@

feat(ai): add agent loop with tool calling, fix connection routing and Gemini compatibility

## 背景说明
将 dbx AI 助手从单次 SQL 文本生成器升级为具备 tool calling 的数据库 Agent。
Phase 1 实现 Agent Loop 框架 + 只读工具（list_tables、get_columns）。
参考：Pi agent-core，.trellis/tasks/06-11-feat-ai-agent-loop-tool-calling-upgrade

## 修改内容
- 新增 agent_loop.rs：Agent Loop 核心，支持 OpenAI/Claude/Gemini Provider 的 tool calling，最大 10 轮，Ollama 自动降级为 text-only
- 新增 agent_tools.rs：read_only_tools (list_tables, get_columns) 的工具定义与执行器
- 新增 agent_events.rs：AgentEvent/ToolCall/ToolResult/ToolDefinition 类型定义
- 修改 ai.rs：AiMessage 新增 tool_call_id 字段
- 修改 lib.rs：注册 agent_events、agent_loop、agent_tools 模块
- 修改 routes/ai.rs & commands/ai.rs：新增 ai_agent_stream SSE 端点
- 修改 tauri.ts & http.ts：新增 AgentEvent 类型与 aiAgentStream 函数
- 修改 api.ts：注册 aiAgentStream 转发
- 修改 ai.ts：新增 runAgentStream，AiContext 增加 connectionId 字段
- 修改 AiAssistant.vue：接入 runAgentStream，解析 AgentEvent 渲染工具调用步骤
- 修改 en.ts / zh-CN.ts：新增 agentSteps 国际化文案

## 修复的问题
- [#I01] connectionName 误传为 connectionId → add connectionId to AiContext
- [#I02] agent loop 取消操作无效 → add cancel checks before each turn
- [#C01] TODO 缺少 issue 引用 → add PRD reference to TODO
- [#C02] stream_with_tools 命名不准确 → add clarifying doc comment
- [#C03] Gemini functionResponse.name 解析错误 → fix tool_call_id format and extraction

## 测试建议
- execute_get_columns：合法表名/空表名/SQL 注入字符拒绝
- call_openai_with_tools：空 choices/空 tool_calls/非 JSON arguments
- call_gemini_with_tools：空 candidates/parts 无 functionCall
- runAgentStream：有效 connectionId 正常调用
@

src-tauri/src/commands/ai.rs

@@ -42,11 +42,53 @@ pub async fn ai_stream(app: AppHandle, session_id: String, request: AiCompletion
     result
 }
 
+use dbx_core::agent_events::AgentEvent;
+use dbx_core::agent_loop::{run_agent_loop, AgentLoopContext};
+use dbx_core::models::connection::DatabaseType;
+
 #[tauri::command]
 pub async fn ai_cancel_stream(session_id: String) -> Result<bool, String> {
     Ok(dbx_core::ai::cancel_stream(&session_id).await)
 }
 
+#[tauri::command]
+pub async fn ai_agent_stream(
+    app: AppHandle,
+    state: State<'_, Arc<AppState>>,
+    session_id: String,
+    request: AiCompletionRequest,
+    connection_id: String,
+    database: String,
+    db_type: String,
+) -> Result<String, String> {
+    let cancelled = dbx_core::ai::register_stream(&session_id).await;
+
+    let parsed_db_type: DatabaseType =
+        serde_json::from_str(&format!("\"{}\"", db_type)).map_err(|_| format!("Unknown database type: {db_type}"))?;
+
+    let agent_ctx = AgentLoopContext { state: state.inner().clone(), connection_id, database, db_type: parsed_db_type };
+
+    let result = run_agent_loop(
+        &request.config,
+        &request.system_prompt,
+        &request.messages,
+        &agent_ctx,
+        {
+            let app = app.clone();
+            move |event: AgentEvent| {
+                let _ = app.emit("ai-agent-event", &event);
+            }
+        },
+        &cancelled,
+        request.max_tokens,
+        request.temperature,
+    )
+    .await;
+
+    dbx_core::ai::unregister_stream(&session_id).await;
+    result
+}
+
 #[tauri::command]
 pub async fn save_ai_conversation(state: State<'_, Arc<AppState>>, conversation: AiConversation) -> Result<(), String> {
     state.storage.save_ai_conversation(&conversation).await