📋 概述

� MPP-Core 改进建议 (2025-10-31)

当前 mpp-core 架构概览

核心问题诊断

1. 缺乏异步通信层⚠️ 高优先级

// 建议实现sealedclassAgentSubmission {data classSendPrompt(valtext:String) : AgentSubmission()data classCancelTask(valtaskId:String) : AgentSubmission()data classApproveToolCall(valcallId:String) : AgentSubmission()}sealedclassAgentEvent {data classStreamUpdate(valtext:String) : AgentEvent()data classToolCallRequest(valtool:String,valparams:Map<String,Any>) : AgentEvent()data classTaskComplete(valresult:String) : AgentEvent()data classError(valmessage:String) : AgentEvent()}classAgentOrchestrator {privateval submissionChannel=Channel<AgentSubmission>(Channel.BUFFERED)privateval eventChannel=Channel<AgentEvent>(Channel.BUFFERED)suspendfunsubmit(submission:AgentSubmission) {        submissionChannel.send(submission)    }funevents():Flow<AgentEvent>= eventChannel.receiveAsFlow()}

2. 缺乏工具调度系统⚠️ 高优先级

// 建议实现sealedclassToolCallState {data classValidating(valcallId:String,valtool:String) : ToolCallState()data classScheduled(valcallId:String,valtool:Tool) : ToolCallState()data classAwaitingApproval(valcallId:String,valtool:Tool) : ToolCallState()data classExecuting(valcallId:String,valtool:Tool,valstartTime:Long) : ToolCallState()data classSuccess(valcallId:String,valresult:String) : ToolCallState()data classError(valcallId:String,valerror:String) : ToolCallState()}classToolScheduler(privatevalregistry:ToolRegistry,privatevalpolicyEngine:PolicyEngine) {privateval toolCalls= mutableMapOf<String,ToolCallState>()privateval queue=Channel<ToolCallState>(Channel.UNLIMITED)suspendfunschedule(toolName:String,params:Map<String,Any>):String {val callId=UUID.randomUUID().toString()val state=ToolCallState.Validating(callId, toolName)        queue.send(state)return callId    }privatesuspendfunprocessQueue() {        queue.receiveAsFlow().collect { state->when (state) {isToolCallState.Validating-> validate(state)isToolCallState.Scheduled-> checkPolicy(state)isToolCallState.Executing-> execute(state)// ...            }        }    }}

3. 缺乏并发执行能力⚠️ 中优先级

// 建议实现interfaceTool {val name:Stringval isReadOnly:Boolean// 新增标志suspendfunexecute(params:ToolParams):ToolResult}classParallelToolExecutor {privateval stateLock=ReentrantReadWriteLock()suspendfunexecuteBatch(calls:List<ToolCall>):List<ToolResult> {// 分类工具val (readCalls, writeCalls)= calls.partition {             registry.getTool(it.name)?.isReadOnly==true         }// 并行执行只读工具val readResults= coroutineScope {            readCalls.map { call->                async {                    stateLock.readLock().withLock {                        executeTool(call)                    }                }            }.awaitAll()        }// 串行执行写工具val writeResults= writeCalls.map { call->            stateLock.writeLock().withLock {                executeTool(call)            }        }return readResults+ writeResults    }}

4. 缺乏子任务机制 (Subagent)⚠️ 中优先级

// 建议实现data classAgentDefinition(valname:String,valsystemPrompt:String,valallowedTools:List<String>,// 只授予必要工具valoutputSchema:JsonSchema?// 强制结构化输出)classAgentExecutor(privatevaldefinition:AgentDefinition,privatevalparentRegistry:ToolRegistry) {// 创建隔离的工具注册表privateval isolatedRegistry=ToolRegistry().apply {        definition.allowedTools.forEach { toolName->            parentRegistry.getTool(toolName)?.let { register(it) }        }// 强制添加 complete_task 工具        register(CompleteTaskTool(definition.outputSchema))    }suspendfunrun(inputs:Map<String,Any>):AgentResult {var turnCount=0val maxTurns=20while (turnCount< maxTurns) {val response= llm.chat(history)// 检查是否调用了 complete_taskval completeCall= response.toolCalls.find { it.name=="complete_task" }if (completeCall!=null) {returnAgentResult.Success(completeCall.output)            }// 执行其他工具调用            processToolCalls(response.toolCalls)            turnCount++        }returnAgentResult.MaxTurnsReached    }}

// 定义代码审查子 Agentval codeReviewer=AgentDefinition(    name="code-reviewer",    systemPrompt="You review code for security and quality issues...",    allowedTools=listOf("read_file","grep","git_diff"),// 只读权限    outputSchema=JsonSchema.of<CodeReviewResult>())// 在主 Agent 中调用val executor=AgentExecutor(codeReviewer, mainRegistry)val result= executor.run(mapOf("filePath" to"src/Auth.kt"))// 返回结构化的审查结果

5. 循环检测和历史压缩缺失⚠️ 低优先级

// 建议实现classLoopDetectionService {privateval toolCallHistory= mutableListOf<ToolCallRecord>()privateval windowSize=10funrecordToolCall(toolName:String,params:Map<String,Any>) {        toolCallHistory.add(ToolCallRecord(toolName, params,System.currentTimeMillis()))    }fundetectLoop():LoopResult {if (toolCallHistory.size< windowSize)returnLoopResult.Noneval recent= toolCallHistory.takeLast(windowSize)val signature= recent.joinToString("|") {"${it.toolName}:${it.params.hashCode()}" }// 检测重复模式val pattern= findRepeatingPattern(signature)if (pattern!=null&& pattern.repetitions>=3) {returnLoopResult.Detected(pattern)        }returnLoopResult.None    }}classChatCompressionService(privatevalllm:LLMService) {suspendfuncompress(history:List<Message>):List<Message> {if (history.size<20)return history// 保留最新 5 轮对话val recent= history.takeLast(10)// 压缩中间历史val middle= history.dropLast(10).drop(2)// 保留开头的 system promptval compressed= llm.summarize(middle, maxTokens=500)return history.take(2)+listOf(Message.system("Previous conversation summary:$compressed")        )+ recent    }}

6. 缺乏权限控制系统⚠️ 高优先级

// 建议实现enumclassPolicyDecision {ALLOW,// 自动允许DENY,// 自动拒绝ASK_USER// 需要用户确认}data classPolicyRule(valtoolPattern:Regex,valdecision:PolicyDecision,valcondition: ((ToolCall)->Boolean)? =null)classPolicyEngine {privateval rules= mutableListOf<PolicyRule>()privateval approvalCache= mutableMapOf<String,Boolean>()// 会话内缓存funcheckToolCall(call:ToolCall):PolicyDecision {// 1. 检查缓存val cacheKey="${call.toolName}:${call.params.hashCode()}"if (approvalCache.containsKey(cacheKey)) {returnif (approvalCache[cacheKey]!!)PolicyDecision.ALLOWelsePolicyDecision.DENY        }// 2. 应用规则for (rulein rules) {if (rule.toolPattern.matches(call.toolName)) {val conditionMet= rule.condition?.invoke(call)?:trueif (conditionMet)return rule.decision            }        }// 3. 默认策略：危险工具需要确认returnif (call.isDangerous())PolicyDecision.ASK_USERelsePolicyDecision.ALLOW    }funaddRule(rule:PolicyRule) {        rules.add(rule)    }}// 预定义规则val defaultPolicy=PolicyEngine().apply {// 只读工具自动允许    addRule(PolicyRule(        toolPattern=Regex("read_file|grep|glob"),        decision=PolicyDecision.ALLOW    ))// 危险工具需要确认    addRule(PolicyRule(        toolPattern=Regex("shell|delete_file|write_file"),        decision=PolicyDecision.ASK_USER    ))// 禁止某些危险操作    addRule(PolicyRule(        toolPattern=Regex("shell"),        decision=PolicyDecision.DENY,        condition= { call->val command= call.params["command"]as?String            command?.contains("rm -rf")==true        }    ))}

实施优先级和路线图

立即可行的改进 (本周内)

1. 添加 Tool 抽象接口

// 文件: mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/tool/Tool.ktinterfaceTool {val name:Stringval description:Stringval isReadOnly:Booleanval parameters:ToolParameterssuspendfunexecute(params:Map<String,Any>,context:ToolContext):ToolResult}data classToolResult(valsuccess:Boolean,valoutput:String,valerror:String? =null,valmetadata:Map<String,Any> = emptyMap())

2. 引入基础通信层

// 文件: mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/communication/AgentChannel.ktclassAgentChannel {privateval_submissions=MutableSharedFlow<AgentSubmission>()privateval_events=MutableSharedFlow<AgentEvent>()val submissions:SharedFlow<AgentSubmission>=_submissionsval events:SharedFlow<AgentEvent>=_eventssuspendfunsubmit(submission:AgentSubmission) {_submissions.emit(submission)    }suspendfunemit(event:AgentEvent) {_events.emit(event)    }}

3. 分离工具注册表

// 文件: mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/tool/ToolRegistry.ktclassToolRegistry {privateval tools= mutableMapOf<String,Tool>()funregister(tool:Tool) {        tools[tool.name]= tool    }fungetTool(name:String):Tool?= tools[name]funlistTools():List<Tool>= tools.values.toList()funfilterTools(predicate: (Tool)->Boolean):List<Tool> {return tools.values.filter(predicate)    }}

与现有代码的整合建议

保持 DevInsCompiler 不变

重构 KoogLLMService

// 当前classKoogLLMService {funstreamPrompt(userPrompt:String):Flow<String>}// 建议重构为classKoogLLMService(privatevalchannel:AgentChannel,// 新增privatevaltoolScheduler:ToolScheduler// 新增) {suspendfunprocessSubmission(submission:AgentSubmission) {when (submission) {isAgentSubmission.SendPrompt-> {val compiled=DevInsCompilerFacade.compile(submission.text)                streamLLM(compiled.output).collect { chunk->                    channel.emit(AgentEvent.StreamUpdate(chunk))                }            }isAgentSubmission.ApproveToolCall-> {                toolScheduler.approveCall(submission.callId)            }        }    }}

增强 ChatHistoryManager

// 添加压缩支持classChatHistoryManager(privatevalcompressionService:ChatCompressionService// 新增) {suspendfunaddMessage(message:Message) {val session= getCurrentSession()        session.messages.add(message)// 自动压缩if (session.messages.size>50) {            session.messages= compressionService.compress(session.messages).toMutableList()        }    }}

关键设计原则

性能目标

测试策略

单元测试 (P0)

// 测试异步通信@Testfun`should emit events when submission processed`()= runTest {val channel=AgentChannel()val events= mutableListOf<AgentEvent>()        launch {        channel.events.take(2).toList(events)    }        channel.submit(AgentSubmission.SendPrompt("test"))// ...        assertEquals(2, events.size)}

集成测试 (P1)

// 测试完整工具调用流程@Testfun`should execute tool with approval`()= runTest {val orchestrator=AgentOrchestrator(        registry= testToolRegistry,        policy= testPolicyEngine    )        orchestrator.submit(AgentSubmission.SendPrompt("/read file.txt"))val event= orchestrator.events().first()    assertTrue(eventisAgentEvent.ToolCallRequest)}

性能测试 (P1)

@Testfun`parallel read should be faster than serial`()= runTest {val files=List(10) {"file$it.txt" }val serialTime= measureTime {        files.forEach { readFile(it) }    }val parallelTime= measureTime {        parallelExecutor.executeBatch(files.map {ReadFileCall(it) })    }        assertTrue(parallelTime< serialTime/3)}

文档更新计划

�🆕 最新文档（推荐优先阅读）

1.gemini-cli-architecture.md ⭐⭐⭐⭐⭐

2.coding-agents-architecture.md ⭐⭐⭐⭐

� 历史参考文档

3. AutoDev Sketch 重构实现指南 (本文件后续内容)

🏗️ 架构设计

三大 Agent 对比分析

核心架构：融合三者优势

原始实现分析 (IDEA 版 AutoDev)

核心组件架构

关键特性

Codex 核心设计模式

1. Queue Pair 通信模式 ⭐

// Kotlin 实现示例classCodingAgent {privateval submissionChannel=Channel<Submission>(capacity=64)privateval eventChannel=Channel<Event>(capacity=Channel.UNLIMITED)suspendfunsubmit(op:Operation):String {val id= generateId()        submissionChannel.send(Submission(id, op))return id    }suspendfunnextEvent():Event= eventChannel.receive()}

2. Tool Orchestrator 模式 ⭐⭐⭐

3. Parallel Tool Execution ⭐⭐

classToolCallRuntime {privateval parallelLock=ReentrantReadWriteLock()suspendfunexecute(tool:Tool,request:Request):Result {val lock=if (tool.supportsParallel) {            parallelLock.readLock()// 允许多个并发        }else {            parallelLock.writeLock()// 独占执行        }                lock.lock()try {return tool.run(request)        }finally {            lock.unlock()        }    }}

Gemini CLI 参考架构

核心设计原则

MPP 重构架构设计

详细架构图

整体架构概览

数据流架构

组件依赖关系

Agent 类型和能力映射

工具系统集成

平台适配策略

安全与沙箱

🧠 提示词与上下文工程

核心设计原则

提示词工程架构

核心系统提示词结构

// 基于 Gemini CLI 的系统提示词设计interfaceSystemPromptConfig{// 核心身份和规则identity:string;// "You are an interactive CLI agent..."coreMandates:string[];// 核心约束和规则workflows:WorkflowConfig[];// 主要工作流程guidelines:string[];// 操作指导原则// 上下文注入contextSources:ContextSource[];// 工具配置availableTools:string[];toolUsageGuidelines:string;}

提示词模板系统

基础模板结构

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/prompt/PromptTemplate.kt@Serializabledata classPromptTemplate(valname:String,valtemplate:String,valvariables:List<TemplateVariable> = emptyList(),valconditionals:List<ConditionalBlock> = emptyList(),valcontextSources:List<ContextSource> = emptyList())@Serializabledata classTemplateVariable(valname:String,valtype:VariableType,valrequired:Boolean =true,valdefaultValue:String? =null,valdescription:String ="")enumclassVariableType {STRING,NUMBER,BOOLEAN,LIST,OBJECT}@Serializabledata classConditionalBlock(valcondition:String,// 条件表达式，如 "agentType == 'CODER'"valcontent:String,// 条件为真时的内容valelseContent:String? =null// 条件为假时的内容)enumclassContextSource {WORKSPACE_STRUCTURE,// 工作空间文件结构PROJECT_CONFIG,// AGENTS.md 等项目配置GIT_INFO,// Git 仓库信息BUILD_SYSTEM,// 构建系统信息USER_MEMORY,// 用户偏好记忆RECENT_FILES,// 最近访问的文件ERROR_CONTEXT// 错误上下文信息}

Agent 特定提示词

代码库调查员 Agent

valCODEBASE_INVESTIGATOR_PROMPT=PromptTemplate(    name="codebase_investigator",    template="""You are **Codebase Investigator**, a hyper-specialized AI agent and expert in reverse-engineering complex software projects.Your **SOLE PURPOSE** is to build a complete mental model of the code relevant to a given investigation.## Your Mission- Find key modules, classes, and functions that are part of the problem and solution- Understand *why* the code is written the way it is- Foresee ripple effects of changes- Provide actionable insights to the main agent## Investigation Objective${objective}## Available ToolsYou have access to read-only tools: ls, read_file, glob, grep## Output FormatProvide a structured JSON report with:- SummaryOfFindings: Key insights and conclusions- ExplorationTrace: Step-by-step investigation actions- RelevantLocations: Important files with reasoning and key symbols**DO NOT** write implementation code yourself. Focus on analysis and insights.""".trimIndent(),    variables=listOf(TemplateVariable("objective",VariableType.STRING, description="调查目标")    ))

代码生成器 Agent

valCODE_GENERATOR_PROMPT=PromptTemplate(    name="code_generator",    template="""You are a **Code Generation Expert** specialized in creating high-quality, maintainable code.## Your Expertise- Following existing project conventions and patterns- Generating clean, readable, and efficient code- Creating appropriate tests and documentation- Ensuring seamless integration with existing codebase## Current Task${task}## Project Context{{CONTEXT}}## Guidelines- Always analyze existing code patterns before generating new code- Include unit tests for new functionality- Follow the project's established coding style- Add meaningful comments only when necessary- Ensure all file paths are absolute## Available Tools${availableTools.join(',')}""".trimIndent(),    variables=listOf(TemplateVariable("task",VariableType.STRING, description="具体的代码生成任务"),TemplateVariable("availableTools",VariableType.LIST, description="可用工具列表")    ),    contextSources=listOf(ContextSource.WORKSPACE_STRUCTURE,ContextSource.PROJECT_CONFIG,ContextSource.RECENT_FILES    ))

上下文工程系统

上下文数据模型

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/context/ContextModels.kt@Serializabledata classContextBundle(valworkspaceContext:WorkspaceContext? =null,valprojectContext:ProjectContext? =null,valenvironmentContext:EnvironmentContext? =null,valfileContents:Map<String,String> = emptyMap(),valgitContext:GitContext? =null,valbuildSystemContext:BuildSystemContext? =null,valuserMemory:UserMemory? =null,valrecentActivity:RecentActivity? =null)@Serializabledata classWorkspaceContext(valrootDirectories:List<String>,valfolderStructure:String,valtotalFiles:Int,valprimaryLanguages:List<String>,vallastUpdated:Long =Clock.System.now().toEpochMilliseconds())// ... and other data classes from prompt_context_engineering_guide.md

**参考 Kode 实现**：Kode 的 `Tool` 接口 (`src/Tool.ts`) 中定义的 `ToolUseContext` 是一个很好的运行时上下文示例，它包含了 `messageId`、`agentId`、`safeMode`、`readFileTimestamps` 和各种 `options` 等实用元素，展示了如何将上下文传递给工具执行。

上下文提供者系统

工作空间上下文提供者

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/context/WorkspaceContextProvider.ktclassWorkspaceContextProvider(privatevalfileSystem:ToolFileSystem) {privatevar cachedContext:WorkspaceContext?=nullprivatevar lastCacheTime:Long=0privateval cacheValidityMs=30_000// 30秒缓存suspendfungetWorkspaceContext(directories:List<String>):WorkspaceContext {// ...    }privatesuspendfunbuildFolderStructure(directories:List<String>):String {// ...    }}

项目上下文加载器

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/context/ProjectContextLoader.ktclassProjectContextLoader(privatevalfileSystem:ToolFileSystem) {suspendfunloadProjectContext(rootDirectory:String):ProjectContext {// ...    }}

动态上下文解析器

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/context/DynamicContextResolver.ktclassDynamicContextResolver(privatevalworkspaceProvider:WorkspaceContextProvider,privatevalprojectLoader:ProjectContextLoader,privatevaltoolRegistry:ToolRegistry) {suspendfunresolveContext(query:String,directories:List<String>,maxTokens:Int = 8000    ):ContextBundle {// ...    }}

核心模块与实现

工具系统架构

模块划分

mpp-core 核心模块

mpp-ui 界面模块

实现示例

核心抽象层

Agent 编排器接口

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/orchestrator/AgentOrchestrator.ktpackagecc.unitmesh.agent.orchestratorimportcc.unitmesh.agent.config.AgentConfigimportcc.unitmesh.agent.sketch.SketchInputimportcc.unitmesh.agent.sketch.SketchOutputimportkotlinx.coroutines.flow.Flow/** * Agent 编排器，负责协调不同 Agent 的执行 * 参考 Gemini CLI 的 Agent 管理模式*/interfaceAgentOrchestrator {/**     * 处理 Sketch 输入，返回流式输出*/suspendfunprocessSketch(input:SketchInput):Flow<SketchOutput>/**     * 根据上下文选择合适的 Agent*/suspendfunselectAgent(context:SketchContext):AgentConfig/**     * 执行指定的 Agent*/suspendfunexecuteAgent(config:AgentConfig,input:String):Flow<String>/**     * 获取可用的 Agent 列表*/suspendfungetAvailableAgents():List<AgentConfig>}/** * Sketch 上下文信息*/data classSketchContext(valprojectPath:String?,valcurrentFile:String?,valselectedText:String?,valopenFiles:List<String>,valuserIntent:String,valtools:List<String> = emptyList())

Sketch 数据模型

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/sketch/SketchModels.ktpackagecc.unitmesh.agent.sketchimportkotlinx.serialization.Serializable/** * Sketch 输入数据*/@Serializabledata classSketchInput(valcontent:String,valtype:SketchInputType =SketchInputType.TEXT,valcontext:Map<String,String> = emptyMap())/** * Sketch 输出数据*/@Serializabledata classSketchOutput(valcontent:String,valtype:SketchOutputType,valmetadata:Map<String,String> = emptyMap(),valtimestamp:Long = kotlinx.datetime.Clock.System.now().toEpochMilliseconds())enumclassSketchInputType {TEXT,CODE,MARKDOWN,DEVIN}enumclassSketchOutputType {TEXT,CODE,PLAN,DIAGRAM,ERROR,TOOL_RESULT}/** * DevIn 代码块*/@Serializabledata classDevInBlock(valcontent:String,vallanguage:String ="devin",valcommands:List<String> = emptyList())/** * Sketch 分析结果*/@Serializabledata classSketchAnalysis(valintent:String,valcomplexity:SketchComplexity,valrequiredTools:List<String>,valsuggestedAgent:String,valdevInBlocks:List<DevInBlock> = emptyList())enumclassSketchComplexity {SIMPLE,MEDIUM,COMPLEX}

Agent 配置

// mpp-core/src/commonMain/kotlin/cc/unitmesh/agent/config/AgentConfig.ktpackagecc.unitmesh.agent.configimportkotlinx.serialization.Serializable/** * Agent 配置，参考原始 CustomAgentConfig 但简化为跨平台版本*/@Serializabledata classAgentConfig(valname:String,valtype:AgentType,valdescription:String ="",valsystemPrompt:String ="",valcapabilities:List<AgentCapability> = emptyList(),valtools:List<String> = emptyList(),valmaxTokens:Int =4096,valtemperature:Double =0.7,valenabled:Boolean =true)enumclassAgentType {CODER,// 代码生成和修改REVIEWER,// 代码审查TESTER,// 测试生成ARCHITECT,// 架构设计ANALYST,// 需求分析HELPER,// 通用助手CUSTOM// 自定义}enumclassAgentCapability {CODE_GENERATION,CODE_REVIEW,TEST_GENERATION,DOCUMENTATION,REFACTORING,DEBUGGING,ARCHITECTURE_DESIGN,REQUIREMENT_ANALYSIS}

UI 组件扩展

Agent 选择器组件

// mpp-ui/src/main/kotlin/cc/unitmesh/devins/ui/compose/agent/AgentSelector.ktpackagecc.unitmesh.devins.ui.compose.agentimportandroidx.compose.foundation.layout.*importandroidx.compose.material3.*importandroidx.compose.runtime.*importandroidx.compose.ui.Modifierimportandroidx.compose.ui.unit.dpimportcc.unitmesh.agent.config.AgentConfig@OptIn(ExperimentalMaterial3Api::class)@ComposablefunAgentSelector(agents:List<AgentConfig>,selectedAgent:AgentConfig?,onAgentSelected: (AgentConfig)->Unit,modifier:Modifier =Modifier) {var expanded by remember { mutableStateOf(false) }ExposedDropdownMenuBox(        expanded= expanded,        onExpandedChange= { expanded=!expanded },        modifier= modifier    ) {OutlinedTextField(            value= selectedAgent?.name?:"选择 Agent",            onValueChange= { },            readOnly=true,            trailingIcon= {ExposedDropdownMenuDefaults.TrailingIcon(expanded= expanded) },            modifier=Modifier.menuAnchor()        )ExposedDropdownMenu(            expanded= expanded,            onDismissRequest= { expanded=false }        ) {            agents.forEach { agent->DropdownMenuItem(                    text= {Column {Text(agent.name)Text(                                text= agent.description,                                style=MaterialTheme.typography.bodySmall,                                color=MaterialTheme.colorScheme.onSurfaceVariant                            )                        }                    },                    onClick= {                        onAgentSelected(agent)                        expanded=false                    }                )            }        }    }}

实施计划与测试

实现步骤

阶段 1: 核心抽象层 (mpp-core)

阶段 2: UI 层重构 (mpp-ui)

阶段 3: 平台集成

测试策略

性能与错误处理

性能考虑

错误处理策略

分层错误处理

sealedclassSketchError :Exception() {data classCompilationError(valdetails:String) : SketchError()data classToolExecutionError(valtoolName:String,valerror:String) : SketchError()data classLLMError(valprovider:String,valerror:String) : SketchError()data classPlatformError(valplatform:String,valerror:String) : SketchError()}