✨ feat(llm): 增强LLM服务，支持图片生成、响应验证与OpenRouter集成 (#2054)

* ✨ feat(llm): 增强LLM服务，支持图片生成、响应验证与OpenRouter集成

- 【新功能】统一图片生成与编辑API `create_image`，支持文生图、图生图及多图输入
- 【新功能】引入LLM响应验证机制，通过 `validation_policy` 和 `response_validator` 确保响应内容符合预期，例如强制返回图片
- 【新功能】适配OpenRouter API，扩展LLM服务提供商支持，并添加OpenRouter特定请求头
- 【重构】将日志净化逻辑重构至 `log_sanitizer` 模块，提供统一的净化入口，并应用于NoneBot消息、LLM请求/响应日志
- 【修复】优化Gemini适配器，正确解析图片生成响应中的Base64图片数据，并更新模型能力注册表

* ✨ feat(image): 优化图片生成响应并返回完整LLMResponse

* ✨ feat(llm): 为 OpenAI 兼容请求体添加日志净化

* 🐛 fix(ui): 截断UI调试HTML日志中的长base64图片数据

---------

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

zhenxun/services/llm/adapters/base.py

@@ -3,6 +3,9 @@ LLM 适配器基类和通用数据结构
 """
 
 from abc import ABC, abstractmethod
+import base64
+import binascii
+import json
 from typing import TYPE_CHECKING, Any
 
 from pydantic import BaseModel
@@ -32,6 +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
     usage_info: dict[str, Any] | None = None
     raw_response: dict[str, Any] | None = None
     tool_calls: list[LLMToolCall] | None = None
@@ -242,6 +246,38 @@ class BaseAdapter(ABC):
             if content:
                 content = content.strip()
 
+            image_bytes: bytes | None = None
+            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"])
+                        content = "[图片已生成]"
+                    elif "data" in content_json and isinstance(
+                        content_json["data"], str
+                    ):
+                        image_bytes = base64.b64decode(content_json["data"])
+                        content = "[图片已生成]"
+
+                except (json.JSONDecodeError, KeyError, binascii.Error):
+                    pass
+            elif (
+                "images" in message
+                and isinstance(message["images"], list)
+                and message["images"]
+            ):
+                image_info = message["images"][0]
+                if image_info.get("type") == "image_url":
+                    image_url_obj = image_info.get("image_url", {})
+                    url_str = image_url_obj.get("url", "")
+                    if url_str.startswith("data:image/png;base64,"):
+                        try:
+                            b64_data = url_str.split(",", 1)[1]
+                            image_bytes = 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}")
+
             parsed_tool_calls: list[LLMToolCall] | None = None
             if message_tool_calls := message.get("tool_calls"):
                 from ..types.models import LLMToolFunction
@@ -280,6 +316,7 @@ class BaseAdapter(ABC):
                 text=final_text,
                 tool_calls=parsed_tool_calls,
                 usage_info=usage_info,
+                image_bytes=image_bytes,
                 raw_response=response_json,
             )
 
@@ -450,6 +487,13 @@ class OpenAICompatAdapter(BaseAdapter):
         """准备高级请求 - OpenAI兼容格式"""
         url = self.get_api_url(model, self.get_chat_endpoint(model))
         headers = self.get_base_headers(api_key)
+        if model.api_type == "openrouter":
+            headers.update(
+                {
+                    "HTTP-Referer": "https://github.com/zhenxun-org/zhenxun_bot",
+                    "X-Title": "Zhenxun Bot",
+                }
+            )
         openai_messages = self.convert_messages_to_openai_format(messages)
 
         body = {

zhenxun/services/llm/adapters/gemini.py

@@ -2,6 +2,7 @@
 Gemini API 适配器
 """
 
+import base64
 from typing import TYPE_CHECKING, Any
 
 from zhenxun.services.log import logger
@@ -373,7 +374,16 @@ class GeminiAdapter(BaseAdapter):
         self.validate_response(response_json)
 
         try:
-            candidates = response_json.get("candidates", [])
+            if "image_generation" in response_json and isinstance(
+                response_json["image_generation"], dict
+            ):
+                candidates_source = response_json["image_generation"]
+            else:
+                candidates_source = response_json
+
+            candidates = candidates_source.get("candidates", [])
+            usage_info = response_json.get("usageMetadata")
+
             if not candidates:
                 logger.debug("Gemini响应中没有candidates。")
                 return ResponseData(text="", raw_response=response_json)
@@ -398,6 +408,7 @@ class GeminiAdapter(BaseAdapter):
             parts = content_data.get("parts", [])
 
             text_content = ""
+            image_bytes: bytes | None = None
             parsed_tool_calls: list["LLMToolCall"] | None = None
             thought_summary_parts = []
             answer_parts = []
@@ -409,6 +420,14 @@ class GeminiAdapter(BaseAdapter):
                     thought_summary_parts.append(part["thought"])
                 elif "thoughtSummary" in part:
                     thought_summary_parts.append(part["thoughtSummary"])
+                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')}]"
+                        )
+
                 elif "functionCall" in part:
                     if parsed_tool_calls is None:
                         parsed_tool_calls = []
@@ -475,6 +494,7 @@ class GeminiAdapter(BaseAdapter):
             return ResponseData(
                 text=text_content,
                 tool_calls=parsed_tool_calls,
+                image_bytes=image_bytes,
                 usage_info=usage_info,
                 raw_response=response_json,
                 grounding_metadata=grounding_metadata_obj,

zhenxun/services/llm/adapters/openai.py

@@ -21,7 +21,14 @@ class OpenAIAdapter(OpenAICompatAdapter):
 
     @property
     def supported_api_types(self) -> list[str]:
-        return ["openai", "deepseek", "zhipu", "general_openai_compat", "ark"]
+        return [
+            "openai",
+            "deepseek",
+            "zhipu",
+            "general_openai_compat",
+            "ark",
+            "openrouter",
+        ]
 
     def get_chat_endpoint(self, model: "LLMModel") -> str:
         """返回聊天完成端点"""