Agent/agent/llm/gemini.py

"""
Gemini Provider (HTTP API)

使用 httpx 直接调用 Gemini REST API，避免 google-generativeai SDK 的兼容性问题

参考：Resonote/llm/providers/gemini.py
"""

import os
import json
import sys
import httpx
from typing import List, Dict, Any, Optional


def _dump_llm_request(endpoint: str, payload: Dict[str, Any], model: str):
    """
    Dump完整的LLM请求用于调试（需要设置 AGENT_DEBUG=1）

    特别处理：
    - 图片base64数据：只显示前50字符 + 长度信息
    - Tools schema：完整显示
    - 输出到stderr，避免污染正常输出
    """
    if not os.getenv("AGENT_DEBUG"):
        return

    def truncate_images(obj):
        """递归处理对象，truncate图片base64数据"""
        if isinstance(obj, dict):
            result = {}
            for key, value in obj.items():
                # 处理 inline_data 中的 base64 图片
                if key == "inline_data" and isinstance(value, dict):
                    mime_type = value.get("mime_type", "unknown")
                    data = value.get("data", "")
                    data_size_kb = len(data) / 1024 if data else 0
                    result[key] = {
                        "mime_type": mime_type,
                        "data": f"<BASE64_IMAGE: {data_size_kb:.1f}KB, preview: {data[:50]}...>"
                    }
                else:
                    result[key] = truncate_images(value)
            return result
        elif isinstance(obj, list):
            return [truncate_images(item) for item in obj]
        else:
            return obj

    # 构造完整的调试信息
    debug_info = {
        "endpoint": endpoint,
        "model": model,
        "payload": truncate_images(payload)
    }

    # 输出到stderr
    print("\n" + "="*80, file=sys.stderr)
    print("[AGENT_DEBUG] LLM Request Dump", file=sys.stderr)
    print("="*80, file=sys.stderr)
    print(json.dumps(debug_info, indent=2, ensure_ascii=False), file=sys.stderr)
    print("="*80 + "\n", file=sys.stderr)


def _convert_messages_to_gemini(messages: List[Dict]) -> tuple[List[Dict], Optional[str]]:
    """
    将 OpenAI 格式消息转换为 Gemini 格式

    Returns:
        (gemini_contents, system_instruction)
    """
    contents = []
    system_instruction = None
    tool_parts_buffer = []

    def flush_tool_buffer():
        """合并连续的 tool 消息为单个 user 消息"""
        if tool_parts_buffer:
            contents.append({
                "role": "user",
                "parts": tool_parts_buffer.copy()
            })
            tool_parts_buffer.clear()

    for msg in messages:
        role = msg.get("role")

        # System 消息 -> system_instruction
        if role == "system":
            system_instruction = msg.get("content", "")
            continue

        # Tool 消息 -> functionResponse
        if role == "tool":
            tool_name = msg.get("name")
            content_text = msg.get("content", "")

            if not tool_name:
                print(f"[WARNING] Tool message missing 'name' field, skipping")
                continue

            # 尝试解析为 JSON
            try:
                parsed = json.loads(content_text) if content_text else {}
                if isinstance(parsed, list):
                    response_data = {"result": parsed}
                else:
                    response_data = parsed
            except (json.JSONDecodeError, ValueError):
                response_data = {"result": content_text}

            # 添加到 buffer
            tool_parts_buffer.append({
                "functionResponse": {
                    "name": tool_name,
                    "response": response_data
                }
            })
            continue

        # 非 tool 消息：先 flush buffer
        flush_tool_buffer()

        content = msg.get("content", "")
        tool_calls = msg.get("tool_calls")

        # Assistant 消息 + tool_calls
        if role == "assistant" and tool_calls:
            parts = []
            if content and (isinstance(content, str) and content.strip()):
                parts.append({"text": content})

            # 转换 tool_calls 为 functionCall
            for tc in tool_calls:
                func = tc.get("function", {})
                func_name = func.get("name", "")
                func_args_str = func.get("arguments", "{}")
                try:
                    func_args = json.loads(func_args_str) if isinstance(func_args_str, str) else func_args_str
                except json.JSONDecodeError:
                    func_args = {}

                parts.append({
                    "functionCall": {
                        "name": func_name,
                        "args": func_args
                    }
                })

            if parts:
                contents.append({
                    "role": "model",
                    "parts": parts
                })
            continue

        # 处理多模态消息（content 为数组）
        if isinstance(content, list):
            parts = []
            for item in content:
                item_type = item.get("type")

                # 文本部分
                if item_type == "text":
                    text = item.get("text", "")
                    if text.strip():
                        parts.append({"text": text})

                # 图片部分（OpenAI format -> Gemini format）
                elif item_type == "image_url":
                    image_url = item.get("image_url", {})
                    url = image_url.get("url", "")

                    # 处理 data URL (data:image/png;base64,...)
                    if url.startswith("data:"):
                        # 解析 MIME type 和 base64 数据
                        # 格式：data:image/png;base64,<base64_data>
                        try:
                            header, base64_data = url.split(",", 1)
                            mime_type = header.split(";")[0].replace("data:", "")

                            parts.append({
                                "inline_data": {
                                    "mime_type": mime_type,
                                    "data": base64_data
                                }
                            })
                        except Exception as e:
                            print(f"[WARNING] Failed to parse image data URL: {e}")

            if parts:
                gemini_role = "model" if role == "assistant" else "user"
                contents.append({
                    "role": gemini_role,
                    "parts": parts
                })
            continue

        # 普通文本消息（content 为字符串）
        if isinstance(content, str):
            # 跳过空消息
            if not content.strip():
                continue

            gemini_role = "model" if role == "assistant" else "user"
            contents.append({
                "role": gemini_role,
                "parts": [{"text": content}]
            })
            continue

    # Flush 剩余的 tool messages
    flush_tool_buffer()

    # 合并连续的 user 消息（Gemini 要求严格交替）
    merged_contents = []
    i = 0
    while i < len(contents):
        current = contents[i]

        if current["role"] == "user":
            merged_parts = current["parts"].copy()
            j = i + 1
            while j < len(contents) and contents[j]["role"] == "user":
                merged_parts.extend(contents[j]["parts"])
                j += 1

            merged_contents.append({
                "role": "user",
                "parts": merged_parts
            })
            i = j
        else:
            merged_contents.append(current)
            i += 1

    return merged_contents, system_instruction


def _convert_tools_to_gemini(tools: List[Dict]) -> List[Dict]:
    """
    将 OpenAI 工具格式转换为 Gemini REST API 格式

    OpenAI: [{"type": "function", "function": {"name": "...", "parameters": {...}}}]
    Gemini API: [{"functionDeclarations": [{"name": "...", "parameters": {...}}]}]
    """
    if not tools:
        return []

    function_declarations = []
    for tool in tools:
        if tool.get("type") == "function":
            func = tool.get("function", {})

            # 清理不支持的字段
            parameters = func.get("parameters", {})
            if "properties" in parameters:
                cleaned_properties = {}
                for prop_name, prop_def in parameters["properties"].items():
                    # 移除 default 字段
                    cleaned_prop = {k: v for k, v in prop_def.items() if k != "default"}
                    cleaned_properties[prop_name] = cleaned_prop

                # Gemini API 需要完整的 schema
                cleaned_parameters = {
                    "type": "object",
                    "properties": cleaned_properties
                }
                if "required" in parameters:
                    cleaned_parameters["required"] = parameters["required"]

                parameters = cleaned_parameters

            function_declarations.append({
                "name": func.get("name"),
                "description": func.get("description", ""),
                "parameters": parameters
            })

    return [{"functionDeclarations": function_declarations}] if function_declarations else []


def create_gemini_llm_call(
    base_url: Optional[str] = None,
    api_key: Optional[str] = None
):
    """
    创建 Gemini LLM 调用函数（HTTP API）

    Args:
        base_url: Gemini API base URL（默认使用 Google 官方）
        api_key: API key（默认从环境变量读取）

    Returns:
        async 函数
    """
    base_url = base_url or "https://generativelanguage.googleapis.com/v1beta"
    api_key = api_key or os.getenv("GEMINI_API_KEY")

    if not api_key:
        raise ValueError("GEMINI_API_KEY not found")

    # 创建 HTTP 客户端
    client = httpx.AsyncClient(
        headers={"x-goog-api-key": api_key},
        timeout=httpx.Timeout(120.0, connect=10.0)
    )

    async def gemini_llm_call(
        messages: List[Dict[str, Any]],
        model: str = "gemini-2.5-pro",
        tools: Optional[List[Dict]] = None,
        **kwargs
    ) -> Dict[str, Any]:
        """
        调用 Gemini REST API

        Args:
            messages: OpenAI 格式消息
            model: 模型名称
            tools: OpenAI 格式工具列表
            **kwargs: 其他参数

        Returns:
            {
                "content": str,
                "tool_calls": List[Dict] | None,
                "prompt_tokens": int,
                "completion_tokens": int,
                "finish_reason": str,
                "cost": float
            }
        """
        # 转换消息
        contents, system_instruction = _convert_messages_to_gemini(messages)

        print(f"\n[Gemini HTTP] Converted {len(contents)} messages: {[c['role'] for c in contents]}")

        # 构建请求
        endpoint = f"{base_url}/models/{model}:generateContent"
        payload = {"contents": contents}

        # 添加 system instruction
        if system_instruction:
            payload["systemInstruction"] = {"parts": [{"text": system_instruction}]}

        # 添加工具
        if tools:
            gemini_tools = _convert_tools_to_gemini(tools)
            if gemini_tools:
                payload["tools"] = gemini_tools

        # Debug: dump完整请求（需要设置 AGENT_DEBUG=1）
        _dump_llm_request(endpoint, payload, model)

        # 调用 API
        try:
            response = await client.post(endpoint, json=payload)
            response.raise_for_status()
            gemini_resp = response.json()

        except httpx.HTTPStatusError as e:
            error_body = e.response.text
            print(f"[Gemini HTTP] Error {e.response.status_code}: {error_body}")
            raise
        except Exception as e:
            print(f"[Gemini HTTP] Request failed: {e}")
            raise

        # Debug: 输出原始响应（如果启用）
        if os.getenv("AGENT_DEBUG"):
            print("\n[AGENT_DEBUG] Gemini Response:", file=sys.stderr)
            print(json.dumps(gemini_resp, ensure_ascii=False, indent=2)[:2000], file=sys.stderr)
            print("\n", file=sys.stderr)

        # 解析响应
        content = ""
        tool_calls = None
        finish_reason = "stop"  # 默认值

        candidates = gemini_resp.get("candidates", [])
        if candidates:
            candidate = candidates[0]

            # 提取 finish_reason（Gemini -> OpenAI 格式映射）
            gemini_finish_reason = candidate.get("finishReason", "STOP")
            if gemini_finish_reason == "STOP":
                finish_reason = "stop"
            elif gemini_finish_reason == "MAX_TOKENS":
                finish_reason = "length"
            elif gemini_finish_reason in ("SAFETY", "RECITATION"):
                finish_reason = "content_filter"
            elif gemini_finish_reason == "MALFORMED_FUNCTION_CALL":
                finish_reason = "stop"  # 映射为 stop，但在 content 中包含错误信息
            else:
                finish_reason = gemini_finish_reason.lower()  # 保持原值，转小写

            # 检查是否有错误
            if gemini_finish_reason == "MALFORMED_FUNCTION_CALL":
                # Gemini 返回了格式错误的函数调用
                # 提取 finishMessage 中的内容作为 content
                finish_message = candidate.get("finishMessage", "")
                print(f"[Gemini HTTP] Warning: MALFORMED_FUNCTION_CALL\n{finish_message}")
                content = f"[模型尝试调用工具但格式错误]\n\n{finish_message}"
            else:
                # 正常解析
                parts = candidate.get("content", {}).get("parts", [])

                # 提取文本
                for part in parts:
                    if "text" in part:
                        content += part.get("text", "")

                # 提取 functionCall
                for i, part in enumerate(parts):
                    if "functionCall" in part:
                        if tool_calls is None:
                            tool_calls = []

                        fc = part["functionCall"]
                        name = fc.get("name", "")
                        args = fc.get("args", {})

                        tool_calls.append({
                            "id": f"call_{i}",
                            "type": "function",
                            "function": {
                                "name": name,
                                "arguments": json.dumps(args, ensure_ascii=False)
                            }
                        })

        # 提取 usage
        usage_meta = gemini_resp.get("usageMetadata", {})
        prompt_tokens = usage_meta.get("promptTokenCount", 0)
        completion_tokens = usage_meta.get("candidatesTokenCount", 0)

        return {
            "content": content,
            "tool_calls": tool_calls,
            "prompt_tokens": prompt_tokens,
            "completion_tokens": completion_tokens,
            "finish_reason": finish_reason,
            "cost": 0.0
        }

    return gemini_llm_call