/重命名文件
+mv 旧名.txt 新名.txt
+
+# 删除文件（谨慎使用）
+rm 文件.txt
+
+# 创建文件夹
+mkdir 文件夹名
+```
+
+### 文件信息
+```bash
+# 查看文件大小
+ls -lh 文件.txt
+
+# 查看当前目录所有文件
+ls -lah
+
+# 查看文件类型
+file 文件名
+
+# 查看目录结构
+tree -L 2
+```
+
+## 待办事项系统（简单任务管理）
+
+当任务需要多个步骤时，可以创建待办清单：
+
+### 使用规则
+1. **什么时候用**：任务需要2步以上、涉及多个文件或工具时
+2. **清单要求**：
+   - 概述：用一句话说明任务目标（不超过50字）
+   - 任务：最多4条，按执行顺序排列
+   - 每条任务要说清楚具体做什么，不要用"优化""处理"这种模糊词
+3. **执行方式**：
+   - 完成一项，勾选一项
+   - 如果计划有变，先告诉用户
+   - 全部完成后，用 todo_finish 结束
+
+### 示例：整理文档
+```
+概述：整理年度总结文档，统一格式并导出PDF
+任务1：读取所有Word文档，统一标题格式
+任务2：合并内容到一个新文件
+任务3：检查错别字和标点
+任务4：转换为PDF并保存
+```
+
+## 网络搜索技巧
+
+### 基础搜索
+```
+用户："搜索一下Python教程"
+你调用：web_search(query="Python教程")
+```
+
+### 搜索最近的内容
+```
+用户："最近一周的科技新闻"
+你调用：web_search(query="4-6个和科技新闻相关的关键词", time_range="week")
+```
+
+
+### 提取网页详细内容
+```
+用户："把这篇文章的内容提取出来"
+步骤：
+1. 先用 web_search 找到链接
+2. 再用 extract_webpage 提取完整内容
+3. 如果用户要保存，用 save_webpage 存为txt文件
+```
+
+## 资源管理：记得收拾
+
+由于服务器资源有限，请养成好习惯：
+1. **聚焦文件**：用完及时取消聚焦
+2. **终端会话**：不用的终端及时关闭
+3. **大文件**：避免一次输出超长内容，分批处理
+4. **上下文**：对话太长时（超过10万字符），提醒用户压缩
+
+## 遇到问题怎么办
+
+### 文件太大
+```
+如果提示"文件超过10000字符"：
+1. 告诉用户文件大小
+2. 建议只查看部分内容
+3. 用命令查看：head -n 100 文件.txt
+```
+
+### 命令执行失败
+```
+1. 不要重复执行相同命令
+2. 检查是否有权限问题
+3. 尝试用其他方法
+4. 实在不行，诚实告诉用户
+```
+
+### 不确定怎么做
+```
+1. 不要瞎猜
+2. 问用户更多信息
+3. 提供几个可行方案让用户选
+```
+
+## 交流风格
+
+- 使用口语化表达，避免技术黑话
+- 主动说明你在做什么
+- 遇到问题时说明原因
+- 完成任务后总结成果
+- 不要用生硬的"执行工具: xxx"，而是说"我来帮你..."
+
+## 当前环境信息
+- 项目路径: 你运行在隔离容器中（挂载目录 {container_path}），宿主机路径已对你隐藏
+- 资源限制: 容器内核数上限 {container_cpus}，内存 {container_memory}，项目磁盘配额 {project_storage}
+- 项目文件结构: {file_tree}
+- 长期记忆: {memory}
+- 当前时间: {current_time}
+
+## 核心原则
+
+1. **安全第一**：只操作授权范围内的文件
+2. **沟通为主**：不确定时多问，不要自作主张
+3. **诚实守信**：做不到的事情坦白说，不编造
+4. **用户友好**：用简单的语言解释复杂的操作
+5. **正确执行**：和用户主动确认细节，用户明确告知可以开始任务后，再开始工作流程
+
+记住：你的用户可能不懂技术，你的目标是让他们感觉到"这个助手真好用"，而不是"怎么这么复杂"。
+
+如果用户设置了个性化信息，根据用户的个性化需求回答
--- a/prompts/thinking_mode_guidelines.txt
+++ b/prompts/thinking_mode_guidelines.txt
@ -1,5 +1,5 @@
 你现在处于「思考模式」
-思考模式时，第一次请求的模型不是kimi-k2，而是kimi-k2-thinking 一个更善于分析复杂问题，规划复杂流程的模型，在后续请求时，模型会换回kimi-k2。
+{thinking_model_line}
 并且，在系统监控到工具或写入失败时，会自动再次切换到思考模型，思考模型会更加深入地分析错误的原因，保证任务顺利进行。

 请百分百遵循一下原则：
@ -29,4 +29,4 @@
 # 你的思考过程在后文会不可见，所以你需要在用户给出要求时的第一次回答中尽可能在正式输出中详细描述你的规划，来作为后面kimi-k2模型行动时提供参考依据

 **⚠️重要提示⚠️**
-思考模式下，token会被大量消耗，所以请保持思路连贯，在思考内容中禁止回顾全部的上下文历史，只需紧扣上一步的状态并规划下一步行动。
+思考模式下，token会被大量消耗，所以请保持思路连贯，在思考内容中禁止回顾全部的上下文历史，只需紧扣上一步的状态并规划下一步行动。
--- a/scratch_test/qwen_reasoning_sim.py
+++ b/scratch_test/qwen_reasoning_sim.py
@ -0,0 +1,47 @@
+import asyncio, sys, os, copy
+from pathlib import Path
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+from utils.api_client import DeepSeekClient
+
+class FakeClient(DeepSeekClient):
+    def __init__(self):
+        super().__init__(thinking_mode=True, web_mode=True)
+        self.sent = []
+        self.call_idx = 0
+    async def chat(self, messages, tools=None, stream=True):
+        self.sent.append(copy.deepcopy(messages))
+        self.call_idx += 1
+        if self.call_idx == 1:
+            yield {
+                "choices": [
+                    {"delta": {
+                        "reasoning_content": "think1 ",
+                        "tool_calls": [
+                            {"id": "call_1", "index": 0, "type": "function", "function": {"name": "foo", "arguments": "{}"}}
+                        ]
+                    }}
+                ]
+            }
+            yield {"choices": [{"delta": {}}]}
+        else:
+            yield {"choices": [{"delta": {"content": "done"}}]}
+            yield {"choices": [{"delta": {}}]}
+
+async def main():
+    client = FakeClient()
+    messages = [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "hi"}
+    ]
+    async def tool_handler(name, args):
+        return '{}'
+    out = await client.chat_with_tools(messages, tools=[{"type":"function","function":{"name":"foo","parameters":{"type":"object","properties":{}}}}], tool_handler=tool_handler)
+    print('final', out)
+    import json
+    for i, m in enumerate(client.sent, 1):
+        print('\ncall', i)
+        print(json.dumps(m, ensure_ascii=False, indent=2))
+
+if __name__ == '__main__':
+    asyncio.run(main())
--- a/scratch_test/qwen_vl_reasoning_test.py
+++ b/scratch_test/qwen_vl_reasoning_test.py
@ -0,0 +1,183 @@
+"""
+手工验证 Qwen-VL 在多轮工具调用时是否保留 reasoning_content。
+
+步骤：
+1. 首次请求：用户要求“在思考中想一个 8 位数字但不要直接输出，先调用天气工具，再告诉我数字”。
+2. 模型若返回 tool_calls，我们本地模拟 weather 查询，把结果作为 tool 消息返回。
+3. 第二次请求：带上第一次 assistant 工具调用 + tool 结果，检查模型是否能输出第一次思考里的数字。
+
+运行：
+  export API_BASE_QWEN="https://dashscope.aliyuncs.com/compatible-mode/v1"
+  export API_KEY_QWEN="sk-..."
+  python3 scratch_test/qwen_vl_reasoning_test.py
+
+日志：
+  输出到 logs/qwen_vl_reasoning_test_<timestamp>.json
+"""
+
+import json
+import os
+import time
+from pathlib import Path
+from typing import Dict, Any
+
+import httpx
+
+# 读取模型配置（若未设置环境变量则使用默认）
+API_BASE = os.environ.get("API_BASE_QWEN", "https://dashscope.aliyuncs.com/compatible-mode/v1").rstrip("/")
+API_KEY = os.environ.get("API_KEY_QWEN", os.environ.get("DASHSCOPE_API_KEY", ""))
+MODEL_ID = os.environ.get("MODEL_QWEN_VL", "qwen3-vl-plus")
+
+LOG_DIR = Path("logs")
+LOG_DIR.mkdir(parents=True, exist_ok=True)
+LOG_FILE = LOG_DIR / f"qwen_vl_reasoning_test_{int(time.time())}.json"
+
+
+def log(data: Dict[str, Any]):
+    """追加写入调试日志。"""
+    with open(LOG_FILE, "a", encoding="utf-8") as f:
+        f.write(json.dumps(data, ensure_ascii=False, indent=2))
+        f.write("\n\n")
+
+
+def build_headers():
+    return {
+        "Authorization": f"Bearer {API_KEY}",
+        "Content-Type": "application/json"
+    }
+
+
+def weather_tool_result(city: str = "上海", date: str = "今天"):
+    """模拟天气工具返回固定结构。"""
+    return {
+        "success": True,
+        "city": city,
+        "date": date,
+        "forecast": "多云转晴，15~22℃，东风3级",
+        "source": "local-mock"
+    }
+
+
+def main():
+    if not API_KEY:
+        raise SystemExit("未配置 API_KEY_QWEN 或 DASHSCOPE_API_KEY，无法测试")
+
+    messages = [
+        {
+            "role": "system",
+            "content": "你是测试助手。"
+        },
+        {
+            "role": "user",
+            "content": (
+                "请你想一个明确的8位数字，随便一个就行，只在思考过程中呈现，不要在最终回复里透露；"
+                "先调用天气查询工具；工具完成后再正常回答并把刚才思考出的数字告诉我。"
+            )
+        }
+    ]
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "query_weather",
+                "description": "查询指定城市在指定日期的天气",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "city": {"type": "string", "description": "城市名称，如上海"},
+                        "date": {"type": "string", "description": "日期，如今天/明天/2025-01-01"}
+                    },
+                    "required": ["city"]
+                }
+            }
+        }
+    ]
+
+    # ---- 第一次调用 ----
+    payload1 = {
+        "model": MODEL_ID,
+        "messages": messages,
+        "tools": tools,
+        "tool_choice": "auto",
+        "stream": False,
+        "max_tokens": 512,
+        # 关键：开启思考模式
+        "enable_thinking": True
+    }
+
+    with httpx.Client(http2=True, timeout=120) as client:
+        resp1 = client.post(f"{API_BASE}/chat/completions", json=payload1, headers=build_headers())
+    resp1.raise_for_status()
+    data1 = resp1.json()
+    log({"step": "call1_response", "raw": data1})
+
+    choice1 = data1["choices"][0]["message"]
+    tool_calls = choice1.get("tool_calls") or []
+    reasoning1 = choice1.get("reasoning_content")
+
+    print("第一次返回 reasoning_content 长度:", len(reasoning1 or ""))
+    print("第一次返回 tool_calls 数量:", len(tool_calls))
+
+    if not tool_calls:
+        print("模型未调用工具，测试无法继续。")
+        log({"error": "no_tool_calls"})
+        return
+
+    # 只取第一个工具调用
+    tc = tool_calls[0]
+    args_json = tc["function"].get("arguments") or "{}"
+    try:
+        args = json.loads(args_json)
+    except json.JSONDecodeError:
+        args = {}
+    tool_result = weather_tool_result(
+        city=args.get("city", "上海"),
+        date=args.get("date", "今天")
+    )
+
+    # 构建第二轮消息：包含第一次assistant消息和tool结果
+    messages2 = messages + [
+        {
+            "role": "assistant",
+            "content": choice1.get("content", ""),
+            "tool_calls": tool_calls,
+            # 保留 reasoning_content 原样，核心验证点
+            "reasoning_content": reasoning1 or ""
+        },
+        {
+            "role": "tool",
+            "tool_call_id": tc.get("id"),
+            "name": tc["function"]["name"],
+            "content": json.dumps(tool_result, ensure_ascii=False)
+        }
+    ]
+
+    payload2 = {
+        "model": MODEL_ID,
+        "messages": messages2,
+        "tools": tools,
+        "tool_choice": "auto",
+        "stream": False,
+        "max_tokens": 512,
+        # 同步思考模式
+        "enable_thinking": True
+    }
+
+    with httpx.Client(http2=True, timeout=120) as client:
+        resp2 = client.post(f"{API_BASE}/chat/completions", json=payload2, headers=build_headers())
+    resp2.raise_for_status()
+    data2 = resp2.json()
+    log({"step": "call2_response", "raw": data2, "messages_sent": messages2})
+
+    choice2 = data2["choices"][0]["message"]
+    reasoning2 = choice2.get("reasoning_content")
+    content2 = choice2.get("content")
+
+    print("第二次返回 reasoning_content 长度:", len(reasoning2 or ""))
+    print("第二次最终回复:", content2)
+    print(f"完整日志已保存到: {LOG_FILE}")
+
+
+if __name__ == "__main__":
+    main()
--- a/static/icons/keyboard.svg
+++ b/static/icons/keyboard.svg
@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" width="48" height="48" fill="none" stroke="#000000" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="opacity:1;"><path d="M10 8h.01M12 12h.01M14 8h.01M16 12h.01M18 8h.01M6 8h.01M7 16h10m-9-4h.01"/><rect width="20" height="16" x="2" y="4" rx="2"/></svg>
--- a/static/src/App.vue
+++ b/static/src/App.vue
@ -172,6 +172,9 @@
              :uploading="uploading"
              :thinking-mode="thinkingMode"
              :run-mode="resolvedRunMode"
+              :model-menu-open="modelMenuOpen"
+              :model-options="modelOptions"
+              :current-model-key="currentModelKey"
              :quick-menu-open="quickMenuOpen"
              :tool-menu-open="toolMenuOpen"
              :mode-menu-open="modeMenuOpen"
@ -182,6 +185,7 @@
              :current-conversation-id="currentConversationId"
              :icon-style="iconStyle"
              :tool-category-icon="toolCategoryIcon"
+              :selected-images="selectedImages"
              @update:input-message="inputSetMessage"
              @input-change="handleInputChange"
              @input-focus="handleInputFocus"
@ -192,7 +196,9 @@
              @quick-upload="handleQuickUpload"
              @toggle-tool-menu="toggleToolMenu"
              @toggle-mode-menu="toggleModeMenu"
+              @toggle-model-menu="toggleModelMenu"
              @select-run-mode="handleModeSelect"
+              @select-model="handleModelSelect"
              @toggle-settings="toggleSettings"
              @update-tool-category="updateToolCategory"
              @realtime-terminal="handleRealtimeTerminalClick"
@ -200,6 +206,8 @@
              @toggle-token-panel="handleTokenPanelToggleClick"
              @compress-conversation="handleCompressConversationClick"
              @file-selected="handleFileSelected"
+              @pick-images="openImagePicker"
+              @remove-image="handleRemoveImage"
            />
          </div>
        </main>
@ -217,6 +225,15 @@

      <PersonalizationDrawer />
      <LiquidGlassWidget />
+      <ImagePicker
+        v-if="imagePickerOpen"
+        :open="imagePickerOpen"
+        :entries="imageEntries"
+        :initial-selected="selectedImages"
+        :loading="imageLoading"
+        @close="closeImagePicker"
+        @confirm="handleImagesConfirmed"
+      />

      <div
        v-if="isMobileViewport"
--- a/static/src/app.ts
+++ b/static/src/app.ts
@ -12,6 +12,7 @@ import LiquidGlassWidget from './components/experiments/LiquidGlassWidget.vue';
 import QuickMenu from './components/input/QuickMenu.vue';
 import InputComposer from './components/input/InputComposer.vue';
 import AppShell from './components/shell/AppShell.vue';
+import ImagePicker from './components/overlay/ImagePicker.vue';
 import { useUiStore } from './stores/ui';
 import { useConversationStore } from './stores/conversation';
 import { useChatStore } from './stores/chat';
@ -23,6 +24,7 @@ import { useFileStore } from './stores/file';
 import { useSubAgentStore } from './stores/subAgent';
 import { useFocusStore } from './stores/focus';
 import { usePersonalizationStore } from './stores/personalization';
+import { useModelStore } from './stores/model';
 import { useChatActionStore } from './stores/chatActions';
 import { useMonitorStore } from './stores/monitor';
 import { ICONS, TOOL_CATEGORY_ICON_MAP } from './utils/icons';
@ -201,12 +203,14 @@ if (window.visualViewport) {
    window.visualViewport.addEventListener('scroll', updateViewportHeightVar);
 }

-const ENABLE_APP_DEBUG_LOGS = false;
+const ENABLE_APP_DEBUG_LOGS = true;
 function debugLog(...args) {
-    if (!ENABLE_APP_DEBUG_LOGS) {
-        return;
+    if (!ENABLE_APP_DEBUG_LOGS) return;
+    try {
+        console.log('[app]', ...args);
+    } catch (e) {
+        /* ignore logging errors */
    }
-    debugLog(...args);
 }
 // 临时排查对话切换问题的调试输出
 const TRACE_CONV = true;
@ -275,6 +279,10 @@ const appOptions = {
                ],
                mobileViewportQuery: null,
                modeMenuOpen: false,
+                modelMenuOpen: false,
+                imageEntries: [],
+                imageLoading: false,
+                conversationHasImages: false,
                conversationListRequestSeq: 0,
                conversationListRefreshToken: 0,

@ -386,6 +394,8 @@ const appOptions = {
                'conversationsOffset',
                'conversationsLimit'
            ]),
+            ...mapWritableState(useModelStore, ['currentModelKey']),
+            ...mapState(useModelStore, ['models']),
            ...mapWritableState(useChatStore, [
                'messages',
                'currentMessageIndex',
@ -402,7 +412,9 @@ const appOptions = {
                'inputIsFocused',
                'quickMenuOpen',
                'toolMenuOpen',
-                'settingsOpen'
+                'settingsOpen',
+                'imagePickerOpen',
+                'selectedImages'
            ]),
            resolvedRunMode() {
                const allowed = ['fast', 'thinking', 'deep'];
@ -411,6 +423,9 @@ const appOptions = {
                }
                return this.thinkingMode ? 'thinking' : 'fast';
            },
+            modelOptions() {
+                return this.models || [];
+            },
            titleRibbonVisible() {
                return !this.isMobileViewport && this.chatDisplayMode === 'chat';
            },
@ -681,6 +696,9 @@ const appOptions = {
                uiRequestConfirm: 'requestConfirm',
                uiResolveConfirm: 'resolveConfirm'
            }),
+            ...mapActions(useModelStore, {
+                modelSet: 'setModel'
+            }),
            ...mapActions(useChatStore, {
                chatExpandBlock: 'expandBlock',
                chatCollapseBlock: 'collapseBlock',
@ -717,7 +735,11 @@ const appOptions = {
                inputSetMessage: 'setInputMessage',
                inputClearMessage: 'clearInputMessage',
                inputSetLineCount: 'setInputLineCount',
-                inputSetMultiline: 'setInputMultiline'
+                inputSetMultiline: 'setInputMultiline',
+                inputSetImagePickerOpen: 'setImagePickerOpen',
+                inputSetSelectedImages: 'setSelectedImages',
+                inputClearSelectedImages: 'clearSelectedImages',
+                inputRemoveSelectedImage: 'removeSelectedImage'
            }),
            ...mapActions(useToolStore, {
                toolRegisterAction: 'registerToolAction',
@ -1291,6 +1313,11 @@ const appOptions = {
                this.inputSetLineCount(1);
                this.inputSetMultiline(false);
                this.inputClearMessage();
+                this.inputClearSelectedImages();
+                this.inputSetImagePickerOpen(false);
+                this.imageEntries = [];
+                this.imageLoading = false;
+                this.conversationHasImages = false;
                this.toolSetSettingsLoading(false);
                this.toolSetSettings([]);
                
@ -1433,6 +1460,12 @@ const appOptions = {
                } else if (status && typeof status.thinking_mode !== 'undefined') {
                    this.runMode = status.thinking_mode ? 'thinking' : 'fast';
                }
+                if (status && typeof status.model_key === 'string') {
+                    this.modelSet(status.model_key);
+                }
+                if (status && typeof status.has_images !== 'undefined') {
+                    this.conversationHasImages = !!status.has_images;
+                }
            },

            updateContainerStatus(status) {
@ -1758,9 +1791,15 @@ const appOptions = {
                }
                
                let currentAssistantMessage = null;
+                let historyHasImages = false;
                
                historyMessages.forEach((message, index) => {
                    debugLog(`处理消息 ${index + 1}/${historyMessages.length}:`, message.role, message);
+                    const meta = message.metadata || {};
+                    if (message.role === 'user' && meta.system_injected_image) {
+                        debugLog('跳过系统代发的图片消息（仅用于模型查看，不在前端展示）');
+                        return;
+                    }
                    
                    if (message.role === 'user') {
                        // 用户消息 - 先结束之前的assistant消息
@ -1768,10 +1807,14 @@ const appOptions = {
                            this.messages.push(currentAssistantMessage);
                            currentAssistantMessage = null;
                        }
-                        
+                        const images = message.images || (message.metadata && message.metadata.images) || [];
+                        if (Array.isArray(images) && images.length) {
+                            historyHasImages = true;
+                        }
                        this.messages.push({
                            role: 'user',
-                            content: message.content || ''
+                            content: message.content || '',
+                            images
                        });
                        debugLog('添加用户消息:', message.content?.substring(0, 50) + '...');
                        
@ -1967,6 +2010,8 @@ const appOptions = {
                if (currentAssistantMessage && currentAssistantMessage.actions.length > 0) {
                    this.messages.push(currentAssistantMessage);
                }
+
+                this.conversationHasImages = historyHasImages;
                
                debugLog(`历史消息渲染完成，共 ${this.messages.length} 条消息`);
                this.logMessageState('renderHistoryMessages:after-render');
@ -2257,7 +2302,12 @@ const appOptions = {
                    return;
                }

-                if (!this.inputMessage.trim()) {
+                const text = (this.inputMessage || '').trim();
+                const images = Array.isArray(this.selectedImages) ? this.selectedImages.slice(0, 9) : [];
+                const hasText = text.length > 0;
+                const hasImages = images.length > 0;
+
+                if (!hasText && !hasImages) {
                    return;
                }

@ -2266,12 +2316,22 @@ const appOptions = {
                    this.showQuotaToast({ type: quotaType });
                    return;
                }
-                
-                const message = this.inputMessage;
-                
-                if (message.startsWith('/')) {
+
+                if (hasImages && this.currentModelKey !== 'qwen3-vl-plus') {
+                    this.uiPushToast({
+                        title: '当前模型不支持图片',
+                        message: '请切换到 Qwen-VL 再发送图片',
+                        type: 'error'
+                    });
+                    return;
+                }
+
+                const message = text;
+                const isCommand = hasText && !hasImages && message.startsWith('/');
+                if (isCommand) {
                    this.socket.emit('send_command', { command: message });
                    this.inputClearMessage();
+                    this.inputClearSelectedImages();
                    this.autoResizeInput();
                    return;
                }
@ -2288,14 +2348,19 @@ const appOptions = {
                
                // 标记任务进行中，直到任务完成或用户手动停止
                this.taskInProgress = true;
-                this.chatAddUserMessage(message);
-                this.socket.emit('send_message', { message: message, conversation_id: this.currentConversationId });
+                this.chatAddUserMessage(message, images);
+                this.socket.emit('send_message', { message: message, images, conversation_id: this.currentConversationId });
                if (typeof this.monitorShowPendingReply === 'function') {
                    this.monitorShowPendingReply();
                }
                this.inputClearMessage();
+                this.inputClearSelectedImages();
+                this.inputSetImagePickerOpen(false);
                this.inputSetLineCount(1);
                this.inputSetMultiline(false);
+                if (hasImages) {
+                    this.conversationHasImages = true;
+                }
                if (this.autoScrollEnabled) {
                    this.scrollToBottom();
                }
@ -2447,6 +2512,7 @@ const appOptions = {
                    return;
                }
                this.modeMenuOpen = false;
+                this.modelMenuOpen = false;
                const nextState = this.inputToggleToolMenu();
                if (nextState) {
                    this.inputSetSettingsOpen(false);
@ -2466,12 +2532,117 @@ const appOptions = {
                const opened = this.inputToggleQuickMenu();
                if (!opened) {
                    this.modeMenuOpen = false;
+                    this.modelMenuOpen = false;
                }
            },

            closeQuickMenu() {
                this.inputCloseMenus();
                this.modeMenuOpen = false;
+                this.modelMenuOpen = false;
+            },
+
+            async openImagePicker() {
+                if (this.currentModelKey !== 'qwen3-vl-plus') {
+                    this.uiPushToast({
+                        title: '当前模型不支持图片',
+                        message: '请选择 Qwen-VL 后再发送图片',
+                        type: 'error'
+                    });
+                    return;
+                }
+                this.closeQuickMenu();
+                this.inputSetImagePickerOpen(true);
+                await this.loadWorkspaceImages();
+            },
+
+            closeImagePicker() {
+                this.inputSetImagePickerOpen(false);
+            },
+
+            async loadWorkspaceImages() {
+                this.imageLoading = true;
+                try {
+                    const entries = await this.fetchAllImageEntries('');
+                    this.imageEntries = entries;
+                    if (!entries.length) {
+                        this.uiPushToast({
+                            title: '未找到图片',
+                            message: '工作区内没有可用的图片文件',
+                            type: 'info'
+                        });
+                    }
+                } catch (error) {
+                    console.error('加载图片列表失败', error);
+                    this.uiPushToast({
+                        title: '加载图片失败',
+                        message: error?.message || '请稍后重试',
+                        type: 'error'
+                    });
+                } finally {
+                    this.imageLoading = false;
+                }
+            },
+
+            async fetchAllImageEntries(startPath = '') {
+                const queue: string[] = [startPath || ''];
+                const visited = new Set<string>();
+                const results: Array<{ name: string; path: string }> = [];
+                const exts = new Set(['.png', '.jpg', '.jpeg', '.webp', '.gif', '.bmp', '.svg']);
+                const maxFolders = 120;
+
+                while (queue.length && visited.size < maxFolders) {
+                    const path = queue.shift() || '';
+                    if (visited.has(path)) {
+                        continue;
+                    }
+                    visited.add(path);
+                    try {
+                        const resp = await fetch(`/api/gui/files/entries?path=${encodeURIComponent(path)}`, {
+                            method: 'GET',
+                            credentials: 'include',
+                            headers: { Accept: 'application/json' }
+                        });
+                        const data = await resp.json().catch(() => null);
+                        if (!data?.success) {
+                            continue;
+                        }
+                        const items = Array.isArray(data?.data?.items) ? data.data.items : [];
+                        for (const item of items) {
+                            const rawPath =
+                                item?.path ||
+                                [path, item?.name].filter(Boolean).join('/').replace(/\\/g, '/').replace(/\/{2,}/g, '/');
+                            const type = String(item?.type || '').toLowerCase();
+                            if (type === 'directory' || type === 'folder') {
+                                queue.push(rawPath);
+                                continue;
+                            }
+                            const ext =
+                                String(item?.extension || '').toLowerCase() ||
+                                (rawPath.includes('.') ? `.${rawPath.split('.').pop()?.toLowerCase()}` : '');
+                            if (exts.has(ext)) {
+                                results.push({
+                                    name: item?.name || rawPath.split('/').pop() || rawPath,
+                                    path: rawPath
+                                });
+                                if (results.length >= 400) {
+                                    return results;
+                                }
+                            }
+                        }
+                    } catch (error) {
+                        console.warn('遍历文件夹失败', path, error);
+                    }
+                }
+                return results;
+            },
+
+            handleImagesConfirmed(list) {
+                this.inputSetSelectedImages(Array.isArray(list) ? list : []);
+                this.inputSetImagePickerOpen(false);
+            },
+            handleRemoveImage(path) {
+                this.inputRemoveSelectedImage(path);
            },

            handleQuickUpload() {
@ -2488,6 +2659,25 @@ const appOptions = {
                const next = !this.modeMenuOpen;
                this.modeMenuOpen = next;
                if (next) {
+                    this.modelMenuOpen = false;
+                }
+                if (next) {
+                    this.inputSetToolMenuOpen(false);
+                    this.inputSetSettingsOpen(false);
+                    if (!this.quickMenuOpen) {
+                        this.inputOpenQuickMenu();
+                    }
+                }
+            },
+
+            toggleModelMenu() {
+                if (!this.isConnected || this.streamingMessage) {
+                    return;
+                }
+                const next = !this.modelMenuOpen;
+                this.modelMenuOpen = next;
+                if (next) {
+                    this.modeMenuOpen = false;
                    this.inputSetToolMenuOpen(false);
                    this.inputSetSettingsOpen(false);
                    if (!this.quickMenuOpen) {
@ -2503,6 +2693,56 @@ const appOptions = {
                await this.setRunMode(mode);
            },

+            async handleModelSelect(key) {
+                if (!this.isConnected || this.streamingMessage) {
+                    return;
+                }
+                if (this.conversationHasImages && key !== 'qwen3-vl-plus') {
+                    this.uiPushToast({
+                        title: '切换失败',
+                        message: '当前对话包含图片，仅支持 Qwen-VL',
+                        type: 'error'
+                    });
+                    return;
+                }
+                const modelStore = useModelStore();
+                const prev = this.currentModelKey;
+                try {
+                    const resp = await fetch('/api/model', {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' },
+                        body: JSON.stringify({ model_key: key })
+                    });
+                    const payload = await resp.json();
+                    if (!resp.ok || !payload.success) {
+                        throw new Error(payload.error || payload.message || '切换失败');
+                    }
+                    const data = payload.data || {};
+                    modelStore.setModel(data.model_key || key);
+                    if (data.run_mode) {
+                        this.runMode = data.run_mode;
+                        this.thinkingMode = data.thinking_mode ?? (data.run_mode !== 'fast');
+                    }
+                    this.uiPushToast({
+                        title: '模型已切换',
+                        message: modelStore.currentModel?.label || key,
+                        type: 'success'
+                    });
+                } catch (error) {
+                    modelStore.setModel(prev);
+                    const msg = error instanceof Error ? error.message : String(error || '切换失败');
+                    this.uiPushToast({
+                        title: '切换模型失败',
+                        message: msg,
+                        type: 'error'
+                    });
+                } finally {
+                    this.modelMenuOpen = false;
+                    this.inputCloseMenus();
+                    this.inputSetQuickMenuOpen(false);
+                }
+            },
+
            async handleCycleRunMode() {
                const modes: Array<'fast' | 'thinking' | 'deep'> = ['fast', 'thinking', 'deep'];
                const currentMode = this.resolvedRunMode;
@ -2511,11 +2751,39 @@ const appOptions = {
                await this.setRunMode(nextMode);
            },

-            async setRunMode(mode) {
+            async setRunMode(mode, options = {}) {
                if (!this.isConnected || this.streamingMessage) {
                    this.modeMenuOpen = false;
                    return;
                }
+                const modelStore = useModelStore();
+                const fastOnly = modelStore.currentModel?.fastOnly;
+                const currentModelKey = modelStore.currentModel?.key;
+                if (fastOnly && mode !== 'fast') {
+                    if (!options.suppressToast) {
+                        this.uiPushToast({
+                            title: '模式不可用',
+                            message: 'Qwen-Max只支持快速模式',
+                            type: 'warning'
+                        });
+                    }
+                    this.modeMenuOpen = false;
+                    this.inputCloseMenus();
+                    return;
+                }
+                // Qwen-VL 不支持深度思考模式
+                if (currentModelKey === 'qwen3-vl-plus' && mode === 'deep') {
+                    if (!options.suppressToast) {
+                        this.uiPushToast({
+                            title: '模式不可用',
+                            message: 'Qwen-VL 不支持深度思考模式，请使用快速或思考模式',
+                            type: 'warning'
+                        });
+                    }
+                    this.modeMenuOpen = false;
+                    this.inputCloseMenus();
+                    return;
+                }
                if (mode === this.resolvedRunMode) {
                    this.modeMenuOpen = false;
                    this.closeQuickMenu();
@ -2774,6 +3042,7 @@ const appOptions = {
                    return;
                }
                this.modeMenuOpen = false;
+                this.modelMenuOpen = false;
                const nextState = this.inputToggleSettingsMenu();
                if (nextState) {
                    this.inputSetToolMenuOpen(false);
@ -2961,7 +3230,8 @@ const appOptions = {
    LiquidGlassWidget,
    QuickMenu,
    InputComposer,
-    AppShell
+    AppShell,
+    ImagePicker
 };

 export default appOptions;
--- a/static/src/components/chat/ChatArea.vue
+++ b/static/src/components/chat/ChatArea.vue
@ -1,13 +1,18 @@
 <template>
  <div class="messages-area" ref="rootEl">
    <div class="messages-flow">
-      <div v-for="(msg, index) in messages" :key="index" class="message-block">
+      <div v-for="(msg, index) in filteredMessages" :key="index" class="message-block">
        <div v-if="msg.role === 'user'" class="user-message">
          <div class="message-header icon-label">
            <span class="icon icon-sm" :style="iconStyleSafe('user')" aria-hidden="true"></span>
            <span>用户</span>
          </div>
-          <div class="message-text">{{ msg.content }}</div>
+          <div class="message-text user-bubble-text">
+            <div v-if="msg.content" class="bubble-text">{{ msg.content }}</div>
+            <div v-if="msg.images && msg.images.length" class="image-inline-row">
+              <span class="image-name" v-for="img in msg.images" :key="img">{{ formatImageName(img) }}</span>
+            </div>
+          </div>
        </div>
        <div v-else-if="msg.role === 'assistant'" class="assistant-message">
          <div class="message-header icon-label">
@ -368,6 +373,9 @@ const props = defineProps<{

 const personalization = usePersonalizationStore();
 const stackedBlocksEnabled = computed(() => personalization.experiments.stackedBlocksEnabled);
+const filteredMessages = computed(() =>
+  (props.messages || []).filter(m => !(m && m.metadata && m.metadata.system_injected_image))
+);

 const DEFAULT_GENERATING_TEXT = '生成中…';
 const rootEl = ref<HTMLElement | null>(null);
@ -403,6 +411,12 @@ function iconStyleSafe(key: string, size?: string) {
  return {};
 }

+function formatImageName(path: string): string {
+  if (!path) return '';
+  const parts = path.split(/[/\\]/);
+  return parts[parts.length - 1] || path;
+}
+
 const isStackable = (action: any) => action && (action.type === 'thinking' || action.type === 'tool');
 const splitActionGroups = (actions: any[] = [], messageIndex = 0) => {
  const result: Array<
--- a/static/src/components/input/InputComposer.vue
+++ b/static/src/components/input/InputComposer.vue
@ -11,39 +11,49 @@
        }"
      >
        <input type="file" ref="fileUploadInput" class="file-input-hidden" @change="onFileChange" />
-        <button
-          type="button"
-          class="stadium-btn add-btn"
-          @click.stop="$emit('toggle-quick-menu')"
-          :disabled="!isConnected"
-        >
-          +
-        </button>
-        <textarea
-          ref="stadiumInput"
-          class="stadium-input"
-          rows="1"
-          :value="inputMessage"
-          :disabled="!isConnected || streamingMessage || inputLocked"
-          placeholder="输入消息... (Ctrl+Enter 发送)"
-          @input="onInput"
-          @focus="$emit('input-focus')"
-          @blur="$emit('input-blur')"
-          @keydown.enter.ctrl.prevent="$emit('send-message')"
-        ></textarea>
-        <button
-          type="button"
-          class="stadium-btn send-btn"
-          @click="$emit('send-or-stop')"
-          :disabled="
-            !isConnected ||
-            (inputLocked && !streamingMessage) ||
-            (!(inputMessage || '').trim() && !streamingMessage)
-          "
-        >
-          <span v-if="streamingMessage" class="stop-icon"></span>
-          <span v-else class="send-icon"></span>
-        </button>
+        <div class="input-stack">
+          <div v-if="selectedImages && selectedImages.length" class="image-inline-row">
+            <span class="image-name" v-for="img in selectedImages" :key="img">
+              {{ formatImageName(img) }}
+              <button type="button" class="image-remove-btn" @click.stop="$emit('remove-image', img)">×</button>
+            </span>
+          </div>
+          <div class="input-row">
+            <button
+              type="button"
+              class="stadium-btn add-btn"
+              @click.stop="$emit('toggle-quick-menu')"
+              :disabled="!isConnected"
+            >
+              +
+            </button>
+            <textarea
+              ref="stadiumInput"
+              class="stadium-input"
+              rows="1"
+              :value="inputMessage"
+              :disabled="!isConnected || streamingMessage || inputLocked"
+              placeholder="输入消息... (Ctrl+Enter 发送)"
+              @input="onInput"
+              @focus="$emit('input-focus')"
+              @blur="$emit('input-blur')"
+              @keydown.enter.ctrl.prevent="$emit('send-message')"
+            ></textarea>
+            <button
+              type="button"
+              class="stadium-btn send-btn"
+              @click="$emit('send-or-stop')"
+              :disabled="
+                !isConnected ||
+                (inputLocked && !streamingMessage) ||
+                ((!(inputMessage || '').trim() && (!selectedImages || !selectedImages.length)) && !streamingMessage)
+              "
+            >
+              <span v-if="streamingMessage" class="stop-icon"></span>
+              <span v-else class="send-icon"></span>
+            </button>
+          </div>
+        </div>
      </div>
      <QuickMenu
        :open="quickMenuOpen"
@ -52,6 +62,9 @@
        :streaming-message="streamingMessage"
        :thinking-mode="thinkingMode"
        :run-mode="runMode"
+        :model-menu-open="modelMenuOpen"
+        :model-options="modelOptions"
+        :current-model-key="currentModelKey"
        :tool-menu-open="toolMenuOpen"
        :tool-settings="toolSettings"
        :tool-settings-loading="toolSettingsLoading"
@ -62,10 +75,13 @@
        :icon-style="iconStyle"
        :tool-category-icon="toolCategoryIcon"
        @quick-upload="triggerQuickUpload"
+        @pick-images="$emit('pick-images')"
        @toggle-tool-menu="$emit('toggle-tool-menu')"
        @toggle-settings="$emit('toggle-settings')"
        @toggle-mode-menu="$emit('toggle-mode-menu')"
        @select-run-mode="(mode) => $emit('select-run-mode', mode)"
+        @toggle-model-menu="$emit('toggle-model-menu')"
+        @select-model="(key) => $emit('select-model', key)"
        @update-tool-category="(id, enabled) => $emit('update-tool-category', id, enabled)"
        @realtime-terminal="$emit('realtime-terminal')"
        @toggle-focus-panel="$emit('toggle-focus-panel')"
@ -92,16 +108,20 @@ const emit = defineEmits([
  'send-message',
  'send-or-stop',
  'quick-upload',
+  'pick-images',
  'toggle-tool-menu',
  'toggle-mode-menu',
+  'toggle-model-menu',
  'select-run-mode',
+  'select-model',
  'toggle-settings',
  'update-tool-category',
  'realtime-terminal',
  'toggle-focus-panel',
  'toggle-token-panel',
  'compress-conversation',
-  'file-selected'
+  'file-selected',
+  'remove-image'
 ]);

 const props = defineProps<{
@ -117,6 +137,7 @@ const props = defineProps<{
  quickMenuOpen: boolean;
  toolMenuOpen: boolean;
  modeMenuOpen: boolean;
+  modelMenuOpen: boolean;
  toolSettings: Array<{ id: string; label: string; enabled: boolean }>;
  toolSettingsLoading: boolean;
  settingsOpen: boolean;
@ -124,6 +145,9 @@ const props = defineProps<{
  currentConversationId: string | null;
  iconStyle: (key: string) => Record<string, string>;
  toolCategoryIcon: (categoryId: string) => string;
+  modelOptions: Array<{ key: string; label: string; description: string }>;
+  currentModelKey: string;
+  selectedImages?: string[];
 }>();

 const inputStore = useInputStore();
@ -132,6 +156,12 @@ const compactInputShell = ref<HTMLElement | null>(null);
 const stadiumInput = ref<HTMLTextAreaElement | null>(null);
 const fileUploadInput = ref<HTMLInputElement | null>(null);

+const formatImageName = (path: string): string => {
+  if (!path) return '';
+  const parts = path.split(/[/\\]/);
+  return parts[parts.length - 1] || path;
+};
+
 const applyLineMetrics = (lines: number, multiline: boolean) => {
  inputStore.setInputLineCount(lines);
  inputStore.setInputMultiline(multiline);
@ -204,3 +234,18 @@ onMounted(() => {
  adjustTextareaSize();
 });
 </script>
+
+<style scoped>
+.image-inline-row {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+  padding: 4px 10px 2px;
+  font-size: 12px;
+  color: var(--text-secondary, #7f8792);
+  line-height: 1.4;
+}
+.image-name {
+  white-space: nowrap;
+}
+</style>
--- a/static/src/components/input/QuickMenu.vue
+++ b/static/src/components/input/QuickMenu.vue
@ -4,6 +4,15 @@
      <button type="button" class="menu-entry" @click="$emit('quick-upload')" :disabled="!isConnected || uploading">
        {{ uploading ? '上传中...' : '上传文件' }}
      </button>
+      <button
+        v-if="currentModelKey === 'qwen3-vl-plus'"
+        type="button"
+        class="menu-entry"
+        @click.stop="$emit('pick-images')"
+        :disabled="!isConnected || streamingMessage"
+      >
+        发送图片
+      </button>
      <button
        type="button"
        class="menu-entry has-submenu"
@ -13,6 +22,15 @@
        <span>运行模式</span>
        <span class="entry-arrow">{{ runModeLabel }}</span>
      </button>
+      <button
+        type="button"
+        class="menu-entry has-submenu"
+        @click.stop="$emit('toggle-model-menu')"
+        :disabled="!isConnected"
+      >
+        <span>切换模型</span>
+        <span class="entry-arrow">{{ currentModelLabel }}</span>
+      </button>
      <button
        type="button"
        class="menu-entry has-submenu"
@ -50,6 +68,28 @@
        </div>
      </transition>

+      <transition name="submenu-slide">
+        <div class="quick-submenu model-submenu" v-if="modelMenuOpen">
+          <div class="submenu-list">
+            <button
+              v-for="option in modelOptions"
+              :key="option.key"
+              type="button"
+              class="menu-entry submenu-entry"
+              :class="{ active: option.key === currentModelKey }"
+              @click.stop="$emit('select-model', option.key)"
+              :disabled="streamingMessage || !isConnected"
+            >
+              <span class="submenu-label">
+                <span>{{ option.label }}</span>
+                <span class="submenu-desc">{{ option.description }}</span>
+              </span>
+              <span v-if="option.key === currentModelKey" class="entry-arrow">✓</span>
+            </button>
+          </div>
+        </div>
+      </transition>
+
      <transition name="submenu-slide">
        <div class="quick-submenu tool-submenu" v-if="toolMenuOpen">
          <div class="submenu-status" v-if="toolSettingsLoading">正在同步工具状态...</div>
@ -137,6 +177,9 @@ const props = defineProps<{
  toolCategoryIcon: (categoryId: string) => string;
  modeMenuOpen: boolean;
  runMode?: 'fast' | 'thinking' | 'deep';
+  modelMenuOpen: boolean;
+  modelOptions: Array<{ key: string; label: string; description: string }>;
+  currentModelKey: string;
 }>();

 defineEmits<{
@ -150,6 +193,8 @@ defineEmits<{
  (event: 'compress-conversation'): void;
  (event: 'toggle-mode-menu'): void;
  (event: 'select-run-mode', mode: 'fast' | 'thinking' | 'deep'): void;
+  (event: 'toggle-model-menu'): void;
+  (event: 'select-model', key: string): void;
 }>();

 const runModeOptions = [
@ -174,4 +219,18 @@ const resolvedRunMode = computed<'fast' | 'thinking' | 'deep'>(() => {
 const runModeLabel = computed(() => runModeLabelMap[resolvedRunMode.value]);

 const getIconStyle = (key: string) => (props.iconStyle ? props.iconStyle(key) : {});
+
+const currentModelLabel = computed(() => {
+  const found = props.modelOptions?.find(m => m.key === props.currentModelKey);
+  return found ? found.label : '未选择';
+});
 </script>
+
+<style scoped>
+.submenu-desc {
+  display: block;
+  font-size: 12px;
+  color: var(--text-secondary, #7f8792);
+  margin-top: 2px;
+}
+</style>
--- a/static/src/components/overlay/ImagePicker.vue
+++ b/static/src/components/overlay/ImagePicker.vue
@ -0,0 +1,211 @@
+<template>
+  <div class="image-picker-backdrop" @click.self="close">
+    <div class="image-picker-panel">
+      <div class="header">
+        <div class="title">选择图片（最多9张）</div>
+        <button class="close-btn" @click="close">×</button>
+      </div>
+      <div class="body">
+        <div v-if="loading" class="loading">加载中...</div>
+        <div v-else-if="!images.length" class="empty">未找到图片文件</div>
+        <div v-else class="grid">
+          <div
+            v-for="item in images"
+            :key="item.path"
+            class="card"
+            :class="{ selected: selectedSet.has(item.path) }"
+            @click="toggle(item.path)"
+            :title="item.path"
+          >
+            <img :src="previewUrl(item.path)" :alt="item.name" />
+            <div class="name">{{ item.name }}</div>
+          </div>
+        </div>
+      </div>
+      <div class="footer">
+        <div class="count">已选 {{ selectedSet.size }} / 9</div>
+        <div class="actions">
+          <button type="button" class="btn secondary" @click="close">取消</button>
+          <button type="button" class="btn primary" :disabled="!selectedSet.size" @click="confirm">确认</button>
+        </div>
+      </div>
+    </div>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { computed, ref, watch, onMounted } from 'vue';
+
+interface ImageEntry {
+  name: string;
+  path: string;
+}
+
+const props = defineProps<{
+  open: boolean;
+  entries: ImageEntry[];
+  initialSelected: string[];
+  loading: boolean;
+}>();
+
+const emit = defineEmits<{
+  (e: 'close'): void;
+  (e: 'confirm', list: string[]): void;
+}>();
+
+const selectedSet = ref<Set<string>>(new Set(props.initialSelected || []));
+
+watch(
+  () => props.initialSelected,
+  (val) => {
+    selectedSet.value = new Set(val || []);
+  }
+);
+
+const images = computed(() => props.entries || []);
+
+const toggle = (path: string) => {
+  if (!path) return;
+  const set = new Set(selectedSet.value);
+  if (set.has(path)) {
+    set.delete(path);
+  } else {
+    if (set.size >= 9) return;
+    set.add(path);
+  }
+  selectedSet.value = set;
+};
+
+const close = () => emit('close');
+
+const confirm = () => emit('confirm', Array.from(selectedSet.value));
+
+const previewUrl = (path: string) => `/api/gui/files/download?path=${encodeURIComponent(path)}`;
+
+onMounted(() => {
+  selectedSet.value = new Set(props.initialSelected || []);
+});
+</script>
+
+<style scoped>
+.image-picker-backdrop {
+  position: fixed;
+  inset: 0;
+  background: rgba(0, 0, 0, 0.45);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  z-index: 1200;
+}
+.image-picker-panel {
+  width: min(980px, 92vw);
+  max-height: 88vh;
+  background: #0f1116;
+  color: #e8ecf2;
+  border: 1px solid #2a2f3a;
+  border-radius: 12px;
+  display: flex;
+  flex-direction: column;
+  box-shadow: 0 16px 40px rgba(0, 0, 0, 0.4);
+}
+.header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 14px 16px;
+  border-bottom: 1px solid #1f2430;
+}
+.title {
+  font-weight: 600;
+}
+.close-btn {
+  background: transparent;
+  color: #9aa3b5;
+  border: none;
+  font-size: 20px;
+  cursor: pointer;
+}
+.body {
+  padding: 12px 16px;
+  overflow: auto;
+  flex: 1;
+}
+.grid {
+  display: grid;
+  grid-template-columns: repeat(auto-fill, minmax(160px, 1fr));
+  gap: 12px;
+}
+.card {
+  border: 1px solid #1f2430;
+  border-radius: 10px;
+  background: #151922;
+  cursor: pointer;
+  overflow: hidden;
+  display: flex;
+  flex-direction: column;
+}
+.card img {
+  width: 100%;
+  height: 120px;
+  object-fit: cover;
+  background: #0c0f14;
+}
+.card .name {
+  padding: 8px 10px;
+  font-size: 12px;
+  color: #c5ccda;
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+}
+.card.selected {
+  border-color: #4ca6ff;
+  box-shadow: 0 0 0 2px rgba(76, 166, 255, 0.2);
+}
+.loading,
+.empty {
+  padding: 40px 0;
+  text-align: center;
+  color: #9aa3b5;
+}
+.footer {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 12px 16px;
+  border-top: 1px solid #1f2430;
+}
+.actions {
+  display: flex;
+  gap: 10px;
+}
+.btn {
+  border: 1px solid #2f3645;
+  padding: 8px 14px;
+  border-radius: 8px;
+  background: #1b202c;
+  color: #e8ecf2;
+  cursor: pointer;
+}
+.btn.primary {
+  background: #4ca6ff;
+  border-color: #4ca6ff;
+  color: #0d1117;
+}
+.btn:disabled {
+  opacity: 0.5;
+  cursor: not-allowed;
+}
+.count {
+  font-size: 13px;
+  color: #9aa3b5;
+}
+@media (max-width: 640px) {
+  .grid {
+    grid-template-columns: repeat(auto-fill, minmax(120px, 1fr));
+  }
+  .card img {
+    height: 90px;
+  }
+}
+</style>
--- a/static/src/composables/useLegacySocket.ts
+++ b/static/src/composables/useLegacySocket.ts
@ -4,7 +4,7 @@ import { renderLatexInRealtime } from './useMarkdownRenderer';

 export async function initializeLegacySocket(ctx: any) {
    try {
-        const SOCKET_DEBUG_LOGS_ENABLED = false;
+        const SOCKET_DEBUG_LOGS_ENABLED = true;
        const socketLog = (...args: any[]) => {
            if (!SOCKET_DEBUG_LOGS_ENABLED) {
                return;
--- a/static/src/stores/chat.ts
+++ b/static/src/stores/chat.ts
@ -164,10 +164,11 @@ export const useChatStore = defineStore('chat', {
      this.currentMessageIndex = this.messages.length - 1;
      return message;
    },
-    addUserMessage(content: string) {
+    addUserMessage(content: string, images: string[] = []) {
      this.messages.push({
        role: 'user',
-        content
+        content,
+        images
      });
      this.currentMessageIndex = -1;
    },
--- a/static/src/stores/input.ts
+++ b/static/src/stores/input.ts
@ -8,6 +8,8 @@ interface InputState {
  quickMenuOpen: boolean;
  toolMenuOpen: boolean;
  settingsOpen: boolean;
+  imagePickerOpen: boolean;
+  selectedImages: string[];
 }

 export const useInputStore = defineStore('input', {
@ -18,7 +20,9 @@ export const useInputStore = defineStore('input', {
    inputIsFocused: false,
    quickMenuOpen: false,
    toolMenuOpen: false,
-    settingsOpen: false
+    settingsOpen: false,
+    imagePickerOpen: false,
+    selectedImages: []
  }),
  actions: {
    setInputMessage(value: string) {
@ -69,6 +73,23 @@ export const useInputStore = defineStore('input', {
    },
    setSettingsOpen(open: boolean) {
      this.settingsOpen = open;
+    },
+    setImagePickerOpen(open: boolean) {
+      this.imagePickerOpen = open;
+    },
+    setSelectedImages(list: string[]) {
+      this.selectedImages = list.slice(0, 9);
+    },
+    addSelectedImage(path: string) {
+      if (!path) return;
+      const next = Array.from(new Set([...this.selectedImages, path]));
+      this.selectedImages = next.slice(0, 9);
+    },
+    removeSelectedImage(path: string) {
+      this.selectedImages = this.selectedImages.filter(item => item !== path);
+    },
+    clearSelectedImages() {
+      this.selectedImages = [];
    }
  }
 });
--- a/static/src/stores/model.ts
+++ b/static/src/stores/model.ts
@ -0,0 +1,66 @@
+import { defineStore } from 'pinia';
+
+export type ModelKey = 'kimi' | 'deepseek' | 'qwen3-max' | 'qwen3-vl-plus';
+
+export interface ModelOption {
+  key: ModelKey;
+  label: string;
+  description: string;
+  fastOnly: boolean;
+  supportsThinking: boolean;
+}
+
+interface ModelState {
+  currentModelKey: ModelKey;
+  models: ModelOption[];
+}
+
+export const useModelStore = defineStore('model', {
+  state: (): ModelState => ({
+    currentModelKey: 'kimi',
+    models: [
+      {
+        key: 'kimi',
+        label: 'Kimi-k2',
+        description: '综合能力较强',
+        fastOnly: false,
+        supportsThinking: true
+      },
+      {
+        key: 'deepseek',
+        label: 'Deepseek-V3.2',
+        description: '数学能力较强',
+        fastOnly: false,
+        supportsThinking: true
+      },
+      {
+        key: 'qwen3-max',
+        label: 'Qwen-Max',
+        description: '仅支持快速模式',
+        fastOnly: true,
+        supportsThinking: false
+      },
+      {
+        key: 'qwen3-vl-plus',
+        label: 'Qwen-VL',
+        description: '支持图片输入',
+        fastOnly: false,
+        supportsThinking: true
+      }
+    ]
+  }),
+  getters: {
+    currentModel(state): ModelOption {
+      return state.models.find(m => m.key === state.currentModelKey) || state.models[0];
+    }
+  },
+  actions: {
+    setModel(key: ModelKey) {
+      if (this.currentModelKey === key) return;
+      const exists = this.models.some(m => m.key === key);
+      if (exists) {
+        this.currentModelKey = key;
+      }
+    }
+  }
+});
--- a/static/src/styles/components/chat/_chat-area.scss
+++ b/static/src/styles/components/chat/_chat-area.scss
@ -192,6 +192,9 @@

 .user-message .message-text {
    background: rgba(255, 255, 255, 0.88);
+    display: flex;
+    flex-direction: column;
+    gap: 10px;
 }

 .assistant-message .message-text {
@ -199,6 +202,20 @@
    border-left: 4px solid var(--claude-accent);
 }

+.user-message .message-text.user-bubble-text .image-inline-row {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 10px;
+    font-size: 12px;
+    color: var(--claude-text-secondary);
+    line-height: 1.4;
+    padding-bottom: 2px;
+}
+
+.user-message .message-text.user-bubble-text .image-name {
+    white-space: nowrap;
+}
+
 .assistant-generating-block {
    width: 100%;
 }
--- a/static/src/styles/components/input/_composer.scss
+++ b/static/src/styles/components/input/_composer.scss
@ -34,7 +34,6 @@
    background: #ffffff;
    box-shadow: 0 18px 46px rgba(15, 23, 42, 0.16);
    display: flex;
-    align-items: center;
    gap: 12px;
    transition:
        padding 0.2s ease,
@ -68,6 +67,52 @@
        0 32px 86px rgba(15, 23, 42, 0.28);
 }

+.input-stack {
+    display: flex;
+    flex-direction: column;
+    flex: 1 1 auto;
+    gap: 6px;
+}
+
+.input-row {
+    display: flex;
+    align-items: center;
+    gap: 12px;
+    width: 100%;
+}
+
+.image-inline-row {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 6px;
+    padding: 0 4px 0;
+    font-size: 12px;
+    color: var(--text-secondary, #7f8792);
+    line-height: 1.4;
+}
+
+.image-name {
+    white-space: nowrap;
+    display: inline-flex;
+    align-items: center;
+    gap: 6px;
+}
+
+.image-remove-btn {
+    border: none;
+    background: transparent;
+    color: var(--text-secondary, #7f8792);
+    cursor: pointer;
+    padding: 0 4px;
+    font-size: 12px;
+    line-height: 1;
+    transition: color 0.15s ease, transform 0.15s ease;
+}
+.image-remove-btn:hover {
+    color: #d14b4b;
+    transform: scale(1.05);
+}
+
 .stadium-input {
    flex: 1 1 auto;
    width: 100%;
--- a/static/src/utils/icons.ts
+++ b/static/src/utils/icons.ts
@ -19,6 +19,7 @@ export const ICONS = Object.freeze({
  info: '/static/icons/info.svg',
  laptop: '/static/icons/laptop.svg',
  layers: '/static/icons/layers.svg',
+  keyboard: '/static/icons/keyboard.svg',
  menu: '/static/icons/menu.svg',
  monitor: '/static/icons/monitor.svg',
  octagon: '/static/icons/octagon.svg',
@ -62,7 +63,7 @@ export const TOOL_ICON_MAP = Object.freeze({
  todo_finish: 'flag',
  todo_finish_confirm: 'circleAlert',
  todo_update_task: 'check',
-  terminal_input: 'terminal',
+  terminal_input: 'keyboard',
  terminal_reset: 'recycle',
  terminal_session: 'monitor',
  terminal_snapshot: 'clipboard',
@ -70,7 +71,8 @@ export const TOOL_ICON_MAP = Object.freeze({
  update_memory: 'brain',
  wait_sub_agent: 'clock',
  web_search: 'search',
-  trigger_easter_egg: 'sparkles'
+  trigger_easter_egg: 'sparkles',
+  view_image: 'camera'
 });

 export const TOOL_CATEGORY_ICON_MAP = Object.freeze({
--- a/sub_agent/core/tool_config.py
+++ b/sub_agent/core/tool_config.py
@ -40,7 +40,7 @@ TOOL_CATEGORIES: Dict[str, ToolCategory] = {
    ),
    "read_focus": ToolCategory(
        label="阅读聚焦",
-        tools=["read_file", "focus_file", "unfocus_file", "ocr_image"],
+        tools=["read_file", "focus_file", "unfocus_file", "ocr_image", "vlm_analyze", "view_image"],
    ),
    "terminal_realtime": ToolCategory(
        label="实时终端",
--- a/utils/api_client.py
+++ b/utils/api_client.py
@ -45,6 +45,10 @@ class DeepSeekClient:
            "api_key": THINKING_API_KEY or API_KEY,
            "model_id": THINKING_MODEL_ID or MODEL_ID
        }
+        self.fast_max_tokens = None
+        self.thinking_max_tokens = None
+        self.fast_extra_params: Dict = {}
+        self.thinking_extra_params: Dict = {}
        self.thinking_mode = thinking_mode  # True=智能思考模式, False=快速模式
        self.deep_thinking_mode = False  # 深度思考模式：整轮都使用思考模型
        self.deep_thinking_session = False  # 当前任务是否处于深度思考会话
@ -53,6 +57,7 @@ class DeepSeekClient:
        self.api_base_url = self.fast_api_config["base_url"]
        self.api_key = self.fast_api_config["api_key"]
        self.model_id = self.fast_api_config["model_id"]
+        self.model_key = None  # 由宿主终端注入，便于做模型兼容处理
        # 每个任务的独立状态
        self.current_task_first_call = True  # 当前任务是否是第一次调用
        self.current_task_thinking = ""  # 当前任务的思考内容
@ -160,6 +165,41 @@ class DeepSeekClient:
            "api_key": config.get("api_key") or fallback["api_key"],
            "model_id": config.get("model_id") or fallback["model_id"]
        }
+
+
+    def apply_profile(self, profile: Dict):
+        """
+        动态应用模型配置
+        profile 示例：
+        {
+            "fast": {"base_url": "...", "api_key": "...", "model_id": "...", "max_tokens": 8192},
+            "thinking": {...} 或 None,
+            "supports_thinking": True/False,
+            "fast_only": True/False
+        }
+        """
+        if not profile or "fast" not in profile:
+            raise ValueError("无效的模型配置")
+        fast = profile["fast"] or {}
+        thinking = profile.get("thinking") or fast
+        self.fast_api_config = {
+            "base_url": fast.get("base_url") or self.fast_api_config.get("base_url"),
+            "api_key": fast.get("api_key") or self.fast_api_config.get("api_key"),
+            "model_id": fast.get("model_id") or self.fast_api_config.get("model_id")
+        }
+        self.thinking_api_config = {
+            "base_url": thinking.get("base_url") or self.thinking_api_config.get("base_url"),
+            "api_key": thinking.get("api_key") or self.thinking_api_config.get("api_key"),
+            "model_id": thinking.get("model_id") or self.thinking_api_config.get("model_id")
+        }
+        self.fast_max_tokens = fast.get("max_tokens")
+        self.thinking_max_tokens = thinking.get("max_tokens")
+        self.fast_extra_params = fast.get("extra_params") or {}
+        self.thinking_extra_params = thinking.get("extra_params") or {}
+        # 同步旧字段
+        self.api_base_url = self.fast_api_config["base_url"]
+        self.api_key = self.fast_api_config["api_key"]
+        self.model_id = self.fast_api_config["model_id"]
    
    def get_current_thinking_mode(self) -> bool:
        """获取当前应该使用的思考模式"""
@ -279,18 +319,26 @@ class DeepSeekClient:
            self.skip_thinking_next_call = False
        
        try:
-            max_tokens = int(DEFAULT_RESPONSE_MAX_TOKENS)
+            override_max = self.thinking_max_tokens if current_thinking_mode else self.fast_max_tokens
+            if override_max is not None:
+                max_tokens = int(override_max)
+            else:
+                max_tokens = int(DEFAULT_RESPONSE_MAX_TOKENS)
            if max_tokens <= 0:
                raise ValueError("max_tokens must be positive")
        except (TypeError, ValueError):
            max_tokens = 4096
-        
+
        payload = {
            "model": api_config["model_id"],
            "messages": messages,
            "stream": stream,
            "max_tokens": max_tokens
        }
+        # 注入模型额外参数（如 Qwen enable_thinking）
+        extra_params = self.thinking_extra_params if current_thinking_mode else self.fast_extra_params
+        if extra_params:
+            payload.update(extra_params)
        if tools:
            payload["tools"] = tools
            payload["tool_choice"] = "auto"
@ -361,7 +409,7 @@ class DeepSeekClient:
        max_iterations = 200  # 最大迭代次数
        iteration = 0
        all_tool_results = []  # 记录所有工具调用结果
-        
+
        while iteration < max_iterations:
            iteration += 1
            
@ -369,11 +417,14 @@ class DeepSeekClient:
            full_response = ""
            tool_calls = []
            current_thinking = ""
+            # 针对 append_to_file / modify_file 的占位结构，防止未定义变量导致异常
+            append_result = {"handled": False}
+            modify_result = {"handled": False}
            
            # 状态标志
            in_thinking = False
            thinking_printed = False
-            
+
            async for chunk in self.chat(messages, tools, stream=True):
                if "choices" not in chunk:
                    continue
@ -435,7 +486,7 @@ class DeepSeekClient:
                            if new_args:  # 只拼接非空内容
                                existing_call["function"]["arguments"] += new_args
            
-            self._print()  # 最终换行
+            self._print("")  # 最终换行
            
            # 如果思考还没结束（只调用工具没有文本），手动结束
            if in_thinking:
@ -603,6 +654,24 @@ class DeepSeekClient:
        full_response = ""
        thinking_content = ""
        in_thinking = False
+
+        # 如果思考模式且已有本任务的思考内容，补充到上下文，确保多次调用时思考不割裂
+        if (
+            self.thinking_mode
+            and not self.current_task_first_call
+            and self.current_task_thinking
+        ):
+            thinking_context = (
+                "\n=== 📋 本次任务的思考 ===\n"
+                f"{self.current_task_thinking}\n"
+                "=== 思考结束 ===\n"
+                "提示：以上是本轮任务先前的思考，请在此基础上继续。"
+            )
+            messages.append({
+                "role": "system",
+                "content": thinking_context
+            })
+            thinking_context_injected = True
        
        try:
            async for chunk in self.chat(messages, tools=None, stream=True):
@ -631,7 +700,7 @@ class DeepSeekClient:
                        full_response += content
                        self._print(content, end="", flush=True)
            
-            self._print()  # 最终换行
+            self._print("")  # 最终换行
            
            # 如果思考还没结束（极少情况），手动结束
            if in_thinking:
--- a/utils/context_manager.py
+++ b/utils/context_manager.py
@ -2,6 +2,8 @@

 import os
 import json
+import base64
+import mimetypes
 from copy import deepcopy
 from typing import Dict, List, Optional, Any
 from pathlib import Path
@ -16,6 +18,7 @@ try:
        TERMINAL_SANDBOX_MEMORY,
        PROJECT_MAX_STORAGE_MB,
    )
+    from config.model_profiles import get_model_prompt_replacements
 except ImportError:
    import sys
    from pathlib import Path
@ -31,6 +34,7 @@ except ImportError:
        TERMINAL_SANDBOX_MEMORY,
        PROJECT_MAX_STORAGE_MB,
    )
+    from config.model_profiles import get_model_prompt_replacements
 from utils.conversation_manager import ConversationManager

 class ContextManager:
@ -47,6 +51,7 @@ class ContextManager:
        self.file_annotations = {}  # 文件备注
        self.conversation_history = []  # 当前对话历史（内存中）
        self.todo_list: Optional[Dict[str, Any]] = None
+        self.has_images: bool = False
        
        # 新增：对话持久化管理器
        self.conversation_manager = ConversationManager(base_dir=self.data_dir)
@ -317,13 +322,16 @@ class ContextManager:
            project_path=project_path,
            thinking_mode=thinking_mode,
            run_mode=run_mode or ("thinking" if thinking_mode else "fast"),
-            initial_messages=[]
+            initial_messages=[],
+            model_key=getattr(self.main_terminal, "model_key", None),
+            has_images=False
        )
        
        # 重置当前状态
        self.current_conversation_id = conversation_id
        self.conversation_history = []
        self.todo_list = None
+        self.has_images = False
        
        print(f"📝 开始新对话: {conversation_id}")
        return conversation_id
@ -372,7 +380,14 @@ class ContextManager:
        self.project_path = resolved_project_path
        
        run_mode = metadata.get("run_mode")
+        model_key = metadata.get("model_key")
+        self.has_images = metadata.get("has_images", False)
        if self.main_terminal:
+            try:
+                if model_key:
+                    self.main_terminal.set_model(model_key)
+            except Exception:
+                pass
            try:
                if run_mode:
                    self.main_terminal.set_run_mode(run_mode)
@ -410,7 +425,9 @@ class ContextManager:
                project_path=str(self.project_path),
                todo_list=self.todo_list,
                thinking_mode=getattr(self.main_terminal, "thinking_mode", None) if hasattr(self, "main_terminal") else None,
-                run_mode=run_mode
+                run_mode=run_mode,
+                model_key=getattr(self.main_terminal, "model_key", None) if hasattr(self, "main_terminal") else None,
+                has_images=self.has_images
            )
            
            if success:
@ -584,12 +601,16 @@ class ContextManager:
        project_path = str(resolved_project_path)
        thinking_mode = metadata.get("thinking_mode", False)
        run_mode = metadata.get("run_mode") or ("thinking" if thinking_mode else "fast")
+        model_key = metadata.get("model_key")
+        has_images = metadata.get("has_images", False)

        compressed_conversation_id = self.conversation_manager.create_conversation(
            project_path=project_path,
            thinking_mode=thinking_mode,
            run_mode=run_mode,
-            initial_messages=compressed_messages
+            initial_messages=compressed_messages,
+            model_key=model_key,
+            has_images=has_images
        )

        return {
@ -615,12 +636,16 @@ class ContextManager:
        project_path = str(resolved_project_path)
        thinking_mode = metadata.get("thinking_mode", False)
        run_mode = metadata.get("run_mode") or ("thinking" if thinking_mode else "fast")
+        model_key = metadata.get("model_key")
+        has_images = metadata.get("has_images", False)

        duplicate_conversation_id = self.conversation_manager.create_conversation(
            project_path=project_path,
            thinking_mode=thinking_mode,
            run_mode=run_mode,
-            initial_messages=original_messages
+            initial_messages=original_messages,
+            model_key=model_key,
+            has_images=has_images
        )

        token_stats = conversation_data.get("token_statistics")
@ -699,7 +724,8 @@ class ContextManager:
        tool_call_id: Optional[str] = None,
        name: Optional[str] = None,
        metadata: Optional[Dict[str, Any]] = None,
-        reasoning_content: Optional[str] = None
+        reasoning_content: Optional[str] = None,
+        images: Optional[List[str]] = None
    ):
        """添加对话记录（改进版：集成自动保存 + 智能token统计）"""
        timestamp = datetime.now().isoformat()
@ -719,6 +745,9 @@ class ContextManager:
        
        if metadata:
            message["metadata"] = metadata
+        if images:
+            message["images"] = images
+            self.has_images = True
        
        # 如果是assistant消息且有工具调用，保存完整格式
        if role == "assistant" and tool_calls:
@ -1152,16 +1181,41 @@ class ContextManager:
            "is_overflow": sizes["total"] > MAX_CONTEXT_SIZE,
            "usage_percent": (sizes["total"] / MAX_CONTEXT_SIZE) * 100
        }
+    def _build_content_with_images(self, text: str, images: List[str]) -> Any:
+        """将文本与图片路径组合成多模态content，图片转换为data URI。"""
+        if not images:
+            return text
+        parts: List[Dict[str, Any]] = []
+        if text:
+            parts.append({"type": "text", "text": text})
+        for path in images:
+            try:
+                abs_path = Path(self.project_path) / path
+                if not abs_path.exists() or not abs_path.is_file():
+                    continue
+                mime, _ = mimetypes.guess_type(abs_path.name)
+                if not mime:
+                    mime = "image/png"
+                data = abs_path.read_bytes()
+                b64 = base64.b64encode(data).decode("utf-8")
+                parts.append({"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}})
+            except Exception:
+                continue
+        return parts if parts else text
+
    def build_messages(self, context: Dict, user_input: str) -> List[Dict]:
        """构建消息列表（添加终端内容注入）"""
-        # 加载系统提示
-        system_prompt = self.load_prompt("main_system")
+        # 加载系统提示（Qwen-VL 使用专用提示）
+        model_key = getattr(self.main_terminal, "model_key", "kimi") if hasattr(self, "main_terminal") else "kimi"
+        prompt_name = "main_system_qwenvl" if model_key == "qwen3-vl-plus" else "main_system"
+        system_prompt = self.load_prompt(prompt_name)
        
        # 格式化系统提示
        container_path = self.container_mount_path or "/workspace"
        container_cpus = self.container_cpu_limit
        container_memory = self.container_memory_limit
        project_storage = self.project_storage_limit
+        prompt_replacements = get_model_prompt_replacements(model_key)
        system_prompt = system_prompt.format(
            project_path=container_path,
            container_path=container_path,
@ -1170,7 +1224,8 @@ class ContextManager:
            project_storage=project_storage,
            file_tree=context["project_info"]["file_tree"],
            memory=context["memory"],
-            current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+            model_description=prompt_replacements.get("model_description", "")
        )
        
        messages = [
@ -1199,9 +1254,11 @@ class ContextManager:
                }
                messages.append(message)
            else:
+                images = conv.get("images") or (conv.get("metadata") or {}).get("images") or []
+                content_payload = self._build_content_with_images(conv["content"], images) if images else conv["content"]
                messages.append({
                    "role": conv["role"],
-                    "content": conv["content"]
+                    "content": content_payload
                })
        
        # 添加聚焦文件内容
--- a/utils/conversation_manager.py
+++ b/utils/conversation_manager.py
@ -30,6 +30,8 @@ class ConversationMetadata:
    total_messages: int
    total_tools: int
    run_mode: str = "fast"
+    model_key: Optional[str] = None
+    has_images: bool = False
    status: str = "active"  # active, archived, error

 class ConversationManager:
@ -86,6 +88,8 @@ class ConversationManager:
                "project_relative_path": metadata.get("project_relative_path"),
                "thinking_mode": metadata.get("thinking_mode", False),
                "run_mode": metadata.get("run_mode") or ("thinking" if metadata.get("thinking_mode") else "fast"),
+                "model_key": metadata.get("model_key"),
+                "has_images": metadata.get("has_images", False),
                "total_messages": metadata.get("total_messages", 0),
                "total_tools": metadata.get("total_tools", 0),
                "status": metadata.get("status", "active"),
@ -271,11 +275,13 @@ class ConversationManager:
        return data
    
    def create_conversation(
-        self, 
-        project_path: str, 
+        self,
+        project_path: str,
        thinking_mode: bool = False,
        run_mode: str = "fast",
-        initial_messages: List[Dict] = None
+        initial_messages: List[Dict] = None,
+        model_key: Optional[str] = None,
+        has_images: bool = False
    ) -> str:
        """
        创建新对话
@ -306,6 +312,8 @@ class ConversationManager:
                "project_relative_path": path_metadata["project_relative_path"],
                "thinking_mode": thinking_mode,
                "run_mode": normalized_mode,
+                "model_key": model_key,
+                "has_images": has_images,
                "total_messages": len(messages),
                "total_tools": self._count_tools_in_messages(messages),
                "status": "active"
@ -393,6 +401,8 @@ class ConversationManager:
                "project_relative_path": metadata.project_relative_path,
                "thinking_mode": metadata.thinking_mode,
                "run_mode": metadata.run_mode,
+                "model_key": conversation_data["metadata"].get("model_key"),
+                "has_images": conversation_data["metadata"].get("has_images", False),
                "total_messages": metadata.total_messages,
                "total_tools": metadata.total_tools,
                "status": metadata.status
@ -409,7 +419,9 @@ class ConversationManager:
        project_path: str = None,
        thinking_mode: bool = None,
        run_mode: Optional[str] = None,
-        todo_list: Optional[Dict] = None
+        todo_list: Optional[Dict] = None,
+        model_key: Optional[str] = None,
+        has_images: Optional[bool] = None
    ) -> bool:
        """
        保存对话（更新现有对话）
@ -457,6 +469,14 @@ class ConversationManager:
                existing_data["metadata"]["run_mode"] = normalized_mode
            elif "run_mode" not in existing_data["metadata"]:
                existing_data["metadata"]["run_mode"] = "thinking" if existing_data["metadata"].get("thinking_mode") else "fast"
+            if model_key is not None:
+                existing_data["metadata"]["model_key"] = model_key
+            elif "model_key" not in existing_data["metadata"]:
+                existing_data["metadata"]["model_key"] = None
+            if has_images is not None:
+                existing_data["metadata"]["has_images"] = bool(has_images)
+            elif "has_images" not in existing_data["metadata"]:
+                existing_data["metadata"]["has_images"] = False
            
            existing_data["metadata"]["total_messages"] = len(messages)
            existing_data["metadata"]["total_tools"] = self._count_tools_in_messages(messages)
--- a/web_server.py
+++ b/web_server.py
@ -20,6 +20,7 @@ from datetime import timedelta
 import time
 from datetime import datetime
 from collections import defaultdict, deque, Counter
+from config.model_profiles import get_model_profile
 from werkzeug.utils import secure_filename
 from werkzeug.routing import BaseConverter
 import secrets
@ -1487,7 +1488,8 @@ def update_thinking_mode(terminal: WebTerminal, workspace: UserWorkspace, userna
                    project_path=str(ctx.project_path),
                    todo_list=ctx.todo_list,
                    thinking_mode=terminal.thinking_mode,
-                    run_mode=terminal.run_mode
+                    run_mode=terminal.run_mode,
+                    model_key=getattr(terminal, "model_key", None)
                )
            except Exception as exc:
                print(f"[API] 保存思考模式到对话失败: {exc}")
@ -1504,11 +1506,64 @@ def update_thinking_mode(terminal: WebTerminal, workspace: UserWorkspace, userna
        })
    except Exception as exc:
        print(f"[API] 切换思考模式失败: {exc}")
+        code = 400 if isinstance(exc, ValueError) else 500
        return jsonify({
            "success": False,
            "error": str(exc),
            "message": "切换思考模式时发生异常"
-        }), 500
+        }), code
+
+
+@app.route('/api/model', methods=['POST'])
+@api_login_required
+@with_terminal
+@rate_limited("model_switch", 10, 60, scope="user")
+def update_model(terminal: WebTerminal, workspace: UserWorkspace, username: str):
+    """切换基础模型（快速/思考模型组合）。"""
+    try:
+        data = request.get_json() or {}
+        model_key = data.get("model_key")
+        if not model_key:
+            return jsonify({"success": False, "error": "缺少 model_key"}), 400
+
+        terminal.set_model(model_key)
+        # fast-only 时 run_mode 可能被强制为 fast
+        session["model_key"] = terminal.model_key
+        session["run_mode"] = terminal.run_mode
+        session["thinking_mode"] = terminal.thinking_mode
+
+        # 更新当前对话元数据
+        ctx = terminal.context_manager
+        if ctx.current_conversation_id:
+            try:
+                ctx.conversation_manager.save_conversation(
+                    conversation_id=ctx.current_conversation_id,
+                    messages=ctx.conversation_history,
+                    project_path=str(ctx.project_path),
+                    todo_list=ctx.todo_list,
+                    thinking_mode=terminal.thinking_mode,
+                    run_mode=terminal.run_mode,
+                    model_key=terminal.model_key,
+                    has_images=getattr(ctx, "has_images", False)
+                )
+            except Exception as exc:
+                print(f"[API] 保存模型到对话失败: {exc}")
+
+        status = terminal.get_status()
+        socketio.emit('status_update', status, room=f"user_{username}")
+
+        return jsonify({
+            "success": True,
+            "data": {
+                "model_key": terminal.model_key,
+                "run_mode": terminal.run_mode,
+                "thinking_mode": terminal.thinking_mode
+            }
+        })
+    except Exception as exc:
+        print(f"[API] 切换模型失败: {exc}")
+        code = 400 if isinstance(exc, ValueError) else 500
+        return jsonify({"success": False, "error": str(exc), "message": str(exc)}), code


@app.route('/api/personalization', methods=['GET'])
@ -2405,9 +2460,13 @@ def handle_message(data):
        return
    
    message = (data.get('message') or '').strip()
-    if not message:
+    images = data.get('images') or []
+    if not message and not images:
        emit('error', {'message': '消息不能为空'})
        return
+    if images and getattr(terminal, "model_key", None) != "qwen3-vl-plus":
+        emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL'})
+        return
    
    print(f"[WebSocket] 收到消息: {message}")
    debug_log(f"\n{'='*80}\n新任务开始: {message}\n{'='*80}")
@ -2469,7 +2528,8 @@ def handle_message(data):
        send_to_client(event_type, data)

    # 传递客户端ID
-    socketio.start_background_task(process_message_task, terminal, message, send_with_activity, client_sid, workspace, username)
+    images = data.get('images') or []
+    socketio.start_background_task(process_message_task, terminal, message, images, send_with_activity, client_sid, workspace, username)


@socketio.on('client_chunk_log')
@ -2952,14 +3012,14 @@ def get_current_conversation(terminal: WebTerminal, workspace: UserWorkspace, us
            "error": str(e)
        }), 500
    
-def process_message_task(terminal: WebTerminal, message: str, sender, client_sid, workspace: UserWorkspace, username: str):
+def process_message_task(terminal: WebTerminal, message: str, images, sender, client_sid, workspace: UserWorkspace, username: str):
    """在后台处理消息任务"""
    try:
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        
        # 创建可取消的任务
-        task = loop.create_task(handle_task_with_sender(terminal, workspace, message, sender, client_sid, username))
+        task = loop.create_task(handle_task_with_sender(terminal, workspace, message, images, sender, client_sid, username))
        
        entry = stop_flags.get(client_sid)
        if not isinstance(entry, dict):
@ -3034,7 +3094,7 @@ def detect_malformed_tool_call(text):
            
    return False

-async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspace, message, sender, client_sid, username: str):
+async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspace, message, images, sender, client_sid, username: str):
    """处理任务并发送消息 - 集成token统计版本"""
    web_terminal = terminal
    conversation_id = getattr(web_terminal.context_manager, "current_conversation_id", None)
@ -3050,7 +3110,7 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
    # 添加到对话历史
    history_len_before = len(getattr(web_terminal.context_manager, "conversation_history", []) or [])
    is_first_user_message = history_len_before == 0
-    web_terminal.context_manager.add_conversation("user", message)
+    web_terminal.context_manager.add_conversation("user", message, images=images)
    
    if is_first_user_message and getattr(web_terminal, "context_manager", None):
        try:
@ -4713,7 +4773,7 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
            except:
                result_data = {'output': tool_result}
            tool_failed = detect_tool_failure(result_data)
-            
+
            action_status = 'completed'
            action_message = None
            awaiting_flag = False
@ -4833,6 +4893,38 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
                "content": tool_result_content
            })

+            # 处理图片注入：必须紧跟在对应的 tool 消息之后，且工具成功时才插入
+            if (
+                function_name == "view_image"
+                and getattr(web_terminal, "pending_image_view", None)
+                and not tool_failed
+                and (isinstance(result_data, dict) and result_data.get("success") is not False)
+            ):
+                inj = web_terminal.pending_image_view
+                web_terminal.pending_image_view = None
+                injected_text = "这是一条系统控制发送的信息，并非用户主动发送，目的是返回你需要查看的图片。"
+                # 记录到对话历史
+                web_terminal.context_manager.add_conversation(
+                    "user",
+                    injected_text,
+                    images=[inj["path"]],
+                    metadata={"system_injected_image": True}
+                )
+                # 同步到当前消息列表（直接带多模态 content），保证顺序为 tool_call -> tool -> (系统代发)user
+                content_payload = web_terminal.context_manager._build_content_with_images(
+                    injected_text,
+                    [inj["path"]]
+                )
+                messages.append({
+                    "role": "user",
+                    "content": content_payload,
+                    "metadata": {"system_injected_image": True}
+                })
+                # 提示前端
+                sender('system_message', {
+                    'content': f'系统已按模型请求插入图片: {inj.get("path")}'
+                })
+
            if function_name != 'write_file_diff':
                await process_sub_agent_updates(messages, inline=True, after_tool_call_id=tool_call_id)
				`@ -0,0 +1 @@`
				`<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" width="48" height="48" fill="none" stroke="#000000" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="opacity:1;"><path d="M10 8h.01M12 12h.01M14 8h.01M16 12h.01M18 8h.01M6 8h.01M7 16h10m-9-4h.01"/><rect width="20" height="16" x="2" y="4" rx="2"/></svg>`