src-tauri/src/proxy/mappers/claude/request.rs

// Claude 请求转换 (Claude → Gemini v1internal)
// 对应 transformClaudeRequestIn

use super::models::*;
use crate::proxy::mappers::signature_store::get_thought_signature; // Deprecated, kept for fallback
use crate::proxy::mappers::tool_result_compressor;
use crate::proxy::session_manager::SessionManager;
use serde_json::{json, Value};
use std::collections::HashMap;

// ===== Safety Settings Configuration =====

/// Safety threshold levels for Gemini API
/// Can be configured via GEMINI_SAFETY_THRESHOLD environment variable
#[derive(Debug, Clone, Copy, PartialEq)]
pub enum SafetyThreshold {
    /// Disable all safety filters (default for proxy compatibility)
    Off,
    /// Block low probability and above
    BlockLowAndAbove,
    /// Block medium probability and above
    BlockMediumAndAbove,
    /// Only block high probability content
    BlockOnlyHigh,
    /// Don't block anything (BLOCK_NONE)
    BlockNone,
}

impl SafetyThreshold {
    /// Get threshold from environment variable or default to Off
    pub fn from_env() -> Self {
        match std::env::var("GEMINI_SAFETY_THRESHOLD").as_deref() {
            Ok("OFF") | Ok("off") => SafetyThreshold::Off,
            Ok("LOW") | Ok("low") => SafetyThreshold::BlockLowAndAbove,
            Ok("MEDIUM") | Ok("medium") => SafetyThreshold::BlockMediumAndAbove,
            Ok("HIGH") | Ok("high") => SafetyThreshold::BlockOnlyHigh,
            Ok("NONE") | Ok("none") => SafetyThreshold::BlockNone,
            _ => SafetyThreshold::Off, // Default: maintain current behavior
        }
    }

    /// Convert to Gemini API threshold string
    pub fn to_gemini_threshold(&self) -> &'static str {
        match self {
            SafetyThreshold::Off => "OFF",
            SafetyThreshold::BlockLowAndAbove => "BLOCK_LOW_AND_ABOVE",
            SafetyThreshold::BlockMediumAndAbove => "BLOCK_MEDIUM_AND_ABOVE",
            SafetyThreshold::BlockOnlyHigh => "BLOCK_ONLY_HIGH",
            SafetyThreshold::BlockNone => "BLOCK_NONE",
        }
    }
}

/// Build safety settings based on configuration
fn build_safety_settings() -> Value {
    let threshold = SafetyThreshold::from_env();
    let threshold_str = threshold.to_gemini_threshold();

    json!([
        { "category": "HARM_CATEGORY_HARASSMENT", "threshold": threshold_str },
        { "category": "HARM_CATEGORY_HATE_SPEECH", "threshold": threshold_str },
        { "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": threshold_str },
        { "category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": threshold_str },
        { "category": "HARM_CATEGORY_CIVIC_INTEGRITY", "threshold": threshold_str },
    ])
}

/// 清理消息中的 cache_control 字段
///
/// 这个函数会深度遍历所有消息内容块,移除 cache_control 字段。
/// 这是必要的,因为:
/// 1. VS Code 等客户端会将历史消息(包含 cache_control)原封不动发回
/// 2. Anthropic API 不接受请求中包含 cache_control 字段
/// 3. 即使是转发到 Gemini,也应该清理以保持协议纯净性
///
/// [FIX #593] 增强版本:添加详细日志用于调试 MCP 工具兼容性问题
pub fn clean_cache_control_from_messages(messages: &mut [Message]) {
    tracing::info!(
        "[DEBUG-593] Starting cache_control cleanup for {} messages",
        messages.len()
    );

    let mut total_cleaned = 0;

    for (idx, msg) in messages.iter_mut().enumerate() {
        if let MessageContent::Array(blocks) = &mut msg.content {
            for (block_idx, block) in blocks.iter_mut().enumerate() {
                match block {
                    ContentBlock::Thinking { cache_control, .. } => {
                        if cache_control.is_some() {
                            tracing::info!(
                                "[ISSUE-744] Found cache_control in Thinking block at message[{}].content[{}]: {:?}",
                                idx,
                                block_idx,
                                cache_control
                            );
                            *cache_control = None;
                            total_cleaned += 1;
                        }
                    }
                    ContentBlock::Image { cache_control, .. } => {
                        if cache_control.is_some() {
                            tracing::debug!(
                                "[Cache-Control-Cleaner] Removed cache_control from Image block at message[{}].content[{}]",
                                idx,
                                block_idx
                            );
                            *cache_control = None;
                            total_cleaned += 1;
                        }
                    }
                    ContentBlock::Document { cache_control, .. } => {
                        if cache_control.is_some() {
                            tracing::debug!(
                                "[Cache-Control-Cleaner] Removed cache_control from Document block at message[{}].content[{}]",
                                idx,
                                block_idx
                            );
                            *cache_control = None;
                            total_cleaned += 1;
                        }
                    }
                    ContentBlock::ToolUse { cache_control, .. } => {
                        if cache_control.is_some() {
                            tracing::debug!(
                                "[Cache-Control-Cleaner] Removed cache_control from ToolUse block at message[{}].content[{}]",
                                idx,
                                block_idx
                            );
                            *cache_control = None;
                            total_cleaned += 1;
                        }
                    }
                    _ => {}
                }
            }
        }
    }

    if total_cleaned > 0 {
        tracing::info!(
            "[DEBUG-593] Cache control cleanup complete: removed {} cache_control fields",
            total_cleaned
        );
    } else {
        tracing::debug!("[DEBUG-593] No cache_control fields found");
    }
}

/// [FIX #593] 递归深度清理 JSON 中的 cache_control 字段
///
/// 用于处理嵌套结构和非标准位置的 cache_control。
/// 这是最后一道防线,确保发送给 Antigravity 的请求中不包含任何 cache_control。
fn deep_clean_cache_control(value: &mut Value) {
    match value {
        Value::Object(map) => {
            if map.remove("cache_control").is_some() {
                tracing::debug!("[DEBUG-593] Removed cache_control from nested JSON object");
            }
            for (_, v) in map.iter_mut() {
                deep_clean_cache_control(v);
            }
        }
        Value::Array(arr) => {
            for item in arr.iter_mut() {
                deep_clean_cache_control(item);
            }
        }
        _ => {}
    }
}

/// [FIX #564] Sort blocks in assistant messages to ensure thinking blocks are first
///
/// When context compression (kilo) reorders message blocks, thinking blocks may appear
/// after text blocks. Claude/Anthropic API requires thinking blocks to be first if
/// any thinking blocks exist in the message. This function pre-sorts blocks to ensure
/// thinking/redacted_thinking blocks always come before other block types.
fn sort_thinking_blocks_first(messages: &mut [Message]) {
    for msg in messages.iter_mut() {
        if msg.role == "assistant" {
            if let MessageContent::Array(blocks) = &mut msg.content {
                // [FIX #709] Triple-stage partition: [Thinking, Text, ToolUse]
                // This ensures protocol compliance while maintaining logical order.

                let mut thinking_blocks: Vec<ContentBlock> = Vec::new();
                let mut text_blocks: Vec<ContentBlock> = Vec::new();
                let mut tool_blocks: Vec<ContentBlock> = Vec::new();
                let mut other_blocks: Vec<ContentBlock> = Vec::new();

                let original_len = blocks.len();
                let mut needs_reorder = false;
                let mut saw_non_thinking = false;

                for (_i, block) in blocks.iter().enumerate() {
                    match block {
                        ContentBlock::Thinking { .. } | ContentBlock::RedactedThinking { .. } => {
                            if saw_non_thinking {
                                needs_reorder = true;
                            }
                        }
                        ContentBlock::Text { .. } => {
                            saw_non_thinking = true;
                        }
                        ContentBlock::ToolUse { .. } => {
                            saw_non_thinking = true;
                            // Check if tool is after text (this is normal, but we want a strict group order)
                        }
                        _ => saw_non_thinking = true,
                    }
                }

                if needs_reorder || original_len > 1 {
                    // For safety, we always perform the triple partition if there's more than one block.
                    // This also handles empty text block filtering.
                    for block in blocks.drain(..) {
                        match &block {
                            ContentBlock::Thinking { .. }
                            | ContentBlock::RedactedThinking { .. } => {
                                thinking_blocks.push(block);
                            }
                            ContentBlock::Text { text } => {
                                // Filter out purely empty or structural text like "(no content)"
                                if !text.trim().is_empty() && text != "(no content)" {
                                    text_blocks.push(block);
                                }
                            }
                            ContentBlock::ToolUse { .. } => {
                                tool_blocks.push(block);
                            }
                            _ => {
                                other_blocks.push(block);
                            }
                        }
                    }

                    // Reconstruct in strict order: Thinking -> Text/Other -> Tool
                    blocks.extend(thinking_blocks);
                    blocks.extend(text_blocks);
                    blocks.extend(other_blocks);
                    blocks.extend(tool_blocks);

                    if needs_reorder {
                        tracing::warn!(
                            "[FIX #709] Reordered assistant messages to [Thinking, Text, Tool] structure."
                        );
                    }
                }
            }
        }
    }
}

/// 合并 ClaudeRequest 中连续的同角色消息
///
/// 场景: 当从 Spec/Plan 模式切换回编码模式时，可能出现连续两条 "user" 消息
/// (一条是 ToolResult，一条是 <system-reminder>)。
/// 这会违反角色交替规则，导致 400 报错。
pub fn merge_consecutive_messages(messages: &mut Vec<Message>) {
    if messages.len() <= 1 {
        return;
    }

    let mut merged: Vec<Message> = Vec::with_capacity(messages.len());
    let old_messages = std::mem::take(messages);
    let mut messages_iter = old_messages.into_iter();

    if let Some(mut current) = messages_iter.next() {
        for next in messages_iter {
            if current.role == next.role {
                // 合并内容
                match (&mut current.content, next.content) {
                    (MessageContent::Array(current_blocks), MessageContent::Array(next_blocks)) => {
                        current_blocks.extend(next_blocks);
                    }
                    (MessageContent::Array(current_blocks), MessageContent::String(next_text)) => {
                        current_blocks.push(ContentBlock::Text { text: next_text });
                    }
                    (MessageContent::String(current_text), MessageContent::String(next_text)) => {
                        *current_text = format!("{}\n\n{}", current_text, next_text);
                    }
                    (MessageContent::String(current_text), MessageContent::Array(next_blocks)) => {
                        let mut new_blocks = vec![ContentBlock::Text {
                            text: current_text.clone(),
                        }];
                        new_blocks.extend(next_blocks);
                        current.content = MessageContent::Array(new_blocks);
                    }
                }
            } else {
                merged.push(current);
                current = next;
            }
        }
        merged.push(current);
    }

    *messages = merged;
}

/// 转换 Claude 请求为 Gemini v1internal 格式

/// [FIX #709] Reorder serialized Gemini parts to ensure thinking blocks are first
fn reorder_gemini_parts(parts: &mut Vec<Value>) {
    if parts.len() <= 1 {
        return;
    }

    let mut thinking_parts = Vec::new();
    let mut text_parts = Vec::new();
    let mut tool_parts = Vec::new();
    let mut other_parts = Vec::new();

    for part in parts.drain(..) {
        if part.get("thought").and_then(|t| t.as_bool()) == Some(true) {
            thinking_parts.push(part);
        } else if part.get("functionCall").is_some() {
            tool_parts.push(part);
        } else if let Some(text) = part.get("text").and_then(|t| t.as_str()) {
            // Filter empty text parts that might have been created during merging
            if !text.trim().is_empty() && text != "(no content)" {
                text_parts.push(part);
            }
        } else {
            other_parts.push(part);
        }
    }

    parts.extend(thinking_parts);
    parts.extend(text_parts);
    parts.extend(other_parts);
    parts.extend(tool_parts);
}

pub fn transform_claude_request_in(
    claude_req: &ClaudeRequest,
    project_id: &str,
    is_retry: bool,
    account_id: Option<&str>,
    _session_id: &str,
    token: Option<&crate::proxy::token_manager::ProxyToken>, // [NEW] 支持动态规格
) -> Result<Value, String> {
    let message_count = claude_req.messages.len();

    // [CRITICAL FIX] 预先清理所有消息中的 cache_control 字段
    // 这解决了 VS Code 插件等客户端在多轮对话中将历史消息的 cache_control 字段
    // 原封不动发回导致的 "Extra inputs are not permitted" 错误
    let mut cleaned_req = claude_req.clone();

    // [FIX #813] 合并连续的同角色消息 (Consecutive User Messages)
    // 确保请求符合 Anthropic 和 Gemini 的角色交替协议
    merge_consecutive_messages(&mut cleaned_req.messages);

    clean_cache_control_from_messages(&mut cleaned_req.messages);

    // [FIX #564] Pre-sort thinking blocks to be first in assistant messages
    // This handles cases where context compression (kilo) incorrectly reorders blocks
    sort_thinking_blocks_first(&mut cleaned_req.messages);

    // [FIX #1747] If thinking is auto-enabled by model default (e.g. Opus) but no
    // ThinkingConfig was provided by the client, inject a default config with a budget
    // to prevent 'thinking requires a budget' errors from upstream APIs.
    if cleaned_req.thinking.is_none() && should_enable_thinking_by_default(&cleaned_req.model) {
        let default_budget = crate::proxy::model_specs::get_thinking_budget(&cleaned_req.model, token);
        tracing::info!(
            "[Thinking-Mode] Injecting default ThinkingConfig (budget={}) for model: {}",
            default_budget,
            cleaned_req.model
        );
        cleaned_req.thinking = Some(ThinkingConfig {
            type_: "enabled".to_string(),
            budget_tokens: Some(default_budget as u32),
            effort: None,
        });
    }

    let claude_req = &cleaned_req; // 后续使用清理后的请求

    // [NEW] Generate session ID for signature tracking
    // This enables session-isolated signature storage, preventing cross-conversation pollution
    let session_id = SessionManager::extract_session_id(claude_req);
    tracing::debug!("[Claude-Request] Session ID: {}", session_id);

    // 检测是否有联网工具 (server tool or built-in tool)
    let has_web_search_tool = claude_req
        .tools
        .as_ref()
        .map(|tools| {
            tools.iter().any(|t| {
                t.is_web_search()
                    || t.name.as_deref() == Some("google_search")
                    || t.name.as_deref() == Some("builtin_web_search")
                    || t.type_.as_deref() == Some("web_search_20250305")
                    || t.type_.as_deref() == Some("builtin_web_search")
            })
        })
        .unwrap_or(false);

    // 用于存储 tool_use id -> name 映射
    let mut tool_id_to_name: HashMap<String, String> = HashMap::new();

    // 检测是否有 mcp__ 开头的工具
    let has_mcp_tools = claude_req
        .tools
        .as_ref()
        .map(|tools| {
            tools.iter().any(|t| {
                t.name
                    .as_deref()
                    .map(|n| n.starts_with("mcp__"))
                    .unwrap_or(false)
            })
        })
        .unwrap_or(false);

    // [New] 预先构建工具名称到原始 Schema 的映射，用于后续参数类型修正
    let mut tool_name_to_schema = HashMap::new();
    if let Some(tools) = &claude_req.tools {
        for tool in tools {
            if let (Some(name), Some(schema)) = (&tool.name, &tool.input_schema) {
                tool_name_to_schema.insert(name.clone(), schema.clone());
            }
        }
    }

    // 1. System Instruction (注入动态身份防护 & MCP XML 协议)
    let system_instruction =
        build_system_instruction(&claude_req.system, &claude_req.model, has_mcp_tools);

    //  Map model name (Use standard mapping)
    // [IMPROVED] 提取 web search 模型为常量，便于维护
    const WEB_SEARCH_FALLBACK_MODEL: &str = "gemini-2.5-flash";

    let mapped_model = crate::proxy::common::model_mapping::map_claude_model_to_gemini(&claude_req.model);

    // 将 Claude 工具转为 Value 数组以便探测联网
    let tools_val: Option<Vec<Value>> = claude_req.tools.as_ref().map(|list| {
        list.iter()
            .map(|t| serde_json::to_value(t).unwrap_or(json!({})))
            .collect()
    });

    // Resolve grounding config
    let config = crate::proxy::mappers::common_utils::resolve_request_config(
        &claude_req.model,
        &mapped_model,
        &tools_val,
        claude_req.size.as_deref(),    // [NEW] Pass size parameter
        claude_req.quality.as_deref(), // [NEW] Pass quality parameter
        None,                          // [NEW] image_size
        None,                          // body
    );

    // [CRITICAL FIX] Disable dummy thought injection for Vertex AI
    // [CRITICAL FIX] Disable dummy thought injection for Vertex AI
    // Vertex AI rejects thinking blocks without valid signatures
    // Even if thinking is enabled, we should NOT inject dummy blocks for historical messages
    let allow_dummy_thought = false;

    // Check if thinking is enabled in the request
    let thinking_type = claude_req.thinking.as_ref().map(|t| t.type_.as_str());
    let mut is_thinking_enabled = thinking_type == Some("enabled") || thinking_type == Some("adaptive") 
        || (thinking_type.is_none() && should_enable_thinking_by_default(&claude_req.model));

    // [NEW FIX] Check if target model supports thinking
    // Only models with "-thinking" suffix or Claude models support thinking
    // Regular Gemini models (gemini-2.5-flash, gemini-2.5-pro) do NOT support thinking
    // [FIX #1557] Allow "pro" models (e.g. gemini-3-pro, gemini-2.0-pro) to be recognized as thinking capable
    let target_model_supports_thinking = mapped_model.contains("-thinking")
        || mapped_model.starts_with("claude-")
        || mapped_model.contains("gemini-2.0-pro")
        || mapped_model.contains("gemini-3-pro")
        || mapped_model.contains("gemini-3.1-pro")
        // [FIX #2167] gemini-3-flash / gemini-3.1-flash 支持 thinking，必须纳入识别范围
        || mapped_model.contains("gemini-3-flash")
        || mapped_model.contains("gemini-3.1-flash");

    if is_thinking_enabled && !target_model_supports_thinking {
        tracing::warn!(
            "[Thinking-Mode] Target model '{}' does not support thinking. Force disabling thinking mode.",
            mapped_model
        );
        is_thinking_enabled = false;
    }

    // [REMOVED] 智能降级检查 (should_disable_thinking_due_to_history)
    // 原因: 该检查过于激进，会导致 Claude Code CLI 在历史记录不完美时永久禁用思考模式 (Issue #2006)
    // 现在的策略是依赖 thinking_utils.rs 中的 Recovery 机制来修复历史，而不是禁用思考。


    // [FIX #295 & #298] If thinking enabled but no signature available,
    // disable thinking to prevent Gemini 3 Pro rejection
    if is_thinking_enabled {
        let global_sig = get_thought_signature();

        // Check if there are any thinking blocks in message history
        let has_thinking_history = claude_req.messages.iter().any(|m| {
            if m.role == "assistant" {
                if let MessageContent::Array(blocks) = &m.content {
                    return blocks
                        .iter()
                        .any(|b| matches!(b, ContentBlock::Thinking { .. }));
                }
            }
            false
        });

        // Check if there are function calls in the request
        let has_function_calls = claude_req.messages.iter().any(|m| {
            if let MessageContent::Array(blocks) = &m.content {
                blocks
                    .iter()
                    .any(|b| matches!(b, ContentBlock::ToolUse { .. }))
            } else {
                false
            }
        });

        // [FIX #298] For first-time thinking requests (no thinking history),
        // we use permissive mode and let upstream handle validation.
        // We only enforce strict signature checks when function calls are involved.
        let needs_signature_check = has_function_calls;

        if !has_thinking_history && is_thinking_enabled {
            tracing::info!(
                "[Thinking-Mode] First thinking request detected. Using permissive mode - \
                 signature validation will be handled by upstream API."
            );
        }

        if needs_signature_check
            && !has_valid_signature_for_function_calls(
                &claude_req.messages,
                &global_sig,
                &session_id,
            )
        {
            // [FIX #2167] Flash 模型无签名时使用哨兵值而不是禁用 thinking
            // 禁用 thinking 会导致模型失去思考能力，哨兵值可让 Gemini 跳过签名校验
            let is_flash_model = mapped_model.contains("gemini-3-flash")
                || mapped_model.contains("gemini-3.1-flash");
            if is_flash_model {
                tracing::info!(
                    "[Thinking-Mode] [FIX #2167] No signature for flash model function calls. \
                     Will rely on sentinel injection in build_contents."
                );
                // 保持 is_thinking_enabled = true，由 build_contents 内的哨兵处理覆盖
            } else {
                tracing::warn!(
                    "[Thinking-Mode] [FIX #295] No valid signature found for function calls. \
                     Disabling thinking to prevent Gemini 3 Pro rejection."
                );
                is_thinking_enabled = false;
            }
        }
    }

    // 4. Generation Config & Thinking (Pass final is_thinking_enabled)
    let generation_config = build_generation_config(
        claude_req,
        &mapped_model,
        has_web_search_tool,
        is_thinking_enabled,
        token, // [NEW] 传递 token 用于动态限额
    );

    // 2. Contents (Messages)
    let contents = build_google_contents(
        &claude_req.messages,
        claude_req,
        &mut tool_id_to_name,
        &tool_name_to_schema,
        is_thinking_enabled,
        allow_dummy_thought,
        &mapped_model,
        &session_id,
        is_retry,
    )?;

    // 3. Tools
    let tools = build_tools(&claude_req.tools, has_web_search_tool, &mapped_model)?;

    // 5. Safety Settings (configurable via GEMINI_SAFETY_THRESHOLD env var)
    let safety_settings = build_safety_settings();

    // Build inner request
    let mut inner_request = json!({
        "contents": contents,
        "safetySettings": safety_settings,
    });

    if let Some(sys_inst) = system_instruction {
        inner_request["systemInstruction"] = sys_inst;
    }

    if !generation_config.is_null() {
        println!("DEBUG: Assigning generation_config: {}", generation_config);
        inner_request["generationConfig"] = generation_config;
    }

    if let Some(tools_val) = tools {
        inner_request["tools"] = tools_val;
        // 显式设置工具配置模式为 VALIDATED
        inner_request["toolConfig"] = json!({
            "functionCallingConfig": {
                "mode": "VALIDATED"
            }
        });
    }


    // 深度清理 [undefined] 字符串 (Cherry Studio 等客户端常见注入)
    crate::proxy::mappers::common_utils::deep_clean_undefined(&mut inner_request, 0);


    if config.inject_google_search && !has_web_search_tool {
        crate::proxy::mappers::common_utils::inject_google_search_tool(&mut inner_request, Some(&mapped_model));
    }

    // Inject imageConfig if present (for image generation models)
    if let Some(image_config) = config.image_config {
        if let Some(obj) = inner_request.as_object_mut() {
            // 1. Remove tools (image generation does not support tools)
            obj.remove("tools");

            // 2. Remove systemInstruction (image generation does not support system prompts)
            obj.remove("systemInstruction");

            // 3. Clean generationConfig (remove responseMimeType, responseModalities etc.)
            let gen_config = obj.entry("generationConfig").or_insert_with(|| json!({}));
            if let Some(gen_obj) = gen_config.as_object_mut() {
                // [RESOLVE #1694] Check image thinking mode
                let image_thinking_mode = crate::proxy::config::get_image_thinking_mode();
                if image_thinking_mode == "disabled" {
                    tracing::debug!(
                        "[Claude-Request] Image thinking mode disabled: enforcing includeThoughts=false for {}",
                        mapped_model
                    );
                    gen_obj.insert(
                        "thinkingConfig".to_string(),
                        json!({
                            "includeThoughts": false
                        }),
                    );
                }

                gen_obj.remove("responseMimeType");
                gen_obj.remove("responseModalities");
                gen_obj.insert("imageConfig".to_string(), image_config);
            }
        }
    }

    // [ADDED v4.1.24] 注入稳定 sessionId 对齐官方规范
    if let Some(account_id) = account_id {
        inner_request["sessionId"] = json!(crate::proxy::common::session::derive_session_id(account_id));
    }

    // 生成 requestId
    // [CHANGED v4.1.24] Structured requestId to match official format
    let request_id = format!("agent/antigravity/{}/{}", &session_id[..session_id.len().min(8)], message_count);

    // 构建最终请求体
    let mut body = json!({
        "project": project_id,
        "requestId": request_id,
        "request": inner_request,
        "model": config.final_model,
        "userAgent": "antigravity",
        // [CHANGED v4.1.24] Use "agent" for all non-image requests
        "requestType": if config.request_type == "image_gen" { "image_gen" } else { "agent" },
    });

    // 如果提供了 metadata.user_id，则复用为 sessionId
    if let Some(metadata) = &claude_req.metadata {
        if let Some(user_id) = &metadata.user_id {
            body["request"]["sessionId"] = json!(user_id);
        }
    }

    // [FIX #593] 最后一道防线: 递归深度清理所有 cache_control 字段
    // 确保发送给 Antigravity 的请求中不包含任何 cache_control
    deep_clean_cache_control(&mut body);
    tracing::debug!("[DEBUG-593] Final deep clean complete, request ready to send");

    Ok(body)
}



/// Check if thinking mode should be enabled by default for a given model
///
/// Claude Code v2.0.67+ enables thinking by default for Opus 4.5 models.
/// This function determines if the model should have thinking enabled
/// when no explicit thinking configuration is provided.
fn should_enable_thinking_by_default(model: &str) -> bool {
    let model_lower = model.to_lowercase();

    // Enable thinking by default for Opus 4.5 and 4.6 variants
    if model_lower.contains("opus-4-5")
        || model_lower.contains("opus-4.5")
        || model_lower.contains("opus-4-6")
        || model_lower.contains("opus-4.6")
    {
        tracing::debug!(
            "[Thinking-Mode] Auto-enabling thinking for Opus model: {}",
            model
        );
        return true;
    }

    // Also enable for explicit thinking model variants
    if model_lower.contains("-thinking") {
        return true;
    }

    // [FIX #1557] Enable thinking by default for Gemini Pro models (gemini-3-pro, gemini-2.0-pro)
    // These models prioritize reasoning but clients might not send thinking config for them
    // unless they have "-thinking" suffix (which they don't in Antigravity mapping)
    if model_lower.contains("gemini-2.0-pro")
        || model_lower.contains("gemini-3-pro")
        || model_lower.contains("gemini-3.1-pro")
    {
        tracing::debug!(
            "[Thinking-Mode] Auto-enabling thinking for Gemini Pro model: {}",
            model
        );
        return true;
    }

    // [FEATURE] 为 gemini-3-flash / gemini-3.1-flash 自动开启 thinking
    // 让 Cherry Studio 等客户端即使未显式传 thinking.type 也能获取思维链内容
    if model_lower.contains("gemini-3-flash") || model_lower.contains("gemini-3.1-flash") {
        tracing::debug!(
            "[Thinking-Mode] Auto-enabling thinking for Flash model: {}",
            model
        );
        return true;
    }

    false
}

/// Minimum length for a valid thought_signature
const MIN_SIGNATURE_LENGTH: usize = 50;

/// [FIX #295] Check if we have any valid signature available for function calls
/// This prevents Gemini 3 Pro from rejecting requests due to missing thought_signature
///
/// [NEW FIX] Now also checks Session Cache to support retry scenarios
fn has_valid_signature_for_function_calls(
    messages: &[Message],
    global_sig: &Option<String>,
    session_id: &str, // NEW: Add session_id parameter
) -> bool {
    // 1. Check global store (deprecated but kept for compatibility)
    if let Some(sig) = global_sig {
        if sig.len() >= MIN_SIGNATURE_LENGTH {
            tracing::debug!(
                "[Signature-Check] Found valid signature in global store (len: {})",
                sig.len()
            );
            return true;
        }
    }

    // 2. [NEW] Check Session Cache - this is critical for retry scenarios
    // When retrying, the signature may not be in messages but exists in Session Cache
    if let Some(sig) = crate::proxy::SignatureCache::global().get_session_signature(session_id) {
        if sig.len() >= MIN_SIGNATURE_LENGTH {
            tracing::info!(
                "[Signature-Check] Found valid signature in SESSION cache (session: {}, len: {})",
                session_id,
                sig.len()
            );
            return true;
        }
    }

    // 3. Check if any message has a thinking block with valid signature
    for msg in messages.iter().rev() {
        if msg.role == "assistant" {
            if let MessageContent::Array(blocks) = &msg.content {
                for block in blocks {
                    if let ContentBlock::Thinking {
                        signature: Some(sig),
                        ..
                    } = block
                    {
                        if sig.len() >= MIN_SIGNATURE_LENGTH {
                            tracing::debug!(
                                "[Signature-Check] Found valid signature in message history (len: {})",
                                sig.len()
                            );
                            return true;
                        }
                    }
                }
            }
        }
    }

    tracing::warn!(
        "[Signature-Check] No valid signature found (session: {}, checked: global store, session cache, message history)",
        session_id
    );
    false
}

/// 构建 System Instruction (支持动态身份映射与 Prompt 隔离)
fn build_system_instruction(
    system: &Option<SystemPrompt>,
    _model_name: &str,
    has_mcp_tools: bool,
) -> Option<Value> {
    let mut parts = Vec::new();

    // [NEW] Antigravity 身份指令 (原始简化版)
    let antigravity_identity = "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.\n\
    You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.\n\
    **Absolute paths only**\n\
    **Proactiveness**";

    // [HYBRID] 检查用户是否已提供 Antigravity 身份
    let mut user_has_antigravity = false;
    if let Some(sys) = system {
        match sys {
            SystemPrompt::String(text) => {
                if text.contains("You are Antigravity") {
                    user_has_antigravity = true;
                }
            }
            SystemPrompt::Array(blocks) => {
                for block in blocks {
                    if block.block_type == "text" && block.text.contains("You are Antigravity") {
                        user_has_antigravity = true;
                        break;
                    }
                }
            }
        }
    }

    // 如果用户没有提供 Antigravity 身份,则注入
    if !user_has_antigravity {
        parts.push(json!({"text": antigravity_identity}));
    }

    // [NEW] 注入全局系统提示词 (紧跟 Antigravity 身份之后)
    let global_prompt_config = crate::proxy::config::get_global_system_prompt();
    if global_prompt_config.enabled && !global_prompt_config.content.trim().is_empty() {
        parts.push(json!({"text": global_prompt_config.content}));
    }

    // 添加用户的系统提示词
    if let Some(sys) = system {
        match sys {
            SystemPrompt::String(text) => {
                // [MODIFIED] No longer filter "You are an interactive CLI tool"
                // We pass everything through to ensure Flash/Lite models get full instructions
                parts.push(json!({"text": text}));
            }
            SystemPrompt::Array(blocks) => {
                for block in blocks {
                    if block.block_type == "text" {
                        // [MODIFIED] No longer filter "You are an interactive CLI tool"
                        parts.push(json!({"text": block.text}));
                    }
                }
            }
        }
    }

    // [NEW] MCP XML Bridge: 如果存在 mcp__ 开头的工具，注入专用的调用协议
    // 这能有效规避部分 MCP 链路在标准的 tool_use 协议下解析不稳的问题
    if has_mcp_tools {
        let mcp_xml_prompt = "\n\
        ==== MCP XML 工具调用协议 (Workaround) ====\n\
        当你需要调用名称以 `mcp__` 开头的 MCP 工具时：\n\
        1) 优先尝试 XML 格式调用：输出 `<mcp__tool_name>{\"arg\":\"value\"}</mcp__tool_name>`。\n\
        2) 必须直接输出 XML 块，无需 markdown 包装，内容为 JSON 格式的入参。\n\
        3) 这种方式具有更高的连通性和容错性，适用于大型结果返回场景。\n\
        ===========================================";
        parts.push(json!({"text": mcp_xml_prompt}));
    }

    // 如果用户没有提供任何系统提示词,添加结束标记
    if !user_has_antigravity {
        parts.push(json!({"text": "\n--- [SYSTEM_PROMPT_END] ---"}));
    }

    Some(json!({
        "role": "user",
        "parts": parts
    }))
}

/// 构建 Contents (Messages)
fn build_contents(
    content: &MessageContent,
    is_assistant: bool,
    _claude_req: &ClaudeRequest,
    is_thinking_enabled: bool,
    session_id: &str,
    allow_dummy_thought: bool,
    is_retry: bool,
    tool_id_to_name: &mut HashMap<String, String>,
    tool_name_to_schema: &HashMap<String, Value>,
    mapped_model: &str,
    last_thought_signature: &mut Option<String>,
    pending_tool_use_ids: &mut Vec<String>,
    last_user_task_text_normalized: &mut Option<String>,
    previous_was_tool_result: &mut bool,
    _existing_tool_result_ids: &std::collections::HashSet<String>,
) -> Result<Vec<Value>, String> {
    let mut parts = Vec::new();
    // Track tool results in the current turn to identify missing ones
    let mut current_turn_tool_result_ids = std::collections::HashSet::new();

    // Track if we have already seen non-thinking content in this message.
    // Anthropic/Gemini protocol: Thinking blocks MUST come first.
    let mut saw_non_thinking = false;

    match content {
        MessageContent::String(text) => {
            if text != "(no content)" {
                let trimmed = text.trim();
                if !trimmed.is_empty() {
                    parts.push(json!({"text": trimmed}));
                }
            }
        }
        MessageContent::Array(blocks) => {
            for item in blocks {
                match item {
                    ContentBlock::Text { text } => {
                        if text != "(no content)" && !text.trim().is_empty() {
                            // [NEW] 任务去重逻辑: 如果当前是 User 消息，且紧跟在 ToolResult 之后，
                            // 检查该文本是否与上一轮任务描述完全一致。
                            if !is_assistant && *previous_was_tool_result {
                                if let Some(last_task) = last_user_task_text_normalized {
                                    let current_normalized =
                                        text.replace(|c: char| c.is_whitespace(), "");
                                    if !current_normalized.is_empty()
                                        && current_normalized == *last_task
                                    {
                                        tracing::info!("[Claude-Request] Dropping duplicated task text echo (len: {})", text.len());
                                        continue;
                                    }
                                }
                            }

                            parts.push(json!({"text": text}));
                            saw_non_thinking = true;

                            // 记录最近一次 User 任务文本用于后续比对
                            if !is_assistant {
                                *last_user_task_text_normalized =
                                    Some(text.replace(|c: char| c.is_whitespace(), ""));
                            }
                            *previous_was_tool_result = false;
                        }
                    }
                    ContentBlock::Thinking {
                        thinking,
                        signature,
                        ..
                    } => {
                        tracing::debug!(
                            "[DEBUG-TRANSFORM] Processing thinking block. Sig: {:?}",
                            signature
                        );

                        // [HOTFIX] Gemini Protocol Enforcement: Thinking block MUST be the first block.
                        // If we already have content (like Text), we must downgrade this thinking block to Text.
                        if saw_non_thinking || !parts.is_empty() {
                            tracing::warn!("[Claude-Request] Thinking block found at non-zero index (prev parts: {}). Downgrading to Text.", parts.len());
                            if !thinking.trim().is_empty() {
                                parts.push(json!({
                                    "text": thinking.trim()
                                }));
                                saw_non_thinking = true;
                            }
                            continue;
                        }

                        // [FIX] If thinking is disabled (smart downgrade), convert ALL thinking blocks to text
                        // to avoid "thinking is disabled but message contains thinking" error
                        if !is_thinking_enabled {
                            tracing::warn!("[Claude-Request] Thinking disabled. Downgrading thinking block to text.");
                            if !thinking.trim().is_empty() {
                                parts.push(json!({
                                    "text": thinking.trim()
                                }));
                                saw_non_thinking = true;
                            }
                            continue;
                        }

                        // [FIX] Empty thinking blocks cause "Field required" errors.
                        // We downgrade them to Text to avoid structural errors and signature mismatch.
                        if thinking.is_empty() {
                            tracing::warn!("[Claude-Request] Empty thinking block detected. Downgrading to Text.");
                            parts.push(json!({
                                "text": "..."
                            }));
                            continue;
                        }

                        // [FIX #752] Strict signature validation
                        // Only use signatures that are cached and compatible with the target model
                        if let Some(sig) = signature {
                            // Check signature length first - if it's too short, it's definitely invalid
                            if sig.len() < MIN_SIGNATURE_LENGTH {
                                tracing::warn!(
                                    "[Thinking-Signature] Signature too short (len: {} < {}), downgrading to text.",
                                    sig.len(), MIN_SIGNATURE_LENGTH
                                );
                                parts.push(json!({"text": thinking}));
                                saw_non_thinking = true;
                                continue;
                            }

                            let cached_family =
                                crate::proxy::SignatureCache::global().get_signature_family(sig);

                            match cached_family {
                                Some(family) => {
                                    // Check compatibility
                                    // [NEW] If is_retry is true, force incompatibility to strip historical signatures
                                    // which likely caused the previous 400 error.
                                    let compatible =
                                        !is_retry && is_model_compatible(&family, mapped_model);

                                    if !compatible {
                                        tracing::warn!(
                                            "[Thinking-Signature] {} signature (Family: {}, Target: {}). Downgrading to text.",
                                            if is_retry { "Stripping historical" } else { "Incompatible" },
                                            family, mapped_model
                                        );
                                        parts.push(json!({"text": thinking}));
                                        saw_non_thinking = true;
                                        continue;
                                    }
                                    // Compatible and not a retry: use signature
                                    *last_thought_signature = Some(sig.clone());
                                    let mut part = json!({
                                        "text": thinking,
                                        "thought": true,
                                        "thoughtSignature": sig
                                    });
                                    crate::proxy::common::json_schema::clean_json_schema(&mut part);
                                    parts.push(part);
                                }
                                None => {
                                    // For JSON tool calling compatibility, if signature is long enough but unknown,
                                    // we should trust it rather than downgrade to text
                                    if sig.len() >= MIN_SIGNATURE_LENGTH {
                                        tracing::debug!(
                                            "[Thinking-Signature] Unknown signature origin but valid length (len: {}), using as-is for JSON tool calling.",
                                            sig.len()
                                        );
                                        *last_thought_signature = Some(sig.clone());
                                        let mut part = json!({
                                            "text": thinking,
                                            "thought": true,
                                            "thoughtSignature": sig
                                        });
                                        crate::proxy::common::json_schema::clean_json_schema(
                                            &mut part,
                                        );
                                        parts.push(part);
                                    } else {
                                        // Unknown and too short: downgrade to text for safety
                                        tracing::warn!(
                                            "[Thinking-Signature] Unknown signature origin and too short (len: {}). Downgrading to text for safety.",
                                            sig.len()
                                        );
                                        parts.push(json!({"text": thinking}));
                                        saw_non_thinking = true;
                                        continue;
                                    }
                                }
                            }
                        } else {
                            // No signature: downgrade to text
                            tracing::warn!(
                                "[Thinking-Signature] No signature provided. Downgrading to text."
                            );
                            parts.push(json!({"text": thinking}));
                            saw_non_thinking = true;
                        }
                    }
                    ContentBlock::RedactedThinking { data } => {
                        // [FIX] 将 RedactedThinking 作为普通文本处理，保留上下文
                        tracing::debug!("[Claude-Request] Degrade RedactedThinking to text");
                        parts.push(json!({
                            "text": format!("[Redacted Thinking: {}]", data)
                        }));
                        saw_non_thinking = true;
                        continue;
                    }
                    ContentBlock::Image { source, .. } => {
                        if source.source_type == "base64" {
                            parts.push(json!({
                                "inlineData": {
                                    "mimeType": source.media_type,
                                    "data": source.data
                                }
                            }));
                            saw_non_thinking = true;
                        }
                    }
                    ContentBlock::Document { source, .. } => {
                        if source.source_type == "base64" {
                            parts.push(json!({
                                "inlineData": {
                                    "mimeType": source.media_type,
                                    "data": source.data
                                }
                            }));
                            saw_non_thinking = true;
                        }
                    }
                    ContentBlock::ToolUse {
                        id,
                        name,
                        input,
                        signature,
                        ..
                    } => {
                        let mut final_input = input.clone();

                        // [New] 利用通用引擎修正参数类型 (替代以前硬编码的 shell 工具修复逻辑)
                        if let Some(original_schema) = tool_name_to_schema.get(name) {
                            crate::proxy::common::json_schema::fix_tool_call_args(
                                &mut final_input,
                                original_schema,
                            );
                        }

                        let mut part = json!({
                            "functionCall": {
                                "name": name,
                                "args": final_input,
                                "id": id
                            }
                        });
                        saw_non_thinking = true;

                        // Track pending tool use
                        if is_assistant {
                            pending_tool_use_ids.push(id.clone());
                        }

                        // 存储 id -> name 映射
                        tool_id_to_name.insert(id.clone(), name.clone());

                        // Signature resolution logic
                        // Priority: Client -> Context -> Session Cache -> Tool Cache -> Global Store (deprecated)
                        // [CRITICAL FIX] Do NOT use skip_thought_signature_validator for Vertex AI
                        // Vertex AI rejects this sentinel value, so we only add thoughtSignature if we have a real one
                        let final_sig = signature.as_ref()
                            .or(last_thought_signature.as_ref())
                            .cloned()
                            .or_else(|| {
                                // [NEW v3.3.17] Try session-based signature cache first (Layer 3)
                                // This provides conversation-level isolation
                                crate::proxy::SignatureCache::global().get_session_signature(session_id)
                                    .map(|s| {
                                        tracing::info!(
                                            "[Claude-Request] Recovered signature from SESSION cache (session: {}, len: {})",
                                            session_id, s.len()
                                        );
                                        s
                                    })
                            })
                            .or_else(|| {
                                // Try tool-specific signature cache (Layer 1)
                                crate::proxy::SignatureCache::global().get_tool_signature(id)
                                    .map(|s| {
                                        tracing::info!("[Claude-Request] Recovered signature from TOOL cache for tool_id: {}", id);
                                        s
                                    })
                            })
                            .or_else(|| {
                                // [DEPRECATED] Global store fallback - kept for backward compatibility
                                let global_sig = get_thought_signature();
                                if global_sig.is_some() {
                                    tracing::warn!(
                                        "[Claude-Request] Using deprecated GLOBAL thought_signature fallback (length: {}). \
                                         This indicates session cache miss.",
                                        global_sig.as_ref().unwrap().len()
                                    );
                                }
                                global_sig
                            });
                        // [FIX #752] Validate signature before using
                        // Only add thoughtSignature if we have a valid and compatible one
                        if let Some(sig) = final_sig {
                            // [NEW] If this is a retry, do NOT backfill signatures to avoid issues.
                            if is_retry && signature.is_none() {
                                tracing::warn!("[Tool-Signature] Skipping signature backfill for tool_use: {} during retry.", id);
                            } else {
                                // Check signature length first - if it's too short, it's definitely invalid
                                if sig.len() < MIN_SIGNATURE_LENGTH {
                                    tracing::warn!(
                                        "[Tool-Signature] Signature too short for tool_use: {} (len: {} < {}), skipping.",
                                        id, sig.len(), MIN_SIGNATURE_LENGTH
                                    );
                                } else {
                                    // Check signature compatibility (optional for tool_use)
                                    let cached_family = crate::proxy::SignatureCache::global()
                                        .get_signature_family(&sig);

                                    let should_use_sig = match cached_family {
                                        Some(family) => {
                                            // For tool_use, check compatibility
                                            if is_model_compatible(&family, mapped_model) {
                                                true
                                            } else {
                                                tracing::warn!(
                                                    "[Tool-Signature] Incompatible signature for tool_use: {} (Family: {}, Target: {})",
                                                    id, family, mapped_model
                                                );
                                                false
                                            }
                                        }
                                        None => {
                                            // For JSON tool calling compatibility, if signature is long enough but unknown,
                                            // we should trust it rather than drop it
                                            if sig.len() >= MIN_SIGNATURE_LENGTH {
                                                tracing::debug!(
                                                    "[Tool-Signature] Unknown signature origin but valid length (len: {}) for tool_use: {}, using as-is for JSON tool calling.",
                                                    sig.len(), id
                                                );
                                                true
                                            } else {
                                                // Unknown and too short: only use in non-thinking mode
                                                if is_thinking_enabled {
                                                    tracing::warn!(
                                                        "[Tool-Signature] Unknown signature origin and too short for tool_use: {} (len: {}). Dropping in thinking mode.",
                                                        id, sig.len()
                                                    );
                                                    false
                                                } else {
                                                    // In non-thinking mode, allow unknown signatures
                                                    true
                                                }
                                            }
                                        }
                                    };
                                    if should_use_sig {
                                        part["thoughtSignature"] = json!(sig);
                                    }
                                }
                            }
                        } else {
                            // [NEW] Handle missing signature for Gemini thinking models
                            // Use skip_thought_signature_validator as a sentinel value
                            let is_google_cloud = mapped_model.starts_with("projects/");
                            if is_thinking_enabled && !is_google_cloud {
                                tracing::debug!("[Tool-Signature] Adding GEMINI_SKIP_SIGNATURE for tool_use: {}", id);
                                part["thoughtSignature"] =
                                    json!("skip_thought_signature_validator");
                            }
                        }
                        parts.push(part);
                    }
                    ContentBlock::ToolResult {
                        tool_use_id,
                        content,
                        is_error,
                        ..
                    } => {
                        // Mark this tool ID as resolved in this turn
                        current_turn_tool_result_ids.insert(tool_use_id.clone());
                        // 优先使用之前记录的 name，否则用 tool_use_id
                        let func_name = tool_id_to_name
                            .get(tool_use_id)
                            .cloned()
                            .unwrap_or_else(|| tool_use_id.clone());

                        // [FIX #593] 工具输出压缩: 处理超大工具输出
                        // 使用智能压缩策略(浏览器快照、大文件提示等)
                        let mut compacted_content = content.clone();
                        if let Some(blocks) = compacted_content.as_array_mut() {
                            tool_result_compressor::sanitize_tool_result_blocks(blocks);
                        }

                        // Smart Truncation: No longer stripping images from Tool Results
                        // Tool results should pass transparency. If images are present, map them to inlineData.
                        let mut extra_parts = Vec::new();

                        let mut merged_content = match &compacted_content {
                            serde_json::Value::String(s) => s.clone(),
                            serde_json::Value::Array(arr) => {
                                let mut texts = Vec::new();
                                for block in arr {
                                    if let Some(text) = block.get("text").and_then(|v| v.as_str()) {
                                        texts.push(text.to_string());
                                    } else if block.get("source").is_some() {
                                        if block.get("type").and_then(|v| v.as_str()) == Some("image") {
                                            let source = block.get("source").unwrap();
                                            if let (Some(media_type), Some(data)) = (
                                                source.get("media_type").and_then(|v| v.as_str()),
                                                source.get("data").and_then(|v| v.as_str())
                                            ) {
                                                extra_parts.push(json!({
                                                    "inlineData": {
                                                        "mimeType": media_type,
                                                        "data": data
                                                    }
                                                }));
                                            }
                                        }
                                    }
                                }
                                texts.join("\n")
                            }
                            _ => content.to_string(),
                        };

                        // Smart Truncation: max chars limit
                        const MAX_TOOL_RESULT_CHARS: usize = 200_000;
                        if merged_content.len() > MAX_TOOL_RESULT_CHARS {
                            tracing::warn!(
                                "Truncating tool result from {} chars to {}",
                                merged_content.len(),
                                MAX_TOOL_RESULT_CHARS
                            );
                            let mut truncated = merged_content
                                .chars()
                                .take(MAX_TOOL_RESULT_CHARS)
                                .collect::<String>();
                            truncated.push_str("\n...[truncated output]");
                            merged_content = truncated;
                        }

                        // [优化] 如果结果为空，注入显式确认信号，防止模型幻觉
                        if merged_content.trim().is_empty() {
                            if is_error.unwrap_or(false) {
                                merged_content =
                                    "Tool execution failed with no output.".to_string();
                            } else {
                                merged_content = "Command executed successfully.".to_string();
                            }
                        }

                        let mut part = json!({
                            "functionResponse": {
                                "name": func_name,
                                "response": {"result": merged_content},
                                "id": tool_use_id
                            }
                        });

                        // [FIX] Tool Result 也需要回填签名（如果上下文中有）
                        if let Some(sig) = last_thought_signature.as_ref() {
                            part["thoughtSignature"] = json!(sig);
                        }

                        parts.push(part);

                        // 追加图片 parts
                        for extra in extra_parts {
                            parts.push(extra);
                        }

                        // 标记状态，用于下一条 User 消息的去重判断
                        *previous_was_tool_result = true;
                    }
                    // ContentBlock::RedactedThinking handled above at line 583
                    ContentBlock::ServerToolUse { .. }
                    | ContentBlock::WebSearchToolResult { .. } => {
                        // 搜索结果 block 不应由客户端发回给上游 (已由 tool_result 替代)
                        continue;
                    }
                }
            }
        }
    }

    // If this is a User message, check if we need to inject missing tool results
    if !is_assistant && !pending_tool_use_ids.is_empty() {
        let missing_ids: Vec<_> = pending_tool_use_ids
            .iter()
            .filter(|id| !current_turn_tool_result_ids.contains(*id))
            .cloned()
            .collect();

        if !missing_ids.is_empty() {
            tracing::warn!("[Elastic-Recovery] Injecting {} missing tool results into User message (IDs: {:?})", missing_ids.len(), missing_ids);
            for id in missing_ids.iter().rev() {
                // Insert in reverse order to maintain order at index 0? No, just insert at 0.
                let name = tool_id_to_name.get(id).cloned().unwrap_or(id.clone());
                let synthetic_part = json!({
                    "functionResponse": {
                        "name": name,
                        "response": {
                            "result": "Tool execution interrupted. No result provided."
                        },
                        "id": id
                    }
                });
                // Prepend to ensure they are present before any text
                parts.insert(0, synthetic_part);
            }
        }
        // All pending IDs are now handled (either present or injected)
        pending_tool_use_ids.clear();
    }

    // Fix for "Thinking enabled, assistant message must start with thinking block" 400 error
    // [Optimization] Apply this to ALL assistant messages in history, not just the last one.
    // Vertex AI requires every assistant message to start with a thinking block when thinking is enabled.
    if allow_dummy_thought && is_assistant && is_thinking_enabled {
        let has_thought_part = parts.iter().any(|p| {
            p.get("thought").and_then(|v| v.as_bool()).unwrap_or(false)
                || p.get("thoughtSignature").is_some()
                || p.get("thought").and_then(|v| v.as_str()).is_some() // 某些情况下可能是 text + thought: true 的组合
        });

        if !has_thought_part {
            // Prepend a dummy thinking block to satisfy Gemini v1internal requirements
            parts.insert(
                0,
                json!({
                    "text": "Thinking...",
                    "thought": true
                }),
            );
            tracing::debug!(
                "Injected dummy thought block for historical assistant message at index {}",
                parts.len()
            );
        } else {
            // [Crucial Check] 即使有 thought 块，也必须保证它位于 parts 的首位 (Index 0)
            // 且必须包含 thought: true 标记
            let first_is_thought = parts.get(0).map_or(false, |p| {
                (p.get("thought").is_some() || p.get("thoughtSignature").is_some())
                    && p.get("text").is_some() // 对于 v1internal，通常 text + thought: true 才是合规的思维块
            });

            if !first_is_thought {
                // 如果首项不符合思维块特征，强制补入一个
                parts.insert(
                    0,
                    json!({
                        "text": "...",
                        "thought": true
                    }),
                );
                tracing::debug!("First part of model message at {} is not a valid thought block. Prepending dummy.", parts.len());
            } else {
                // 确保首项包含了 thought: true (防止只有 signature 的情况)
                if let Some(p0) = parts.get_mut(0) {
                    if p0.get("thought").is_none() {
                        p0.as_object_mut()
                            .map(|obj| obj.insert("thought".to_string(), json!(true)));
                    }
                }
            }
        }
    }

    Ok(parts)
}

/// 构建 Contents (Messages)
fn build_google_content(
    msg: &Message,
    claude_req: &ClaudeRequest,
    is_thinking_enabled: bool,
    session_id: &str,
    allow_dummy_thought: bool,
    is_retry: bool,
    tool_id_to_name: &mut HashMap<String, String>,
    tool_name_to_schema: &HashMap<String, Value>,
    mapped_model: &str,
    last_thought_signature: &mut Option<String>,
    pending_tool_use_ids: &mut Vec<String>,
    last_user_task_text_normalized: &mut Option<String>,
    previous_was_tool_result: &mut bool,
    existing_tool_result_ids: &std::collections::HashSet<String>,
) -> Result<Value, String> {
    let role = if msg.role == "assistant" {
        "model"
    } else {
        &msg.role
    };

    // Proactive Tool Chain Repair:
    // If we are about to process an Assistant message, but we still have pending tool_use_ids,
    // it means the previous turn was interrupted or the user ignored the tool.
    // We MUST inject a synthetic User message with error results to close the loop.
    if role == "model" && !pending_tool_use_ids.is_empty() {
        tracing::warn!("[Elastic-Recovery] Detected interrupted tool chain (Assistant -> Assistant). Injecting synthetic User message for IDs: {:?}", pending_tool_use_ids);

        let synthetic_parts: Vec<serde_json::Value> = pending_tool_use_ids
            .iter()
            .filter(|id| !existing_tool_result_ids.contains(*id)) // [FIX #632] Only inject if ID is truly missing
            .map(|id| {
                let name = tool_id_to_name.get(id).cloned().unwrap_or(id.clone());
                json!({
                    "functionResponse": {
                        "name": name,
                        "response": {
                            "result": "Tool execution interrupted. No result provided."
                        },
                        "id": id
                    }
                })
            })
            .collect();

        if !synthetic_parts.is_empty() {
            return Ok(json!({
                "role": "user",
                "parts": synthetic_parts
            }));
        }
        // Clear pending IDs as we have handled them
        pending_tool_use_ids.clear();
    }

    let parts = build_contents(
        &msg.content,
        msg.role == "assistant",
        claude_req,
        is_thinking_enabled,
        session_id,
        allow_dummy_thought,
        is_retry,
        tool_id_to_name,
        tool_name_to_schema,
        mapped_model,
        last_thought_signature,
        pending_tool_use_ids,
        last_user_task_text_normalized,
        previous_was_tool_result,
        existing_tool_result_ids,
    )?;

    if parts.is_empty() {
        return Ok(json!(null)); // Indicate no content to add
    }

    Ok(json!({
        "role": role,
        "parts": parts
    }))
}

/// 构建 Contents (Messages)
fn build_google_contents(
    messages: &[Message],
    claude_req: &ClaudeRequest,
    tool_id_to_name: &mut HashMap<String, String>,
    tool_name_to_schema: &HashMap<String, Value>,
    is_thinking_enabled: bool,
    allow_dummy_thought: bool,
    mapped_model: &str,
    session_id: &str, // [NEW v3.3.17] Session ID for signature caching
    is_retry: bool,
) -> Result<Value, String> {
    let mut contents = Vec::new();
    let mut last_thought_signature: Option<String> = None;
    let mut _accumulated_usage: Option<Value> = None;
    // Track pending tool_use IDs for recovery
    let mut pending_tool_use_ids: Vec<String> = Vec::new();

    // [NEW] 用于识别并过滤 Claude Code 重复回显的任务指令
    let mut last_user_task_text_normalized: Option<String> = None;
    let mut previous_was_tool_result = false;

    let _msg_count = messages.len();

    // [FIX #632] Pre-scan all messages to identify all tool_result IDs that ALREADY exist in the conversation.
    // This prevents Elastic-Recovery from injecting duplicate results if they are present later in the chain.
    let mut existing_tool_result_ids = std::collections::HashSet::new();
    for msg in messages {
        if let MessageContent::Array(blocks) = &msg.content {
            for block in blocks {
                if let ContentBlock::ToolResult { tool_use_id, .. } = block {
                    existing_tool_result_ids.insert(tool_use_id.clone());
                }
            }
        }
    }

    for (_i, msg) in messages.iter().enumerate() {
        let google_content = build_google_content(
            msg,
            claude_req,
            is_thinking_enabled,
            session_id,
            allow_dummy_thought,
            is_retry,
            tool_id_to_name,
            tool_name_to_schema,
            mapped_model,
            &mut last_thought_signature,
            &mut pending_tool_use_ids,
            &mut last_user_task_text_normalized,
            &mut previous_was_tool_result,
            &existing_tool_result_ids,
        )?;

        if !google_content.is_null() {
            contents.push(google_content);
        }
    }

    // [Removed] ensure_last_assistant_has_thinking
    // Corrupted signature issues proved we cannot fake thinking blocks.
    // Instead we rely on should_disable_thinking_due_to_history to prevent this state.

    // [FIX P3-3] Strict Role Alternation (Message Merging)
    // Merge adjacent messages with the same role to satisfy Gemini's strict alternation rule
    let mut merged_contents = merge_adjacent_roles(contents);

    // [FIX P3-4] Deep "Un-thinking" Cleanup
    // If thinking is disabled (e.g. smart downgrade), recursively remove any stray 'thought'/'thoughtSignature'
    // This is critical because converting Thinking->Text isn't enough; metadata must be gone.
    if !is_thinking_enabled {
        for msg in &mut merged_contents {
            clean_thinking_fields_recursive(msg);
        }
    }

    Ok(json!(merged_contents))
}

/// Merge adjacent messages with the same role
fn merge_adjacent_roles(mut contents: Vec<Value>) -> Vec<Value> {
    if contents.is_empty() {
        return contents;
    }

    let mut merged = Vec::new();
    let mut current_msg = contents.remove(0);

    for msg in contents {
        let current_role = current_msg["role"].as_str().unwrap_or_default();
        let next_role = msg["role"].as_str().unwrap_or_default();

        if current_role == next_role {
            // Merge parts
            if let Some(current_parts) = current_msg.get_mut("parts").and_then(|p| p.as_array_mut())
            {
                if let Some(next_parts) = msg.get("parts").and_then(|p| p.as_array()) {
                    current_parts.extend(next_parts.clone());

                    // [FIX #709] Core Fix: After merging parts from adjacent messages,
                    // we must RE-SORT them to ensure any thinking blocks from the
                    // second message are moved to the very front of the combined array.
                    reorder_gemini_parts(current_parts);
                }
            }
        } else {
            merged.push(current_msg);
            current_msg = msg;
        }
    }
    merged.push(current_msg);
    merged
}

/// 构建 Tools
fn build_tools(
    tools: &Option<Vec<Tool>>,
    has_web_search: bool,
    mapped_model: &str,
) -> Result<Option<Value>, String> {
    if let Some(tools_list) = tools {
        let mut function_declarations: Vec<Value> = Vec::new();
        let mut has_google_search = has_web_search;

        for tool in tools_list {
            // 1. Detect server tools / built-in tools like web_search
            if tool.is_web_search() {
                has_google_search = true;
                continue;
            }

            if let Some(t_type) = &tool.type_ {
                if t_type == "web_search_20250305" {
                    has_google_search = true;
                    continue;
                }
            }

            // 2. Detect by name
            if let Some(name) = &tool.name {
                if name == "web_search"
                    || name == "google_search"
                    || name == "builtin_web_search"
                {
                    has_google_search = true;
                    continue;
                }

                // 3. Client tools require input_schema
                let mut input_schema = tool.input_schema.clone().unwrap_or(json!({
                    "type": "object",
                    "properties": {}
                }));
                crate::proxy::common::json_schema::clean_json_schema(&mut input_schema);

                function_declarations.push(json!({
                    "name": name,
                    "description": tool.description,
                    "parameters": input_schema
                }));
            }
        }

        let mut tool_list = Vec::new();

        // [优化] Gemini 2.0+ 及 3.0 系列模型通常支持混合工具调用 (Function Calling + Google Search)
        // 只有针对老旧模型或特定受限环境才需要互斥。
        let model_lower = mapped_model.to_lowercase();
        let supports_mixed_tools = model_lower.contains("gemini-2.0")
            || model_lower.contains("gemini-2.5")
            || model_lower.contains("gemini-3");

        if !function_declarations.is_empty() {
            let mut func_obj = serde_json::Map::new();
            func_obj.insert(
                "functionDeclarations".to_string(),
                json!(function_declarations),
            );
            tool_list.push(json!(func_obj));

            if has_google_search {
                if supports_mixed_tools {
                    tracing::info!(
                        "[Claude-Request] Enabling MIXED tool calling for {}: Function Calling + Google Search.",
                        mapped_model
                    );
                    let mut search_obj = serde_json::Map::new();
                    search_obj.insert("googleSearch".to_string(), json!({}));
                    tool_list.push(json!(search_obj));
                } else {
                    tracing::info!(
                        "[Claude-Request] Skipping googleSearch injection for {} due to existing function declarations. \
                         Older Gemini models may not support mixed tool types.",
                        mapped_model
                    );
                }
            }
        } else if has_google_search {
            let mut search_obj = serde_json::Map::new();
            search_obj.insert("googleSearch".to_string(), json!({}));
            tool_list.push(json!(search_obj));
        }

        if !tool_list.is_empty() {
            return Ok(Some(json!(tool_list)));
        }
    }

    Ok(None)
}

/// 构建 Generation Config
fn build_generation_config(
    claude_req: &ClaudeRequest,
    mapped_model: &str,
    _has_web_search: bool,
    is_thinking_enabled: bool,
    token: Option<&crate::proxy::token_manager::ProxyToken>, // [NEW]
) -> Value {
    let mut config = json!({});

    // Thinking 配置
    if is_thinking_enabled {
        let mut thinking_config = json!({"includeThoughts": true});
        let user_thinking_type = claude_req.thinking.as_ref().map(|t| t.type_.as_str());
        let user_is_adaptive = user_thinking_type == Some("adaptive");

        let budget_tokens = claude_req
            .thinking
            .as_ref()
            .and_then(|t| t.budget_tokens)
            .unwrap_or_else(|| crate::proxy::model_specs::get_thinking_budget(mapped_model, token) as u32);

        let thinking_budget_cap = crate::proxy::model_specs::get_thinking_budget(mapped_model, token);

        let tb_config = crate::proxy::config::get_thinking_budget_config();
        let budget = match tb_config.mode {
            crate::proxy::config::ThinkingBudgetMode::Passthrough => budget_tokens as u64,
            crate::proxy::config::ThinkingBudgetMode::Custom => {
                let mut custom_value = tb_config.custom_value as u64;
                // [FIX #1602] 针对 Gemini 系列模型，在自定义模式下也强制执行动态限额
                let model_lower = mapped_model.to_lowercase();
                let is_gemini_limited = (model_lower.contains("gemini") && !model_lower.contains("-image"))
                    || model_lower.contains("flash")
                    || model_lower.ends_with("-thinking");

                if is_gemini_limited && custom_value > thinking_budget_cap {
                    tracing::warn!(
                        "[Claude-Request] Custom mode: capping thinking_budget from {} to {} for Gemini model {}",
                        custom_value, thinking_budget_cap, mapped_model
                    );
                    custom_value = thinking_budget_cap;
                }
                custom_value
            }
            crate::proxy::config::ThinkingBudgetMode::Auto => {
                // [FIX #1592] Use mapped model for robust detection, same as OpenAI protocol
                let model_lower = mapped_model.to_lowercase();
                let is_gemini_limited = (model_lower.contains("gemini") && !model_lower.contains("-image"))
                    || model_lower.contains("flash")
                    || model_lower.ends_with("-thinking");
                if is_gemini_limited && budget_tokens as u64 > thinking_budget_cap {
                    tracing::info!(
                        "[Claude-Request] Auto mode: capping thinking_budget from {} to {} for Gemini model {}", 
                        budget_tokens, thinking_budget_cap, mapped_model
                    );
                    thinking_budget_cap
                } else {
                    budget_tokens as u64
                }
            }
            crate::proxy::config::ThinkingBudgetMode::Adaptive => budget_tokens as u64, // Adaptive 模式透传原始预算（但不作为限制），用于后续逻辑判断
        };

        let global_mode_is_adaptive = matches!(tb_config.mode, crate::proxy::config::ThinkingBudgetMode::Adaptive);
        // 只要用户指定 adaptive 或者全局配置为 adaptive，且是支持的思维模型，就启用自适应
        let should_use_adaptive = (user_is_adaptive || global_mode_is_adaptive) && (mapped_model.to_lowercase().contains("claude") || mapped_model.to_lowercase().contains("gemini-3"));

        let effort = claude_req.output_config.as_ref().and_then(|c| c.effort.as_ref())
            .or_else(|| claude_req.thinking.as_ref().and_then(|t| t.effort.as_ref()));

        if should_use_adaptive {
            // [FIX #2208] thinkingLevel is ONLY supported by Claude models via Vertex AI native protocol.
            // Gemini models (including gemini-3.x) use v1internal which only accepts thinkingBudget.
            // Previous code incorrectly used contains("gemini-3") as the condition, causing 400 INVALID_ARGUMENT
            // for gemini-3.1-pro-high / gemini-3.1-pro-low in adaptive mode.
            let lower_mapped = mapped_model.to_lowercase();
            if lower_mapped.contains("claude") {
                // Claude 系列走 Vertex AI 原生协议，支持 thinkingLevel 分级参数
                let mapped_level = match effort.map(|e| e.to_lowercase()).as_deref() {
                    Some("low") => "low",
                    Some("medium") => "medium",
                    Some("high") | Some("max") => "high",
                    _ => "high",
                };
                tracing::debug!("[Claude-Request] Mapping adaptive mode to thinkingLevel: {} for Claude model", mapped_level);
                thinking_config["thinkingLevel"] = json!(mapped_level);
                // Claude using thinkingLevel must NOT have thinkingBudget to avoid conflict
                thinking_config.as_object_mut().unwrap().remove("thinkingBudget");
            } else {
                // Gemini 系列（含 gemini-3.x）走 v1internal 协议，只接受 thinkingBudget，不支持 thinkingLevel
                // [FIX #2007] Cherry Studio / Claude Protocol 400 Error Fix
                // Gemini 1.5/2.0 models via Vertex AI often reject thinkingBudget: -1 (Adaptive) with 400 Invalid Argument
                // especially when maxOutputTokens is high.
                // We align with OpenAI mapper behavior: use 24576 as safe adaptive budget.
                tracing::debug!("[Claude-Request] Mapping adaptive mode to safe budget (24576) for Gemini model (thinkingLevel not supported)");
                thinking_config["thinkingBudget"] = json!(24576);
            }
            
            // 针对自适应模式，如果没有显式设置，确保 maxOutputTokens 给足空间
            // OpenAI mapper uses 57344 (24576 + 32768), we normally use 64k limit.
            if config.get("maxOutputTokens").is_none() {
                config["maxOutputTokens"] = json!(64000);
            }
        } else {
            // [FIX #2007] Opus 4.6 Thinking Alignment (OpenAI Protocol Recipe)
            // Explicitly set fixed budget for Opus 4.6 to match successful OpenAI pattern
            if mapped_model.to_lowercase().contains("claude-opus-4-6-thinking") {
                tracing::debug!("[Opus-Alignment] Enforcing fixed thinkingBudget 24576 for Opus 4.6");
                thinking_config["thinkingBudget"] = json!(24576);
            } else {
                thinking_config["thinkingBudget"] = json!(budget);
            }
        }
        
        config["thinkingConfig"] = thinking_config;
    }

    // 其他参数
    if let Some(temp) = claude_req.temperature {
        config["temperature"] = json!(temp);
    }
    if let Some(top_p) = claude_req.top_p {
        config["topP"] = json!(top_p);
    } else {
        config["topP"] = json!(1.0); // [CHANGED v4.1.24] Default topP=1.0 to match official client
    }
    if let Some(top_k) = claude_req.top_k {
        config["topK"] = json!(top_k);
    } else {
        config["topK"] = json!(40); // [ADDED v4.1.24] Default topK=40 to match official client
    }


    // web_search 强制 candidateCount=1
    /*if has_web_search {
        config["candidateCount"] = json!(1);
    }*/

    // max_tokens 映射为 maxOutputTokens
    // [FIX] 不再默认设置 81920，防止非思维模型 (如 claude-sonnet-4-6) 报 400 Invalid Argument
    let mut final_max_tokens: Option<i64> = claude_req.max_tokens.map(|t| t as i64);

    // [NEW] 确保 maxOutputTokens 大于 thinkingBudget (API 强约束)
    // [NEW] 确保 maxOutputTokens 大于 thinkingBudget (API 强约束)
    let model_lower = mapped_model.to_lowercase();
    // 重新计算 should_use_adaptive (因为上面定义的作用域仅在其 if 块内有效，或者我们可以假设在这里也需要同样的逻辑)
    // 但为了简洁和解耦，我们这里重新从 config 读取
    let tb_config_chk = crate::proxy::config::get_thinking_budget_config();
    let global_adaptive = matches!(tb_config_chk.mode, crate::proxy::config::ThinkingBudgetMode::Adaptive);
    let req_adaptive = claude_req.thinking.as_ref().map(|t| t.type_ == "adaptive").unwrap_or(false);
    
    let is_adaptive_effective = (req_adaptive || global_adaptive) && model_lower.contains("claude");
    // [FIX] Lower default overhead to keep total under 65536
    let final_overhead = if is_adaptive_effective { 64000 } else { 32768 };

    // [FIX #2007] Opus 4.6 Thinking Alignment
    // OpenAI logs show maxOutputTokens = 57344 (24576 + 32768)
    if model_lower.contains("claude-opus-4-6-thinking") && is_thinking_enabled {
        final_max_tokens = Some(57344);
        tracing::debug!("[Opus-Alignment] Enforcing maxOutputTokens 57344 for Opus 4.6");
    }

    if let Some(thinking_config) = config.get("thinkingConfig") {
        if let Some(budget) = thinking_config
            .get("thinkingBudget")
            .and_then(|t| t.as_u64())
        {
            let current = final_max_tokens.unwrap_or(0);
            if current <= budget as i64 {
                // [FIX #1675] 针对图像模型使用更小的增量 (2048)
                let overhead = if mapped_model.contains("-image") { 2048 } else { 8192 };
                let boosted = (budget + overhead).min(65536); // [FIX] Never exceed hard limit
                final_max_tokens = Some(boosted as i64);
                tracing::info!(
                    "[Generation-Config] Bumping maxOutputTokens to {} due to thinking budget of {}", 
                    boosted, budget
                );
            }
        } else if is_adaptive_effective {
             // [FIX] Adaptive mode (no budget set in thinkingConfig), apply default maxOutputTokens
             if final_max_tokens.is_none() {
                  final_max_tokens = Some(final_overhead as i64);
             }
        }
    } else {
        // No thinkingConfig
        if final_max_tokens.is_none() && is_adaptive_effective {
            final_max_tokens = Some(final_overhead as i64);
        }
    }


    if let Some(val) = final_max_tokens {
        // [FIX] Cap maxOutputTokens to 65536 to avoid INVALID_ARGUMENT (Cherry Studio sends 128000)
        // Gemini models typically support max 8192 or 65536 output tokens. 128k is usually invalid.
        let safe_limit = 65536;
        if val > safe_limit {
            tracing::warn!(
                "[Generation-Config] Capping maxOutputTokens from {} to {} to prevent 400 Invalid Argument",
                val, safe_limit
            );
            config["maxOutputTokens"] = json!(safe_limit);
        } else {
            config["maxOutputTokens"] = json!(val);
        }
    }

    // [优化] 设置全局停止序列,防止模型幻觉出对话标记
    // [FIX #2007] Opus 4.6 Thinking Alignment
    // Successful OpenAI logs show NO stop sequences were sent for Opus 4.6 Thinking.
    if !(model_lower.contains("claude-opus-4-6-thinking") && is_thinking_enabled) {
        config["stopSequences"] = json!(["<|user|>", "<|end_of_turn|>", "\n\nHuman:"]);
    } else {
        tracing::debug!("[Opus-Alignment] Skipping stopSequences for Opus 4.6 to match OpenAI protocol");
    }

    config
}

/// Recursively remove 'thought' and 'thoughtSignature' fields
/// Used when downgrading thinking (e.g. during 400 retry)
pub fn clean_thinking_fields_recursive(val: &mut Value) {
    match val {
        Value::Object(map) => {
            map.remove("thought");
            map.remove("thoughtSignature");
            for (_, v) in map.iter_mut() {
                clean_thinking_fields_recursive(v);
            }
        }
        Value::Array(arr) => {
            for v in arr.iter_mut() {
                clean_thinking_fields_recursive(v);
            }
        }
        _ => {}
    }
}

/// Check if two model strings are compatible (same family)
fn is_model_compatible(cached: &str, target: &str) -> bool {
    // Simple heuristic: check if they share the same base prefix
    // e.g. "gemini-1.5-pro" vs "gemini-1.5-pro-002" -> Compatible
    // "gemini-1.5-pro" vs "gemini-2.0-flash" -> Incompatible

    // Normalize
    let c = cached.to_lowercase();
    let t = target.to_lowercase();

    if c == t {
        return true;
    }

    // Check specific families
    // Vertex AI signatures are very strict. 1.5-pro vs 1.5-flash are NOT cross-compatible.
    // 2.0-flash vs 2.0-pro are also NOT cross-compatible.

    // Exact model string match (already handled by c == t)

    // Grouped family match (Claude models are more permissive)
    if c.contains("claude-3-5") && t.contains("claude-3-5") {
        return true;
    }
    if c.contains("claude-3-7") && t.contains("claude-3-7") {
        return true;
    }

    // Gemini models: strict family match required for signatures
    if c.contains("gemini-1.5-pro") && t.contains("gemini-1.5-pro") {
        return true;
    }
    if c.contains("gemini-1.5-flash") && t.contains("gemini-1.5-flash") {
        return true;
    }
    if c.contains("gemini-2.0-flash") && t.contains("gemini-2.0-flash") {
        return true;
    }
    if c.contains("gemini-2.0-pro") && t.contains("gemini-2.0-pro") {
        return true;
    }

    // Fallback: strict match required
    false
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::proxy::common::json_schema::clean_json_schema;
    use crate::proxy::config::{ThinkingBudgetConfig, update_thinking_budget_config};

    #[test]
    fn test_ephemeral_injection_debug() {
        // This test simulates the issue where cache_control might be injected
        let json_with_null = json!({
            "model": "claude-3-5-sonnet-20241022",
            "messages": [
                {
                    "role": "assistant",
                    "content": [
                        {
                            "type": "thinking",
                            "thinking": "test",
                            "signature": "sig_1234567890",
                            "cache_control": null
                        }
                    ]
                }
            ]
        });

        let req: ClaudeRequest = serde_json::from_value(json_with_null).unwrap();
        if let MessageContent::Array(blocks) = &req.messages[0].content {
            if let ContentBlock::Thinking { cache_control, .. } = &blocks[0] {
                assert!(
                    cache_control.is_none(),
                    "Deserialization should result in None for null cache_control"
                );
            }
        }

        // Now test serialization
        let serialized = serde_json::to_value(&req).unwrap();
        println!("Serialized: {}", serialized);
        assert!(serialized["messages"][0]["content"][0]
            .get("cache_control")
            .is_none());
    }

    #[test]
    fn test_simple_request() {
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Hello".to_string()),
            }],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok());

        let body = result.unwrap();
        assert_eq!(body["project"], "test-project");
        assert!(body["requestId"].as_str().unwrap().starts_with("agent/"));
    }

    #[test]
    fn test_clean_json_schema() {
        let mut schema = json!({
            "$schema": "http://json-schema.org/draft-07/schema#",
            "type": "object",
            "additionalProperties": false,
            "properties": {
                "location": {
                    "type": "string",
                    "description": "The city and state, e.g. San Francisco, CA",
                    "minLength": 1,
                    "exclusiveMinimum": 0
                },
                "unit": {
                    "type": ["string", "null"],
                    "enum": ["celsius", "fahrenheit"],
                    "default": "celsius"
                },
                "date": {
                    "type": "string",
                    "format": "date"
                }
            },
            "required": ["location"]
        });

        clean_json_schema(&mut schema);

        // Check removed fields
        assert!(schema.get("$schema").is_none());
        assert!(schema.get("additionalProperties").is_none());
        assert!(schema["properties"]["location"].get("minLength").is_none());
        assert!(schema["properties"]["unit"].get("default").is_none());
        assert!(schema["properties"]["date"].get("format").is_none());

        // Check union type handling ["string", "null"] -> "string"
        assert_eq!(schema["properties"]["unit"]["type"], "string");

        // Check types are lowercased
        assert_eq!(schema["type"], "object");
        assert_eq!(schema["properties"]["location"]["type"], "string");
        assert_eq!(schema["properties"]["date"]["type"], "string");
    }

    #[test]
    fn test_complex_tool_result() {
        let req = ClaudeRequest {
            model: "claude-3-5-sonnet-20241022".to_string(),
            messages: vec![
                Message {
                    role: "user".to_string(),
                    content: MessageContent::String("Run command".to_string()),
                },
                Message {
                    role: "assistant".to_string(),
                    content: MessageContent::Array(vec![ContentBlock::ToolUse {
                        id: "call_1".to_string(),
                        name: "run_command".to_string(),
                        input: json!({"command": "ls"}),
                        signature: None,
                        cache_control: None,
                    }]),
                },
                Message {
                    role: "user".to_string(),
                    content: MessageContent::Array(vec![ContentBlock::ToolResult {
                        tool_use_id: "call_1".to_string(),
                        content: json!([
                            {"type": "text", "text": "file1.txt\n"},
                            {"type": "text", "text": "file2.txt"}
                        ]),
                        is_error: Some(false),
                    }]),
                },
            ],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok());

        let body = result.unwrap();
        let contents = body["request"]["contents"].as_array().unwrap();

        // Check the tool result message (last message)
        let tool_resp_msg = &contents[2];
        let parts = tool_resp_msg["parts"].as_array().unwrap();
        let func_resp = &parts[0]["functionResponse"];

        assert_eq!(func_resp["name"], "run_command");
        assert_eq!(func_resp["id"], "call_1");

        // Verify merged content
        let resp_text = func_resp["response"]["result"].as_str().unwrap();
        assert!(resp_text.contains("file1.txt"));
        assert!(resp_text.contains("file2.txt"));
        assert!(resp_text.contains("\n"));
    }

    #[test]
    fn test_cache_control_cleanup() {
        // 模拟 VS Code 插件发送的包含 cache_control 的历史消息
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![
                Message {
                    role: "user".to_string(),
                    content: MessageContent::String("Hello".to_string()),
                },
                Message {
                    role: "assistant".to_string(),
                    content: MessageContent::Array(vec![
                        ContentBlock::Thinking {
                            thinking: "Let me think...".to_string(),
                            signature: Some("sig123".to_string()),
                            cache_control: Some(json!({"type": "ephemeral"})), // 这个应该被清理
                        },
                        ContentBlock::Text {
                            text: "Here is my response".to_string(),
                        },
                    ]),
                },
                Message {
                    role: "user".to_string(),
                    content: MessageContent::Array(vec![ContentBlock::Image {
                        source: ImageSource {
                            source_type: "base64".to_string(),
                            media_type: "image/png".to_string(),
                            data: "iVBORw0KGgo=".to_string(),
                        },
                        cache_control: Some(json!({"type": "ephemeral"})), // 这个也应该被清理
                    }]),
                },
            ],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok());

        // 验证请求成功转换
        let body = result.unwrap();
        assert_eq!(body["project"], "test-project");

        // 注意: cache_control 的清理发生在内部,我们无法直接从 JSON 输出验证
        // 但如果没有清理,后续发送到 Anthropic API 时会报错
        // 这个测试主要确保清理逻辑不会导致转换失败
    }

    #[test]
    fn test_thinking_mode_auto_disable_on_tool_use_history() {
        // [场景] 历史消息中有一个工具调用链，且 Assistant 消息没有 Thinking 块
        // 期望: 系统自动降级，禁用 Thinking 模式，以避免 400 错误
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![
                Message {
                    role: "user".to_string(),
                    content: MessageContent::String("Check files".to_string()),
                },
                // Assistant 使用工具，但在非 Thinking 模式下
                Message {
                    role: "assistant".to_string(),
                    content: MessageContent::Array(vec![
                        ContentBlock::Text {
                            text: "Checking...".to_string(),
                        },
                        ContentBlock::ToolUse {
                            id: "tool_1".to_string(),
                            name: "list_files".to_string(),
                            input: json!({}),
                            cache_control: None,
                            signature: None,
                        },
                    ]),
                },
                // 用户返回工具结果
                Message {
                    role: "user".to_string(),
                    content: MessageContent::Array(vec![ContentBlock::ToolResult {
                        tool_use_id: "tool_1".to_string(),
                        content: serde_json::Value::String("file1.txt\nfile2.txt".to_string()),
                        is_error: Some(false),
                        // cache_control: None, // removed
                    }]),
                },
            ],
            system: None,
            tools: Some(vec![Tool {
                name: Some("list_files".to_string()),
                description: Some("List files".to_string()),
                input_schema: Some(json!({"type": "object"})),
                type_: None,
                // cache_control: None, // removed
            }]),
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: Some(ThinkingConfig {
                type_: "enabled".to_string(),
                budget_tokens: Some(1024),
                effort: None,
            }),
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok());

        let body = result.unwrap();
        let request = &body["request"];

        // 验证: generationConfig 中不应包含 thinkingConfig (因为被降级了)
        // 即使请求中明确启用了 thinking
        if let Some(gen_config) = request.get("generationConfig") {
            assert!(
                gen_config.get("thinkingConfig").is_none(),
                "thinkingConfig should be removed due to downgrade"
            );
        }

        // 验证: 依然能生成有效的请求体
        assert!(request.get("contents").is_some());
    }

    #[test]
    fn test_thinking_block_not_prepend_when_disabled() {
        // 验证当 thinking 未启用时,不会补全 thinking 块
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![
                Message {
                    role: "user".to_string(),
                    content: MessageContent::String("Hello".to_string()),
                },
                Message {
                    role: "assistant".to_string(),
                    content: MessageContent::Array(vec![ContentBlock::Text {
                        text: "Response".to_string(),
                    }]),
                },
            ],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None, // 未启用 thinking
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok());

        let body = result.unwrap();
        let contents = body["request"]["contents"].as_array().unwrap();

        let last_model_msg = contents
            .iter()
            .rev()
            .find(|c| c["role"] == "model")
            .unwrap();

        let parts = last_model_msg["parts"].as_array().unwrap();

        // 验证没有补全 thinking 块
        assert_eq!(parts.len(), 1, "Should only have the original text block");
        assert_eq!(parts[0]["text"], "Response");
    }

    #[test]
    fn test_thinking_block_empty_content_fix() {
        // [场景] 客户端发送了一个内容为空的 thinking 块
        // 期望: 自动填充 "..."
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![Message {
                role: "assistant".to_string(),
                content: MessageContent::Array(vec![
                    ContentBlock::Thinking {
                        thinking: "".to_string(), // 空内容
                        signature: Some("sig".to_string()),
                        cache_control: None,
                    },
                    ContentBlock::Text {
                        text: "Hi".to_string(),
                    },
                ]),
            }],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: Some(ThinkingConfig {
                type_: "enabled".to_string(),
                budget_tokens: Some(1024),
                effort: None,
            }),
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok(), "Transformation failed");
        let body = result.unwrap();
        let contents = body["request"]["contents"].as_array().unwrap();
        let parts = contents[0]["parts"].as_array().unwrap();

        // 验证 thinking 块
        assert_eq!(
            parts[0]["text"], "...",
            "Empty thinking should be filled with ..."
        );
        assert!(
            parts[0].get("thought").is_none(),
            "Empty thinking should be downgraded to text"
        );
    }

    #[test]
    fn test_redacted_thinking_degradation() {
        // [场景] 客户端包含 RedactedThinking
        // 期望: 降级为普通文本，不带 thought: true
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![Message {
                role: "assistant".to_string(),
                content: MessageContent::Array(vec![
                    ContentBlock::RedactedThinking {
                        data: "some data".to_string(),
                    },
                    ContentBlock::Text {
                        text: "Hi".to_string(),
                    },
                ]),
            }],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-project", false, None, "test_session", None);
        assert!(result.is_ok());
        let body = result.unwrap();
        let parts = body["request"]["contents"][0]["parts"].as_array().unwrap();

        // 验证 RedactedThinking -> Text
        let text = parts[0]["text"].as_str().unwrap();
        assert!(text.contains("[Redacted Thinking: some data]"));
        assert!(
            parts[0].get("thought").is_none(),
            "Redacted thinking should NOT have thought: true"
        );
    }

    // ==================================================================================
    // [FIX #564] Test: Thinking blocks are sorted to be first after context compression
    // ==================================================================================
    #[test]
    fn test_thinking_blocks_sorted_first_after_compression() {
        // Simulate kilo context compression reordering: text BEFORE thinking
        let mut messages = vec![Message {
            role: "assistant".to_string(),
            content: MessageContent::Array(vec![
                // Wrong order: Text before Thinking (simulates kilo compression)
                ContentBlock::Text {
                    text: "Some regular text".to_string(),
                },
                ContentBlock::Thinking {
                    thinking: "My thinking process".to_string(),
                    signature: Some(
                        "valid_signature_1234567890_abcdefghij_klmnopqrstuvwxyz_test".to_string(),
                    ),
                    cache_control: None,
                },
                ContentBlock::Text {
                    text: "More text".to_string(),
                },
            ]),
        }];

        // Apply the fix
        sort_thinking_blocks_first(&mut messages);

        // Verify thinking is now first
        if let MessageContent::Array(blocks) = &messages[0].content {
            assert_eq!(blocks.len(), 3, "Should still have 3 blocks");
            assert!(
                matches!(blocks[0], ContentBlock::Thinking { .. }),
                "Thinking should be first"
            );
            assert!(
                matches!(blocks[1], ContentBlock::Text { .. }),
                "Text should be second"
            );
            assert!(
                matches!(blocks[2], ContentBlock::Text { .. }),
                "Text should be third"
            );

            // Verify content preserved
            if let ContentBlock::Thinking { thinking, .. } = &blocks[0] {
                assert_eq!(thinking, "My thinking process");
            }
        } else {
            panic!("Expected Array content");
        }
    }

    #[test]
    fn test_thinking_blocks_no_reorder_when_already_first() {
        // Correct order: Thinking already first - should not trigger reorder
        let mut messages = vec![Message {
            role: "assistant".to_string(),
            content: MessageContent::Array(vec![
                ContentBlock::Thinking {
                    thinking: "My thinking".to_string(),
                    signature: Some("sig123".to_string()),
                    cache_control: None,
                },
                ContentBlock::Text {
                    text: "Some text".to_string(),
                },
            ]),
        }];

        // Apply the fix (should be no-op)
        sort_thinking_blocks_first(&mut messages);

        // Verify order unchanged
        if let MessageContent::Array(blocks) = &messages[0].content {
            assert!(
                matches!(blocks[0], ContentBlock::Thinking { .. }),
                "Thinking should still be first"
            );
            assert!(
                matches!(blocks[1], ContentBlock::Text { .. }),
                "Text should still be second"
            );
        }
    }

    #[test]
    fn test_merge_consecutive_messages() {
        let mut messages = vec![
            Message {
                role: "user".to_string(),
                content: MessageContent::String("Hello".to_string()),
            },
            Message {
                role: "user".to_string(),
                content: MessageContent::Array(vec![ContentBlock::Text {
                    text: "World".to_string(),
                }]),
            },
            Message {
                role: "assistant".to_string(),
                content: MessageContent::String("Hi".to_string()),
            },
            Message {
                role: "user".to_string(),
                content: MessageContent::Array(vec![ContentBlock::ToolResult {
                    tool_use_id: "test_id".to_string(),
                    content: serde_json::json!("result"),
                    is_error: None,
                }]),
            },
            Message {
                role: "user".to_string(),
                content: MessageContent::Array(vec![ContentBlock::Text {
                    text: "System Reminder".to_string(),
                }]),
            },
        ];

        merge_consecutive_messages(&mut messages);

        assert_eq!(messages.len(), 3);
        assert_eq!(messages[0].role, "user");
        if let MessageContent::Array(blocks) = &messages[0].content {
            assert_eq!(blocks.len(), 2);
            match &blocks[0] {
                ContentBlock::Text { text } => assert_eq!(text, "Hello"),
                _ => panic!("Expected text block"),
            }
            match &blocks[1] {
                ContentBlock::Text { text } => assert_eq!(text, "World"),
                _ => panic!("Expected text block"),
            }
        } else {
            panic!("Expected array content at index 0");
        }

        assert_eq!(messages[1].role, "assistant");

        assert_eq!(messages[2].role, "user");
        if let MessageContent::Array(blocks) = &messages[2].content {
            assert_eq!(blocks.len(), 2);
            match &blocks[0] {
                ContentBlock::ToolResult { tool_use_id, .. } => assert_eq!(tool_use_id, "test_id"),
                _ => panic!("Expected tool_result block"),
            }
            match &blocks[1] {
                ContentBlock::Text { text } => assert_eq!(text, "System Reminder"),
                _ => panic!("Expected text block"),
            }
        } else {
            panic!("Expected array content at index 2");
        }
    }
    #[test]
    fn test_default_max_tokens() {
        let req = ClaudeRequest {
            model: "claude-3-opus".to_string(),
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Hello".to_string()),
            }],
            system: None,
            tools: None,
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "test-v", false, None, "test_session", None).unwrap();
        // [FIX] Since we removed the default 81920, maxOutputTokens should NOT be present
        // when max_tokens is None and thinking is disabled
        let gen_config = &result["request"]["generationConfig"];
        assert!(
            gen_config.get("maxOutputTokens").is_none(),
            "maxOutputTokens should not be set when max_tokens is None"
        );
    }
    #[test]
    fn test_claude_flash_thinking_budget_capping() {
        // Use full path or ensure import of ThinkingConfig
        // transform_claude_request and models are needed.
        // Assuming models are available via super imports, but let's be explicit if needed.

        // Setup request with high budget
        let req = ClaudeRequest {
            model: "gemini-2.0-flash-thinking-exp".to_string(), // Contains "flash"
            messages: vec![],
            thinking: Some(ThinkingConfig {
                type_: "enabled".to_string(),
                budget_tokens: Some(32000),
                effort: None,
            }),
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None, // Added missing field
            stream: false,
            system: None,
            tools: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        let result = transform_claude_request_in(&req, "proj", false, None, "test_session", None).unwrap();
        let budget = result["request"]["generationConfig"]["thinkingConfig"]["thinkingBudget"]
            .as_u64()
            .unwrap();
        assert_eq!(budget, 24576); // capped by model_specs.get_thinking_budget("gemini-2.0-flash-thinking-exp")

        // Setup request for Pro thinking model (mock name for testing)
        let req_pro = ClaudeRequest {
            model: "gemini-2.0-pro-thinking-exp".to_string(), // Contains "thinking" but not "flash"
            messages: vec![],
            thinking: Some(ThinkingConfig {
                type_: "enabled".to_string(),
                budget_tokens: Some(32000),
                effort: None,
            }),
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None, // Added missing field
            stream: false,
            system: None,
            tools: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        // Should cap
        let result_pro = transform_claude_request_in(&req_pro, "proj", false, None, "test_session", None).unwrap();
        assert_eq!(result_pro["request"]["generationConfig"]["thinkingConfig"]["thinkingBudget"], 24576);
    }

    #[test]
    fn test_gemini_pro_thinking_support() {
        // Setup request for Gemini Pro (no -thinking suffix)
        let req = ClaudeRequest {
            model: "gemini-3-pro-preview".to_string(),
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Hello".to_string()),
            }],
            thinking: Some(ThinkingConfig {
                type_: "enabled".to_string(),
                budget_tokens: Some(16000),
                effort: None,
            }),
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            stream: false,
            system: None,
            tools: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        // Transform
        let result = transform_claude_request_in(&req, "proj", false, None, "test_session", None).unwrap();
        let gen_config = &result["request"]["generationConfig"];

        // thinkingConfig should be present (not forced disabled)
        assert!(
            gen_config.get("thinkingConfig").is_some(),
            "thinkingConfig should be preserved for gemini-3-pro"
        );

        let budget = gen_config["thinkingConfig"]["thinkingBudget"]
            .as_u64()
            .unwrap();
        // [FIX #1592] Since it's < 24576, it should be kept as 16000
        assert_eq!(budget, 16000);
    }

    #[test]
    fn test_gemini_pro_default_thinking() {
        // Setup request for Gemini Pro WITHOUT thinking config
        let req = ClaudeRequest {
            model: "gemini-3-pro-preview".to_string(),
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Hello".to_string()),
            }],
            thinking: None, // No thinking config provided by client
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            stream: false,
            system: None,
            tools: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        // Transform
        let result = transform_claude_request_in(&req, "proj", false, None, "test_session", None).unwrap();
        let gen_config = &result["request"]["generationConfig"];

        // thinkingConfig SHOULD be injected because of default-on logic
        assert!(
            gen_config.get("thinkingConfig").is_some(),
            "thinkingConfig should be auto-enabled for gemini-3-pro"
        );
    }

    #[test]
    fn test_claude_image_thinking_mode_disabled() {
        // 1. Force image thinking mode to "disabled"
        crate::proxy::config::update_image_thinking_mode(Some("disabled".to_string()));

        // 2. Setup Claude request for an image model (mapped to gemini-3-pro-image)
        let req = ClaudeRequest {
            model: "gemini-3-pro-image".to_string(), // Explicitly use recognized image model
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Draw a cat".to_string()),
            }],
            thinking: None,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            stream: false,
            system: None,
            tools: None,
            metadata: None,
            output_config: None,
            size: Some("1024x1024".to_string()),
            quality: Some("hd".to_string()),
        };

        // 3. Transform request
        let result = transform_claude_request_in(&req, "test-proj", false, None, "test_session", None).unwrap();

        // 4. Verify thinkingConfig has includeThoughts: false
        let gen_config = result["request"]["generationConfig"].as_object().expect("Should have generationConfig");
        let thinking_config = gen_config.get("thinkingConfig").and_then(|t| t.as_object()).expect("Should have thinkingConfig (explicitly disabled)");
        
        assert_eq!(thinking_config["includeThoughts"], false);
        
        // 5. Reset global mode
        crate::proxy::config::update_image_thinking_mode(Some("enabled".to_string()));
    }

    #[test]
    fn test_claude_adaptive_global_config() {
        // Set global config to Adaptive + High effort
        let config = ThinkingBudgetConfig {
            mode: crate::proxy::config::ThinkingBudgetMode::Adaptive,
            custom_value: 0,
            effort: Some("high".to_string()),
        };
        crate::proxy::config::update_thinking_budget_config(config);

        let req = ClaudeRequest {
            model: "claude-3-7-sonnet-thinking".to_string(), // thinking capable
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("test".to_string()),
            }],
            thinking: None, // No client thinking config
            stream: false,
            // ... minimal fields
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            system: None,
            tools: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        // Transform
        let result = transform_claude_request_in(&req, "test-proj", false, None, "test_session", None).unwrap();
        
        let gen_config = result["request"]["generationConfig"].as_object().unwrap();
        let thinking_config = gen_config["thinkingConfig"].as_object().unwrap();

        // Check injection
        assert_eq!(thinking_config["includeThoughts"], true);
        assert_eq!(thinking_config["thinkingBudget"], -1);
        assert!(thinking_config.get("thinkingType").is_none());
        assert!(thinking_config.get("effort").is_none());

        // Check maxOutputTokens default for adaptive
        let max_output_tokens = gen_config["maxOutputTokens"].as_i64().unwrap();
        assert_eq!(max_output_tokens, 131072);

        // Reset global config
        crate::proxy::config::update_thinking_budget_config(ThinkingBudgetConfig::default());
    }

    #[test]
    fn test_mixed_tools_injection_for_gemini_2_0() {
        // [场景] 使用 Gemini 2.0 模型，同时提供自定义工具和启用全网搜索
        // 期望: 转换后的请求应同时包含 googleSearch 和 functionDeclarations
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(), // 映射到 gemini-2.0-flash-exp
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Help me search and use tools".to_string()),
            }],
            system: None,
            tools: Some(vec![Tool {
                type_: None,
                name: Some("get_weather".to_string()),
                description: Some("Get weather".to_string()),
                input_schema: Some(serde_json::json!({
                    "type": "object",
                    "properties": {
                        "location": {"type": "string"}
                    }
                })),
            }]),
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        // 模拟映射到 Gemini 2.0
        let mapped_model = "gemini-2.0-flash-exp";
        
        // 这里我们直接测试 build_tools 函数 (它是 pub(crate) 且在同模块下)
        let result = build_tools(&req.tools, true, mapped_model);
        assert!(result.is_ok());
        
        let tools_val = result.unwrap().expect("Should have tools");
        let tools_arr = tools_val.as_array().expect("Tools should be an array");
        
        let has_google_search = tools_arr.iter().any(|t| t.get("googleSearch").is_some());
        let has_functions = tools_arr.iter().any(|t| t.get("functionDeclarations").is_some());
        
        assert!(has_google_search, "Gemini 2.0 should support mixed Google Search");
        assert!(has_functions, "Gemini 2.0 should support mixed function declarations");
    }

    #[test]
    fn test_no_mixed_tools_for_older_gemini() {
        // [场景] 使用 Gemini 1.5 模型，同时提供自定义工具和启用全网搜索
        // 期望: 转换后的请求应只包含 functionDeclarations，googleSearch 被跳过以避免 400
        let req = ClaudeRequest {
            model: "claude-sonnet-4-6".to_string(),
            messages: vec![Message {
                role: "user".to_string(),
                content: MessageContent::String("Help me search and use tools".to_string()),
            }],
            system: None,
            tools: Some(vec![Tool {
                type_: None,
                name: Some("get_weather".to_string()),
                description: Some("Get weather".to_string()),
                input_schema: Some(serde_json::json!({
                    "type": "object",
                    "properties": {
                        "location": {"type": "string"}
                    }
                })),
            }]),
            stream: false,
            max_tokens: None,
            temperature: None,
            top_p: None,
            top_k: None,
            thinking: None,
            metadata: None,
            output_config: None,
            size: None,
            quality: None,
        };

        // 模拟映射到 Gemini 1.5
        let mapped_model = "gemini-1.5-flash-002";
        
        // 测试 build_tools 函数
        let result = build_tools(&req.tools, true, mapped_model);
        assert!(result.is_ok());
        
        let tools_val = result.unwrap().expect("Should have tools");
        let tools_arr = tools_val.as_array().expect("Tools should be an array");
        
        let has_google_search = tools_arr.iter().any(|t| t.get("googleSearch").is_some());
        let has_functions = tools_arr.iter().any(|t| t.get("functionDeclarations").is_some());
        
        assert!(!has_google_search, "Older Gemini models should NOT have mixed tools");
        assert!(has_functions);
    }
}