@yuaone/core 0.6.1 → 0.7.1

package/dist/agent-loop.js

@@ -8,6 +8,8 @@
  */
 import { EventEmitter } from "node:events";
 import { randomUUID } from "node:crypto";
+import { execSync } from "node:child_process";
+import { basename, join as pathJoin } from "node:path";
 import { BYOKClient } from "./llm-client.js";
 import { Governor } from "./governor.js";
 import { ContextManager } from "./context-manager.js";
@@ -50,6 +52,10 @@ import { DependencyAnalyzer } from "./dependency-analyzer.js";
 import { CrossFileRefactor } from "./cross-file-refactor.js";
 import { ContextBudgetManager } from "./context-budget.js";
 import { QAPipeline } from "./qa-pipeline.js";
+import { PersonaManager } from "./persona.js";
+import { InMemoryVectorStore, OllamaEmbeddingProvider } from "./vector-store.js";
+import { StateStore, TransitionModel, SimulationEngine, StateUpdater, } from "./world-model/index.js";
+import { MilestoneChecker, RiskEstimator, PlanEvaluator, ReplanningEngine, } from "./planner/index.js";
 /**
  * AgentLoop — YUAN 에이전트의 핵심 실행 루프.
  *
@@ -103,6 +109,7 @@ export class AgentLoop extends EventEmitter {
     tokenBudgetManager;
     allToolResults = [];
     currentTaskIndex = 0;
+    _lastInjectedTaskIndex = -1; // track when plan progress was last injected
     changedFiles = [];
     aborted = false;
     initialized = false;
@@ -150,6 +157,8 @@ export class AgentLoop extends EventEmitter {
     iterationSystemMsgCount = 0;
     /** Task 1: ContextBudgetManager for LLM-based summarization at 60-70% context usage */
     contextBudgetManager = null;
+    /** Task 1: Flag to ensure LLM summarization only runs once per agent run (non-blocking guard) */
+    _contextSummarizationDone = false;
     /** Task 2: Track whether write tools ran this iteration for QA triggering */
     iterationWriteToolPaths = [];
     /** Task 2: Last QA result (surfaced to LLM on issues) */
@@ -158,6 +167,25 @@ export class AgentLoop extends EventEmitter {
     iterationTsFilesModified = [];
     /** Task 3: Whether tsc was run in the previous iteration (skip cooldown) */
     tscRanLastIteration = false;
+    /** PersonaManager — learns user communication style, injects persona into system prompt */
+    personaManager = null;
+    /** InMemoryVectorStore — RAG: semantic code context retrieval for relevant snippets */
+    vectorStore = null;
+    /** Last user message — used for task-specific memory retrieval */
+    lastUserMessage = "";
+    // World Model
+    worldModel = null;
+    transitionModel = null;
+    simulationEngine = null;
+    stateUpdater = null;
+    // Proactive Replanning
+    planEvaluator = null;
+    riskEstimator = null;
+    replanningEngine = null;
+    milestoneChecker = null;
+    activeMilestones = [];
+    completedPlanTaskIds = new Set();
+    allToolResultsSinceLastReplan = [];
     tokenUsage = {
         input: 0,
         output: 0,
@@ -322,6 +350,27 @@ export class AgentLoop extends EventEmitter {
                 // MemoryManager (structured learnings)
                 this.memoryManager = new MemoryManager(projectPath);
                 await this.memoryManager.load();
+                // PersonaManager — user communication style learning + persona injection
+                const personaUserId = basename(projectPath) || "default";
+                this.personaManager = new PersonaManager({
+                    userId: personaUserId,
+                    profilePath: pathJoin(projectPath, ".yuan", `persona-${personaUserId}.json`),
+                    enableLearning: true,
+                });
+                await this.personaManager.loadProfile().catch(() => { });
+                // InMemoryVectorStore — RAG semantic code context (TF-IDF fallback if Ollama unavailable)
+                this.vectorStore = new InMemoryVectorStore({
+                    projectId: personaUserId,
+                    projectPath,
+                    embeddingProvider: new OllamaEmbeddingProvider(),
+                });
+                await this.vectorStore.load().catch(() => { });
+                // Background indexing — non-blocking, fires and forgets
+                const vectorStoreRef = this.vectorStore;
+                import("./code-indexer.js").then(({ CodeIndexer }) => {
+                    const indexer = new CodeIndexer({});
+                    indexer.indexProject(projectPath, vectorStoreRef).catch(() => { });
+                }).catch(() => { });
                 // 프로젝트 구조 분석
                 projectStructure = await this.yuanMemory.analyzeProject();
             }
@@ -362,6 +411,32 @@ export class AgentLoop extends EventEmitter {
         catch {
             // WorldState 수집 실패는 치명적이지 않음
         }
+        // Initialize World Model
+        if (this.worldState && projectPath) {
+            try {
+                this.transitionModel = new TransitionModel();
+                this.worldModel = StateStore.fromSnapshot(this.worldState, projectPath);
+                this.simulationEngine = new SimulationEngine(this.transitionModel, this.worldModel);
+                this.stateUpdater = new StateUpdater(this.worldModel, projectPath);
+            }
+            catch {
+                // World Model initialization failure is non-fatal
+            }
+        }
+        // Capture last known good git commit for FailureRecovery rollback
+        try {
+            const headHash = execSync("git rev-parse HEAD", {
+                cwd: projectPath,
+                stdio: "pipe",
+                timeout: 5_000,
+            }).toString().trim();
+            if (headHash) {
+                this.failureRecovery.setLastGoodCommit(headHash, projectPath);
+            }
+        }
+        catch {
+            // Not a git repo or git unavailable — FailureRecovery will use file-level rollback only
+        }
         // ImpactAnalyzer 생성
         this.impactAnalyzer = new ImpactAnalyzer({ projectPath });
         // ContinuationEngine 생성
@@ -479,6 +554,18 @@ export class AgentLoop extends EventEmitter {
         if (this.skillLearner) {
             this.planner.setSkillLearner(this.skillLearner);
         }
+        // Initialize Proactive Replanning (requires planner + impactAnalyzer + worldModel)
+        try {
+            if (this.impactAnalyzer && this.worldModel && this.simulationEngine && this.transitionModel) {
+                this.milestoneChecker = new MilestoneChecker();
+                this.riskEstimator = new RiskEstimator(this.transitionModel, this.impactAnalyzer);
+                this.planEvaluator = new PlanEvaluator(this.worldModel, this.simulationEngine);
+                this.replanningEngine = new ReplanningEngine(this.planner, this.planEvaluator, this.riskEstimator, this.milestoneChecker);
+            }
+        }
+        catch {
+            // Proactive replanning initialization failure is non-fatal
+        }
         if (this.skillLearner) {
             const learnedSkills = this.skillLearner.getAllSkills();
             if (learnedSkills.length > 0) {
@@ -516,12 +603,15 @@ export class AgentLoop extends EventEmitter {
                     const bgAgent = this.backgroundAgentManager.get(agent.id);
                     if (bgAgent) {
                         bgAgent.on("event", (event) => {
-                            if (event.type === "error" || event.type === "warning") {
-                                this.emitEvent({
-                                    kind: "agent:thinking",
-                                    content: `[Background: ${event.agentId}] ${event.message}`,
-                                });
-                            }
+                            // Emit structured bg_update for TUI task panel
+                            this.emitEvent({
+                                kind: "agent:bg_update",
+                                agentId: event.agentId,
+                                agentLabel: event.agentId.replace(/-/g, " "),
+                                eventType: event.type,
+                                message: event.message,
+                                timestamp: event.timestamp,
+                            });
                         });
                     }
                 }
@@ -670,6 +760,8 @@ export class AgentLoop extends EventEmitter {
             // Task 3: reset tsc tracking per run
             this.iterationTsFilesModified = [];
             this.tscRanLastIteration = false;
+            // Task 1: reset context summarization guard per run
+            this._contextSummarizationDone = false;
         }
         this.checkpointSaved = false;
         this.failureRecovery.reset();
@@ -719,8 +811,58 @@ export class AgentLoop extends EventEmitter {
             role: "user",
             content: userMessage,
         });
+        // PersonaManager — 유저 메시지로 커뮤니케이션 스타일 학습
+        this.lastUserMessage = userMessage;
+        if (this.personaManager) {
+            this.personaManager.analyzeUserMessage(userMessage);
+        }
         this.emitEvent({ kind: "agent:start", goal: userMessage });
         try {
+            // Persona injection — 유저 선호도/언어/스타일 어댑테이션을 시스템 메시지로 주입
+            if (this.personaManager) {
+                const personaSection = this.personaManager.buildPersonaPrompt();
+                if (personaSection) {
+                    this.contextManager.addMessage({ role: "system", content: personaSection });
+                }
+            }
+            // MemoryManager.getRelevant() — 현재 태스크와 관련된 conventions/patterns/warnings 주입
+            if (this.memoryManager) {
+                const relevant = this.memoryManager.getRelevant(userMessage);
+                const parts = [];
+                if (relevant.conventions.length > 0) {
+                    parts.push(`## Project Conventions\n${relevant.conventions.slice(0, 5).map((c) => `- ${c}`).join("\n")}`);
+                }
+                if (relevant.warnings.length > 0) {
+                    parts.push(`## Relevant Warnings\n${relevant.warnings.slice(0, 3).map((w) => `⚠ ${w}`).join("\n")}`);
+                }
+                if (relevant.patterns.length > 0) {
+                    parts.push(`## Relevant Code Patterns\n${relevant.patterns.slice(0, 3).map((p) => `- **${p.name}**: ${p.description}`).join("\n")}`);
+                }
+                if (parts.length > 0) {
+                    this.contextManager.addMessage({
+                        role: "system",
+                        content: `[Task Memory]\n${parts.join("\n\n")}`,
+                    });
+                }
+            }
+            // VectorStore RAG — 태스크와 의미적으로 유사한 코드 컨텍스트 검색·주입
+            if (this.vectorStore) {
+                try {
+                    const hits = await this.vectorStore.search(userMessage, 3, 0.2);
+                    if (hits.length > 0) {
+                        const ragCtx = hits
+                            .map((h) => `**${h.id}** (relevance: ${(h.similarity * 100).toFixed(0)}%)\n${h.text.slice(0, 400)}`)
+                            .join("\n\n---\n\n");
+                        this.contextManager.addMessage({
+                            role: "system",
+                            content: `[RAG Context — semantically relevant code snippets]\n${ragCtx}`,
+                        });
+                    }
+                }
+                catch {
+                    // VectorStore search failure is non-fatal
+                }
+            }
             // Reflexion: 과거 실행에서 배운 가이던스 주입
             if (this.reflexionEngine) {
                 try {
@@ -984,6 +1126,10 @@ export class AgentLoop extends EventEmitter {
             }
             // 메모리 저장
             await this.memoryManager.save();
+            // PersonaManager — 유저 프로필 저장 (학습된 커뮤니케이션 스타일 유지)
+            if (this.personaManager) {
+                await this.personaManager.saveProfile().catch(() => { });
+            }
         }
         catch {
             // 메모리 저장 실패는 치명적이지 않음
@@ -1152,7 +1298,7 @@ export class AgentLoop extends EventEmitter {
     async maybeCreatePlan(userMessage) {
         if (!this.planner || !this.enablePlanning)
             return;
-        const complexity = this.detectComplexity(userMessage);
+        const complexity = await this.detectComplexity(userMessage);
         this.currentComplexity = complexity;
         // 임계값 미만이면 플래닝 스킵
         // Bug 4 fix: extend thresholdOrder to include "massive" (4), so that when planningThreshold
@@ -1172,6 +1318,20 @@ export class AgentLoop extends EventEmitter {
             const plan = await this.planner.createHierarchicalPlan(userMessage, this.llmClient);
             this.activePlan = plan;
             this.currentTaskIndex = 0;
+            // Run plan simulation + extract milestones
+            if (this.simulationEngine && this.milestoneChecker && this.activePlan) {
+                const capturedPlan = this.activePlan;
+                this.simulationEngine.simulate(capturedPlan).then((simResult) => {
+                    if (simResult.criticalSteps.length > 0 || simResult.overallSuccessProbability < 0.6) {
+                        this.contextManager.addMessage({
+                            role: "system",
+                            content: this.simulationEngine.formatForPrompt(simResult),
+                        });
+                    }
+                }).catch(() => { });
+                this.activeMilestones = this.milestoneChecker.extractMilestones(this.activePlan);
+                this.completedPlanTaskIds.clear();
+            }
             const planTokenEstimate = plan.tactical.length * 500;
             this.tokenBudgetManager.recordUsage("planner", planTokenEstimate, planTokenEstimate);
             const planContext = this.formatPlanForContext(plan);
@@ -1227,7 +1387,7 @@ export class AgentLoop extends EventEmitter {
         }
         return "pnpm build";
     }
-    detectComplexity(message) {
+    _detectComplexityHeuristic(message) {
         const lower = message.toLowerCase();
         const len = message.length;
         // 복잡도 점수 계산
@@ -1264,6 +1424,48 @@ export class AgentLoop extends EventEmitter {
             score += 2;
         else if (filePaths && filePaths.length >= 2)
             score += 1;
+        // 야망형 태스크 키워드 — 짧아도 대형 작업 (score +4 → 즉시 "complex" 이상)
+        // "OS 만들어", "설계해줘", "컴파일러 구현" 같은 짧지만 massive한 요청을 잡기 위함
+        const ambitiousSystemKeywords = [
+            "운영체제", "os ", " os", "kernel", "커널",
+            "compiler", "컴파일러", "인터프리터", "interpreter",
+            "database", "데이터베이스", "dbms",
+            "framework", "프레임워크",
+            "vm ", "virtual machine", "가상머신", "hypervisor",
+            "distributed", "분산 시스템", "분산시스템",
+            "blockchain", "블록체인",
+            "게임 엔진", "game engine",
+        ];
+        for (const kw of ambitiousSystemKeywords) {
+            if (lower.includes(kw)) {
+                score += 4;
+                break;
+            }
+        }
+        // 설계/아키텍처 요청 키워드 (score +3)
+        const designKeywords = [
+            "설계", "디자인해", "아키텍처 만들", "구조 설계",
+            "design the", "architect the", "design a ", "design an ",
+            "from scratch", "처음부터", "새로 만들", "전부 만들",
+        ];
+        for (const kw of designKeywords) {
+            if (lower.includes(kw)) {
+                score += 3;
+                break;
+            }
+        }
+        // 전체 구현 요청 키워드 (score +2)
+        const buildKeywords = [
+            "만들어줘", "만들어 줘", "만들어봐", "구현해줘", "개발해줘",
+            "build me", "create a full", "implement a full", "make me a",
+            "전체 구현", "full implementation", "complete implementation",
+        ];
+        for (const kw of buildKeywords) {
+            if (lower.includes(kw)) {
+                score += 2;
+                break;
+            }
+        }
         // 간단한 작업 키워드 (감점)
         const simpleKeywords = [
             "fix", "고쳐", "수정해",
@@ -1288,6 +1490,37 @@ export class AgentLoop extends EventEmitter {
             return "complex";
         return "massive";
     }
+    /**
+     * Hybrid complexity detection: keyword heuristic for clear cases,
+     * LLM single-word classification for ambiguous borderline cases (score 1-3).
+     * This prevents both over-planning (trivial tasks) and under-planning
+     * (ambitious short requests that keywords miss across any language).
+     */
+    async detectComplexity(message) {
+        const heuristic = this._detectComplexityHeuristic(message);
+        // Clear extremes — trust heuristic, skip LLM call cost
+        if (heuristic === "trivial" || heuristic === "massive")
+            return heuristic;
+        // Borderline ambiguous range: ask LLM for one-word verdict
+        // Short cheap call: no tools, 1-word response, ~50 tokens total
+        try {
+            const resp = await this.llmClient.chat([
+                {
+                    role: "user",
+                    content: `Rate this software task complexity in ONE word only (trivial/simple/moderate/complex/massive). Task: "${message.slice(0, 300)}"`,
+                },
+            ], []);
+            const word = (resp.content ?? "").trim().toLowerCase().replace(/[^a-z]/g, "");
+            const valid = ["trivial", "simple", "moderate", "complex", "massive"];
+            if (valid.includes(word)) {
+                return word;
+            }
+        }
+        catch {
+            // LLM classification failure — fall back to heuristic
+        }
+        return heuristic;
+    }
     /**
      * HierarchicalPlan을 LLM이 따라갈 수 있는 컨텍스트 메시지로 포맷.
      */
@@ -1426,6 +1659,47 @@ export class AgentLoop extends EventEmitter {
             const iterationStart = Date.now();
             this.emitReasoning(`iteration ${iteration}: preparing context`);
             this.iterationSystemMsgCount = 0; // Reset per-iteration (prevents accumulation across iterations)
+            // Proactive replanning check (every 10 iterations when plan is active)
+            if (this.replanningEngine &&
+                this.activePlan &&
+                this.activeMilestones.length > 0 &&
+                iteration > 0 &&
+                iteration % 10 === 0) {
+                try {
+                    const tokenBudget = this.config?.loop?.totalTokenBudget ?? 200_000;
+                    const tokensUsed = this.tokenUsage.total;
+                    const replanResult = await this.replanningEngine.evaluate(this.activePlan, this.worldModel?.getState() ?? {}, [...this.completedPlanTaskIds], this.allToolResultsSinceLastReplan, tokensUsed, tokenBudget, [...this.changedFiles], iteration, this.activeMilestones, this.llmClient);
+                    if (replanResult.triggered) {
+                        this.emitEvent({ kind: "agent:thinking", content: `[Proactive Replan] ${replanResult.message}` });
+                        if (replanResult.newPlan) {
+                            this.activePlan = replanResult.newPlan;
+                            this.activeMilestones = this.milestoneChecker.extractMilestones(this.activePlan);
+                            this.completedPlanTaskIds.clear();
+                        }
+                        else if (replanResult.modifiedTasks && replanResult.modifiedTasks.length > 0) {
+                            for (const modified of replanResult.modifiedTasks) {
+                                const idx = this.activePlan.tactical.findIndex(t => t.id === modified.id);
+                                if (idx >= 0)
+                                    this.activePlan.tactical[idx] = modified;
+                            }
+                        }
+                        // Reset tool results accumulator after replan
+                        this.allToolResultsSinceLastReplan = [];
+                        // Inject replan notice into context
+                        this.contextManager.addMessage({
+                            role: "system",
+                            content: `[Proactive Replan] ${replanResult.message}\nScope: ${replanResult.decision.scope}, Risk: ${replanResult.decision.urgency}`,
+                        });
+                    }
+                }
+                catch {
+                    // Non-blocking — proactive replanning failures should not crash the agent
+                }
+            }
+            // Plan progress injection — every 3 iterations or when task advances
+            if (this.activePlan) {
+                this.injectPlanProgress(iteration);
+            }
             // Policy validation — check cost limits from ExecutionPolicyEngine
             if (this.policyEngine) {
                 try {
@@ -1446,25 +1720,26 @@ export class AgentLoop extends EventEmitter {
             // Task 1: ContextBudgetManager LLM summarization at 60-70% — runs BEFORE ContextCompressor
             // Summarizes old "medium" priority conversation turns into a compact summary message,
             // freeing tokens before the heavier ContextCompressor kicks in at 70%.
-            if (contextUsageRatio >= 0.60 && contextUsageRatio < 0.70 && this.contextBudgetManager) {
-                try {
-                    // Sync current messages into the budget manager so it knows what to summarize
-                    this.contextBudgetManager.importMessages(this.contextManager.getMessages());
-                    if (this.contextBudgetManager.needsSummarization()) {
-                        const summary = await this.contextBudgetManager.summarize(async (prompt) => {
-                            const resp = await this.llmClient.chat([{ role: "user", content: prompt }], []);
-                            return typeof resp.content === "string" ? resp.content : "";
-                        });
+            if (contextUsageRatio >= 0.60 && contextUsageRatio < 0.70 && this.contextBudgetManager && !this._contextSummarizationDone) {
+                this._contextSummarizationDone = true; // run at most once per agent turn
+                // Non-blocking: fire-and-forget so the main iteration is not stalled
+                this.contextBudgetManager.importMessages(this.contextManager.getMessages());
+                if (this.contextBudgetManager.needsSummarization()) {
+                    const budgetMgr = this.contextBudgetManager;
+                    const ratio = contextUsageRatio;
+                    budgetMgr.summarize(async (prompt) => {
+                        const resp = await this.llmClient.chat([{ role: "user", content: prompt }], []);
+                        return typeof resp.content === "string" ? resp.content : "";
+                    }).then(summary => {
                         if (summary) {
                             this.emitEvent({
                                 kind: "agent:thinking",
-                                content: `Context at ${Math.round(contextUsageRatio * 100)}%: summarized ${summary.originalIds.length} old messages (${summary.originalTokens} → ${summary.summarizedTokens} tokens, ${Math.round(summary.compressionRatio * 100)}% ratio).`,
+                                content: `Context at ${Math.round(ratio * 100)}%: summarized ${summary.originalIds.length} old messages (${summary.originalTokens} → ${summary.summarizedTokens} tokens, ${Math.round(summary.compressionRatio * 100)}% ratio).`,
                             });
                         }
-                    }
-                }
-                catch {
-                    // ContextBudgetManager summarization failure is non-fatal
+                    }).catch(() => {
+                        // summarization failure is non-fatal
+                    });
                 }
             }
             // Bug 5 fix: use ContextCompressor as an alternative when context pressure is high (>70%)
@@ -1768,6 +2043,8 @@ export class AgentLoop extends EventEmitter {
             const { results: toolResults, deferredFixPrompts } = await this.executeTools(response.toolCalls);
             // Reflexion: 도구 결과 수집
             this.allToolResults.push(...toolResults);
+            // Accumulate tool results for proactive replanning evaluation
+            this.allToolResultsSinceLastReplan.push(...toolResults);
             // Tool plan tracking: 실행된 도구 이름 기록
             for (const result of toolResults) {
                 this.executedToolNames.push(result.name);
@@ -1898,6 +2175,10 @@ export class AgentLoop extends EventEmitter {
                     this.tscRanLastIteration = false;
                 }
             }
+            // Plan task advancement — check if current task's target files were modified
+            if (this.activePlan) {
+                this.tryAdvancePlanTask();
+            }
             // iteration 이벤트
             this.emitEvent({
                 kind: "agent:iteration",
@@ -2216,7 +2497,7 @@ export class AgentLoop extends EventEmitter {
         let usage = { input: 0, output: 0 };
         let finishReason = "stop";
         const allTools = [...this.config.loop.tools, ...this.mcpToolDefinitions];
-        const stream = this.llmClient.chatStream(messages, allTools);
+        const stream = this.llmClient.chatStream(messages, allTools, this.abortSignal ?? undefined);
         // 텍스트 버퍼링 — 1토큰씩 emit하지 않고 청크 단위로 모아서 emit
         let textBuffer = "";
         let flushTimer = null;
@@ -2521,6 +2802,10 @@ export class AgentLoop extends EventEmitter {
                     });
                 }
             }
+            // StateUpdater: sync world model with actual tool execution result
+            if (this.stateUpdater && result.success) {
+                this.stateUpdater.applyToolResult(toolCall.name, args, result).catch(() => { });
+            }
             const fixPrompt = await this.validateAndFeedback(toolCall.name, result);
             return { result, deferredFixPrompt: fixPrompt ?? null };
         }
@@ -2577,136 +2862,172 @@ export class AgentLoop extends EventEmitter {
         }
     }
     async executeTools(toolCalls) {
-        // Reorder write tool calls using DependencyAnalyzer so files with no deps run first
-        const writeToolNames = new Set(['file_write', 'file_edit']);
-        const writeToolCalls = toolCalls.filter((tc) => writeToolNames.has(tc.name));
+        if (toolCalls.length === 0)
+            return { results: [], deferredFixPrompts: [] };
+        // ─── Step 1: Build execution plan ───────────────────────────────────────
+        // Strategy:
+        //   • Read-only tools (low-risk)      → batch together, run all in parallel
+        //   • Write tools (independent files) → use DependencyAnalyzer to group,
+        //                                       run each independent group in parallel
+        //   • Write tools (dependent files)   → run sequentially after their deps
+        //   • shell_exec / git_ops / etc.     → always sequential (side-effects)
+        const READ_ONLY = new Set(['file_read', 'grep', 'glob', 'code_search', 'security_scan']);
+        const WRITE_TOOLS = new Set(['file_write', 'file_edit']);
+        // Separate reads, writes, and heavy side-effect tools
+        const writeToolCalls = toolCalls.filter((tc) => WRITE_TOOLS.has(tc.name));
+        // ─── Step 2: Dependency-aware write tool batching ────────────────────────
+        // Map each write tool call to a "wave index" (0 = can run first, 1 = needs wave-0 done, etc.)
+        const writeBatchMap = new Map(); // tc.id → wave index
         if (writeToolCalls.length > 1 && this.config.loop.projectPath) {
             try {
                 const depAnalyzer = new DependencyAnalyzer();
                 const depGraph = await depAnalyzer.analyze(this.config.loop.projectPath);
-                const writeFilePaths = writeToolCalls
-                    .map((tc) => {
+                // Collect target file paths from write tool args
+                const writeFilePaths = writeToolCalls.flatMap((tc) => {
                     const args = this.parseToolArgs(tc.arguments);
-                    return typeof args.path === "string" ? args.path : null;
-                })
-                    .filter((p) => p !== null);
+                    const p = typeof args.path === "string" ? args.path
+                        : typeof args.file_path === "string" ? args.file_path
+                            : null;
+                    return p ? [p] : [];
+                });
                 if (writeFilePaths.length > 1) {
                     const groups = depAnalyzer.groupIndependentFiles(depGraph, writeFilePaths);
-                    // Build ordered list of file paths: independent groups first, dependent files after
-                    const orderedPaths = [];
+                    // Assign wave indices: independent groups get wave 0,
+                    // dependent groups get wave = max(their dep waves) + 1
+                    // For simplicity: canParallelize=true → wave 0, else sequential waves
+                    let wave = 0;
                     for (const group of groups) {
-                        for (const f of group.files) {
-                            if (!orderedPaths.includes(f))
-                                orderedPaths.push(f);
-                        }
-                    }
-                    // Reorder writeToolCalls according to orderedPaths
-                    const reordered = [];
-                    for (const filePath of orderedPaths) {
-                        const tc = writeToolCalls.find((c) => {
-                            const args = this.parseToolArgs(c.arguments);
-                            return args.path === filePath;
-                        });
-                        if (tc)
-                            reordered.push(tc);
-                    }
-                    // Add any write calls that didn't match a path (shouldn't happen, but be safe)
-                    for (const tc of writeToolCalls) {
-                        if (!reordered.includes(tc))
-                            reordered.push(tc);
-                    }
-                    // Rebuild toolCalls preserving non-write tools in their original positions,
-                    // replacing write tools with dependency-ordered sequence
-                    const reorderedAll = [];
-                    let writeIdx = 0;
-                    for (const tc of toolCalls) {
-                        if (writeToolNames.has(tc.name)) {
-                            reorderedAll.push(reordered[writeIdx++] ?? tc);
-                        }
-                        else {
-                            reorderedAll.push(tc);
+                        if (!group.canParallelize)
+                            wave++;
+                        for (const filePath of group.files) {
+                            const tc = writeToolCalls.find((c) => {
+                                const args = this.parseToolArgs(c.arguments);
+                                const p = args.path ?? args.file_path;
+                                return p === filePath;
+                            });
+                            if (tc)
+                                writeBatchMap.set(tc.id, wave);
                         }
+                        if (group.canParallelize)
+                            wave = 0; // reset: next independent group is also wave 0
                     }
-                    toolCalls = reorderedAll;
                 }
             }
             catch {
-                // DependencyAnalyzer failure is non-fatal — fall through to original ordering
+                // DependencyAnalyzer failure is non-fatal — all writes run sequentially
             }
         }
-        // Group tool calls: read-only can run in parallel, write tools run sequentially
-        const readOnlyTools = new Set(['file_read', 'grep', 'glob', 'code_search']);
+        // ─── Step 3: Build ordered batch list ────────────────────────────────────
+        // Final structure: array of batches, each batch runs in parallel.
+        // Reads accumulate until interrupted by a non-read tool.
+        // Writes are grouped by wave (same wave → parallel, different wave → sequential).
         const batches = [];
-        let currentBatch = [];
-        for (const toolCall of toolCalls) {
-            if (readOnlyTools.has(toolCall.name)) {
-                currentBatch.push(toolCall);
+        let readBatch = [];
+        const writeBatchGroups = new Map(); // wave → calls
+        const flushReadBatch = () => {
+            if (readBatch.length > 0) {
+                batches.push({ calls: [...readBatch], label: `${readBatch.length} read-only` });
+                readBatch = [];
+            }
+        };
+        const flushWriteBatches = () => {
+            if (writeBatchGroups.size === 0)
+                return;
+            const waves = [...writeBatchGroups.keys()].sort((a, b) => a - b);
+            for (const w of waves) {
+                const wCalls = writeBatchGroups.get(w);
+                batches.push({
+                    calls: wCalls,
+                    label: wCalls.length > 1
+                        ? `${wCalls.length} independent writes (wave ${w})`
+                        : `write: ${wCalls[0].name}`,
+                });
+            }
+            writeBatchGroups.clear();
+        };
+        for (const tc of toolCalls) {
+            if (READ_ONLY.has(tc.name)) {
+                // Reads accumulate; don't flush write batches yet (they don't conflict)
+                readBatch.push(tc);
+            }
+            else if (WRITE_TOOLS.has(tc.name)) {
+                // Flush any pending reads first (reads before writes)
+                flushReadBatch();
+                const wave = writeBatchMap.get(tc.id) ?? 99; // unknown dep → run last
+                if (!writeBatchGroups.has(wave))
+                    writeBatchGroups.set(wave, []);
+                writeBatchGroups.get(wave).push(tc);
             }
             else {
-                if (currentBatch.length > 0) {
-                    batches.push(currentBatch);
-                    currentBatch = [];
-                }
-                batches.push([toolCall]); // write tools run solo
+                // Heavy tool (shell_exec, git_ops, etc.) → flush everything, run solo
+                flushReadBatch();
+                flushWriteBatches();
+                batches.push({ calls: [tc], label: tc.name });
             }
         }
-        if (currentBatch.length > 0)
-            batches.push(currentBatch);
-        if (toolCalls.length > 1)
-            this.emitReasoning(`parallel tool batch started (${toolCalls.length} tools)`);
+        flushReadBatch();
+        flushWriteBatches();
+        // ─── Step 4: Execute batches ─────────────────────────────────────────────
+        if (toolCalls.length > 1) {
+            const parallelCount = batches.filter(b => b.calls.length > 1).length;
+            this.emitReasoning(parallelCount > 0
+                ? `executing ${toolCalls.length} tools in ${batches.length} batches (${parallelCount} parallel)`
+                : `executing ${toolCalls.length} tools sequentially`);
+        }
         const results = [];
         const deferredFixPrompts = [];
+        let interrupted = false;
         for (const batch of batches) {
-            if (batch.length === 1) {
-                // Sequential single tool execution
-                const { result, deferredFixPrompt } = await this.executeSingleTool(batch[0], toolCalls);
+            if (interrupted) {
+                // Fill remaining tools with SKIPPED placeholders
+                for (const tc of batch.calls) {
+                    results.push({
+                        tool_call_id: tc.id,
+                        name: tc.name,
+                        output: '[SKIPPED] Execution interrupted.',
+                        success: false,
+                        durationMs: 0,
+                    });
+                }
+                continue;
+            }
+            if (batch.calls.length === 1) {
+                // Single tool — sequential execution
+                const { result, deferredFixPrompt } = await this.executeSingleTool(batch.calls[0], toolCalls);
                 if (result)
                     results.push(result);
                 if (deferredFixPrompt)
                     deferredFixPrompts.push(deferredFixPrompt);
-                // Check for interrupt result — stop processing remaining batches
-                if (result && result.output.startsWith('[INTERRUPTED]')) {
-                    // Add placeholder results for remaining unexecuted tool calls
-                    const executedIds = new Set(results.map(r => r.tool_call_id));
-                    for (const tc of toolCalls) {
-                        if (!executedIds.has(tc.id)) {
-                            results.push({
-                                tool_call_id: tc.id,
-                                name: tc.name,
-                                output: `[SKIPPED] Previous tool was interrupted.`,
-                                success: false,
-                                durationMs: 0,
-                            });
-                        }
-                    }
-                    break;
-                }
+                if (result?.output.startsWith('[INTERRUPTED]'))
+                    interrupted = true;
             }
             else {
-                // Parallel execution for read-only tools in this batch
-                this.emitReasoning(`running ${batch.length} read-only tools in parallel`);
-                const batchResults = await Promise.allSettled(batch.map(tc => this.executeSingleTool(tc, toolCalls)));
-                for (const settled of batchResults) {
-                    if (settled.status === 'fulfilled') {
-                        if (settled.value.result)
-                            results.push(settled.value.result);
-                        if (settled.value.deferredFixPrompt)
-                            deferredFixPrompts.push(settled.value.deferredFixPrompt);
+                // Multi-tool — parallel execution
+                this.emitReasoning(`⚡ running ${batch.label} in parallel`);
+                const settled = await Promise.allSettled(batch.calls.map((tc) => this.executeSingleTool(tc, toolCalls)));
+                for (let i = 0; i < settled.length; i++) {
+                    const s = settled[i];
+                    const tc = batch.calls[i];
+                    if (s.status === 'fulfilled') {
+                        if (s.value.result)
+                            results.push(s.value.result);
+                        if (s.value.deferredFixPrompt)
+                            deferredFixPrompts.push(s.value.deferredFixPrompt);
+                        if (s.value.result?.output.startsWith('[INTERRUPTED]'))
+                            interrupted = true;
                     }
                     else {
-                        // Parallel tool failure — record as error result
-                        const tc = batch[batchResults.indexOf(settled)];
                         results.push({
-                            tool_call_id: tc?.id ?? 'unknown',
-                            name: tc?.name ?? 'unknown',
-                            output: `Error: ${settled.reason instanceof Error ? settled.reason.message : String(settled.reason)}`,
+                            tool_call_id: tc.id,
+                            name: tc.name,
+                            output: `Error: ${s.reason instanceof Error ? s.reason.message : String(s.reason)}`,
                             success: false,
                             durationMs: 0,
                         });
                     }
-                } // end for (const settled of batchResults)
-            } // end else (parallel batch)
-        } // end for (const batch of batches)
+                }
+            }
+        }
         return { results, deferredFixPrompts };
     }
     /**
@@ -2841,6 +3162,100 @@ export class AgentLoop extends EventEmitter {
             // 체크포인트 저장 실패는 치명적이지 않음
         }
     }
+    /**
+     * 매 iteration 시작 시 현재 플랜 진행 상황을 컨텍스트에 주입.
+     * 같은 태스크 인덱스라면 3 iteration마다만 주입 (컨텍스트 bloat 방지).
+     */
+    injectPlanProgress(iteration) {
+        if (!this.activePlan)
+            return;
+        const tasks = this.activePlan.tactical;
+        if (tasks.length === 0)
+            return;
+        const idx = this.currentTaskIndex;
+        const total = tasks.length;
+        // 같은 태스크면 3iteration마다, 태스크 전진 시 즉시 주입
+        const taskAdvanced = idx !== this._lastInjectedTaskIndex;
+        if (!taskAdvanced && iteration > 1 && iteration % 3 !== 1)
+            return;
+        this._lastInjectedTaskIndex = idx;
+        const lines = [
+            `## Plan Progress [${idx}/${total} done]`,
+        ];
+        for (let i = 0; i < tasks.length; i++) {
+            const task = tasks[i];
+            const marker = i < idx ? "✓" : i === idx ? "→" : "○";
+            lines.push(`${marker} Task ${i + 1}: ${task.description}`);
+            if (i === idx && task.targetFiles.length > 0) {
+                lines.push(`   Files: ${task.targetFiles.join(", ")}`);
+            }
+        }
+        if (idx < total) {
+            const cur = tasks[idx];
+            lines.push(`\nCurrent task: **${cur.description}**`);
+            if (cur.toolStrategy.length > 0) {
+                lines.push(`Suggested tools: ${cur.toolStrategy.join(", ")}`);
+            }
+            if (cur.readFiles.length > 0) {
+                lines.push(`Read first: ${cur.readFiles.join(", ")}`);
+            }
+        }
+        else {
+            lines.push(`\nAll tasks complete — verify and wrap up.`);
+        }
+        this.contextManager.addMessage({
+            role: "system",
+            content: lines.join("\n"),
+        });
+        this.iterationSystemMsgCount++;
+    }
+    /**
+     * 현재 태스크의 targetFiles가 changedFiles에 포함됐는지 확인해
+     * 완료 감지 시 다음 태스크로 자동 전진.
+     */
+    tryAdvancePlanTask() {
+        if (!this.activePlan)
+            return;
+        const tasks = this.activePlan.tactical;
+        if (this.currentTaskIndex >= tasks.length)
+            return;
+        const currentTask = tasks[this.currentTaskIndex];
+        if (!currentTask)
+            return;
+        // targetFiles가 없으면 tool call이 있었던 것만으로 완료 간주
+        if (currentTask.targetFiles.length === 0) {
+            if (this.allToolResults.length > 0) {
+                this.completedPlanTaskIds.add(currentTask.id);
+                this.currentTaskIndex++;
+                this._emitPlanAdvance(tasks);
+            }
+            return;
+        }
+        // targetFiles 중 하나라도 changedFiles에 있으면 완료
+        const changedBasenames = new Set(this.changedFiles.map((f) => f.split("/").pop().toLowerCase()));
+        const targetBasenames = currentTask.targetFiles.map((f) => f.split("/").pop().toLowerCase());
+        const hit = targetBasenames.some((b) => changedBasenames.has(b));
+        if (hit) {
+            this.completedPlanTaskIds.add(currentTask.id);
+            this.currentTaskIndex++;
+            this._emitPlanAdvance(tasks);
+        }
+    }
+    _emitPlanAdvance(tasks) {
+        const idx = this.currentTaskIndex;
+        if (idx < tasks.length) {
+            this.emitEvent({
+                kind: "agent:thinking",
+                content: `✓ Task ${idx}/${tasks.length} done. Next: ${tasks[idx].description}`,
+            });
+        }
+        else {
+            this.emitEvent({
+                kind: "agent:thinking",
+                content: `✓ All ${tasks.length} tasks completed.`,
+            });
+        }
+    }
     /**
      * 현재 plan에서 진행 상황을 추출한다.
      */