✨ feat(core): 支持LLM多图片响应，增强UI主题皮肤系统及优化JSON/Markdown处理 (#2062)

- 【LLM服务】
  - `LLMResponse` 模型现在支持 `images: list[bytes]`，允许模型返回多张图片。
  - LLM适配器 (`base.py`, `gemini.py`) 和 API 层 (`api.py`, `service.py`) 已更新以处理多图片响应。
  - 响应验证逻辑已调整，以检查 `images` 列表而非单个 `image_bytes`。
- 【UI渲染服务】
  - 引入组件“皮肤”（variant）概念，允许为同一组件提供不同视觉风格。
  - 改进了 `manifest.json` 的加载、合并和缓存机制，支持基础清单与皮肤清单的递归合并。
  - `ThemeManager` 现在会缓存已加载的清单，并在主题重载时清除缓存。
  - 增强了资源解析器 (`ResourceResolver`)，支持 `@` 命名空间路径和更健壮的相对路径处理。
  - 独立模板现在会继承主 Jinja 环境的过滤器。
- 【工具函数】
  - 引入 `dump_json_safely` 工具函数，用于更安全地序列化包含 Pydantic 模型、枚举等复杂类型的对象为 JSON。
  - LLM 服务中的请求体和缓存键生成已改用 `dump_json_safely`。
  - 优化了 `format_usage_for_markdown` 函数，改进了 Markdown 文本的格式化，确保块级元素前有正确换行，并正确处理段落内硬换行。

Co-authored-by: webjoin111 <455457521@qq.com>

zhenxun/services/llm/adapters/base.py

@@ -35,7 +35,7 @@ class ResponseData(BaseModel):
     """响应数据封装 - 支持所有高级功能"""
 
     text: str
-    image_bytes: bytes | None = None
+    images: list[bytes] | None = None
     usage_info: dict[str, Any] | None = None
     raw_response: dict[str, Any] | None = None
     tool_calls: list[LLMToolCall] | None = None
@@ -246,17 +246,17 @@ class BaseAdapter(ABC):
             if content:
                 content = content.strip()
 
-            image_bytes: bytes | None = None
+            images_bytes: list[bytes] = []
             if content and content.startswith("{") and content.endswith("}"):
                 try:
                     content_json = json.loads(content)
                     if "b64_json" in content_json:
-                        image_bytes = base64.b64decode(content_json["b64_json"])
+                        images_bytes.append(base64.b64decode(content_json["b64_json"]))
                         content = "[图片已生成]"
                     elif "data" in content_json and isinstance(
                         content_json["data"], str
                     ):
-                        image_bytes = base64.b64decode(content_json["data"])
+                        images_bytes.append(base64.b64decode(content_json["data"]))
                         content = "[图片已生成]"
 
                 except (json.JSONDecodeError, KeyError, binascii.Error):
@@ -273,7 +273,7 @@ class BaseAdapter(ABC):
                     if url_str.startswith("data:image/png;base64,"):
                         try:
                             b64_data = url_str.split(",", 1)[1]
-                            image_bytes = base64.b64decode(b64_data)
+                            images_bytes.append(base64.b64decode(b64_data))
                             content = content if content else "[图片已生成]"
                         except (IndexError, binascii.Error) as e:
                             logger.warning(f"解析OpenRouter Base64图片数据失败: {e}")
@@ -316,7 +316,7 @@ class BaseAdapter(ABC):
                 text=final_text,
                 tool_calls=parsed_tool_calls,
                 usage_info=usage_info,
-                image_bytes=image_bytes,
+                images=images_bytes if images_bytes else None,
                 raw_response=response_json,
             )
 

zhenxun/services/llm/adapters/gemini.py

@@ -408,7 +408,7 @@ class GeminiAdapter(BaseAdapter):
             parts = content_data.get("parts", [])
 
             text_content = ""
-            image_bytes: bytes | None = None
+            images_bytes: list[bytes] = []
             parsed_tool_calls: list["LLMToolCall"] | None = None
             thought_summary_parts = []
             answer_parts = []
@@ -423,10 +423,7 @@ class GeminiAdapter(BaseAdapter):
                 elif "inlineData" in part:
                     inline_data = part["inlineData"]
                     if "data" in inline_data:
-                        image_bytes = base64.b64decode(inline_data["data"])
-                        answer_parts.append(
-                            f"[图片已生成: {inline_data.get('mimeType', 'image')}]"
-                        )
+                        images_bytes.append(base64.b64decode(inline_data["data"]))
 
                 elif "functionCall" in part:
                     if parsed_tool_calls is None:
@@ -494,7 +491,7 @@ class GeminiAdapter(BaseAdapter):
             return ResponseData(
                 text=text_content,
                 tool_calls=parsed_tool_calls,
-                image_bytes=image_bytes,
+                images=images_bytes if images_bytes else None,
                 usage_info=usage_info,
                 raw_response=response_json,
                 grounding_metadata=grounding_metadata_obj,

zhenxun/services/llm/api.py

@@ -339,7 +339,7 @@ async def _generate_image_from_message(
 
             response = await model_instance.generate_response(messages, config=config)
 
-            if not response.image_bytes:
+            if not response.images:
                 error_text = response.text or "模型未返回图片数据。"
                 logger.warning(f"图片生成调用未返回图片，返回文本内容: {error_text}")
 

zhenxun/services/llm/manager.py

@@ -5,12 +5,12 @@ LLM 模型管理器
 """
 
 import hashlib
-import json
 import time
 from typing import Any
 
 from zhenxun.configs.config import Config
 from zhenxun.services.log import logger
+from zhenxun.utils.pydantic_compat import dump_json_safely
 
 from .config import validate_override_params
 from .config.providers import AI_CONFIG_GROUP, PROVIDERS_CONFIG_KEY, get_ai_config
@@ -43,7 +43,7 @@ def _make_cache_key(
 ) -> str:
     """生成缓存键"""
     config_str = (
-        json.dumps(override_config, sort_keys=True) if override_config else "None"
+        dump_json_safely(override_config, sort_keys=True) if override_config else "None"
     )
     key_data = f"{provider_model_name}:{config_str}"
     return hashlib.md5(key_data.encode()).hexdigest()

zhenxun/services/llm/service.py

@@ -13,6 +13,7 @@ from pydantic import BaseModel
 
 from zhenxun.services.log import logger
 from zhenxun.utils.log_sanitizer import sanitize_for_logging
+from zhenxun.utils.pydantic_compat import dump_json_safely
 
 from .adapters.base import RequestData
 from .config import LLMGenerationConfig
@@ -194,13 +195,15 @@ class LLMModel(LLMModelBase):
             sanitized_body = sanitize_for_logging(
                 request_data.body, context=sanitizer_req_context
             )
-            request_body_str = json.dumps(sanitized_body, ensure_ascii=False, indent=2)
+            request_body_str = dump_json_safely(
+                sanitized_body, ensure_ascii=False, indent=2
+            )
             logger.debug(f"📦 请求体: {request_body_str}")
 
             http_response = await http_client.post(
                 request_data.url,
                 headers=request_data.headers,
-                json=request_data.body,
+                content=dump_json_safely(request_data.body, ensure_ascii=False),
             )
 
             logger.debug(f"📥 响应状态码: {http_response.status_code}")
@@ -394,7 +397,7 @@ class LLMModel(LLMModelBase):
             return LLMResponse(
                 text=response_data.text,
                 usage_info=response_data.usage_info,
-                image_bytes=response_data.image_bytes,
+                images=response_data.images,
                 raw_response=response_data.raw_response,
                 tool_calls=response_tool_calls if response_tool_calls else None,
                 code_executions=response_data.code_executions,
@@ -424,7 +427,7 @@ class LLMModel(LLMModelBase):
 
             policy = config.validation_policy
             if policy:
-                if policy.get("require_image") and not parsed_data.image_bytes:
+                if policy.get("require_image") and not parsed_data.images:
                     if self.api_type == "gemini" and parsed_data.raw_response:
                         usage_metadata = parsed_data.raw_response.get(
                             "usageMetadata", {}

zhenxun/services/llm/types/content.py

@@ -425,7 +425,7 @@ class LLMResponse(BaseModel):
     """LLM 响应"""
 
     text: str
-    image_bytes: bytes | None = None
+    images: list[bytes] | None = None
     usage_info: dict[str, Any] | None = None
     raw_response: dict[str, Any] | None = None
     tool_calls: list[Any] | None = None