feat: release v1.0.0

src/config.ts

@@ -213,11 +213,12 @@ export interface OffloadConfig {
    * LLM execution mode for L1/L1.5/L2 tasks.
    * - "local": call LLM directly via AI SDK (uses offload.model or main agent model)
    * - "backend": route through remote backend service (requires backendUrl)
+   * - "client": stateless client mode — all compression delegated to offload server v2
    * - "collect": data collection only — runs L1/L1.5/L2 asynchronously but disables
    *   L3 compression and does NOT occupy the contextEngine slot (uses legacy compaction)
    * Default: "local" (auto-detects based on backendUrl presence for backward compat)
    */
-  mode: "local" | "backend" | "collect";
+  mode: "local" | "backend" | "client" | "collect";
   /** LLM model for offload tasks, format: "provider/model-id". Falls back to agents.defaults.model when omitted. */
   model?: string;
   /** LLM temperature (default: 0.2) */
@@ -265,6 +266,22 @@ export interface OffloadConfig {
    * primary non-loopback IPv4 address.
    */
   userId?: string;
+
+  // ── Client mode fields (used when mode === "client") ──────────────
+  /** Offload server v2 base URL (e.g. "http://localhost:9100"). */
+  serverUrl?: string;
+  /** Bearer token for offload server v2 Authorization header. */
+  apiKey?: string;
+  /** X-TDAI-Service-Id header value. */
+  serviceId?: string;
+  /** Agent name used in storage path (default: "default"). */
+  agentName?: string;
+  /** Client-side threshold: skip compaction when ratio < this value (default: 0.5). */
+  compactionRatio?: number;
+  /** Ingest request timeout in ms (default: 5000). */
+  ingestTimeoutMs?: number;
+  /** Compaction request timeout in ms (default: 30000). */
+  compactionTimeoutMs?: number;
 }
 
 /** Fully resolved plugin configuration (v3). */
@@ -439,10 +456,20 @@ export function parseConfig(raw: Record<string, unknown> | undefined): MemoryTda
   // --- Offload ---
   const offloadGroup = obj(c, "offload");
 
-  const offloadMode: "local" | "backend" | "collect" = (() => {
+  // Auto-derive offload serverUrl/apiKey/serviceId from top-level server config
+  // when offload client fields are not explicitly set.
+  const serverGroup = obj(c, "server");
+  const serverUrl = optStr(serverGroup, "url");
+  const serverApiKey = optStr(serverGroup, "apiKey");
+  const serverInstanceId = optStr(serverGroup, "instanceId");
+
+  const offloadMode: "local" | "backend" | "client" | "collect" = (() => {
     const raw = optStr(offloadGroup, "mode");
-    if (raw === "local" || raw === "backend" || raw === "collect") return raw;
-    return optStr(offloadGroup, "backendUrl") ? "backend" : "local";
+    if (raw === "local" || raw === "backend" || raw === "client" || raw === "collect") return raw;
+    // Auto-derive: if backendUrl is set → "backend"; if server.url is set → "client"; else "local"
+    if (optStr(offloadGroup, "backendUrl")) return "backend";
+    if (serverUrl) return "client";
+    return "local";
   })();
 
   const offload: OffloadConfig = {
@@ -465,6 +492,13 @@ export function parseConfig(raw: Record<string, unknown> | undefined): MemoryTda
     offloadRetentionDays: normalizeOffloadRetentionDays(num(offloadGroup, "offloadRetentionDays") ?? 0),
     logMaxSizeMb: num(offloadGroup, "logMaxSizeMb") ?? 50,
     userId: optStr(offloadGroup, "userId"),
+    // Client mode fields — fall back to top-level server config when not explicitly set
+    serverUrl: optStr(offloadGroup, "serverUrl") ?? serverUrl,
+    apiKey: optStr(offloadGroup, "apiKey") ?? serverApiKey,
+    serviceId: optStr(offloadGroup, "serviceId") ?? serverInstanceId,
+    compactionRatio: num(offloadGroup, "compactionRatio") ?? 0.5,
+    ingestTimeoutMs: num(offloadGroup, "ingestTimeoutMs") ?? 5000,
+    compactionTimeoutMs: num(offloadGroup, "compactionTimeoutMs") ?? 30000,
   };
 
   return {

src/core/conversation/l0-recorder.ts

@@ -18,6 +18,7 @@ import crypto from "node:crypto";
 import { sanitizeText, stripCodeBlocks, shouldCaptureL0 } from "../../utils/sanitize.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
+import type { Logger } from "../types.js";
 
 // ============================
 // Types
@@ -63,13 +64,6 @@ export interface L0ConversationRecord {
   messages: ConversationMessage[];
 }
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 const TAG = "[memory-tdai][l0]";
 
 // ============================

src/core/hooks/auto-capture.ts

@@ -20,14 +20,9 @@ import type { IMemoryStore, L0Record } from "../store/types.js";
 import type { EmbeddingService } from "../store/embedding.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 
-const TAG = "[memory-tdai] [capture]";
+import type { Logger } from "../types.js";
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+const TAG = "[memory-tdai] [capture]";
 
 export interface AutoCaptureResult {
   /** Whether the scheduler was notified (conversation count incremented) */

src/core/hooks/auto-recall.ts

@@ -21,6 +21,7 @@ import type { EmbeddingService, EmbeddingCallOptions } from "../store/embedding.
 import { sanitizeText } from "../../utils/sanitize.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
+import type { Logger } from "../types.js";
 
 const TAG = "[memory-tdai] [recall]";
 const RECALL_TRUNCATION_SUFFIX = "…（已截断；可用 tdai_memory_search 或 tdai_conversation_search 查看详情）";
@@ -46,13 +47,6 @@ const MEMORY_TOOLS_GUIDE = `<memory-tools-guide>
 - 若 3 次搜索后仍无结果，说明该信息不在记忆中，请直接根据已有信息回复用户，不要继续搜索。
 </memory-tools-guide>`
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 /** A single recalled L1 memory with its search score and type. */
 export interface RecalledMemory {
   content: string;
@@ -890,11 +884,15 @@ function normalizeBudgetLimit(value: number | undefined): number | undefined {
 }
 
 function truncateRecallLine(line: string, maxChars: number): string {
-  if (line.length <= maxChars) return line;
+  // Count and slice by code point, not UTF-16 code unit, so a cut never lands
+  // between the halves of a surrogate pair (which would corrupt a non-BMP
+  // character to U+FFFD when the line is UTF-8 encoded for the request).
+  const cps = Array.from(line);
+  if (cps.length <= maxChars) return line;
   if (maxChars <= RECALL_TRUNCATION_SUFFIX.length) {
-    return line.slice(0, maxChars);
+    return cps.slice(0, maxChars).join("");
   }
-  return `${line.slice(0, maxChars - RECALL_TRUNCATION_SUFFIX.length).trimEnd()}${RECALL_TRUNCATION_SUFFIX}`;
+  return `${cps.slice(0, maxChars - RECALL_TRUNCATION_SUFFIX.length).join("").trimEnd()}${RECALL_TRUNCATION_SUFFIX}`;
 }
 
 /**

src/core/persona/persona-generator.ts

@@ -12,19 +12,12 @@ import { BackupManager } from "../../utils/backup.js";
 import { escapeXmlTags } from "../../utils/sanitize.js";
 import { report } from "../report/reporter.js";
 import { reportL3LatencyMetrics } from "../report/metric-tracking-l3-latency.js";
-import type { LLMRunner } from "../types.js";
+import type { LLMRunner, Logger } from "../types.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
 
 const TAG = "[memory-tdai] [persona]";
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 export class PersonaGenerator {
   private dataDir: string;
   private runner: LLMRunner;

src/core/persona/persona-trigger.ts

@@ -8,14 +8,11 @@ import { stripSceneNavigation } from "../scene/scene-navigation.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
 
+import type { Logger } from "../types.js";
+
 const TAG = "[memory-tdai] [trigger]";
 
-interface TriggerLogger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+type TriggerLogger = Logger;
 
 export interface TriggerResult {
   should: boolean;

src/core/profile/profile-sync.ts

@@ -6,6 +6,7 @@ import { readSceneIndex, syncSceneIndex } from "../scene/scene-index.js";
 import { generateSceneNavigation, stripSceneNavigation } from "../scene/scene-navigation.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
+import type { Logger } from "../types.js";
 
 const PROFILE_SCOPE = "global";
 
@@ -15,13 +16,6 @@ function isRenameRaceError(err: unknown): boolean {
   return code === "ENOTEMPTY" || code === "EEXIST";
 }
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 export interface ProfileBaseline {
   version: number;
   contentMd5: string;

src/core/record/l1-dedup.ts

@@ -19,14 +19,7 @@ import { sanitizeJsonForParse } from "../../utils/sanitize.js";
 import type { IMemoryStore } from "../store/types.js";
 import { buildFtsQuery } from "../store/sqlite.js";
 import type { EmbeddingService } from "../store/embedding.js";
-import type { LLMRunner } from "../types.js";
-
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+import type { LLMRunner, Logger } from "../types.js";
 
 const TAG = "[memory-tdai][l1-dedup]";
 

src/core/record/l1-extractor.ts

@@ -24,16 +24,9 @@ import type { EmbeddingService } from "../store/embedding.js";
 import { report } from "../report/reporter.js";
 import { metricProducer } from "../report/kafka-metric-producer.js";
 import { reportL1LatencyMetrics } from "../report/metric-tracking-l1-latency.js";
-import type { LLMRunner } from "../types.js";
+import type { LLMRunner, Logger } from "../types.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 const TAG = "[memory-tdai][l1-extractor]";
 
 // ============================

src/core/record/l1-reader.ts

@@ -19,13 +19,7 @@ import { StoragePaths } from "../storage/types.js";
 // Re-export types that readers need
 export type { MemoryRecord, MemoryType, EpisodicMetadata } from "./l1-writer.js";
 export type { L1QueryFilter } from "../store/types.js";
-
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+import type { Logger } from "../types.js";
 
 const TAG = "[memory-tdai] [l1-reader]";
 

src/core/record/l1-writer.ts

@@ -21,6 +21,7 @@ import type { IMemoryStore } from "../store/types.js";
 import type { EmbeddingService } from "../store/embedding.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
+import type { Logger } from "../types.js";
 
 // ============================
 // Types
@@ -110,13 +111,6 @@ export interface DedupDecision {
   merged_timestamps?: string[];
 }
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 const TAG = "[memory-tdai][l1-writer]";
 
 // ============================

src/core/report/metric-tracking-runner.ts

@@ -198,14 +198,19 @@ export class MetricTrackingRunner implements LLMRunner {
   }
 
   async run(params: LLMRunParams): Promise<string> {
+    // 0. 注入 instanceId（如果调用方没传，从 getInstanceId 回调获取）
+    const enrichedParams = params.instanceId
+      ? params
+      : { ...params, instanceId: this.getInstanceId() };
+
     // 1. 先执行原方法，拿到结果（异常直接 re-throw）
-    const text = await this.inner.run(params);
+    const text = await this.inner.run(enrichedParams);
 
     // 2. 原方法成功后，try-catch 做上报（静默失败）
     try {
       const metricName = taskIdToMetricName(params.taskId);
       if (metricName) {
-        const instanceId = params.instanceId ?? this.getInstanceId();
+        const instanceId = enrichedParams.instanceId ?? this.getInstanceId();
         if (instanceId) {
           // 优先从 inner runner 的 lastUsage side-channel 读取精确 token 数
           const innerWithUsage = this.inner as LLMRunnerWithUsage;

src/core/report/traced-task-executor.ts

@@ -62,6 +62,24 @@ export class TracedTaskExecutor implements TaskExecutor {
     return this.executeL1(task);
   }
 
+  async executeOffloadL1?(task: TaskPayload, signal?: AbortSignal): Promise<void> {
+    if (this.inner.executeOffloadL1) {
+      return this.executeWithTrace("offload-l1", task, () => this.inner.executeOffloadL1!(task, signal));
+    }
+  }
+
+  async executeOffloadL15?(task: TaskPayload, signal?: AbortSignal): Promise<void> {
+    if (this.inner.executeOffloadL15) {
+      return this.executeWithTrace("offload-l15", task, () => this.inner.executeOffloadL15!(task, signal));
+    }
+  }
+
+  async executeOffloadL2?(task: TaskPayload, signal?: AbortSignal): Promise<void> {
+    if (this.inner.executeOffloadL2) {
+      return this.executeWithTrace("offload-l2", task, () => this.inner.executeOffloadL2!(task, signal));
+    }
+  }
+
   /**
    * 核心方法：在 Trace Context 中执行任务。
    *

src/core/scene/scene-extractor.ts

@@ -27,18 +27,13 @@ import { normalizeSceneFilenames } from "./filename-normalizer.js";
 import { buildSceneExtractionPrompt } from "../prompts/scene-extraction.js";
 import { report } from "../report/reporter.js";
 import { reportL2LatencyMetrics } from "../report/metric-tracking-l2-latency.js";
-import type { LLMRunner } from "../types.js";
+import type { LLMRunner, Logger } from "../types.js";
 import type { StorageAdapter } from "../storage/adapter.js";
 import { StoragePaths } from "../storage/types.js";
 
 const TAG = "[memory-tdai] [extractor]";
 
-interface ExtractorLogger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+type ExtractorLogger = Logger;
 
 export interface ExtractionResult {
   memoriesProcessed: number;
@@ -470,24 +465,24 @@ export class SceneExtractor {
         success: true,
         error: null,
       });
-    }
 
-    // ── 评测指标：L2 延迟 + 场景变化 ──
-    try {
-      const postSceneCount = (await readSceneIndex(this.dataDir, this.storage)).length;
-      reportL2LatencyMetrics({
-        instanceId: this.instanceId ?? "",
-        extractionLatencyMs: totalMs,
-        llmDurationMs: llmDurationMs > 0 ? llmDurationMs : null,
-        sceneCountBefore: preExtractIndex.size,
-        sceneCountAfter: postSceneCount,
-        scenesCreated,
-        scenesUpdated,
-        scenesDeleted,
-        hasError: false,
-      });
-    } catch {
-      // 静默忽略
+      // ── 评测指标：L2 延迟 + 场景变化 ──
+      try {
+        const postSceneCount = (await readSceneIndex(this.dataDir, this.storage)).length;
+        reportL2LatencyMetrics({
+          instanceId: this.instanceId ?? "",
+          extractionLatencyMs: totalMs,
+          llmDurationMs: llmDurationMs > 0 ? llmDurationMs : null,
+          sceneCountBefore: preExtractIndex.size,
+          sceneCountAfter: postSceneCount,
+          scenesCreated,
+          scenesUpdated,
+          scenesDeleted,
+          hasError: false,
+        });
+      } catch {
+        // 静默忽略
+      }
     }
 
     // Detect empty extraction: pre and post scene_index both empty means LLM didn't write anything

src/core/state/types.ts

@@ -56,7 +56,7 @@ export interface TimerEntry {
 
 export interface TaskPayload {
   id: string;
-  type: "L1" | "L2" | "L3" | "flush";
+  type: "L1" | "L2" | "L3" | "flush" | "offload-l1" | "offload-l15" | "offload-l2";
   instanceId: string;
   sessionId: string;
   priority: number; // 0=high, 1=normal, 2=low

src/core/store/bm25-client.ts

@@ -13,6 +13,8 @@
  * check health to dynamically downgrade to pure semantic search.
  */
 
+import type { Logger } from "../types.js";
+
 // ============================
 // Types
 // ============================
@@ -20,13 +22,6 @@
 /** Sparse vector: array of [token_hash, weight] pairs. */
 export type SparseVector = Array<[number, number]>;
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 export interface BM25ClientConfig {
   /** Sidecar service URL (default: "http://127.0.0.1:8084") */
   serviceUrl: string;

src/core/store/bm25-local.ts

@@ -11,16 +11,10 @@
 
 import { BM25Encoder } from "@tencentdb-agent-memory/tcvdb-text";
 import type { SparseVector } from "@tencentdb-agent-memory/tcvdb-text";
+import type { Logger } from "../types.js";
 
 export type { SparseVector };
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 export interface BM25LocalConfig {
   /** Whether BM25 sparse encoding is enabled (default: true) */
   enabled: boolean;

src/core/store/embedding.ts

@@ -13,6 +13,8 @@
  * - Throws on failure; callers decide fallback strategy.
  */
 
+import type { Logger } from "../types.js";
+
 // ============================
 // Types
 // ============================
@@ -104,17 +106,6 @@ export class EmbeddingNotReadyError extends Error {
   }
 }
 
-// ============================
-// Logger interface
-// ============================
-
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 const TAG = "[memory-tdai][embedding]";
 
 // ============================

src/core/store/sqlite.ts

@@ -21,6 +21,8 @@
  */
 
 import { createRequire } from "node:module";
+import { mkdirSync, existsSync } from "node:fs";
+import path from "node:path";
 import type { DatabaseSync, StatementSync, SQLInputValue } from "node:sqlite";
 import type { MemoryRecord } from "../record/l1-writer.js";
 import type { EmbeddingProviderInfo } from "./embedding.js";
@@ -40,6 +42,9 @@ import type {
   L1PaginatedFilter,
   L1PaginatedResult,
 } from "./types.js";
+import type { Logger } from "../types.js";
+
+export type { L1RecordRow } from "./types.js";
 
 // ============================
 // Types
@@ -86,13 +91,6 @@ export interface L0RecordRow {
   timestamp: number;
 }
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 const TAG = "[memory-tdai][sqlite]";
 
 /** Persisted metadata about the embedding provider used to generate stored vectors. */
@@ -394,6 +392,12 @@ export class VectorStore implements IMemoryStore {
     this.dimensions = dimensions;
     this.logger = logger;
 
+    // Ensure parent directory exists (for non-default instance paths)
+    const dbDir = path.dirname(dbPath);
+    if (!existsSync(dbDir)) {
+      mkdirSync(dbDir, { recursive: true });
+    }
+
     // Open database with extension support enabled
     const { DatabaseSync: DbSync } = requireNodeSqlite();
     this.db = new DbSync(dbPath, { allowExtension: true });

src/core/store/store-pool.ts

@@ -16,6 +16,7 @@
  */
 
 import path from "node:path";
+import { existsSync, mkdirSync } from "node:fs";
 import type { MemoryTdaiConfig } from "../../config.js";
 import type { IMemoryStore, StoreLogger } from "./types.js";
 import type { EmbeddingService } from "./embedding.js";
@@ -303,6 +304,11 @@ export class StorePool {
 
     const dims = embCfg.dimensions ?? 0;
     const dbPath = this.getSqlitePath(instanceId);
+    // 确保数据库目录存在（对于非 default instance）
+    const dbDir = path.dirname(dbPath);
+    if (!existsSync(dbDir)) {
+      mkdirSync(dbDir, { recursive: true });
+    }
     const store = new VectorStore(dbPath, dims, this.logger as StoreLogger);
 
     return {

src/core/store/types.ts

@@ -17,6 +17,7 @@
 
 import type { MemoryRecord } from "../record/l1-writer.js";
 import type { EmbeddingProviderInfo } from "./embedding.js";
+import type { Logger } from "../types.js";
 
 // Re-export so consumers can import everything from types.ts
 export type { MemoryRecord, EmbeddingProviderInfo };
@@ -26,12 +27,7 @@ export type { MemoryRecord, EmbeddingProviderInfo };
 // ============================
 
 /** Minimal logger interface accepted by store implementations. */
-export interface StoreLogger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+export type StoreLogger = Logger;
 
 // ============================
 // L1 Types (Structured Memories)

src/core/tools/conversation-search.ts

@@ -13,18 +13,12 @@
 import type { IMemoryStore, L0SearchResult } from "../store/types.js";
 import { buildFtsQuery } from "../store/sqlite.js";
 import type { EmbeddingService } from "../store/embedding.js";
+import type { Logger } from "../types.js";
 
 // ============================
 // Types
 // ============================
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 export interface ConversationSearchResultItem {
   id: string;
   session_key: string;

src/core/tools/memory-search.ts

@@ -13,18 +13,12 @@
 import type { IMemoryStore, L1SearchResult } from "../store/types.js";
 import { buildFtsQuery } from "../store/sqlite.js";
 import type { EmbeddingService } from "../store/embedding.js";
+import type { Logger } from "../types.js";
 
 // ============================
 // Types
 // ============================
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 export interface MemorySearchResultItem {
   id: string;
   content: string;

src/core/types.ts

@@ -15,10 +15,10 @@
 // ============================
 
 /**
- * Minimal logger interface used throughout TDAI Core.
+ * Canonical logger interface used across all TDAI modules.
  *
- * Matches the existing `StoreLogger` and `RunnerLogger` interfaces
- * already used in the codebase — no migration needed for existing callers.
+ * Named variants (StoreLogger, PluginLogger, etc.) are type aliases
+ * of this interface, kept for backward compatibility.
  */
 export interface Logger {
   debug?: (message: string) => void;

src/gateway/config.ts

@@ -292,6 +292,25 @@ export interface GatewayConfig {
   cos: CosExtraConfig;
   /** 可观测性配置 (yaml: observability, env: KAFKA_METRIC_*) */
   observability: ObservabilityConfig;
+  /** Offload server executor 配置 (yaml: offload) */
+  offload: {
+    forceTriggerThreshold: number;
+    pendingMaxAgeSeconds: number;
+    l1Temperature: number;
+    l1MaxTokens: number;
+    l1TimeoutMs: number;
+    l15Temperature: number;
+    l15MaxTokens: number;
+    l15TimeoutMs: number;
+    l2Temperature: number;
+    l2MaxTokens: number;
+    l2TimeoutMs: number;
+    l2NullThreshold: number;
+    mildOffloadRatio: number;
+    aggressiveCompressRatio: number;
+    emergencyCompressRatio: number;
+    maxRetries: number;
+  };
 }
 
 // ============================
@@ -491,6 +510,27 @@ export function loadGatewayConfig(overrides?: Partial<GatewayConfig>): GatewayCo
 
   const observability: ObservabilityConfig = { otel, clickhouse, kafka, langfuse };
 
+  // Offload executor config (yaml: offload)
+  const offloadConfig = obj(fileConfig, "offload");
+  const offload = {
+    forceTriggerThreshold: num(offloadConfig, "forceTriggerThreshold") ?? 4,
+    pendingMaxAgeSeconds: num(offloadConfig, "pendingMaxAgeSeconds") ?? 30,
+    l1Temperature: num(offloadConfig, "l1Temperature") ?? 0.3,
+    l1MaxTokens: num(offloadConfig, "l1MaxTokens") ?? 8000,
+    l1TimeoutMs: num(offloadConfig, "l1TimeoutMs") ?? 120_000,
+    l15Temperature: num(offloadConfig, "l15Temperature") ?? 0.2,
+    l15MaxTokens: num(offloadConfig, "l15MaxTokens") ?? 3000,
+    l15TimeoutMs: num(offloadConfig, "l15TimeoutMs") ?? 120_000,
+    l2Temperature: num(offloadConfig, "l2Temperature") ?? 0.4,
+    l2MaxTokens: num(offloadConfig, "l2MaxTokens") ?? 16000,
+    l2TimeoutMs: num(offloadConfig, "l2TimeoutMs") ?? 120_000,
+    l2NullThreshold: num(offloadConfig, "l2NullThreshold") ?? 6,
+    mildOffloadRatio: num(offloadConfig, "mildOffloadRatio") ?? 0.5,
+    aggressiveCompressRatio: num(offloadConfig, "aggressiveCompressRatio") ?? 0.85,
+    emergencyCompressRatio: num(offloadConfig, "emergencyCompressRatio") ?? 0.95,
+    maxRetries: num(offloadConfig, "maxRetries") ?? 3,
+  };
+
   const base: GatewayConfig = {
     deployMode,
     stateBackend,
@@ -505,6 +545,7 @@ export function loadGatewayConfig(overrides?: Partial<GatewayConfig>): GatewayCo
     worker,
     cos,
     observability,
+    offload,
   };
 
   // Merge overrides one level deep so partial `server`/`data`/`llm` patches

src/gateway/error-handler.ts

@@ -46,6 +46,7 @@ export interface ClassifiedError {
  *
  * Recognized:
  *   - PayloadTooLargeError (CR-7) — detected via duck-typing on statusCode === 413
+ *   - Invalid JSON body — parseJsonBody throws Error("Invalid JSON body") on malformed input
  *   - RecallFailure (H-15) — code from RecallError taxonomy
  *   - SeedValidationError — 400 with generic message
  *   - Anything else — 500 Internal server error
@@ -64,6 +65,37 @@ export function classifyError(err: unknown): ClassifiedError {
     };
   }
 
+  // 1b. Invalid JSON body / decompression error — parseJsonBody throws this fixed message on
+  // JSON.parse failure or gzip/deflate decompression failure. Client error, not server fault.
+  if (err instanceof Error && err.message === "Invalid JSON body") {
+    return {
+      status: 400,
+      client: { code: 400, message: "Invalid JSON body", trace_id, retryable: false },
+      logLine: `[${trace_id}] InvalidJsonBody: request body could not be parsed as JSON`,
+    };
+  }
+
+  // 1c. COS AppendPositionErr — concurrent append conflict. The client should retry.
+  // This is a transient conflict, not a permanent error.
+  if (err instanceof Error && /AppendPositionErr|Position not equal object length/i.test(err.message)) {
+    return {
+      status: 409,
+      client: { code: 409, message: "Concurrent write conflict, please retry", trace_id, retryable: true },
+      logLine: `[${trace_id}] CosAppendConflict: ${err.message}`,
+    };
+  }
+
+  // 1d. Unsupported Content-Encoding — parseJsonBody rejects with this message when the
+  // client sends an encoding the server does not support (not gzip/deflate/identity).
+  if (err instanceof Error && err.message.startsWith("Unsupported Content-Encoding:")) {
+    const safeMsg = err.message.replace(/[^\w\s:/-]/g, "");
+    return {
+      status: 415,
+      client: { code: 415, message: safeMsg, trace_id, retryable: false },
+      logLine: `[${trace_id}] UnsupportedEncoding: ${err.message}`,
+    };
+  }
+
   // 2. RecallFailure (H-15) — the recall layer has already produced a safe message.
   // (Gateway handlers normally translate this into RecallResult.error before reaching
   // the global catch, but if a code path forgets, this fallback ensures no raw leak.)

src/gateway/server.ts

@@ -17,6 +17,7 @@
 import http from "node:http";
 import { URL } from "node:url";
 import { timingSafeEqual } from "node:crypto";
+import zlib from "node:zlib";
 import dayjs from "dayjs";
 import { TdaiCore } from "../core/tdai-core.js";
 import { StandaloneHostAdapter } from "../adapters/standalone/host-adapter.js";
@@ -53,6 +54,9 @@ import { executeSeed } from "../core/seed/seed-runtime.js";
 import type { SeedProgress } from "../core/seed/types.js";
 import { handleV2Route, errorEnvelope, makeRequestId } from "./v2-router.js";
 import type { V2RouterDeps } from "./v2-router.js";
+import { handleOffloadV2Route } from "../offload_server/router.js";
+import type { OffloadV2Deps } from "../offload_server/router.js";
+import { initServerOpikTracer } from "../offload_server/opik-tracer.js";
 import { classifyError } from "./error-handler.js";
 import { LocalStorageBackend } from "../core/storage/local-backend.js";
 import { StorageAdapter } from "../core/storage/adapter.js";
@@ -142,11 +146,25 @@ export async function parseJsonBody<T>(req: http.IncomingMessage): Promise<T> {
       return;
     }
 
+    // Determine if the body is compressed (Content-Encoding header).
+    // Support gzip and deflate; reject unsupported encodings with 400.
+    const encoding = (req.headers["content-encoding"] ?? "").toLowerCase().trim();
+    let source: NodeJS.ReadableStream = req;
+    if (encoding === "gzip" || encoding === "x-gzip") {
+      source = req.pipe(zlib.createGunzip());
+    } else if (encoding === "deflate") {
+      source = req.pipe(zlib.createInflate());
+    } else if (encoding !== "" && encoding !== "identity") {
+      req.resume(); // drain
+      reject(new Error(`Unsupported Content-Encoding: ${encoding}`));
+      return;
+    }
+
     const chunks: Buffer[] = [];
     let received = 0;
     let aborted = false;
 
-    req.on("data", (chunk: Buffer) => {
+    source.on("data", (chunk: Buffer) => {
       if (aborted) return;
       received += chunk.length;
       if (received > MAX_BODY_BYTES) {
@@ -157,18 +175,19 @@ export async function parseJsonBody<T>(req: http.IncomingMessage): Promise<T> {
       }
       chunks.push(chunk);
     });
-    req.on("end", () => {
+    source.on("end", () => {
       if (aborted) return;
       try {
         const body = Buffer.concat(chunks).toString("utf-8");
         resolve(JSON.parse(body) as T);
-      } catch (err) {
+      } catch {
         reject(new Error("Invalid JSON body"));
       }
     });
-    req.on("error", (err) => {
+    source.on("error", (_err) => {
       if (aborted) return;  // already rejected with PayloadTooLargeError
-      reject(err);
+      // Decompression errors (e.g. truncated gzip) are client-side faults
+      reject(new Error("Invalid JSON body"));
     });
   });
 }
@@ -333,6 +352,9 @@ export class TdaiGateway {
     // Initialize core
     await this.core.initialize();
 
+    // ── Initialize Opik tracer for offload server ──
+    await initServerOpikTracer(this.logger);
+
     // ── Initialize StorageAdapter for v2 API ──
     // In standalone mode, use LocalStorageBackend pointing to dataDir.
     // In service mode, CosStorageBackend is injected externally.
@@ -578,6 +600,17 @@ export class TdaiGateway {
         }
       }
 
+      // ── Offload V2 routes (async ingest + mmd query) ──
+      const offloadDeps: OffloadV2Deps = {
+        resolveStorage: v2Deps.resolveStorage,
+        getStorage: v2Deps.getStorage ?? (() => undefined),
+        logger: this.logger,
+        stateBackend: this.stateBackend,
+        config: { ...this.config.offload, l1Model: "", l15Model: "", l2Model: "" },
+      };
+      const offloadHandled = await handleOffloadV2Route(req, res, pathname, method, parseJsonBody, sendJson, offloadDeps);
+      if (offloadHandled) return;
+
       const handled = await handleV2Route(req, res, pathname, method, parseJsonBody, sendJson, v2Deps);
       if (handled) return;
 
@@ -1101,26 +1134,65 @@ export class TdaiGateway {
       type: backendType,
       local: backendType === "local" ? {
         onTimerExpired: (entry) => {
-          // Parse timer member: "sessionId:L2_schedule" or "sessionId:L1_idle"
-          const parts = entry.member.split(":");
-          const sessionId = parts.slice(0, -1).join(":");
-          const timerType = parts[parts.length - 1];
-          const taskType = timerType === "L2_schedule" ? "L2" : timerType === "L1_idle" ? "L1" : "L3";
-          const instanceId = this.config.instanceId ?? "default";
+          // Parse timer member by prefix: "offload-{type}:{instanceId}:{sessionId}[:{extra}]"
+          // or legacy "session:L2_schedule"
+          const member = entry.member;
+          let taskType: string;
+          let instanceId: string;
+          let sessionId: string;
+
+          const firstColon = member.indexOf(":");
+          const prefix = firstColon > 0 ? member.slice(0, firstColon) : member;
+
+          if (prefix === "offload-l1" || prefix === "offload-l15" || prefix === "offload-l2") {
+            taskType = prefix;
+            // Format: "offload-{type}:{instanceId}:{sessionId}[:{mmdFile}]"
+            // instanceId is the segment right after the prefix
+            const rest = member.slice(firstColon + 1);
+            const instanceEnd = rest.indexOf(":");
+            if (instanceEnd > 0) {
+              instanceId = rest.slice(0, instanceEnd);
+              sessionId = rest.slice(instanceEnd + 1);
+            } else {
+              instanceId = this.config.instanceId ?? "default";
+              sessionId = rest;
+            }
+            // For offload-l2: strip trailing ":{mmdFile}" from sessionId
+            // (mmdFile is extracted separately from timerMember in the executor)
+            if (prefix === "offload-l2" && sessionId.endsWith(".mmd")) {
+              const lastColon = sessionId.lastIndexOf(":");
+              if (lastColon > 0) {
+                sessionId = sessionId.slice(0, lastColon);
+              }
+            }
+          } else {
+            // Legacy format: "sessionId:L2_schedule" or "sessionId:L1_idle"
+            const lastColon = member.lastIndexOf(":");
+            const suffix = lastColon >= 0 ? member.slice(lastColon + 1) : "";
+            sessionId = lastColon >= 0 ? member.slice(0, lastColon) : member;
+            taskType = suffix === "L2_schedule" ? "offload-l2" : suffix === "L1_idle" ? "offload-l1" : "L3";
+            instanceId = this.config.instanceId ?? "default";
+          }
           const now = Date.now();
+          // Extract targetMmdFile from member for offload-l2 (needed by pipeline-worker lockKey)
+          let targetMmdFile: string | undefined;
+          if (taskType === "offload-l2") {
+            const mmdMatch = member.match(/(\d+-[^:]+\.mmd)$/);
+            if (mmdMatch) targetMmdFile = mmdMatch[1];
+          }
           const task = {
             id: `${taskType}-${sessionId}-${now}`,
-            type: taskType,
+            type: taskType as any,
             instanceId,
             sessionId,
             priority: 0,
             createdAt: now,
-            data: { triggeredBy: "timer_scanner", timerMember: entry.member, instanceId },
+            data: { triggeredBy: "timer_scanner", timerMember: member, instanceId, targetMmdFile },
           };
           this.stateBackend!.enqueueTask(task).then(() => {
-            this.logger.info(`[local-timer] Timer fired: ${entry.member} → enqueued ${taskType} task`);
+            this.logger.info(`[local-timer] Timer fired: ${member} → enqueued ${taskType} task`);
           }).catch((err) => {
-            this.logger.error(`[local-timer] Failed to enqueue task for ${entry.member}: ${err instanceof Error ? err.message : String(err)}`);
+            this.logger.error(`[local-timer] Failed to enqueue task for ${member}: ${err instanceof Error ? err.message : String(err)}`);
           });
         },
       } : undefined,
@@ -1571,6 +1643,120 @@ export class TdaiGateway {
       async executeFlush(task: TaskPayload) {
         await core.handleSessionEnd(task.sessionId);
       },
+
+      // ── Offload executors (L1 summary, L1.5 task judgment, L2 MMD update) ──
+      async executeOffloadL1(task: TaskPayload, signal?: AbortSignal) {
+        if (signal?.aborted) return;
+        const { OffloadTaskExecutor } = await import("../offload_server/offload-task-executor.js");
+        const storage = await resolveStorage(task);
+        if (!storage) return;
+        const llmClient = gateway.buildOffloadLlmClient();
+        if (!llmClient) {
+          gateway.logger.warn(`[executor] offload-l1 skipped: no LLM client available`);
+          return;
+        }
+        const executor = new OffloadTaskExecutor({
+          resolveStorage: async () => storage,
+          llmClient,
+          stateBackend: gateway.stateBackend!,
+          config: { ...gateway.config.offload, l1Model: "", l15Model: "", l2Model: "" },
+          logger: gateway.logger,
+        });
+        await executor.executeOffloadL1(task, signal);
+      },
+      async executeOffloadL15(task: TaskPayload, signal?: AbortSignal) {
+        if (signal?.aborted) return;
+        const { OffloadTaskExecutor } = await import("../offload_server/offload-task-executor.js");
+        const storage = await resolveStorage(task);
+        if (!storage) return;
+        const llmClient = gateway.buildOffloadLlmClient();
+        if (!llmClient) {
+          gateway.logger.warn(`[executor] offload-l15 skipped: no LLM client available`);
+          return;
+        }
+        const executor = new OffloadTaskExecutor({
+          resolveStorage: async () => storage,
+          llmClient,
+          stateBackend: gateway.stateBackend!,
+          config: { ...gateway.config.offload, l1Model: "", l15Model: "", l2Model: "" },
+          logger: gateway.logger,
+        });
+        await executor.executeOffloadL15(task, signal);
+      },
+      async executeOffloadL2(task: TaskPayload, signal?: AbortSignal) {
+        if (signal?.aborted) return;
+        const { OffloadTaskExecutor } = await import("../offload_server/offload-task-executor.js");
+        const storage = await resolveStorage(task);
+        if (!storage) return;
+        const llmClient = gateway.buildOffloadLlmClient();
+        if (!llmClient) {
+          gateway.logger.warn(`[executor] offload-l2 skipped: no LLM client available`);
+          return;
+        }
+        const executor = new OffloadTaskExecutor({
+          resolveStorage: async () => storage,
+          llmClient,
+          stateBackend: gateway.stateBackend!,
+          config: { ...gateway.config.offload, l1Model: "", l15Model: "", l2Model: "" },
+          logger: gateway.logger,
+        });
+        await executor.executeOffloadL2(task, signal);
+      },
+    };
+  }
+
+  /**
+   * Build a simple LLM client for offload executors using gateway's LLM config.
+   */
+  private buildOffloadLlmClient() {
+    const llmCfg = this.config.llm;
+    if (!llmCfg.baseUrl || !llmCfg.apiKey || !llmCfg.model) return null;
+    const logger = this.logger;
+
+    return {
+      async chat(params: {
+        model: string;
+        messages: Array<{ role: "system" | "user"; content: string }>;
+        temperature: number;
+        max_tokens: number;
+        timeoutMs?: number;
+      }): Promise<string> {
+        const controller = new AbortController();
+        const timer = setTimeout(() => controller.abort(), params.timeoutMs ?? 30000);
+        try {
+          const response = await fetch(`${llmCfg.baseUrl}/chat/completions`, {
+            method: "POST",
+            headers: {
+              "Content-Type": "application/json",
+              Authorization: `Bearer ${llmCfg.apiKey}`,
+            },
+            body: JSON.stringify({
+              model: llmCfg.model || params.model,
+              messages: params.messages,
+              temperature: params.temperature,
+              max_tokens: params.max_tokens,
+            }),
+            signal: controller.signal,
+          });
+          clearTimeout(timer);
+          if (!response.ok) {
+            throw new Error(`LLM API returned ${response.status}: ${await response.text()}`);
+          }
+          const json = (await response.json()) as any;
+          const finishReason = json.choices?.[0]?.finish_reason;
+          if (finishReason === "length") {
+            const content = json.choices?.[0]?.message?.content ?? "";
+            logger.warn(
+              `[offload-llm] Response truncated (finish_reason=length, max_tokens=${params.max_tokens}), ` +
+              `content=${content.length} chars`,
+            );
+          }
+          return json.choices?.[0]?.message?.content ?? "";
+        } catch (err) {
+          clearTimeout(timer);
+          throw err;
+        }
+      },
     };
   }
 }

src/offload-client/context-engine.ts

@@ -0,0 +1,526 @@
+/**
+ * offload-client — OffloadContextEngine.
+ * Occupies the Context Engine slot and delegates compression to the server.
+ */
+import type { OffloadClientConfig, RecentMessage, Logger } from "./types.js";
+import type { OffloadApiClient } from "./offload-api-client.js";
+import { estimateAllTokens, estimateMessageTokens } from "./token-estimator.js";
+
+const DEFAULT_CONTEXT_WINDOW = 128000;
+/** compact target: keep messages until total <= contextWindow * TARGET_RATIO */
+const COMPACT_TARGET_RATIO = 0.5;
+/** When truncating a single large tool_result, keep at most this many chars */
+const TOOL_RESULT_TRUNCATE_CHARS = 2000;
+
+// ─── Message role helpers (handle multiple formats) ─────────────────────────
+
+function getMsgRole(msg: any): string {
+  return msg?.role ?? msg?.message?.role ?? msg?.type ?? "";
+}
+
+function isToolResult(msg: any): boolean {
+  const role = getMsgRole(msg);
+  if (role === "tool" || role === "toolResult" || role === "tool_result") return true;
+  // Anthropic: user message with tool_result content blocks
+  if (role === "user" && Array.isArray(msg?.content)) {
+    return msg.content.some((b: any) => b?.type === "tool_result");
+  }
+  return false;
+}
+
+function isAssistantWithToolUse(msg: any): boolean {
+  const role = getMsgRole(msg);
+  if (role !== "assistant") return false;
+  const content = msg?.type === "message" ? msg?.message?.content : msg?.content;
+  if (!Array.isArray(content)) return false;
+  return content.some((b: any) => b?.type === "tool_use" || b?.type === "toolCall");
+}
+
+/**
+ * Truncate tool_result content in-place, returning a shallow-cloned message.
+ */
+function truncateToolResult(msg: any, maxChars: number): any {
+  const clone = JSON.parse(JSON.stringify(msg));
+  const content = clone.type === "message" ? clone.message?.content : clone.content;
+  if (typeof content === "string" && content.length > maxChars) {
+    const truncated = content.slice(0, maxChars) + "\n...[truncated]";
+    if (clone.type === "message") clone.message.content = truncated;
+    else clone.content = truncated;
+    return clone;
+  }
+  if (Array.isArray(content)) {
+    for (const block of content) {
+      if (typeof block?.text === "string" && block.text.length > maxChars) {
+        block.text = block.text.slice(0, maxChars) + "\n...[truncated]";
+      }
+      if (typeof block?.content === "string" && block.content.length > maxChars) {
+        block.content = block.content.slice(0, maxChars) + "\n...[truncated]";
+      }
+    }
+    return clone;
+  }
+  return clone;
+}
+
+export class OffloadContextEngine {
+  /** Per-session state cache. */
+  private sessions = new Map<string, {
+    lastAccessMs: number;  // ← NEW: track last access time for cleanup
+    lastKnownTotalTokens: number;
+    lastKnownMsgCount: number;
+    lastL15PromptHash: string;
+    cachedPrompt?: string;
+    cachedRecentMessages: RecentMessage[];
+    cachedRecentContext?: string;
+  }>();
+
+  /** Get or create per-session state. */
+  private getSession(sessionKey: string) {
+    let s = this.sessions.get(sessionKey);
+    if (!s) {
+      s = {
+        lastAccessMs: Date.now(),  // ← NEW
+        lastKnownTotalTokens: 0,
+        lastKnownMsgCount: 0,
+        lastL15PromptHash: "",
+        cachedRecentMessages: [],
+      };
+      this.sessions.set(sessionKey, s);
+    } else {
+      s.lastAccessMs = Date.now();  // ← NEW: update on access
+    }
+    return s;
+  }
+
+  /**
+   * Reset session state (call when session is /new'd or destroyed).
+   */
+  resetSession(sessionKey: string): void {
+    this.sessions.delete(sessionKey);
+  }
+
+  /**
+   * Clear all session states (emergency shutdown).
+   */
+  clearAllSessions(): void {
+    const n = this.sessions.size;
+    this.sessions.clear();
+    if (n > 0) {
+      this.logger.info(`[offload-client] cleared ${n} session states`);
+    }
+  }
+
+  /**
+   * Get the cached context for a session (for after_tool_call hook to send with ingest).
+   */
+  getContext(sessionKey?: string): { prompt?: string; recentMessages?: RecentMessage[] } | undefined {
+    if (!sessionKey) return undefined;
+    const s = this.sessions.get(sessionKey);
+    if (!s || (!s.cachedPrompt && s.cachedRecentMessages.length === 0)) return undefined;
+    return { prompt: s.cachedPrompt, recentMessages: s.cachedRecentMessages };
+  }
+
+  /**
+   * Get the cached formatted context string for a session (for legacy compatibility).
+   */
+  getRecentContext(sessionKey?: string): string | undefined {
+    if (!sessionKey) return undefined;
+    return this.sessions.get(sessionKey)?.cachedRecentContext;
+  }
+
+  constructor(
+    private client: OffloadApiClient,
+    private config: OffloadClientConfig,
+    private logger: Logger,
+  ) {}
+
+  get info() {
+    return {
+      id: "memory-tencentdb",
+      name: "Offload Client Context Engine",
+      version: "2.0.0",
+      ownsCompaction: true,
+    };
+  }
+
+  /**
+   * bootstrap — called when a new session starts (e.g. /new command).
+   * Resets per-session cached state.
+   */
+  async bootstrap(params: { sessionKey?: string; sessionId?: string }) {
+    const sk = params.sessionKey ?? params.sessionId;
+    if (sk) this.resetSession(sk);
+    return { bootstrapped: true };
+  }
+
+  /**
+   * ingest — no-op for client mode (ingest is handled by after_tool_call hook).
+   * Required by the framework ContextEngine interface.
+   */
+  async ingest(_params: {
+    sessionId: string;
+    sessionKey?: string;
+    message: any;
+    isHeartbeat?: boolean;
+  }) {
+    return { ingested: true };
+  }
+
+  /**
+   * compact — record the framework's authoritative token count for calibration,
+   * then defer actual compaction to assemble().
+   */
+  async compact(params: {
+    sessionId: string;
+    sessionKey?: string;
+    sessionFile: string;
+    tokenBudget?: number;
+    force?: boolean;
+    currentTokenCount?: number;
+    compactionTarget?: "budget" | "threshold";
+    customInstructions?: string;
+    runtimeContext?: any;
+  }) {
+    // Record framework token count for calibration
+    if (params.currentTokenCount && params.currentTokenCount > 0) {
+      const sk = params.sessionKey ?? params.sessionId;
+      const s = this.getSession(sk);
+      s.lastKnownTotalTokens = params.currentTokenCount;
+      this.logger.debug?.(
+        `[offload-client] compact: calibration updated, knownTokens=${s.lastKnownTotalTokens}`,
+      );
+    }
+
+    const contextWindow = params.tokenBudget ?? DEFAULT_CONTEXT_WINDOW;
+    const targetTokens = Math.floor(contextWindow * COMPACT_TARGET_RATIO);
+
+    this.logger.info(
+      `[offload-client] compact: sessionKey=${params.sessionKey ?? params.sessionId}, ` +
+      `budget=${contextWindow}, target=${targetTokens}, ` +
+      `currentTokens=${params.currentTokenCount}, force=${params.force}`,
+    );
+
+    return {
+      ok: true,
+      compacted: false,
+      reason: "offload-client: compaction deferred to assemble()",
+    };
+  }
+
+  /**
+   * Resolve the best known token total for calibration.
+   * Uses lastKnownTotalTokens only if message count hasn't changed significantly.
+   */
+  private resolveCalibrationTokens(sessionKey: string, msgCount: number): number | undefined {
+    const s = this.sessions.get(sessionKey);
+    if (!s || s.lastKnownTotalTokens <= 0) return undefined;
+    // If message count changed by >20%, the cached total is stale — skip calibration
+    if (s.lastKnownMsgCount > 0 && Math.abs(msgCount - s.lastKnownMsgCount) / s.lastKnownMsgCount > 0.2) {
+      return undefined;
+    }
+    return s.lastKnownTotalTokens;
+  }
+
+  /**
+   * Local brute-force compaction: keep tail messages up to target budget,
+   * respecting tool pairs and truncating oversized tool_results.
+   * Used as fallback when server compaction is unavailable.
+   */
+  private localCompact(messages: any[], contextWindow: number, sessionKey: string): any[] {
+    const targetTokens = Math.floor(contextWindow * COMPACT_TARGET_RATIO);
+    const knownTokens = this.resolveCalibrationTokens(sessionKey, messages.length);
+    const { perMessage } = estimateAllTokens(messages, knownTokens);
+    const n = messages.length;
+
+    // Step 1: scan from tail, find the cut index
+    let cumTokens = 0;
+    let cutIdx = n; // everything before cutIdx gets deleted
+    for (let i = n - 1; i >= 0; i--) {
+      cumTokens += perMessage[i];
+      if (cumTokens > targetTokens) {
+        cutIdx = i + 1;
+        break;
+      }
+      cutIdx = i;
+    }
+
+    // Never delete the very first user message (index 0)
+    if (cutIdx <= 0) cutIdx = 0;
+
+    // Step 2: expand cut boundary to respect tool pairs
+    // If cutIdx lands inside a tool pair, move it to include the full pair.
+
+    // 2a: If msg at cutIdx is a tool_result, its paired assistant+tool_use is
+    //     before cutIdx (would be deleted). Move cutIdx back to include the pair.
+    while (cutIdx < n && isToolResult(messages[cutIdx])) {
+      cutIdx++;
+    }
+
+    // 2b: If msg at cutIdx-1 (last deleted) is assistant+tool_use, its tool_result
+    //     at cutIdx would be orphaned. Pull cutIdx back to keep the pair.
+    while (cutIdx > 0 && cutIdx < n && isAssistantWithToolUse(messages[cutIdx - 1])) {
+      cutIdx--;
+    }
+
+    // Step 3: build retained array
+    const retained = messages.slice(cutIdx);
+
+    if (retained.length === 0) {
+      return [...messages]; // safety: don't delete everything
+    }
+
+    const deletedCount = cutIdx;
+    let retainedTokens = 0;
+    for (let i = cutIdx; i < n; i++) retainedTokens += perMessage[i];
+
+    this.logger.info(
+      `[offload-client] localCompact: deleted ${deletedCount}/${n} msgs, ` +
+      `retained ${retained.length} msgs, tokens=${retainedTokens}/${targetTokens} target`,
+    );
+
+    // Step 4: if still over target and there's a large tool_result, truncate it
+    if (retainedTokens > targetTokens) {
+      let maxTrIdx = -1;
+      let maxTrTokens = 0;
+      for (let i = 0; i < retained.length; i++) {
+        if (isToolResult(retained[i])) {
+          const t = estimateMessageTokens(retained[i]);
+          if (t > maxTrTokens) {
+            maxTrTokens = t;
+            maxTrIdx = i;
+          }
+        }
+      }
+      if (maxTrIdx >= 0 && maxTrTokens > TOOL_RESULT_TRUNCATE_CHARS / 4) {
+        retained[maxTrIdx] = truncateToolResult(retained[maxTrIdx], TOOL_RESULT_TRUNCATE_CHARS);
+        const newTokens = estimateMessageTokens(retained[maxTrIdx]);
+        retainedTokens = retainedTokens - maxTrTokens + newTokens;
+        this.logger.info(
+          `[offload-client] localCompact: truncated tool_result[${maxTrIdx}] ` +
+          `${maxTrTokens}→${newTokens} tokens, total now=${retainedTokens}`,
+        );
+      }
+    }
+
+    return retained;
+  }
+
+  // ─── L1.5 Trigger ───────────────────────────────────────────────────────────
+
+  /**
+   * Trigger L1.5 task judgment via ingest API when a new user prompt is detected.
+   * Fire-and-forget — does not block the assemble flow.
+   */
+  private triggerL15IfNeeded(prompt: string | undefined, messages: any[], sessionKey: string): void {
+    if (!prompt || typeof prompt !== "string" || prompt.length === 0) return;
+
+    // Skip system/internal prompts that are not user-initiated
+    if (this.isInternalPrompt(prompt)) {
+      this.logger.debug?.(`[offload-client] L1.5 skipped: internal prompt (${prompt.slice(0, 60)})`);
+      return;
+    }
+
+    // Always update cached context for after_tool_call hook (L1 needs it)
+    const recentMsgs = this.buildRecentMessages(prompt, messages);
+    const s = this.getSession(sessionKey);
+    s.cachedPrompt = prompt.slice(0, 500);
+    s.cachedRecentMessages = recentMsgs;
+    s.cachedRecentContext = this.formatContextForL1(prompt, recentMsgs);
+
+    // Dedup: skip L1.5 if same prompt as last trigger for this session
+    const hash = this.simpleHash(prompt);
+    if (s.lastL15PromptHash === hash) {
+      this.logger.debug?.(`[offload-client] L1.5 skipped: same prompt hash (${hash})`);
+      return;
+    }
+    s.lastL15PromptHash = hash;
+
+    this.logger.info(
+      `[offload-client] L1.5 triggered: promptHash=${hash}, recentMsgs=${recentMsgs.length}`,
+    );
+
+    // Fire-and-forget L1.5
+    this.client.ingestL15(sessionKey, prompt.slice(0, 500), recentMsgs).catch((err) => {
+      this.logger.warn(`[offload-client] L1.5 ingestL15 failed: ${err}`);
+    });
+  }
+
+  /**
+   * Detect internal/system prompts that should not trigger L1.5.
+   * These are framework-generated messages, not user-initiated conversations.
+   */
+  private isInternalPrompt(prompt: string): boolean {
+    // Compaction flush prompts
+    if (prompt.startsWith("Pre-compaction")) return true;
+    // Inter-session routing messages
+    if (prompt.startsWith("[Inter-session message]")) return true;
+    // Heartbeat/keepalive
+    if (prompt.includes("HEARTBEAT") || prompt.includes("heartbeat")) return true;
+    return false;
+  }
+
+  /**
+   * Build structured RecentMessage[] for ingest API.
+   * Filters: user/assistant text only, no tool calls, no heartbeats.
+   * Max 5 recent turns, 400 chars per message.
+   */
+  private buildRecentMessages(prompt: string, messages: any[]): RecentMessage[] {
+    const normalizedPrompt = prompt.trim().slice(0, 200).toLowerCase();
+
+    // Scan messages, collect user/assistant pairs
+    const pairs: RecentMessage[] = [];
+    for (const msg of messages) {
+      const role = getMsgRole(msg);
+
+      // Skip tool messages entirely
+      if (isToolResult(msg) || isAssistantWithToolUse(msg)) continue;
+      if (role === "tool" || role === "toolResult" || role === "tool_result") continue;
+
+      if (role === "user") {
+        const text = this.extractMsgText(msg);
+        if (!text || text.length <= 5) continue;
+        if (text.includes("HEARTBEAT") || text.includes("heartbeat")) continue;
+        const trimmed = text.slice(0, 400);
+        // Skip if it matches current prompt
+        const normalizedText = trimmed.slice(0, 200).toLowerCase();
+        if (normalizedPrompt && (normalizedText === normalizedPrompt || normalizedText.startsWith(normalizedPrompt) || normalizedPrompt.startsWith(normalizedText))) continue;
+        pairs.push({ role: "user", content: trimmed });
+      } else if (role === "assistant") {
+        const text = this.extractMsgText(msg);
+        if (!text || text.length <= 10) continue;
+        if (text.includes("HEARTBEAT") || text.includes("heartbeat")) continue;
+        pairs.push({ role: "assistant", content: text.slice(0, 400) });
+      }
+    }
+
+    // Keep last N messages (max 10 messages ≈ 5 turns)
+    const recent = pairs.slice(-10);
+    return recent;
+  }
+
+  /**
+   * Format context string for L1 executor (recent-context.txt).
+   */
+  private formatContextForL1(prompt: string, recentMsgs: RecentMessage[]): string {
+    const parts: string[] = [];
+    if (recentMsgs.length > 0) {
+      parts.push("历史消息，可作为参考：");
+      for (const m of recentMsgs) {
+        parts.push(`[${m.role === "user" ? "User" : "Assistant"}]: ${m.content}`);
+      }
+    }
+    parts.push(`\n最新user message：\n[User]: ${prompt.slice(0, 500)}`);
+    return parts.join("\n");
+  }
+
+  /**
+   * Extract text content from a message.
+   */
+  private extractMsgText(msg: any): string {
+    const content = msg?.content ?? msg?.message?.content ?? "";
+    if (typeof content === "string") return content;
+    if (Array.isArray(content)) {
+      return content
+        .map((b: any) => (typeof b === "string" ? b : b?.text ?? ""))
+        .join("");
+    }
+    return "";
+  }
+
+  /**
+   * Simple string hash for prompt deduplication.
+   */
+  private simpleHash(str: string): string {
+    let hash = 0;
+    for (let i = 0; i < str.length; i++) {
+      hash = ((hash << 5) - hash + str.charCodeAt(i)) | 0;
+    }
+    return hash.toString(36);
+  }
+
+  /**
+   * assemble — estimate ratio → call server compaction → fallback to localCompact.
+   * Framework calls this to build the model context for each turn.
+   */
+  async assemble(params: {
+    sessionId: string;
+    sessionKey?: string;
+    messages?: any[];
+    tokenBudget?: number;
+    prompt?: string;
+    availableTools?: Set<string>;
+    citationsMode?: string;
+    model?: string;
+  }) {
+    const { messages, sessionKey, sessionId } = params;
+    if (!messages || messages.length === 0) {
+      return { messages: messages ? [...messages] : [], estimatedTokens: 0 };
+    }
+
+    const sk = sessionKey ?? sessionId ?? "unknown";
+
+    // ── L1.5 trigger: fire-and-forget on new user prompt ──
+    this.triggerL15IfNeeded(params.prompt, messages, sk);
+
+    const contextWindow = params.tokenBudget ?? DEFAULT_CONTEXT_WINDOW;
+    // Don't use framework's knownTokens for calibration — our tiktoken is already precise.
+    // Framework's currentTokenCount may use a different calculation method (e.g. chars/4).
+    const { total, perMessage } = estimateAllTokens(messages);
+    const ratio = total / contextWindow;
+
+    // Update calibration state for future calls
+    const s = this.getSession(sk);
+    s.lastKnownTotalTokens = total;
+    s.lastKnownMsgCount = messages.length;
+
+    // Below client threshold — skip compaction
+    if (ratio < this.config.compactionRatio) {
+      this.logger.debug?.(
+        `[offload-client] assemble: ratio=${(ratio * 100).toFixed(1)}% < ${(this.config.compactionRatio * 100).toFixed(0)}%, skip`,
+      );
+      return { messages: [...messages], estimatedTokens: total };
+    }
+
+    this.logger.info(
+      `[offload-client] assemble: ratio=${(ratio * 100).toFixed(1)}%, msgs=${messages.length}, calling compaction...`,
+    );
+
+    // Try server-side compaction first
+    const result = await this.client.compaction({
+      sessionId: sessionKey ?? sessionId ?? "unknown",
+      messages,
+      ratio,
+      contextWindow,
+      totalTokens: total,
+      messageTokens: perMessage,
+    });
+
+    if (result) {
+      const compactedTokens = result.messages.reduce(
+        (sum: number, msg: any) => sum + estimateMessageTokens(msg),
+        0,
+      );
+      this.logger.info(
+        `[offload-client] server compaction done: level=${result.report.resolvedLevel}, ` +
+        `${result.report.originalCount}→${result.report.compactedCount} msgs, ` +
+        `mild=${result.report.mildReplacements}, agg=${result.report.aggressiveDeleted}, ` +
+        `em=${result.report.emergencyDeleted}, mmd=${result.report.mmdInjected}`,
+      );
+      return { messages: result.messages, estimatedTokens: compactedTokens };
+    }
+
+    // Fallback: local brute-force compaction
+    this.logger.warn("[offload-client] server compaction failed, falling back to local compact");
+    const compacted = this.localCompact(messages, contextWindow, sk);
+    const compactedTokens = compacted.reduce(
+      (sum: number, msg: any) => sum + estimateMessageTokens(msg),
+      0,
+    );
+    return { messages: compacted, estimatedTokens: compactedTokens };
+  }
+
+  /**
+   * afterTurn — no-op (ingest is handled by after_tool_call hook).
+   */
+  async afterTurn() {}
+}

src/offload-client/hooks/after-tool-call.ts

@@ -0,0 +1,53 @@
+/**
+ * offload-client — after_tool_call hook handler.
+ * Fire-and-forget: sends tool pair + context to ingest API for L1 processing.
+ */
+import type { OffloadApiClient } from "../offload-api-client.js";
+import type { OffloadClientConfig, ToolPairPayload, RecentMessage, Logger } from "../types.js";
+
+export interface AfterToolCallEvent {
+  toolName: string;
+  toolCallId: string;
+  params?: unknown;
+  result?: unknown;
+  error?: string;
+  durationMs?: number;
+}
+
+/**
+ * Create the after_tool_call hook handler.
+ * Sends each tool call result to the server for L1 processing.
+ *
+ * @param getContext Optional getter for { prompt, recentMessages } context per session.
+ */
+export function createAfterToolCallHandler(
+  client: OffloadApiClient,
+  config: OffloadClientConfig,
+  logger: Logger,
+  getContext?: (sessionKey: string) => { prompt?: string; recentMessages?: RecentMessage[] } | undefined,
+) {
+  return (event: AfterToolCallEvent, ctx: { sessionKey?: string; sessionId?: string }) => {
+    const sessionId = ctx.sessionKey ?? ctx.sessionId;
+    logger.debug?.(
+      `[offload-client] after_tool_call: tool=${event.toolName}, session=${sessionId ?? "(none)"}, callId=${event.toolCallId ?? "(none)"}`,
+    );
+    if (!sessionId) return;
+
+    const toolPair: ToolPairPayload = {
+      toolName: event.toolName,
+      toolCallId: event.toolCallId,
+      params: event.params ?? {},
+      result: event.result,
+      error: event.error,
+      timestamp: new Date().toISOString(),
+      durationMs: event.durationMs,
+    };
+
+    const context = getContext?.(sessionId);
+
+    // Fire-and-forget — do not block the LLM flow
+    client.ingestWithContext(sessionId, [toolPair], context?.prompt, context?.recentMessages).catch((err) => {
+      logger.warn(`[offload-client] ingest fire-and-forget error: ${err}`);
+    });
+  };
+}

src/offload-client/index.ts

@@ -0,0 +1,101 @@
+/**
+ * offload-client — Plugin registration entry point.
+ * Stateless, server-delegated offload client.
+ * 1 hook (after_tool_call) + 1 Context Engine (assemble → compaction API).
+ */
+import type { OffloadClientConfig, Logger } from "./types.js";
+import { defaultOffloadClientConfig } from "./types.js";
+import { OffloadApiClient } from "./offload-api-client.js";
+import { OffloadContextEngine } from "./context-engine.js";
+import { createAfterToolCallHandler } from "./hooks/after-tool-call.js";
+
+export interface OpenClawPluginApi {
+  on: (hookName: string, handler: (...args: any[]) => any) => void;
+  registerContextEngine: (id: string, factoryOrInstance: any) => any;
+  logger?: Logger;
+}
+
+/**
+ * Register the offload-client plugin.
+ * Call this from the main plugin's register() when offload-client config is enabled.
+ */
+export function registerOffloadClient(api: OpenClawPluginApi, userConfig: Partial<OffloadClientConfig>): void {
+  const config: OffloadClientConfig = { ...defaultOffloadClientConfig(), ...userConfig };
+  const logger: Logger = api.logger ?? { info: console.log, warn: console.warn, error: console.error, debug: console.debug };
+
+  if (!config.enabled) {
+    logger.info("[offload-client] disabled by config");
+    return;
+  }
+
+  if (!config.serverUrl || !config.apiKey || !config.serviceId) {
+    logger.error("[offload-client] missing required config: serverUrl, apiKey, or serviceId");
+    return;
+  }
+
+  const client = new OffloadApiClient(config, logger);
+
+  // Context Engine: occupies slot, assemble() calls compaction API
+  const engine = new OffloadContextEngine(client, config, logger);
+
+  // Hook: fire-and-forget ingest on every tool call (with context from engine per session)
+  const afterToolCallHandler = createAfterToolCallHandler(
+    client, config, logger,
+    (sessionKey) => engine.getContext(sessionKey),
+    config.agentName,  // ← NEW: pass agentName for sessionId construction
+  );
+  api.on("after_tool_call", afterToolCallHandler);
+
+  // ── Memory management hooks ──
+
+  // agent_end: clear token cache after each agent turn
+  api.on("agent_end", (_event: any, ctx: { sessionKey?: string; sessionId?: string }) => {
+    const sk = ctx.sessionKey ?? ctx.sessionId;
+    if (sk) {
+      engine.resetSession(sk);
+      logger.debug?.(`[offload-client] reset session state: ${sk}`);
+    }
+  });
+
+  // gateway_stop: emergency cleanup on shutdown
+  api.on("gateway_stop", async () => {
+    engine.clearAllSessions();
+    logger.info("[offload-client] all session states cleared on gateway_stop");
+  });
+
+  try {
+    const result = api.registerContextEngine("memory-tencentdb", () => engine) as any;
+    if (result?.ok === false) {
+      logger.error(
+        `[offload-client] Context Engine slot occupied by "${result.existingOwner ?? "unknown"}". ` +
+        `Compaction disabled — only ingest will work.`,
+      );
+    } else {
+      logger.info("[offload-client] Context Engine registered");
+    }
+  } catch (err) {
+    logger.error(`[offload-client] registerContextEngine failed: ${err}. Compaction disabled.`);
+  }
+
+  // ── Health check (async, non-blocking) ──
+  client.checkHealth().then((ok) => {
+    if (!ok) {
+      logger.warn(
+        `[offload-client] ⚠️  Server ${config.serverUrl} unreachable! ` +
+        `Ingest calls will fail silently until server becomes available.`
+      );
+    } else {
+      logger.info(`[offload-client] Server health OK: ${config.serverUrl}`);
+    }
+  }).catch((_err) => {
+    // ignore
+  });
+
+  logger.info(`[offload-client] registered (server=${config.serverUrl})`);
+}
+
+export { OffloadApiClient } from "./offload-api-client.js";
+export { OffloadContextEngine } from "./context-engine.js";
+export { createAfterToolCallHandler } from "./hooks/after-tool-call.js";
+export { estimateTokens, estimateMessageTokens, estimateAllTokens } from "./token-estimator.js";
+export type { OffloadClientConfig, ToolPairPayload, CompactionResult, CompactionReport, Logger } from "./types.js";

src/offload-client/offload-api-client.ts

@@ -0,0 +1,177 @@
+/**
+ * offload-client — HTTP client for Offload Server v2 API.
+ */
+import type { OffloadClientConfig, ToolPairPayload, RecentMessage, CompactionResult, Logger } from "./types.js";
+
+export class OffloadApiClient {
+  constructor(
+    private config: OffloadClientConfig,
+    private logger: Logger,
+  ) {}
+
+  /** Health check: GET /v2/offload/health. Returns true if server is reachable (any HTTP response = reachable). */
+  async checkHealth(): Promise<boolean> {
+    try {
+      const controller = new AbortController();
+      const timeoutId = setTimeout(() => controller.abort(), 5000);
+      const res = await fetch(`${this.config.serverUrl}/v2/offload/health`, {
+        method: "GET",
+        headers: { Authorization: `Bearer ${this.config.apiKey}` },
+        signal: controller.signal,
+      });
+      clearTimeout(timeoutId);
+      // Any HTTP response (including 401/403) means server is reachable
+      return res.status < 500;
+    } catch {
+      return false;
+    }
+  }
+
+  /**
+   * Fire-and-forget: send tool pairs to ingest endpoint.
+   * Does not throw — failures are logged as warnings.
+   */
+  async ingest(sessionId: string, toolPairs: ToolPairPayload[]): Promise<void> {
+    return this.ingestWithContext(sessionId, toolPairs, undefined, undefined);
+  }
+
+  /**
+   * Fire-and-forget: send tool pairs + optional context to ingest endpoint.
+   * When prompt/recentMessages are provided, server triggers L1 with context (skip L1.5).
+   */
+  async ingestWithContext(
+    sessionId: string,
+    toolPairs: ToolPairPayload[],
+    prompt?: string,
+    recentMessages?: RecentMessage[],
+  ): Promise<void> {
+    const url = `${this.config.serverUrl}/v2/offload/ingest`;
+    const payload: Record<string, unknown> = {
+      session_id: sessionId,
+      tool_pairs: toolPairs.map((tp) => ({
+        tool_name: tp.toolName,
+        tool_call_id: tp.toolCallId,
+        params: tp.params,
+        result: tp.result,
+        error: tp.error,
+        timestamp: tp.timestamp,
+        duration_ms: tp.durationMs,
+      })),
+    };
+    if (prompt) payload.prompt = prompt;
+    if (recentMessages && recentMessages.length > 0) payload.recent_messages = recentMessages;
+    const body = JSON.stringify(payload);
+
+    try {
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(), this.config.ingestTimeoutMs);
+
+      await fetch(url, {
+        method: "POST",
+        headers: this.buildHeaders(),
+        body,
+        signal: controller.signal,
+      });
+
+      clearTimeout(timer);
+    } catch (err) {
+      this.logger.warn(`[offload-client] ingest failed: ${err}`);
+    }
+  }
+
+  /**
+   * Fire-and-forget: trigger L1.5 task judgment via ingest endpoint.
+   * Sends prompt + recentMessages (empty toolPairs) to activate the L1.5 path on the server.
+   */
+  async ingestL15(sessionId: string, prompt: string, recentMessages?: RecentMessage[]): Promise<void> {
+    const url = `${this.config.serverUrl}/v2/offload/ingest`;
+    const payload: Record<string, unknown> = {
+      session_id: sessionId,
+      tool_pairs: [],
+      prompt,
+    };
+    if (recentMessages && recentMessages.length > 0) payload.recent_messages = recentMessages;
+    const body = JSON.stringify(payload);
+
+    try {
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(), this.config.ingestTimeoutMs);
+
+      const response = await fetch(url, {
+        method: "POST",
+        headers: this.buildHeaders(),
+        body,
+        signal: controller.signal,
+      });
+
+      clearTimeout(timer);
+
+      if (!response.ok) {
+        this.logger.warn(`[offload-client] ingestL15 returned ${response.status}`);
+      }
+    } catch (err) {
+      this.logger.warn(`[offload-client] ingestL15 failed: ${err}`);
+    }
+  }
+
+  /**
+   * Synchronous compaction call. Returns compressed messages + report.
+   * Returns null on timeout/failure (caller should keep original messages).
+   */
+  async compaction(req: {
+    sessionId: string;
+    messages: any[];
+    ratio: number;
+    contextWindow: number;
+    totalTokens: number;
+    messageTokens?: number[];
+  }): Promise<CompactionResult | null> {
+    const url = `${this.config.serverUrl}/v2/offload/compact`;
+    const body = JSON.stringify({
+      session_id: req.sessionId,
+      messages: req.messages,
+      ratio: req.ratio,
+      context_window: req.contextWindow,
+      total_tokens: req.totalTokens,
+      message_tokens: req.messageTokens,
+    });
+
+    try {
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(), this.config.compactionTimeoutMs);
+
+      const response = await fetch(url, {
+        method: "POST",
+        headers: this.buildHeaders(),
+        body,
+        signal: controller.signal,
+      });
+
+      clearTimeout(timer);
+
+      if (!response.ok) {
+        this.logger.warn(`[offload-client] compaction returned ${response.status}`);
+        return null;
+      }
+
+      const json = (await response.json()) as any;
+      if (json.code !== 0 || !json.data) {
+        this.logger.warn(`[offload-client] compaction error: ${json.message ?? "unknown"}`);
+        return null;
+      }
+
+      return { messages: json.data.messages, report: json.data.report };
+    } catch (err) {
+      this.logger.warn(`[offload-client] compaction failed: ${err}`);
+      return null;
+    }
+  }
+
+  private buildHeaders(): Record<string, string> {
+    return {
+      "Content-Type": "application/json",
+      Authorization: `Bearer ${this.config.apiKey}`,
+      "X-TDAI-Service-Id": this.config.serviceId,
+    };
+  }
+}

src/offload-client/token-estimator.ts

@@ -0,0 +1,227 @@
+/**
+ * offload-client — Token estimation using tiktoken (precise).
+ * Aligned with server-side preciseMessageTokens: only counts LLM-visible content.
+ *
+ * Strategy:
+ *   - Primary: tiktoken BPE encoding (o200k_base) on role + content
+ *   - Fallback: CJK-aware heuristic if tiktoken fails
+ *
+ * `estimateAllTokens` still supports optional calibration from framework-reported
+ * totalTokens, but with tiktoken the drift should be minimal.
+ */
+import { getEncoding, type Tiktoken } from "js-tiktoken";
+
+// ─── Tiktoken Encoder (lazy singleton) ──────────────────────────────────────
+
+let _encoder: Tiktoken | null = null;
+function getEncoder(): Tiktoken {
+  if (!_encoder) _encoder = getEncoding("o200k_base");
+  return _encoder;
+}
+
+// ─── Calibration constants ──────────────────────────────────────────────────
+
+/** If heuristic drifts > 15% from known total, apply linear scaling. */
+const CALIBRATION_THRESHOLD = 0.15;
+/** Clamp calibration factor to prevent extreme scaling from noisy estimates. */
+const CALIBRATION_FACTOR_MIN = 0.5;
+const CALIBRATION_FACTOR_MAX = 3.0;
+
+// ─── LLM-visible text extraction (must match server-side extractLlmVisibleText) ──
+
+/**
+ * Extract the LLM-visible portion of a message (role + content only).
+ * Matches server-side preciseMessageTokens logic exactly.
+ */
+function extractLlmVisibleText(msg: any): string {
+  const role: string = msg?.role ?? msg?.message?.role ?? "";
+  const rawContent = msg?.content ?? msg?.message?.content ?? "";
+
+  let contentStr: string;
+  if (typeof rawContent === "string") {
+    contentStr = rawContent;
+  } else if (Array.isArray(rawContent)) {
+    const parts: string[] = [];
+    for (const block of rawContent) {
+      if (typeof block === "string") {
+        parts.push(block);
+      } else if (block?.type === "text" && typeof block.text === "string") {
+        parts.push(block.text);
+      } else if (block?.type === "tool_use" || block?.type === "toolCall") {
+        parts.push(block.name ?? block.toolName ?? "");
+        if (block.arguments) {
+          parts.push(typeof block.arguments === "string" ? block.arguments : JSON.stringify(block.arguments));
+        }
+        if (block.input) {
+          parts.push(typeof block.input === "string" ? block.input : JSON.stringify(block.input));
+        }
+      } else if (block?.type === "tool_result") {
+        if (typeof block.content === "string") parts.push(block.content);
+        else if (block.content) parts.push(JSON.stringify(block.content));
+      } else {
+        parts.push(JSON.stringify(block));
+      }
+    }
+    contentStr = parts.join("\n");
+  } else {
+    contentStr = JSON.stringify(rawContent);
+  }
+
+  return `${role}\n${contentStr}`;
+}
+
+// ─── Per-message token cache (WeakMap — auto GC when msg object is released) ──
+
+const _tokenCache = new WeakMap<object, number>();
+
+/** Max messages to tiktoken precisely per call. Beyond this, old messages use heuristic. */
+const PRECISE_BUDGET = 200;
+
+// ─── Core estimation ────────────────────────────────────────────────────────
+
+/**
+ * Estimate tokens for a text string using tiktoken.
+ * Falls back to CJK-aware heuristic on error.
+ */
+export function estimateTokens(text: string): number {
+  if (!text) return 0;
+  try {
+    return getEncoder().encode(text).length;
+  } catch {
+    return heuristicTokens(text);
+  }
+}
+
+/**
+ * Estimate tokens for a single message using tiktoken (with WeakMap cache).
+ * Counts only LLM-visible content (role + content), aligned with server.
+ */
+export function estimateMessageTokens(msg: any): number {
+  if (msg == null) return 0;
+  // Cache hit
+  if (typeof msg === "object" && _tokenCache.has(msg)) {
+    return _tokenCache.get(msg)!;
+  }
+  const tokens = _computeMessageTokens(msg);
+  // Cache store (only for objects)
+  if (typeof msg === "object" && msg !== null) {
+    _tokenCache.set(msg, tokens);
+  }
+  return tokens;
+}
+
+function _computeMessageTokens(msg: any): number {
+  try {
+    const text = extractLlmVisibleText(msg);
+    return getEncoder().encode(text).length + 4; // +4 for message framing overhead
+  } catch {
+    return heuristicTokens(extractLlmVisibleText(msg));
+  }
+}
+
+/**
+ * Fast heuristic estimate for a message (no tiktoken, ~0.01ms per msg).
+ * Used for old messages when total count exceeds PRECISE_BUDGET.
+ */
+function heuristicMessageTokens(msg: any): number {
+  if (msg == null) return 0;
+  return heuristicTokens(extractLlmVisibleText(msg));
+}
+
+// ─── Calibrated batch estimation ────────────────────────────────────────────
+
+/**
+ * Estimate total tokens and per-message tokens for a message array.
+ *
+ * Performance strategy:
+ *   - If messages.length <= PRECISE_BUDGET (200): tiktoken all (with cache)
+ *   - If messages.length > PRECISE_BUDGET: heuristic for old, tiktoken for recent N
+ *     Then calibrate old estimates using recent tiktoken as reference.
+ *
+ * @param knownTotalTokens  Optional authoritative total from the framework.
+ */
+export function estimateAllTokens(
+  messages: any[],
+  knownTotalTokens?: number,
+): { total: number; perMessage: number[] } {
+  const n = messages.length;
+  if (n === 0) return { total: 0, perMessage: [] };
+
+  let raw: number[];
+
+  if (n <= PRECISE_BUDGET) {
+    // Small batch: tiktoken all (cache makes repeat calls fast)
+    raw = messages.map((msg) => estimateMessageTokens(msg));
+  } else {
+    // Large batch: tiktoken recent, heuristic old, calibrate
+    const preciseStart = n - PRECISE_BUDGET;
+    raw = new Array(n);
+
+    // Recent messages: precise (also populates cache for next call)
+    let preciseSum = 0;
+    let heuristicSumForRecent = 0;
+    for (let i = preciseStart; i < n; i++) {
+      raw[i] = estimateMessageTokens(messages[i]);
+      preciseSum += raw[i];
+      heuristicSumForRecent += heuristicMessageTokens(messages[i]);
+    }
+
+    // Calibration factor from recent messages
+    const calibFactor = heuristicSumForRecent > 0 ? preciseSum / heuristicSumForRecent : 1;
+
+    // Old messages: heuristic × calibration factor (or cache hit if available)
+    for (let i = 0; i < preciseStart; i++) {
+      if (typeof messages[i] === "object" && _tokenCache.has(messages[i])) {
+        raw[i] = _tokenCache.get(messages[i])!;
+      } else {
+        raw[i] = Math.max(1, Math.round(heuristicMessageTokens(messages[i]) * calibFactor));
+      }
+    }
+  }
+
+  let rawTotal = raw.reduce((s, v) => s + v, 0);
+
+  // External calibration (from framework totalTokens)
+  if (!knownTotalTokens || knownTotalTokens <= 0 || rawTotal <= 0) {
+    return { total: rawTotal, perMessage: raw };
+  }
+
+  const drift = Math.abs(knownTotalTokens - rawTotal) / knownTotalTokens;
+  if (drift <= CALIBRATION_THRESHOLD) {
+    return { total: rawTotal, perMessage: raw };
+  }
+
+  // Apply linear calibration factor, clamped
+  const factor = Math.max(
+    CALIBRATION_FACTOR_MIN,
+    Math.min(CALIBRATION_FACTOR_MAX, knownTotalTokens / rawTotal),
+  );
+  const calibrated = raw.map((v) => Math.max(1, Math.round(v * factor)));
+  const calibratedTotal = calibrated.reduce((s, v) => s + v, 0);
+
+  return { total: calibratedTotal, perMessage: calibrated };
+}
+
+// ─── Heuristic fallback ─────────────────────────────────────────────────────
+
+function countCjkChars(text: string): number {
+  let n = 0;
+  for (const ch of text) {
+    const c = ch.codePointAt(0)!;
+    if (
+      (c >= 0x4e00 && c <= 0x9fff) ||
+      (c >= 0x3400 && c <= 0x4dbf) ||
+      (c >= 0xf900 && c <= 0xfaff)
+    ) {
+      n++;
+    }
+  }
+  return n;
+}
+
+function heuristicTokens(text: string): number {
+  if (!text) return 0;
+  const cjk = countCjkChars(text);
+  const rest = Math.max(0, text.length - cjk);
+  return Math.max(1, Math.ceil(cjk / 1.7 + rest / 4));
+}

src/offload-client/types.ts

@@ -0,0 +1,80 @@
+/**
+ * offload-client — Type definitions.
+ */
+
+// ─── Plugin Configuration ────────────────────────────────────────────────────
+
+export interface OffloadClientConfig {
+  enabled: boolean;
+  /** Offload server base URL (e.g. "http://localhost:9100"). */
+  serverUrl: string;
+  /** Bearer token for Authorization header. */
+  apiKey: string;
+  /** X-TDAI-Service-Id header value. */
+  serviceId: string;
+  /** Agent name for sessionId construction. Default "default". */
+  agentName?: string;
+  /** Client-side threshold: skip compaction request when ratio < this value. Default 0.5. */
+  compactionRatio: number;
+  /** Ingest request timeout in ms. Default 5000. */
+  ingestTimeoutMs: number;
+  /** Compaction request timeout in ms. Default 30000. */
+  compactionTimeoutMs: number;
+}
+
+export function defaultOffloadClientConfig(): OffloadClientConfig {
+  return {
+    enabled: false,
+    serverUrl: "http://localhost:9100",
+    apiKey: "",
+    serviceId: "",
+    compactionRatio: 0.5,
+    ingestTimeoutMs: 5000,
+    compactionTimeoutMs: 30000,
+  };
+}
+
+// ─── Ingest Payload ──────────────────────────────────────────────────────────
+
+export interface ToolPairPayload {
+  toolName: string;
+  toolCallId: string;
+  params: unknown;
+  result: unknown;
+  error?: string;
+  timestamp: string;
+  durationMs?: number;
+}
+
+export interface RecentMessage {
+  role: "user" | "assistant";
+  content: string;
+}
+
+// ─── Compaction Response ─────────────────────────────────────────────────────
+
+export interface CompactionReport {
+  resolvedLevel: string;
+  originalCount: number;
+  compactedCount: number;
+  fastPathReplaced: number;
+  fastPathDeleted: number;
+  mildReplacements: number;
+  aggressiveDeleted: number;
+  emergencyDeleted: number;
+  mmdInjected: number;
+}
+
+export interface CompactionResult {
+  messages: any[];
+  report: CompactionReport;
+}
+
+// ─── Logger ──────────────────────────────────────────────────────────────────
+
+export interface Logger {
+  info: (...args: unknown[]) => void;
+  warn: (...args: unknown[]) => void;
+  error: (...args: unknown[]) => void;
+  debug?: (...args: unknown[]) => void;
+}

src/offload/local-llm/llm-caller.ts

@@ -27,6 +27,8 @@ export interface CallLlmOpts {
   timeoutMs?: number;
   /** Label for logging (e.g. "L1", "L1.5", "L2") */
   label?: string;
+  /** Instance ID for telemetry metadata */
+  instanceId?: string;
 }
 
 /**
@@ -64,6 +66,7 @@ export async function callLlm(
       experimental_telemetry: {
         isEnabled: true,
         functionId: opts.label ?? "offload-llm",
+        metadata: { instanceId: opts.instanceId ?? "unknown" },
       },
     });
 

src/offload/types.ts

@@ -3,6 +3,8 @@
  * Ported from context-offload-plugin with updated runtime defaults.
  */
 
+import type { Logger } from "../core/types.js";
+
 // ============================
 // Data types
 // ============================
@@ -212,12 +214,7 @@ export interface PluginConfig {
 // ============================
 
 /** Logger interface used by offload plugin components */
-export interface PluginLogger {
-  info: (msg: string) => void;
-  warn: (msg: string) => void;
-  error: (msg: string) => void;
-  debug?: (msg: string) => void;
-}
+export type PluginLogger = Logger;
 
 // ============================
 // Plugin defaults

src/offload_server/compact/compaction-handler.ts

@@ -0,0 +1,328 @@
+/**
+ * L3 Compaction Handler — orchestrates fast-path, MMD injection, and compression.
+ * Synchronous API: client awaits the response.
+ *
+ * Uses compact-state.json (independent from state.json) to avoid lock contention
+ * with L1/L1.5/L2 executors that write state.json concurrently.
+ */
+import type http from "node:http";
+import type { StorageAdapter } from "../../core/storage/adapter.js";
+import type { OffloadEntry, OffloadState, OffloadExecutorConfig, CompactState } from "../types.js";
+import { defaultOffloadState, defaultCompactState } from "../types.js";
+import { parseJsonl } from "../parsers/json-utils.js";
+import { CompactionRequestSchemaV2 } from "../schemas.js";
+import { buildOffloadBasePath } from "../session-utils.js";
+import { applyFastPath } from "./fast-path.js";
+import { injectActiveMmd, injectHistoryMmds } from "./mmd-injector.js";
+import { resolveLevel, mildCompress, aggressiveCompress, emergencyCompress } from "./compressor.js";
+import { estimateMessageTokens, extractToolResultId } from "./helpers.js";
+import type { Message } from "./helpers.js";
+import { traceServerCompaction } from "../opik-tracer.js";
+
+export interface CompactionDeps {
+  storage: StorageAdapter;
+  config: OffloadExecutorConfig;
+  logger: { info: (...args: unknown[]) => void; warn: (...args: unknown[]) => void; error: (...args: unknown[]) => void };
+}
+
+export interface CompactionReport {
+  resolvedLevel: string;
+  originalCount: number;
+  compactedCount: number;
+  fastPathReplaced: number;
+  fastPathDeleted: number;
+  mildReplacements: number;
+  aggressiveDeleted: number;
+  emergencyDeleted: number;
+  mmdInjected: number;
+}
+
+/**
+ * Handle POST /v2/offload/compact.
+ */
+export async function handleCompaction(
+  req: http.IncomingMessage,
+  res: http.ServerResponse,
+  auth: { serviceId: string },
+  deps: CompactionDeps,
+  requestId: string,
+  parseJsonBody: <T>(req: http.IncomingMessage) => Promise<T>,
+  sendJson: (res: http.ServerResponse, status: number, body: unknown) => void,
+  successEnvelope: <T>(data: T, requestId: string) => unknown,
+  errorEnvelope: (code: number, message: string, requestId: string) => unknown,
+): Promise<void> {
+  const body = await parseJsonBody(req);
+  const parsed = CompactionRequestSchemaV2.safeParse(body);
+  if (!parsed.success) {
+    sendJson(res, 400, errorEnvelope(400, parsed.error.message, requestId));
+    return;
+  }
+
+  const { session_id: sessionId, messages, ratio, context_window: contextWindow, message_tokens: messageTokens } = parsed.data;
+  let { total_tokens: totalTokens } = parsed.data;
+  const { storage, config } = deps;
+  const basePath = buildOffloadBasePath(sessionId);
+  const originalCount = messages.length;
+  const compactionStartMs = Date.now();
+
+  const { preciseMessageTokens } = await import("./compressor.js");
+  const originalTotalTokens = totalTokens;
+
+  // Compute fixed overhead: system prompt + tool schemas + message framing.
+  // These are included in clientTotalTokens but NOT in messages.
+  // overhead = clientTotal - sum(preciseMessageTokens for each message)
+  const messagesTokenSum = (messages as any[]).reduce(
+    (s: number, msg: any) => s + preciseMessageTokens(msg), 0,
+  );
+  const fixedOverhead = Math.max(0, totalTokens - messagesTokenSum);
+
+  deps.logger.info(
+    `[offload-server] compaction: clientTotal=${totalTokens}, msgsTokens=${messagesTokenSum}, ` +
+    `overhead=${fixedOverhead}, msgs=${originalCount}, ratio=${ratio.toFixed(2)}`,
+  );
+
+  // Read offload state (L1/L1.5/L2 managed, read-only here) and compact state (L3 owned)
+  const state = await readOffloadState(storage, basePath);
+  const compactState = await readCompactState(storage, basePath);
+  const entriesRaw = await storage.readFile(`${basePath}/entries.jsonl`);
+  const entries = entriesRaw
+    ? parseJsonl<OffloadEntry>(entriesRaw, (line, err) => {
+        deps.logger.warn(`[offload-server] compaction: bad JSONL line: ${line}`, err);
+      })
+    : [];
+
+  // Merge node-mapping.jsonl into entries (L2 writes node_id to a separate file)
+  const nodeMappingRaw = await storage.readFile(`${basePath}/node-mapping.jsonl`);
+  if (nodeMappingRaw) {
+    const mappings = parseJsonl<{ tool_call_id: string; node_id: string }>(nodeMappingRaw);
+    const nodeMap = new Map(mappings.map((m) => [m.tool_call_id, m.node_id]));
+    for (const entry of entries) {
+      if (!entry.node_id && nodeMap.has(entry.tool_call_id)) {
+        entry.node_id = nodeMap.get(entry.tool_call_id)!;
+      }
+    }
+  }
+
+  // Token array: use tiktoken values directly (no calibration against clientTotal).
+  // clientTotal includes fixedOverhead which is not in messages, so calibration
+  // would inflate per-message tokens incorrectly.
+  const tokenArray = buildTokenArray(messages as Message[], messagesTokenSum, messageTokens);
+
+  const report: CompactionReport = {
+    resolvedLevel: "fastpath",
+    originalCount,
+    compactedCount: 0,
+    fastPathReplaced: 0,
+    fastPathDeleted: 0,
+    mildReplacements: 0,
+    aggressiveDeleted: 0,
+    emergencyDeleted: 0,
+    mmdInjected: 0,
+  };
+
+  // Step 1: Fast-path re-apply (uses compactState for confirmed/deleted IDs)
+  const fp = applyFastPath(messages, entries, compactState);
+  report.fastPathReplaced = fp.replacedCount;
+  report.fastPathDeleted = fp.deletedCount;
+
+  // Recalculate totalTokens after fast-path:
+  // totalTokens = tiktoken(remaining messages) + fixedOverhead
+  if (fp.deletedCount > 0) {
+    const postFpMsgsTokens = (messages as any[]).reduce(
+      (s: number, msg: any) => s + preciseMessageTokens(msg), 0,
+    );
+    totalTokens = postFpMsgsTokens + fixedOverhead;
+    // Rebuild tokenArray for remaining messages
+    tokenArray.length = 0;
+    tokenArray.push(...buildTokenArray(messages as Message[], postFpMsgsTokens, undefined));
+  }
+  const effectiveRatio = contextWindow > 0 ? totalTokens / contextWindow : ratio;
+
+  // Step 2: Resolve compression level using post-fast-path ratio
+  const level = resolveLevel(effectiveRatio, {
+    mildRatio: config.mildOffloadRatio,
+    aggressiveRatio: config.aggressiveCompressRatio,
+    emergencyRatio: config.emergencyCompressRatio,
+  });
+  report.resolvedLevel = level;
+
+  deps.logger.info(
+    `[offload-server] compaction: level=${level}, ratio=${effectiveRatio.toFixed(2)} (pre-fp=${ratio.toFixed(2)}), msgs=${messages.length}, entries=${entries.length}`,
+  );
+
+  // Step 3: Mild compression
+  if (level === "mild" || level === "aggressive" || level === "emergency") {
+    const mild = mildCompress(messages, entries);
+    report.mildReplacements = mild.replacedCount;
+    compactState.confirmedOffloadIds.push(...mild.confirmedIds);
+
+    // Sync tokenArray after mild replacements (content changed)
+    // Incremental update: only recalculate tokens for messages that were actually replaced
+    if (mild.replacedCount > 0) {
+      const confirmedSet = new Set(mild.confirmedIds);
+      let tokenDelta = 0;
+      for (let i = 0; i < messages.length; i++) {
+        const msg = messages[i] as any;
+        // Use extractToolResultId which handles all formats (OpenAI, Anthropic content blocks)
+        const tid = extractToolResultId(msg);
+        if (!tid || !confirmedSet.has(tid)) continue;
+        const newTokens = preciseMessageTokens(msg);
+        tokenDelta += newTokens - tokenArray[i];
+        tokenArray[i] = newTokens;
+      }
+      totalTokens += tokenDelta;
+    }
+  }
+
+  // Step 4: Aggressive compression
+  // Target: just below the aggressive trigger threshold (leave ~5% headroom)
+  let aggRemainingTokens = totalTokens; // track for emergency
+  if (level === "aggressive" || level === "emergency") {
+    const aggTargetTokens = Math.floor(contextWindow * (config.aggressiveCompressRatio - 0.05));
+    // Skip aggressive if mild already brought tokens below target
+    if (totalTokens <= aggTargetTokens) {
+      deps.logger.info(
+        `[offload-server] aggressive skipped: mild already reduced tokens to ${totalTokens} (target=${aggTargetTokens})`,
+      );
+    } else {
+      const agg = aggressiveCompress(messages, aggTargetTokens, tokenArray, totalTokens);
+      report.aggressiveDeleted = agg.deletedCount;
+      aggRemainingTokens = agg.remainingTokens;
+      compactState.deletedOffloadIds.push(...agg.deletedIds);
+
+      // Inject history MMDs for deleted entries
+      if (agg.deletedIds.length > 0) {
+        const mmdBudget = Math.floor(contextWindow * 0.1 / 4); // 10% of context, in estimated tokens
+        const hist = await injectHistoryMmds(
+          messages, agg.deletedIds, entries, state, storage, basePath, mmdBudget,
+        );
+        report.mmdInjected += hist.injectedCount;
+        // Sync tokenArray for any injected MMD messages (calibrate against messages-only tokens)
+        if (hist.injectedCount > 0) {
+          const postAggMsgsTokens = (messages as any[]).reduce(
+            (s: number, msg: any) => s + preciseMessageTokens(msg), 0,
+          );
+          tokenArray.length = 0;
+          tokenArray.push(...buildTokenArray(messages as Message[], postAggMsgsTokens, undefined));
+        }
+      }
+    }
+  }
+
+  // Step 6: Emergency compression
+  // Target: just below the aggressive threshold (so next turn won't immediately re-trigger)
+  if (level === "emergency") {
+    const emTargetTokens = Math.floor(contextWindow * (config.aggressiveCompressRatio - 0.10));
+    const em = emergencyCompress(messages, emTargetTokens, tokenArray, aggRemainingTokens);
+    report.emergencyDeleted = em.deletedCount;
+    aggRemainingTokens = em.remainingTokens;
+    compactState.deletedOffloadIds.push(...em.deletedIds);
+  }
+
+  // Step 7: Inject active MMD (after all compression, so position is correct)
+  const mmdResult = await injectActiveMmd(messages, state, storage, basePath);
+  report.mmdInjected += mmdResult.injectedCount;
+
+  // Step 8: Write compact-state.json (independent file, no lock needed)
+  report.compactedCount = messages.length;
+  compactState.lastCompactedAt = new Date().toISOString();
+  await writeCompactState(storage, basePath, compactState);
+
+  // Compute remaining tokens (tracked from aggressive/emergency, no full re-scan)
+  const remainingTokens = level === "fastpath" || level === "mild"
+    ? totalTokens  // no deletion happened
+    : aggRemainingTokens;  // tracked through aggressive → emergency chain
+  const remainingRatio = contextWindow > 0 ? (remainingTokens / contextWindow).toFixed(2) : "N/A";
+
+  // Opik trace: compaction decision
+  traceServerCompaction({
+    sessionId,
+    level,
+    ratio,
+    contextWindow,
+    totalTokensBefore: originalTotalTokens,
+    totalTokensAfter: remainingTokens,
+    originalMsgCount: originalCount,
+    compactedMsgCount: messages.length,
+    report: report as unknown as Record<string, unknown>,
+    messages: messages as unknown[],
+    durationMs: Date.now() - compactionStartMs,
+    logger: deps.logger,
+  });
+
+  // Step 8: Return
+  sendJson(res, 200, successEnvelope({ messages, report }, requestId));
+
+  deps.logger.info(
+    `[offload-server] compaction done: ${originalCount}→${messages.length} msgs, level=${level}, ` +
+    `tokens=${originalTotalTokens}→${remainingTokens} (${remainingRatio}), ` +
+    `fp=${fp.replacedCount}r/${fp.deletedCount}d, mild=${report.mildReplacements}, ` +
+    `agg=${report.aggressiveDeleted}, em=${report.emergencyDeleted}, mmd=${report.mmdInjected}`,
+  );
+}
+
+// ─── Token Array Builder ─────────────────────────────────────────────────────
+
+/** Calibration threshold: if heuristic estimate drifts >15% from API totalTokens, apply linear scaling. */
+const CALIBRATION_THRESHOLD = 0.15;
+/** Clamp calibration factor to prevent extreme scaling from noisy estimates. */
+const CALIBRATION_FACTOR_MIN = 0.5;
+const CALIBRATION_FACTOR_MAX = 3.0;
+
+/**
+ * Build a pre-computed token array for all messages, with optional linear calibration.
+ * - If messageTokens[i] is available, use it directly (precise value).
+ * - Otherwise, use estimateMessageTokens (CJK-aware heuristic).
+ * - If totalTokens is provided and drift > 15%, apply a calibration factor to estimated items.
+ */
+export function buildTokenArray(
+  messages: Message[],
+  totalTokens: number,
+  messageTokens?: number[],
+): number[] {
+  const raw = messages.map((msg, i) =>
+    messageTokens && i < messageTokens.length
+      ? messageTokens[i]
+      : estimateMessageTokens(msg),
+  );
+
+  const rawTotal = raw.reduce((s, v) => s + v, 0);
+  if (rawTotal <= 0 || totalTokens <= 0) return raw;
+
+  const drift = Math.abs(totalTokens - rawTotal) / totalTokens;
+  if (drift <= CALIBRATION_THRESHOLD) return raw;
+
+  // Linear calibration: only scale estimated items, keep precise items unchanged
+  const factor = Math.max(CALIBRATION_FACTOR_MIN, Math.min(CALIBRATION_FACTOR_MAX, totalTokens / rawTotal));
+  return raw.map((v, i) =>
+    messageTokens && i < messageTokens.length
+      ? v
+      : Math.max(1, Math.round(v * factor)),
+  );
+}
+
+// ─── State Helpers ───────────────────────────────────────────────────────────
+
+async function readOffloadState(storage: StorageAdapter, basePath: string): Promise<OffloadState> {
+  const raw = await storage.readFile(`${basePath}/state.json`);
+  if (!raw) return defaultOffloadState();
+  try {
+    return { ...defaultOffloadState(), ...JSON.parse(raw) };
+  } catch {
+    return defaultOffloadState();
+  }
+}
+
+async function readCompactState(storage: StorageAdapter, basePath: string): Promise<CompactState> {
+  const raw = await storage.readFile(`${basePath}/compact-state.json`);
+  if (!raw) return defaultCompactState();
+  try {
+    return { ...defaultCompactState(), ...JSON.parse(raw) };
+  } catch {
+    return defaultCompactState();
+  }
+}
+
+async function writeCompactState(storage: StorageAdapter, basePath: string, state: CompactState): Promise<void> {
+  await storage.writeFile(`${basePath}/compact-state.json`, JSON.stringify(state));
+}

src/offload_server/compact/fast-path.ts

@@ -0,0 +1,189 @@
+/**
+ * L3 Fast-path — re-apply confirmed/deleted offload state to messages.
+ * Runs on every compaction call as the first step.
+ */
+import type { OffloadEntry, CompactState } from "../types.js";
+import type { Message } from "./helpers.js";
+import {
+  extractToolResultId,
+  isToolResultMessage,
+  isOnlyToolUseAssistant,
+  isAssistantWithToolUse,
+  extractAllToolUseIds,
+  replaceWithSummary,
+  buildOffloadMap,
+} from "./helpers.js";
+
+export interface FastPathResult {
+  replacedCount: number;
+  deletedCount: number;
+}
+
+/**
+ * Apply fast-path: replace confirmed tool results with summary,
+ * delete messages whose tool_call_id is in deletedOffloadIds,
+ * strip orphaned toolCall blocks from mixed assistant messages.
+ * Mutates messages in place.
+ */
+export function applyFastPath(
+  messages: Message[],
+  entries: OffloadEntry[],
+  compactState: CompactState,
+): FastPathResult {
+  const offloadMap = buildOffloadMap(entries);
+  const confirmedSet = new Set(compactState.confirmedOffloadIds);
+  const deletedSet = new Set(compactState.deletedOffloadIds);
+
+  if (confirmedSet.size === 0 && deletedSet.size === 0) {
+    return { replacedCount: 0, deletedCount: 0 };
+  }
+
+  const indicesToDelete: number[] = [];
+  let replacedCount = 0;
+
+  for (let i = 0; i < messages.length; i++) {
+    const msg = messages[i];
+    const tid = extractToolResultId(msg);
+
+    // 1. Deleted tool_result — handle Anthropic multi-block user messages carefully
+    if (tid && deletedSet.has(tid)) {
+      // Anthropic format: user message may have multiple tool_result blocks
+      if (msg.role === "user" && Array.isArray(msg.content)) {
+        const allToolResultIds = msg.content
+          .filter((b: any) => b?.type === "tool_result" && b.tool_use_id)
+          .map((b: any) => b.tool_use_id as string);
+
+        if (allToolResultIds.length > 1) {
+          // Multi-block: only remove the blocks that are in deletedSet, keep the rest
+          const allDeleted = allToolResultIds.every((id: string) => deletedSet.has(id));
+          if (allDeleted) {
+            indicesToDelete.push(i);
+          } else {
+            // Partial: strip only deleted tool_result blocks
+            for (let j = msg.content.length - 1; j >= 0; j--) {
+              const block = msg.content[j];
+              if (block?.type === "tool_result" && block.tool_use_id && deletedSet.has(block.tool_use_id)) {
+                msg.content.splice(j, 1);
+              }
+            }
+          }
+          continue;
+        }
+      }
+      // Single tool_result or non-Anthropic format: delete entire message
+      indicesToDelete.push(i);
+      continue;
+    }
+
+    // 2. Confirmed tool_result → replace with summary
+    if (tid && confirmedSet.has(tid) && isToolResultMessage(msg) && !msg._offloaded) {
+      const entry = offloadMap.get(tid);
+      if (entry) {
+        replaceWithSummary(msg, entry);
+        replacedCount++;
+      }
+    }
+
+    // 3. Pure toolCall assistant — all IDs deleted → mark for deletion
+    if (!tid && isOnlyToolUseAssistant(msg)) {
+      const tuIds = extractAllToolUseIds(msg);
+      if (tuIds.length > 0 && tuIds.every((id) => deletedSet.has(id))) {
+        indicesToDelete.push(i);
+      }
+    }
+
+    // 4. Mixed assistant (text + toolCall) — strip deleted toolCall blocks
+    //    to prevent orphaned tool_use without matching tool_result (provider 400)
+    if (!tid && isAssistantWithToolUse(msg) && !isOnlyToolUseAssistant(msg)) {
+      const content = msg.type === "message" ? msg.message?.content : msg.content;
+      if (Array.isArray(content)) {
+        for (let j = content.length - 1; j >= 0; j--) {
+          const block = content[j];
+          if ((block?.type === "tool_use" || block?.type === "toolCall") && block.id) {
+            if (deletedSet.has(block.id)) {
+              content.splice(j, 1);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // Delete in reverse order to preserve indices
+  const uniqueIndices = [...new Set(indicesToDelete)].sort((a, b) => b - a);
+  for (const idx of uniqueIndices) {
+    messages.splice(idx, 1);
+  }
+
+  // Post-pass: ensure no orphaned tool_use or tool_result after deletion
+  // Collect all remaining tool_use IDs and tool_result IDs
+  const remainingToolUseIds = new Set<string>();
+  const remainingToolResultIds = new Set<string>();
+  for (const msg of messages) {
+    for (const id of extractAllToolUseIds(msg)) {
+      remainingToolUseIds.add(id);
+    }
+    if (msg.role === "user" && Array.isArray(msg.content)) {
+      for (const block of msg.content) {
+        if (block?.type === "tool_result" && block.tool_use_id) {
+          remainingToolResultIds.add(block.tool_use_id);
+        }
+      }
+    }
+    const singleTid = extractToolResultId(msg);
+    if (singleTid) remainingToolResultIds.add(singleTid);
+  }
+
+  // Remove orphaned tool_use blocks from assistant messages (no matching tool_result)
+  for (const msg of messages) {
+    if (!isAssistantWithToolUse(msg)) continue;
+    const content = msg.type === "message" ? msg.message?.content : msg.content;
+    if (!Array.isArray(content)) continue;
+    for (let j = content.length - 1; j >= 0; j--) {
+      const block = content[j];
+      if ((block?.type === "tool_use" || block?.type === "toolCall") && block.id) {
+        if (!remainingToolResultIds.has(block.id)) {
+          content.splice(j, 1);
+        }
+      }
+    }
+  }
+
+  // Remove orphaned tool_result messages/blocks (no matching tool_use)
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (msg.role === "user" && Array.isArray(msg.content)) {
+      for (let j = msg.content.length - 1; j >= 0; j--) {
+        const block = msg.content[j];
+        if (block?.type === "tool_result" && block.tool_use_id && !remainingToolUseIds.has(block.tool_use_id)) {
+          msg.content.splice(j, 1);
+        }
+      }
+      // If user message is now empty, remove it
+      if (msg.content.length === 0) {
+        messages.splice(i, 1);
+        uniqueIndices.push(i); // count it
+      }
+    } else {
+      const singleTid = extractToolResultId(msg);
+      if (singleTid && !remainingToolUseIds.has(singleTid) && isToolResultMessage(msg)) {
+        messages.splice(i, 1);
+        uniqueIndices.push(i);
+      }
+    }
+  }
+
+  // Remove empty assistant messages (all tool_use blocks were stripped)
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i];
+    const role = msg.role ?? msg.message?.role ?? msg.type;
+    if (role !== "assistant") continue;
+    const content = msg.type === "message" ? msg.message?.content : msg.content;
+    if (Array.isArray(content) && content.length === 0) {
+      messages.splice(i, 1);
+      uniqueIndices.push(i);
+    }
+  }
+
+  return { replacedCount, deletedCount: uniqueIndices.length };
+}

src/offload_server/compact/helpers.ts

@@ -0,0 +1,297 @@
+/**
+ * L3 Compaction — Message helper utilities.
+ * Handles multiple message formats (OpenAI, Anthropic, OpenClaw).
+ *
+ * ═══════════════════════════════════════════════════════════════════════════
+ * OpenClaw 消息格式说明（来源：@mariozechner/pi-ai 类型定义）
+ * ═══════════════════════════════════════════════════════════════════════════
+ *
+ * OpenClaw 内部使用统一的 AgentMessage 联合类型，包含以下几种消息：
+ *
+ * ── 1. UserMessage ──
+ * {
+ *   role: "user",
+ *   content: string | ContentBlock[],  // string 或 [{type:"text",text:"..."}] 或含图片块
+ *   timestamp: number,
+ * }
+ *
+ * ── 2. AssistantMessage（纯文本回复）──
+ * {
+ *   role: "assistant",
+ *   content: [{ type: "text", text: "..." }],
+ *   model: "gpt-5.2",
+ *   stopReason: "stop",
+ *   timestamp: number,
+ *   api: "messages" | "chat" | "responses",
+ *   provider: "anthropic" | "openai" | "google",
+ *   usage: { input, output, totalTokens, ... },
+ * }
+ *
+ * ── 3. AssistantMessage（含 tool_use / toolCall）──
+ * {
+ *   role: "assistant",
+ *   content: [
+ *     { type: "text", text: "I'll read the file..." },     // 可选文本块
+ *     { type: "toolCall", id: "call_abc123", name: "read_file", arguments: { path: "..." } },
+ *     { type: "toolCall", id: "call_def456", name: "exec", arguments: { cmd: "..." } },
+ *   ],
+ *   stopReason: "toolUse",
+ *   ...同上
+ * }
+ * 注: Anthropic 原生格式使用 { type: "tool_use", id, name, input }
+ *     OpenClaw 内部统一为 { type: "toolCall", id, name, arguments }
+ *     但消息到达后端时**两种格式都可能出现**（取决于客户端是否已转换），
+ *     因此本模块所有判断都同时匹配 "tool_use" 和 "toolCall"。
+ *
+ * ── 4. ToolResultMessage ──
+ * {
+ *   role: "toolResult",
+ *   toolCallId: "call_abc123",           // 对应 AssistantMessage 中 toolCall 的 id
+ *   toolName: "read_file",
+ *   content: [{ type: "text", text: "文件内容..." }],
+ *   isError: false,
+ *   timestamp: number,
+ *   details?: any,                       // 可选的详细信息（不发给 LLM）
+ * }
+ *
+ * ── 5. 消息配对规则 ──
+ * - 每个 AssistantMessage 中的 toolCall 必须有对应的 ToolResultMessage
+ * - toolCallId 是配对的唯一标识
+ * - 删除 tool_result 时必须同时删除对应的 assistant toolCall（否则 provider 返回 400）
+ * - 一个 assistant 消息可包含多个 toolCall（并行工具调用）
+ *
+ * ── 6. 转换到 LLM Provider 原生格式 ──
+ * OpenAI:     toolResult → { role: "tool", tool_call_id: "...", content: "..." }
+ * Anthropic:  toolResult → { role: "user", content: [{ type: "tool_result", tool_use_id: "...", content: "..." }] }
+ * 注: 本模块处理的是 **转换后** 的消息（发送给 provider 前的格式），
+ *     因此需要兼容所有上述格式。
+ *
+ * ── 7. 特殊标记字段（本模块使用）──
+ * - _offloaded: boolean          — 已被 summary 替换的 tool_result
+ * - _mmdContextMessage: string   — MMD 注入消息标记（"active" | "history"）
+ * - _mmdInjection: boolean       — 历史 MMD 注入消息标记
+ * - _mmdVersion: string          — MMD 内容哈希，用于版本去重
+ * ═══════════════════════════════════════════════════════════════════════════
+ */
+import type { OffloadEntry } from "../types.js";
+
+// ─── Message type aliases ────────────────────────────────────────────────────
+
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+export type Message = Record<string, any>;
+
+// ─── Tool Result ID Extraction ───────────────────────────────────────────────
+
+/**
+ * Extract the linked tool_use_id from a tool-result message (supports multiple formats).
+ */
+export function extractToolResultId(msg: Message): string | null {
+  // OpenAI tool result: { role: "tool", tool_call_id: "..." }
+  if (msg.tool_call_id) return msg.tool_call_id;
+  // Anthropic tool_result: { type: "tool_result", tool_use_id: "..." }
+  if (msg.tool_use_id) return msg.tool_use_id;
+  // OpenClaw wrapped: { type: "message", message: { id: "...", role: "toolResult" } }
+  if (msg.type === "message" && msg.message?.id) return msg.message.id;
+  // Anthropic content block tool_result: { role: "user", content: [{type:"tool_result", tool_use_id}] }
+  if (msg.role === "user" && Array.isArray(msg.content)) {
+    for (const block of msg.content) {
+      if (block?.type === "tool_result" && block.tool_use_id) return block.tool_use_id;
+    }
+  }
+  return null;
+}
+
+// ─── Message Type Checks ─────────────────────────────────────────────────────
+
+/**
+ * Check if a message is a tool result (contains tool output).
+ */
+export function isToolResultMessage(msg: Message): boolean {
+  if (msg.role === "tool") return true;
+  if (msg.type === "tool_result") return true;
+  const innerRole = msg.message?.role ?? msg.type;
+  if (innerRole === "toolResult" || innerRole === "tool_result") return true;
+  // Anthropic user message with tool_result content blocks
+  if (msg.role === "user" && Array.isArray(msg.content)) {
+    return msg.content.some((b: any) => b?.type === "tool_result");
+  }
+  return false;
+}
+
+/**
+ * Check if a message is an assistant message containing only tool_use blocks.
+ */
+export function isOnlyToolUseAssistant(msg: Message): boolean {
+  const role = msg.role ?? msg.message?.role;
+  if (role !== "assistant") return false;
+  const content = msg.type === "message" ? msg.message?.content : msg.content;
+  if (!Array.isArray(content)) return false;
+  if (content.length === 0) return false;
+  return content.every((block: any) =>
+    block?.type === "tool_use" || block?.type === "toolCall",
+  );
+}
+
+/**
+ * Check if a message is an assistant message that contains tool_use (may also have text).
+ * Supports both Anthropic format (content blocks) and OpenAI format (tool_calls field).
+ */
+export function isAssistantWithToolUse(msg: Message): boolean {
+  const role = msg.role ?? msg.message?.role;
+  if (role !== "assistant") return false;
+  // OpenAI format: tool_calls field
+  if (msg.tool_calls && Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0) return true;
+  // Anthropic format: content blocks
+  const content = msg.type === "message" ? msg.message?.content : msg.content;
+  if (!Array.isArray(content)) return false;
+  return content.some((block: any) =>
+    block?.type === "tool_use" || block?.type === "toolCall",
+  );
+}
+
+/**
+ * Extract all tool_use IDs from an assistant message.
+ * Supports both Anthropic format (content blocks) and OpenAI format (tool_calls field).
+ */
+export function extractAllToolUseIds(msg: Message): string[] {
+  const ids: string[] = [];
+  // OpenAI format: tool_calls field
+  if (msg.tool_calls && Array.isArray(msg.tool_calls)) {
+    for (const tc of msg.tool_calls) {
+      if (tc.id) ids.push(tc.id);
+    }
+  }
+  // Anthropic format: content blocks
+  const content = msg.type === "message" ? msg.message?.content : msg.content;
+  if (Array.isArray(content)) {
+    for (const block of content) {
+      if ((block?.type === "tool_use" || block?.type === "toolCall") && block.id) {
+        ids.push(block.id);
+      }
+    }
+  }
+  return ids;
+}
+
+// ─── Replace With Summary ────────────────────────────────────────────────────
+
+/**
+ * Replace a tool result message's content with the offload summary.
+ */
+export function replaceWithSummary(msg: Message, entry: OffloadEntry): void {
+  const parts = [
+    `[Offloaded Tool Result | node: ${entry.node_id ?? "N/A"}]`,
+    `Summary: ${entry.summary}`,
+  ];
+  if (entry.result_ref) {
+    parts.push(`原始工具结果已存档，如需查看完整内容请调用 tdai_read_cos(path="${entry.result_ref}")`);
+  }
+  const summaryContent = parts.join("\n");
+
+  if (msg.type === "message" && msg.message) {
+    if (Array.isArray(msg.message.content)) {
+      msg.message.content = [{ type: "text", text: summaryContent }];
+    } else {
+      msg.message.content = summaryContent;
+    }
+  } else if (msg.role === "user" && Array.isArray(msg.content)) {
+    // Anthropic tool_result in user content blocks
+    for (let i = 0; i < msg.content.length; i++) {
+      if (msg.content[i]?.type === "tool_result") {
+        msg.content[i].content = summaryContent;
+        break;
+      }
+    }
+  } else {
+    if (Array.isArray(msg.content)) {
+      msg.content = [{ type: "text", text: summaryContent }];
+    } else {
+      msg.content = summaryContent;
+    }
+  }
+  msg._offloaded = true;
+}
+
+// ─── Offload Map ─────────────────────────────────────────────────────────────
+
+/**
+ * Build a lookup map from tool_call_id → OffloadEntry.
+ */
+export function buildOffloadMap(entries: OffloadEntry[]): Map<string, OffloadEntry> {
+  const map = new Map<string, OffloadEntry>();
+  for (const entry of entries) {
+    if (entry.tool_call_id) {
+      map.set(entry.tool_call_id, entry);
+    }
+  }
+  return map;
+}
+
+// ─── Token Estimation ────────────────────────────────────────────────────────
+
+/**
+ * Count CJK characters (CJK Unified Ideographs + Extension A + Compatibility).
+ */
+function countCjkChars(text: string): number {
+  let n = 0;
+  for (const ch of text) {
+    const c = ch.codePointAt(0)!;
+    if (
+      (c >= 0x4e00 && c <= 0x9fff) ||
+      (c >= 0x3400 && c <= 0x4dbf) ||
+      (c >= 0xf900 && c <= 0xfaff)
+    ) {
+      n++;
+    }
+  }
+  return n;
+}
+
+/**
+ * Estimate tokens for a text string using CJK-aware heuristic.
+ * CJK characters ≈ 1 token / 1.7 chars, non-CJK ≈ 1 token / 4 chars.
+ * Aligned with plugin-side estimateL3MixedTokensHeuristic.
+ */
+function estimateTextTokens(text: string): number {
+  const cjk = countCjkChars(text);
+  const rest = Math.max(0, text.length - cjk);
+  return Math.max(1, Math.ceil(cjk / 1.7 + rest / 4));
+}
+
+/**
+ * Estimate tokens for a message using CJK-aware heuristic (中文/1.7 + 非中文/4).
+ */
+export function estimateMessageTokens(msg: Message): number {
+  const content = msg.content ?? msg.message?.content ?? "";
+  let text: string;
+  if (typeof content === "string") {
+    text = content;
+  } else if (Array.isArray(content)) {
+    text = content.map((b: any) => (typeof b === "string" ? b : b?.text ?? JSON.stringify(b) ?? "")).join("");
+  } else {
+    text = JSON.stringify(content);
+  }
+
+  // Include tool_calls arguments (OpenAI format)
+  const toolCalls = (msg as any).tool_calls;
+  if (toolCalls && Array.isArray(toolCalls)) {
+    for (const tc of toolCalls) {
+      if (tc.function?.name) text += tc.function.name;
+      if (tc.function?.arguments) text += tc.function.arguments;
+    }
+  }
+
+  return estimateTextTokens(text);
+}
+
+// ─── MMD Message Markers ─────────────────────────────────────────────────────
+
+export const MMD_CONTEXT_MARKER = "_mmdContextMessage";
+export const MMD_INJECTION_MARKER = "_mmdInjection";
+
+/**
+ * Check if a message is a MMD-related message (should be preserved during compression).
+ */
+export function isMmdMessage(msg: Message): boolean {
+  return !!msg[MMD_CONTEXT_MARKER] || !!msg[MMD_INJECTION_MARKER];
+}

src/offload_server/compact/mmd-injector.ts

@@ -0,0 +1,430 @@
+/**
+ * L3 MMD Injector — inject active/history MMD into messages.
+ * Handles version dedup and insertion point calculation.
+ */
+import type { StorageAdapter } from "../../core/storage/adapter.js";
+import type { OffloadEntry, OffloadState } from "../types.js";
+import type { Message } from "./helpers.js";
+import { MMD_CONTEXT_MARKER, MMD_INJECTION_MARKER, buildOffloadMap } from "./helpers.js";
+
+// ─── Active MMD Injection ────────────────────────────────────────────────────
+
+export interface MmdInjectionResult {
+  injectedCount: number;
+  mmdTokensEstimate: number;
+}
+
+/**
+ * Inject the active MMD into messages. Performs version dedup.
+ */
+export async function injectActiveMmd(
+  messages: Message[],
+  state: OffloadState,
+  storage: StorageAdapter,
+  basePath: string,
+): Promise<MmdInjectionResult> {
+  // Resolve the MMD file to inject: use activeMmdFile, or fallback to the latest .mmd
+  let mmdFile = state.activeMmdFile;
+  if (!mmdFile) {
+    // Fallback: find the most recent .mmd file (sorted by prefix number, highest = latest)
+    const allMmds = await storage.readdirNames(`${basePath}/mmds/`, ".mmd");
+    if (allMmds.length > 0) {
+      // Sort descending by prefix number (e.g., "006-current-task.mmd" → 6)
+      allMmds.sort((a, b) => {
+        const numA = parseInt(a.split("-")[0], 10) || 0;
+        const numB = parseInt(b.split("-")[0], 10) || 0;
+        return numB - numA;
+      });
+      // Pick the first non-empty one
+      for (const candidate of allMmds) {
+        const content = await storage.readFile(`${basePath}/mmds/${candidate}`);
+        if (content && content.trim().length > 0) {
+          mmdFile = candidate;
+          break;
+        }
+      }
+    }
+  }
+
+  if (!mmdFile) {
+    removeMmdMessages(messages, "active");
+    return { injectedCount: 0, mmdTokensEstimate: 0 };
+  }
+
+  const mmdContent = await storage.readFile(`${basePath}/mmds/${mmdFile}`);
+  if (!mmdContent) {
+    removeMmdMessages(messages, "active");
+    return { injectedCount: 0, mmdTokensEstimate: 0 };
+  }
+
+  const newVersion = hashContent(mmdContent);
+
+  // Version dedup: check existing MMD message version and position
+  const existingIdx = messages.findIndex((m) => m[MMD_CONTEXT_MARKER] === "active");
+  if (existingIdx >= 0 && messages[existingIdx]._mmdVersion === newVersion) {
+    // Content unchanged — but still need to ensure correct position
+    // (aggressive compression may have shifted it)
+    const correctIdx = findActiveMmdInsertionPoint(
+      messages.filter((m) => m[MMD_CONTEXT_MARKER] !== "active"),
+    );
+    // Check if position is already correct (accounting for removal offset)
+    const effectiveCorrectIdx = correctIdx <= existingIdx ? correctIdx : correctIdx;
+    if (existingIdx === effectiveCorrectIdx) {
+      return { injectedCount: 0, mmdTokensEstimate: 0 };
+    }
+    // Reposition: remove and re-insert at correct point
+    const [mmdMsg] = messages.splice(existingIdx, 1);
+    const newInsertIdx = findActiveMmdInsertionPoint(messages);
+    messages.splice(newInsertIdx, 0, mmdMsg);
+    return { injectedCount: 0, mmdTokensEstimate: 0 };
+  }
+
+  // Remove old active MMD message
+  removeMmdMessages(messages, "active");
+
+  // Build MMD text
+  const mmdText = buildActiveMmdText(mmdFile, mmdContent);
+  const mmdMsg: Message = {
+    role: "user",
+    content: mmdText,
+    [MMD_CONTEXT_MARKER]: "active",
+    _mmdVersion: newVersion,
+    _mmdFilename: mmdFile,
+  };
+
+  // Insert at calculated point
+  const insertIdx = findActiveMmdInsertionPoint(messages);
+  messages.splice(insertIdx, 0, mmdMsg);
+
+  return { injectedCount: 1, mmdTokensEstimate: Math.ceil(mmdContent.length / 4) };
+}
+
+// ─── History MMD Injection ───────────────────────────────────────────────────
+
+export interface HistoryMmdResult {
+  injectedCount: number;
+  mmdFiles: string[];
+  totalTokensEstimate: number;
+}
+
+/**
+ * Inject history MMD files for entries that were deleted during aggressive compression.
+ */
+export async function injectHistoryMmds(
+  messages: Message[],
+  deletedIds: string[],
+  entries: OffloadEntry[],
+  state: OffloadState,
+  storage: StorageAdapter,
+  basePath: string,
+  tokenBudget: number,
+): Promise<HistoryMmdResult> {
+  if (deletedIds.length === 0) {
+    return { injectedCount: 0, mmdFiles: [], totalTokensEstimate: 0 };
+  }
+
+  // 1. Find MMD prefixes from deleted entries' node_ids
+  const offloadMap = buildOffloadMap(entries);
+  const mmdPrefixes = new Set<string>();
+  for (const id of deletedIds) {
+    const entry = offloadMap.get(id);
+    if (entry?.node_id) {
+      const prefix = entry.node_id.split("-")[0];
+      if (prefix) mmdPrefixes.add(prefix);
+    }
+  }
+  if (mmdPrefixes.size === 0) {
+    return { injectedCount: 0, mmdFiles: [], totalTokensEstimate: 0 };
+  }
+
+  // 2. Find matching history MMD files (exclude active)
+  const allMmds = await storage.readdirNames(`${basePath}/mmds/`, ".mmd");
+  const candidates = allMmds.filter((f) => {
+    const prefix = f.split("-")[0];
+    return mmdPrefixes.has(prefix) && f !== state.activeMmdFile;
+  });
+  if (candidates.length === 0) {
+    return { injectedCount: 0, mmdFiles: [], totalTokensEstimate: 0 };
+  }
+
+  // 3. Read and inject (respecting token budget), most recent first
+  candidates.reverse();
+  const injected: Message[] = [];
+  const mmdFiles: string[] = [];
+  let usedTokens = 0;
+
+  for (const filename of candidates) {
+    const content = await storage.readFile(`${basePath}/mmds/${filename}`);
+    if (!content) continue;
+
+    const text = buildHistoryMmdText(filename, content);
+    const tokens = Math.ceil(text.length / 4);
+    if (usedTokens + tokens > tokenBudget) continue;
+
+    injected.push({
+      role: "user",
+      content: text,
+      [MMD_INJECTION_MARKER]: true,
+      _mmdFilename: filename,
+    });
+    mmdFiles.push(filename);
+    usedTokens += tokens;
+  }
+
+  if (injected.length === 0) {
+    return { injectedCount: 0, mmdFiles: [], totalTokensEstimate: 0 };
+  }
+
+  // 4. Remove old history MMD injections and insert new ones
+  removeExistingMmdInjections(messages);
+  const insertIdx = findHistoryMmdInsertionPoint(messages);
+  // Reverse back to chronological order (oldest first)
+  injected.reverse();
+  mmdFiles.reverse();
+  messages.splice(insertIdx, 0, ...injected);
+
+  return { injectedCount: injected.length, mmdFiles, totalTokensEstimate: usedTokens };
+}
+
+// ─── Insertion Point Calculation ─────────────────────────────────────────────
+
+/**
+ * Find insertion point for active MMD.
+ * Strategy: insert after the latest user message in the second half.
+ * Guard: don't split tool_call / tool_result pairs.
+ */
+export function findActiveMmdInsertionPoint(messages: Message[]): number {
+  if (messages.length <= 2) {
+    const idx = Math.min(1, messages.length);
+    return adjustForToolCallPair(messages, idx);
+  }
+
+  let latestUserIdx = -1;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const role = messages[i].role ?? messages[i].message?.role ?? messages[i].type;
+    if (role === "user" && !messages[i][MMD_CONTEXT_MARKER] && !messages[i][MMD_INJECTION_MARKER]) {
+      latestUserIdx = i;
+      break;
+    }
+  }
+
+  let insertIdx: number;
+  if (latestUserIdx >= 0) {
+    // Insert relative to latest user message:
+    // - If it's the last message → insert before it (user prompt stays last)
+    // - Otherwise → insert after it (between user and tool loop, matching plugin behavior)
+    if (latestUserIdx === messages.length - 1) {
+      insertIdx = latestUserIdx;
+    } else {
+      insertIdx = latestUserIdx + 1;
+    }
+  } else {
+    // No user message found — fallback: before the trailing tool loop
+    let loopStart = messages.length;
+    for (let i = messages.length - 1; i >= 0; i--) {
+      const role = messages[i].role ?? messages[i].message?.role ?? messages[i].type;
+      if (messages[i][MMD_CONTEXT_MARKER] || messages[i][MMD_INJECTION_MARKER]) continue;
+      if (role === "toolResult" || role === "tool" || role === "assistant") {
+        loopStart = i;
+      } else {
+        break;
+      }
+    }
+    const maxDistFromTail = 30;
+    const minInsertIdx = Math.max(1, messages.length - maxDistFromTail);
+    insertIdx = Math.max(loopStart, minInsertIdx);
+  }
+
+  // Guard: don't insert between assistant(tool_use) and its tool_result
+  insertIdx = adjustForToolCallPair(messages, insertIdx);
+
+  // Hard guard: never insert before system message
+  if (insertIdx === 0 && messages.length > 0) {
+    const firstRole = messages[0].role ?? messages[0].message?.role ?? messages[0].type;
+    if (firstRole === "system") {
+      insertIdx = 1;
+    }
+  }
+
+  return insertIdx;
+}
+
+/**
+ * Find insertion point for history MMD (before active MMD).
+ */
+export function findHistoryMmdInsertionPoint(messages: Message[]): number {
+  for (let i = 0; i < messages.length; i++) {
+    if (messages[i][MMD_CONTEXT_MARKER] === "active") return i;
+  }
+  return findActiveMmdInsertionPoint(messages);
+}
+
+// ─── MMD Message Management ─────────────────────────────────────────────────
+
+/**
+ * Remove MMD messages by type ("active", "history", or all).
+ */
+export function removeMmdMessages(messages: Message[], type?: "active" | "history"): number {
+  let removed = 0;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (type === "active" && messages[i][MMD_CONTEXT_MARKER] === "active") {
+      messages.splice(i, 1);
+      removed++;
+    } else if (type === "history" && messages[i][MMD_INJECTION_MARKER]) {
+      messages.splice(i, 1);
+      removed++;
+    } else if (!type && (messages[i][MMD_CONTEXT_MARKER] || messages[i][MMD_INJECTION_MARKER])) {
+      messages.splice(i, 1);
+      removed++;
+    }
+  }
+  return removed;
+}
+
+/**
+ * Remove existing history MMD injection messages.
+ */
+export function removeExistingMmdInjections(messages: Message[]): number {
+  let removed = 0;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i][MMD_INJECTION_MARKER]) {
+      messages.splice(i, 1);
+      removed++;
+    }
+  }
+  return removed;
+}
+
+// ─── Text Builders ───────────────────────────────────────────────────────────
+
+/**
+ * Build active MMD injection text.
+ */
+export function buildActiveMmdText(filename: string, mmdContent: string): string {
+  let taskGoal = "";
+  const metaMatch = mmdContent.match(/^%%\{\s*(.*?)\s*\}%%/);
+  if (metaMatch) {
+    try {
+      const meta = JSON.parse(`{${metaMatch[1]}}`);
+      taskGoal = meta.taskGoal || "";
+    } catch { /* ignore */ }
+  }
+
+  return [
+    `<current_task_context>`,
+    `【当前活跃任务的mermaid流程图】这是你最近正在执行的任务的阶段性记录。`,
+    taskGoal ? `**任务目标:** ${taskGoal}` : "",
+    `**任务文件:** ${filename}`,
+    "```mermaid",
+    mmdContent,
+    "```",
+    `标记为 "doing" 的节点是近期焦点，"done" 的已完成。请参考此保持方向感，避免重复已完成的工作。`,
+    `</current_task_context>`,
+  ]
+    .filter((line) => line !== "")
+    .join("\n");
+}
+
+/**
+ * Build history MMD injection text.
+ */
+export function buildHistoryMmdText(filename: string, mmdContent: string): string {
+  let taskGoal = "";
+  const metaMatch = mmdContent.match(/^%%\{\s*(.*?)\s*\}%%/);
+  if (metaMatch) {
+    try {
+      const meta = JSON.parse(`{${metaMatch[1]}}`);
+      taskGoal = meta.taskGoal || "";
+    } catch { /* ignore */ }
+  }
+
+  return [
+    `<history_task_context>`,
+    `【历史任务记录】以下是此前完成的任务的概要。`,
+    taskGoal ? `**任务目标:** ${taskGoal}` : "",
+    `**任务文件:** ${filename}`,
+    "```mermaid",
+    mmdContent,
+    "```",
+    `</history_task_context>`,
+  ]
+    .filter((line) => line !== "")
+    .join("\n");
+}
+
+// ─── Internal Helpers ────────────────────────────────────────────────────────
+
+function hasToolResultContent(msg: Message): boolean {
+  const content = msg?.content;
+  if (!Array.isArray(content)) return false;
+  return content.some(
+    (block: any) =>
+      typeof block === "object" &&
+      block !== null &&
+      (block.type === "tool_result" || block.type === "toolResult"),
+  );
+}
+
+function hasToolUseContent(msg: Message): boolean {
+  const content = msg?.content;
+  if (!Array.isArray(content)) return false;
+  return content.some(
+    (block: any) =>
+      typeof block === "object" &&
+      block !== null &&
+      (block.type === "tool_use" || block.type === "toolUse"),
+  );
+}
+
+function adjustForToolCallPair(messages: Message[], insertIdx: number): number {
+  if (insertIdx <= 0 || insertIdx >= messages.length) return insertIdx;
+  const msgAtIdx = messages[insertIdx];
+  const role = msgAtIdx?.role ?? msgAtIdx?.message?.role ?? msgAtIdx?.type;
+
+  // Check both role-level (OpenAI: role="tool") and content-level (Anthropic: content[].type="tool_result")
+  const isToolResult =
+    role === "tool" ||
+    role === "toolResult" ||
+    role === "tool_result" ||
+    (role === "user" && hasToolResultContent(msgAtIdx));
+
+  if (isToolResult) {
+    // Walk back to before the assistant tool_use
+    let i = insertIdx - 1;
+    while (i >= 0) {
+      const r = messages[i].role ?? messages[i].message?.role ?? messages[i].type;
+      if (r === "assistant" && hasToolUseContent(messages[i])) {
+        return i;
+      }
+      if (r === "assistant") {
+        return i;
+      }
+      const prevIsToolResult =
+        r === "tool" ||
+        r === "toolResult" ||
+        r === "tool_result" ||
+        (r === "user" && hasToolResultContent(messages[i]));
+      if (!prevIsToolResult) break;
+      i--;
+    }
+  }
+
+  // Also check: don't insert right after an assistant with tool_use (before its tool_result)
+  if (insertIdx > 0) {
+    const prevMsg = messages[insertIdx - 1];
+    const prevRole = prevMsg?.role ?? prevMsg?.message?.role ?? prevMsg?.type;
+    if (prevRole === "assistant" && hasToolUseContent(prevMsg)) {
+      return insertIdx - 1;
+    }
+  }
+
+  return insertIdx;
+}
+
+function hashContent(s: string): string {
+  let hash = 0;
+  for (let i = 0; i < s.length; i++) {
+    hash = ((hash << 5) - hash + s.charCodeAt(i)) | 0;
+  }
+  return Math.abs(hash).toString(36).padStart(6, "0").slice(0, 6);
+}

src/offload_server/ingest-handler.ts

@@ -0,0 +1,262 @@
+/**
+ * Offload Ingest Handler — fast path: COS write + task enqueue.
+ * Branch A: toolPairs only → write pending.jsonl, enqueue async L1
+ * Branch B: recentMessages only (no toolPairs) → L1.5 path
+ * Branch C: toolPairs + recentMessages → write pending + cache recentMessages, trigger L1 (skip L1.5)
+ */
+import type http from "node:http";
+import type { StorageAdapter } from "../core/storage/adapter.js";
+import type { IStateBackend, TaskPayload } from "../core/state/types.js";
+import type { OffloadExecutorConfig, OffloadState } from "./types.js";
+import { defaultOffloadState } from "./types.js";
+import { IngestRequestSchema } from "./schemas.js";
+import { serializeJsonl } from "./parsers/json-utils.js";
+import { buildOffloadBasePath } from "./session-utils.js";
+
+// ─── Per-session in-process mutex for COS append serialization ───
+// Layer 1 (local): ensures at most one inflight appendFile per session within
+// the same Node.js process (no I/O overhead, zero-latency queue).
+// Layer 2 (distributed): stateBackend.acquireLock across multiple offload server
+// instances prevents concurrent AppendObject calls to the same COS key.
+const pendingMutexes = new Map<string, Promise<void>>();
+
+async function withSessionMutex<T>(sessionKey: string, fn: () => Promise<T>): Promise<T> {
+  const prev = pendingMutexes.get(sessionKey) ?? Promise.resolve();
+  let resolve!: () => void;
+  const next = new Promise<void>((r) => { resolve = r; });
+  pendingMutexes.set(sessionKey, next);
+
+  try {
+    await prev;
+    return await fn();
+  } finally {
+    resolve();
+    if (pendingMutexes.get(sessionKey) === next) {
+      pendingMutexes.delete(sessionKey);
+    }
+  }
+}
+
+/** Max attempts to acquire the distributed lock. */
+const APPEND_LOCK_MAX_ATTEMPTS = 10;
+/** TTL for the distributed append lock (ms). Short-lived since append is fast. */
+const APPEND_LOCK_TTL_MS = 5000;
+/** Delay between lock acquisition retries (ms). Uses exponential backoff: base * 2^attempt. */
+const APPEND_LOCK_RETRY_BASE_MS = 50;
+
+export interface IngestDeps {
+  storage: StorageAdapter;
+  stateBackend?: IStateBackend;
+  config: OffloadExecutorConfig;
+  logger: { info: (...args: unknown[]) => void; warn: (...args: unknown[]) => void };
+}
+
+export async function handleIngest(
+  req: http.IncomingMessage,
+  res: http.ServerResponse,
+  auth: { serviceId: string },
+  deps: IngestDeps,
+  requestId: string,
+  parseJsonBody: <T>(req: http.IncomingMessage) => Promise<T>,
+  sendJson: (res: http.ServerResponse, status: number, body: unknown) => void,
+  successEnvelope: <T>(data: T, requestId: string) => unknown,
+  errorEnvelope: (code: number, message: string, requestId: string) => unknown,
+): Promise<void> {
+  const body = await parseJsonBody(req);
+  const parsed = IngestRequestSchema.safeParse(body);
+  if (!parsed.success) {
+    sendJson(res, 400, errorEnvelope(400, parsed.error.message, requestId));
+    return;
+  }
+
+  const { session_id: sessionId, tool_pairs: toolPairs, prompt, recent_messages: recentMessages } = parsed.data;
+  const { storage, stateBackend, config } = deps;
+  const basePath = buildOffloadBasePath(sessionId);
+
+  // Build context string from structured prompt + recentMessages (for L1/L1.5)
+  let contextText: string | undefined;
+  if (prompt || (recentMessages && recentMessages.length > 0)) {
+    const parts: string[] = [];
+    if (recentMessages && recentMessages.length > 0) {
+      parts.push("历史消息，可作为参考：");
+      for (const m of recentMessages) {
+        parts.push(`[${m.role === "user" ? "User" : "Assistant"}]: ${m.content}`);
+      }
+    }
+    if (prompt) {
+      parts.push(`\n最新user message：\n[User]: ${prompt}`);
+    }
+    contextText = parts.join("\n");
+  }
+
+  // ─── Session skip filter (shared by L1 and L1.5) ───
+  const INTERNAL_SESSION_RE = /memory-.*-session-\d+/;
+  const shouldSkipSession = INTERNAL_SESSION_RE.test(sessionId) || sessionId.includes("subagent");
+
+  // ─── Branch L1: toolPairs non-empty → write pending + trigger L1 ───
+  if (toolPairs.length > 0) {
+    if (shouldSkipSession) {
+      deps.logger.info(`[offload-server] ingest: L1 skipped (session=${sessionId})`);
+      sendJson(res, 200, successEnvelope({}, requestId));
+      return;
+    }
+    // Save context for L1 executor (if available)
+    if (contextText) {
+      await storage.writeFile(`${basePath}/recent-context.txt`, contextText);
+    }
+
+    const pendingPath = `${basePath}/pending.jsonl`;
+    // Map API snake_case → internal camelCase for JSONL storage
+    const camelPairs = toolPairs.map((tp: Record<string, unknown>) => ({
+      toolName: tp.tool_name,
+      toolCallId: tp.tool_call_id,
+      params: tp.params,
+      result: tp.result,
+      error: tp.error,
+      timestamp: tp.timestamp,
+      durationMs: tp.duration_ms,
+    }));
+    const lines = serializeJsonl(camelPairs);
+
+    // Two-layer serialization to prevent COS AppendPositionErr:
+    // Layer 1 (local mutex): queues concurrent requests within this process.
+    // Layer 2 (distributed lock via stateBackend): prevents races across server instances.
+    // If the lock cannot be acquired, return 409 so the client retries (NOT proceed without lock).
+    const lockAcquired = await withSessionMutex(pendingPath, async () => {
+      const lockKey = `offload-pending:${auth.serviceId}:${sessionId}`;
+      const lockOwner = requestId;
+      let locked = false;
+
+      if (stateBackend) {
+        for (let attempt = 0; attempt < APPEND_LOCK_MAX_ATTEMPTS; attempt++) {
+          locked = await stateBackend.acquireLock(lockKey, lockOwner, APPEND_LOCK_TTL_MS);
+          if (locked) break;
+          // Exponential backoff: 50, 100, 200, 400, 800, 1600... capped at 2000ms
+          const delay = Math.min(APPEND_LOCK_RETRY_BASE_MS * 2 ** attempt, 2000);
+          await new Promise((r) => setTimeout(r, delay));
+        }
+        if (!locked) {
+          deps.logger.warn(`[offload-server] ingest: append lock failed after ${APPEND_LOCK_MAX_ATTEMPTS} attempts (session=${sessionId}), returning 409`);
+          return false;
+        }
+      }
+
+      try {
+        await storage.appendFile(pendingPath, lines);
+
+        if (stateBackend) {
+          const raw = await storage.readFile(pendingPath);
+          const lineCount = raw ? raw.split("\n").filter(Boolean).length : 0;
+
+          if (lineCount >= config.forceTriggerThreshold) {
+            const task: TaskPayload = {
+              id: `offload-l1-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
+              type: "offload-l1" as TaskPayload["type"],
+              instanceId: auth.serviceId,
+              sessionId,
+              priority: 0,
+              data: { sessionId, instanceId: auth.serviceId },
+              createdAt: Date.now(),
+            };
+            await stateBackend.enqueueTask(task);
+            deps.logger.info(`[offload-server] ingest: ${toolPairs.length} pairs, triggered=true (lines=${lineCount})`);
+          } else {
+            await stateBackend.setTimerIfEarlier(
+              auth.serviceId,
+              `offload-l1:${auth.serviceId}:${sessionId}`,
+              Date.now() + config.pendingMaxAgeSeconds * 1000,
+            );
+            deps.logger.info(`[offload-server] ingest: ${toolPairs.length} pairs, lines=${lineCount}/${config.forceTriggerThreshold}, timer set`);
+          }
+        }
+        return true;
+      } finally {
+        if (locked && stateBackend) {
+          await stateBackend.releaseLock(lockKey, lockOwner);
+        }
+      }
+    });
+
+    if (!lockAcquired) {
+      sendJson(res, 409, errorEnvelope(409, "Concurrent write conflict, please retry", requestId));
+      return;
+    }
+
+    sendJson(res, 200, successEnvelope({}, requestId));
+    return;
+  }
+
+  // ─── Branch L1.5: toolPairs empty + prompt → task judgment ───
+  // Skip L1.5 for inter-session messages, internal sessions, and system prompts
+  const shouldSkipL15 = shouldSkipSession || !!(
+    prompt && (
+      prompt.startsWith("[Inter-session message]") ||
+      prompt.startsWith("Pre-compaction")
+    )
+  );
+  if (prompt && stateBackend && !shouldSkipL15) {
+    const lockKey = `offload-state:${auth.serviceId}:${sessionId}`;
+    const lockOwner = requestId;
+    let locked = false;
+
+    // Use exponential backoff, same policy as L1 append lock.
+    for (let attempt = 0; attempt < APPEND_LOCK_MAX_ATTEMPTS; attempt++) {
+      locked = await stateBackend.acquireLock(lockKey, lockOwner, APPEND_LOCK_TTL_MS);
+      if (locked) break;
+      const delay = Math.min(APPEND_LOCK_RETRY_BASE_MS * 2 ** attempt, 2000);
+      await new Promise((r) => setTimeout(r, delay));
+    }
+
+    if (!locked) {
+      // Do NOT proceed without the lock: a concurrent writer would overwrite state.json
+      // and silently drop this boundary, causing the L1.5 executor to skip the task
+      // (findIndex returns -1 → return without error).
+      deps.logger.warn(`[offload-server] ingest: L1.5 state lock failed after ${APPEND_LOCK_MAX_ATTEMPTS} attempts (session=${sessionId}), returning 409`);
+      sendJson(res, 409, errorEnvelope(409, "Concurrent write conflict, please retry", requestId));
+      return;
+    }
+
+    try {
+      const state = await readState(storage, basePath);
+      const boundaryTimestamp = new Date().toISOString();
+      state.boundaries.push({
+        targetMmd: "_pending",
+        timestamp: boundaryTimestamp,
+      });
+      await writeState(storage, basePath, state);
+
+      const task: TaskPayload = {
+        id: `offload-l15-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
+        type: "offload-l15" as TaskPayload["type"],
+        instanceId: auth.serviceId,
+        sessionId,
+        priority: 0,
+        data: { sessionId, recentMessages: contextText, boundaryTimestamp, instanceId: auth.serviceId },
+        createdAt: Date.now(),
+      };
+      await stateBackend.enqueueTask(task);
+      deps.logger.info(`[offload-server] ingest: L1.5 triggered, boundaryTs=${boundaryTimestamp}, prompt=${prompt.slice(0, 200)}`);
+    } finally {
+      await stateBackend.releaseLock(lockKey, lockOwner);
+    }
+  }
+
+  // ─── Fast return ───
+  sendJson(res, 200, successEnvelope({}, requestId));
+}
+
+// ─── State helpers (duplicated intentionally to keep ingest-handler self-contained) ──
+
+async function readState(storage: StorageAdapter, basePath: string): Promise<OffloadState> {
+  const raw = await storage.readFile(`${basePath}/state.json`);
+  if (!raw) return defaultOffloadState();
+  try {
+    return { ...defaultOffloadState(), ...JSON.parse(raw) };
+  } catch {
+    return defaultOffloadState();
+  }
+}
+
+async function writeState(storage: StorageAdapter, basePath: string, state: OffloadState): Promise<void> {
+  await storage.writeFile(`${basePath}/state.json`, JSON.stringify(state));
+}

src/offload_server/mmd-handler.ts

@@ -0,0 +1,90 @@
+/**
+ * Offload MMD Query Handler — returns MMD files.
+ * Body params: sessionId (required), limit (optional, default=all)
+ * When limit=1, only returns the active MMD.
+ */
+import type http from "node:http";
+import type { StorageAdapter } from "../core/storage/adapter.js";
+import type { OffloadState } from "./types.js";
+import { createHash } from "node:crypto";
+import { buildOffloadBasePath } from "./session-utils.js";
+
+function hashContent(content: string): string {
+  return createHash("md5").update(content).digest("hex").slice(0, 12);
+}
+
+export async function handleMmdQuery(
+  _req: http.IncomingMessage,
+  res: http.ServerResponse,
+  _auth: { serviceId: string },
+  storage: StorageAdapter,
+  requestId: string,
+  sendJson: (res: http.ServerResponse, status: number, body: unknown) => void,
+  successEnvelope: <T>(data: T, requestId: string) => unknown,
+  errorEnvelope: (code: number, message: string, requestId: string) => unknown,
+  sessionId: string,
+  limit?: number,
+): Promise<void> {
+  if (!sessionId) {
+    sendJson(res, 400, errorEnvelope(400, "missing sessionId", requestId));
+    return;
+  }
+
+  const basePath = buildOffloadBasePath(sessionId);
+
+  // Read state
+  const stateRaw = await storage.readFile(`${basePath}/state.json`);
+  let state: Partial<OffloadState> = {};
+  if (stateRaw) {
+    try {
+      state = JSON.parse(stateRaw);
+    } catch {
+      // ignore
+    }
+  }
+
+  const mmdsPrefix = `${basePath}/mmds/`;
+  const mmds: Array<{
+    filename: string;
+    content: string;
+    version: string;
+  }> = [];
+
+  if (limit === 1 && state.activeMmdFile) {
+    // Fast path: only return the active MMD
+    const content = await storage.readFile(`${mmdsPrefix}${state.activeMmdFile}`) ?? "";
+    mmds.push({
+      filename: state.activeMmdFile,
+      content,
+      version: content ? hashContent(content) : "",
+    });
+  } else {
+    // Return all (or up to limit)
+    const mmdFiles = await storage.readdirNames(mmdsPrefix, ".mmd");
+    const filesToRead = limit && limit > 0 ? mmdFiles.slice(0, limit) : mmdFiles;
+
+    const readResults = await Promise.all(
+      filesToRead.map(async (filename) => {
+        const content = await storage.readFile(`${mmdsPrefix}${filename}`) ?? "";
+        return {
+          filename,
+          content,
+          version: content ? hashContent(content) : "",
+        };
+      }),
+    );
+    mmds.push(...readResults);
+  }
+
+  sendJson(
+    res,
+    200,
+    successEnvelope(
+      {
+        mmds,
+        currentMmd: state.activeMmdFile ?? null,
+      },
+      requestId,
+    ),
+  );
+}

src/offload_server/offload-task-executor.ts

@@ -0,0 +1,745 @@
+/**
+ * Offload Task Executor — async L1/L1.5/L2 execution via PipelineWorker.
+ */
+import type { StorageAdapter } from "../core/storage/adapter.js";
+import type { IStateBackend, TaskPayload } from "../core/state/types.js";
+import type {
+  OffloadEntry,
+  OffloadState,
+  OffloadExecutorConfig,
+  L2ParsedResponse,
+  MmdMeta,
+  ToolPair,
+} from "./types.js";
+import { defaultOffloadState, defaultOffloadConfig } from "./types.js";
+import { parseJsonl, serializeJsonl } from "./parsers/json-utils.js";
+import { parseL1Response } from "./parsers/l1-parser.js";
+import { parseL15Response } from "./parsers/l15-parser.js";
+import { parseL2Response } from "./parsers/l2-parser.js";
+import { L1_SYSTEM_PROMPT, buildL1UserPrompt } from "./prompts/l1-prompt.js";
+import { L15_SYSTEM_PROMPT, buildL15UserPrompt } from "./prompts/l15-prompt.js";
+import { L2_SYSTEM_PROMPT, buildL2UserPrompt } from "./prompts/l2-prompt.js";
+import { handleTaskTransition, extractMmdMeta } from "./task-transition.js";
+import { buildOffloadBasePath } from "./session-utils.js";
+import { traceServerModelIo, traceServerTaskDecision } from "./opik-tracer.js";
+
+// ─── LLM Client Interface ────────────────────────────────────────────────────
+
+export interface LlmClient {
+  chat(params: {
+    model: string;
+    messages: Array<{ role: "system" | "user"; content: string }>;
+    temperature: number;
+    max_tokens: number;
+    timeoutMs?: number;
+  }): Promise<string>;
+}
+
+// ─── Logger Interface ────────────────────────────────────────────────────────
+
+export interface ExecutorLogger {
+  info: (...args: unknown[]) => void;
+  warn: (...args: unknown[]) => void;
+  error: (...args: unknown[]) => void;
+}
+
+// ─── Executor Deps ───────────────────────────────────────────────────────────
+
+export interface OffloadExecutorDeps {
+  resolveStorage: (instanceId: string) => Promise<StorageAdapter | undefined>;
+  llmClient: LlmClient;
+  stateBackend: IStateBackend;
+  config?: OffloadExecutorConfig;
+  logger: ExecutorLogger;
+}
+
+// ─── Executor Class ──────────────────────────────────────────────────────────
+
+export class OffloadTaskExecutor {
+  private deps: OffloadExecutorDeps;
+  private config: OffloadExecutorConfig;
+
+  constructor(deps: OffloadExecutorDeps) {
+    this.deps = deps;
+    this.config = deps.config ?? defaultOffloadConfig();
+  }
+
+  // ═══════════════════════════════════════════════════════════════════════════
+  // L1: Summarize pending ToolPairs → OffloadEntry[]
+  // ═══════════════════════════════════════════════════════════════════════════
+
+  async executeOffloadL1(task: TaskPayload, _signal?: AbortSignal): Promise<void> {
+    const startMs = Date.now();
+    const sessionId = this.extractSessionId(task);
+    if (!sessionId) return;
+
+    const storage = await this.resolveStorageOrThrow(task.instanceId);
+    const basePath = buildOffloadBasePath(sessionId);
+    const pendingPath = `${basePath}/pending.jsonl`;
+
+    // 1. Claim pending via rename (atomic: new ingest appends to fresh pending.jsonl)
+    const processingPath = `${basePath}/pending-processing-${task.id}.jsonl`;
+    const pendingRaw = await storage.readFile(pendingPath);
+    if (!pendingRaw || !pendingRaw.trim()) return;
+
+    try {
+      await storage.rename(pendingPath, processingPath);
+    } catch {
+      // rename failed (file gone — another L1 claimed it, or doesn't exist)
+      return;
+    }
+
+    // Re-read from processing file (rename succeeded, this is our exclusive copy)
+    const claimedRaw = await storage.readFile(processingPath);
+    if (!claimedRaw || !claimedRaw.trim()) {
+      await storage.unlink(processingPath);
+      return;
+    }
+
+    let toolPairs = parseJsonl<Record<string, unknown>>(claimedRaw, (line, err) => {
+      this.deps.logger.warn(`[offload-server] L1: bad JSONL line in pending: ${line}`, err);
+    });
+    if (toolPairs.length === 0) {
+      await storage.unlink(processingPath);
+      return;
+    }
+
+    // 2. Dedup: remove toolPairs whose toolCallId already exists in entries.jsonl
+    const entriesPath = `${basePath}/entries.jsonl`;
+    const existingEntriesRaw = await storage.readFile(entriesPath);
+    if (existingEntriesRaw) {
+      const existingIds = new Set<string>();
+      for (const e of parseJsonl<OffloadEntry>(existingEntriesRaw)) {
+        if (e.tool_call_id) existingIds.add(e.tool_call_id);
+      }
+      const before = toolPairs.length;
+      toolPairs = toolPairs.filter((tp) => {
+        const id = tp.toolCallId as string;
+        return !id || !existingIds.has(id);
+      });
+      if (toolPairs.length < before) {
+        this.deps.logger.info(`[offload-server] L1: dedup removed ${before - toolPairs.length}/${before} duplicate toolPairs`);
+      }
+      if (toolPairs.length === 0) {
+        await storage.unlink(processingPath);
+        return;
+      }
+    }
+
+    // 3. Build prompt & call LLM
+    const recentContext = await storage.readFile(`${basePath}/recent-context.txt`) ?? "";
+    const userPrompt = buildL1UserPrompt(recentContext, toolPairs as unknown as ToolPair[]);
+
+    let newEntries: OffloadEntry[];
+    const l1LlmStart = Date.now();
+    let l1RawResponse: string | undefined;
+    try {
+      const response = await this.deps.llmClient.chat({
+        model: this.config.l1Model,
+        messages: [
+          { role: "system", content: L1_SYSTEM_PROMPT },
+          { role: "user", content: userPrompt },
+        ],
+        temperature: this.config.l1Temperature,
+        max_tokens: this.config.l1MaxTokens,
+        timeoutMs: this.config.l1TimeoutMs,
+      });
+      l1RawResponse = response;
+      newEntries = parseL1Response(response);
+      traceServerModelIo({
+        sessionId,
+        stage: "L1",
+        model: this.config.l1Model,
+        systemPrompt: L1_SYSTEM_PROMPT,
+        userPrompt,
+        responseContent: response,
+        status: "ok",
+        durationMs: Date.now() - l1LlmStart,
+        logger: this.deps.logger,
+      });
+    } catch (err) {
+      traceServerModelIo({
+        sessionId,
+        stage: "L1",
+        model: this.config.l1Model,
+        systemPrompt: L1_SYSTEM_PROMPT,
+        userPrompt,
+        responseContent: l1RawResponse ?? "",
+        status: "error",
+        errorMessage: String(err),
+        durationMs: Date.now() - l1LlmStart,
+        logger: this.deps.logger,
+      });
+      this.deps.logger.error(`[offload-server] L1 LLM failed:`, err);
+      throw err; // Let Worker retry (processing file remains for re-claim)
+    }
+
+    // 4. Ensure all toolPairs are covered — fallback for missing entries
+    const parsedIds = new Set(newEntries.map((e) => e.tool_call_id));
+    // Build toolCallId → original timestamp map for reliable boundary matching
+    const tpTimestampMap = new Map<string, string>();
+    for (const tp of toolPairs) {
+      const id = tp.toolCallId as string;
+      if (id && tp.timestamp) tpTimestampMap.set(id, tp.timestamp as string);
+    }
+    // Overwrite entry timestamps with original tool pair timestamps (don't trust LLM output)
+    for (const entry of newEntries) {
+      const origTs = tpTimestampMap.get(entry.tool_call_id);
+      if (origTs) entry.timestamp = origTs;
+    }
+    for (const tp of toolPairs) {
+      const id = tp.toolCallId as string;
+      if (id && !parsedIds.has(id)) {
+        newEntries.push({
+          tool_call_id: id,
+          tool_call: (tp.toolName as string) ?? "",
+          summary: "[L1 parse incomplete]",
+          timestamp: (tp.timestamp as string) ?? new Date().toISOString(),
+          score: 2,
+          node_id: null,
+        });
+      }
+    }
+
+    // 5. Write refs: store original tool result content for each pair
+    for (const tp of toolPairs) {
+      const id = tp.toolCallId as string;
+      if (!id) continue;
+      const result = tp.result ?? tp.error ?? "";
+      const resultStr = typeof result === "string" ? result : JSON.stringify(result, null, 2);
+      if (!resultStr || resultStr.length < 20) continue; // skip trivially short results
+      const toolName = (tp.toolName as string) ?? "unknown";
+      const timestamp = (tp.timestamp as string) ?? new Date().toISOString();
+      const header = `# Tool Result: ${toolName}\n\n**tool_call_id:** ${id}\n**Timestamp:** ${timestamp}\n\n---\n\n`;
+      const refPath = `${basePath}/refs/${id}.md`;
+      await storage.writeFile(refPath, header + resultStr);
+      // Set result_ref on the matching entry (full relative path for tdai_read_cos)
+      const entry = newEntries.find((e) => e.tool_call_id === id);
+      if (entry) entry.result_ref = `${basePath}/refs/${id}.md`;
+    }
+
+    // 6. Write entries (atomic append) + delete processing file
+    await storage.appendFile(entriesPath, serializeJsonl(newEntries));
+    await storage.unlink(processingPath);
+
+    // 7. Check L2 trigger: group null entries by boundary targetMmd
+    const state = await this.readState(storage, basePath);
+    const allEntriesRaw = await storage.readFile(entriesPath);
+    if (allEntriesRaw && state.boundaries.length > 0) {
+      const allEntries = parseJsonl<OffloadEntry>(allEntriesRaw, (line, err) => {
+        this.deps.logger.warn(`[offload-server] L1: bad JSONL line in entries: ${line}`, err);
+      });
+      const nodeMapping = await this.readNodeMapping(storage, basePath);
+
+      // Group null entries by resolved targetMmd
+      const nullByMmd = new Map<string, number>();
+      for (const e of allEntries) {
+        if (this.getEffectiveNodeId(e, nodeMapping) !== null || !e.timestamp) continue;
+        const boundary = this.findBoundaryByTimestamp(state.boundaries, e.timestamp);
+        if (!boundary) continue; // no boundary → ignore
+        if (boundary.targetMmd === "_pending" || !boundary.targetMmd) continue; // pending → ignore
+        nullByMmd.set(boundary.targetMmd, (nullByMmd.get(boundary.targetMmd) ?? 0) + 1);
+      }
+
+      // Trigger L2 for each MMD that reached threshold
+      for (const [mmdFile, count] of nullByMmd) {
+        if (count >= this.config.l2NullThreshold) {
+          // Threshold met: short delay (1s) so concurrent L1s merge into one L2
+          await this.deps.stateBackend.setTimerIfEarlier(
+            task.instanceId,
+            `offload-l2:${task.instanceId}:${sessionId}:${mmdFile}`,
+            Date.now() + 1_000,
+          );
+          this.deps.logger.info(`[offload-server] L2 timer set (fast, mmd=${mmdFile}, nullCount=${count})`);
+        } else if (count > 0) {
+          await this.deps.stateBackend.setTimerIfEarlier(
+            task.instanceId,
+            `offload-l2:${task.instanceId}:${sessionId}:${mmdFile}`,
+            Date.now() + 30_000,
+          );
+          this.deps.logger.info(`[offload-server] L2 timer set (mmd=${mmdFile}, nullCount=${count})`);
+        }
+      }
+    }
+
+    this.deps.logger.info(
+      `[offload-server] L1 complete: ${newEntries.length} entries produced (${Date.now() - startMs}ms)`,
+    );
+  }
+
+  // ═══════════════════════════════════════════════════════════════════════════
+  // L1.5: Task judgment — triggered by user new message
+  // ═══════════════════════════════════════════════════════════════════════════
+
+  async executeOffloadL15(task: TaskPayload, _signal?: AbortSignal): Promise<void> {
+    const startMs = Date.now();
+    const { sessionId, recentMessages, boundaryTimestamp } = task.data as {
+      sessionId: string;
+      recentMessages?: string;
+      boundaryTimestamp: string;
+    };
+    const storage = await this.resolveStorageOrThrow(task.instanceId);
+    const basePath = buildOffloadBasePath(sessionId);
+
+    // ─── Phase 1: Read-only snapshot + LLM call (NO LOCK) ─────────────────
+    // Multiple L1.5 tasks can execute this phase concurrently. Each operates
+    // on its own boundary and the LLM call (3-10s) does not block others.
+
+    // 1. Read current active MMD (snapshot for prompt building)
+    const preState = await this.readState(storage, basePath);
+    let currentMmd: { filename: string; content: string } | null = null;
+    if (preState.activeMmdFile) {
+      const content = await storage.readFile(`${basePath}/mmds/${preState.activeMmdFile}`);
+      if (content) {
+        currentMmd = { filename: preState.activeMmdFile, content };
+      }
+    }
+
+    // 2. List available MMDs → extract metas
+    const mmdsPrefix = `${basePath}/mmds/`;
+    const mmdFiles = await storage.readdirNames(mmdsPrefix, ".mmd");
+    const metas: MmdMeta[] = [];
+    for (const f of mmdFiles) {
+      const content = await storage.readFile(`${mmdsPrefix}${f}`);
+      if (content) {
+        metas.push(extractMmdMeta(f, content));
+      }
+    }
+
+    // 3. Build prompt & call LLM (no lock held — this is the expensive part)
+    const userPrompt = buildL15UserPrompt(
+      recentMessages ?? "",
+      currentMmd,
+      metas,
+    );
+
+    let judgment;
+    let rawResponse: string | undefined;
+    const l15LlmStart = Date.now();
+    try {
+      const response = await this.deps.llmClient.chat({
+        model: this.config.l15Model,
+        messages: [
+          { role: "system", content: L15_SYSTEM_PROMPT },
+          { role: "user", content: userPrompt },
+        ],
+        temperature: this.config.l15Temperature,
+        max_tokens: this.config.l15MaxTokens,
+        timeoutMs: this.config.l15TimeoutMs,
+      });
+      rawResponse = response;
+      judgment = parseL15Response(response);
+      traceServerModelIo({
+        sessionId,
+        stage: "L1.5",
+        model: this.config.l15Model,
+        systemPrompt: L15_SYSTEM_PROMPT,
+        userPrompt,
+        responseContent: response,
+        status: "ok",
+        durationMs: Date.now() - l15LlmStart,
+        logger: this.deps.logger,
+      });
+    } catch (err) {
+      traceServerModelIo({
+        sessionId,
+        stage: "L1.5",
+        model: this.config.l15Model,
+        systemPrompt: L15_SYSTEM_PROMPT,
+        userPrompt,
+        responseContent: rawResponse ?? "",
+        status: "error",
+        errorMessage: String(err),
+        durationMs: Date.now() - l15LlmStart,
+        logger: this.deps.logger,
+      });
+      this.deps.logger.error(`[offload-server] L1.5 LLM failed:`, err);
+      throw err;
+    }
+
+    if (!judgment) {
+      this.deps.logger.warn(`[offload-server] L1.5: null response (parse failed), raw=${rawResponse?.slice(0, 500) ?? "(empty)"}`);
+      return;
+    }
+
+    this.deps.logger.info(
+      `[offload-server] L1.5: completed=${judgment.taskCompleted}, long=${judgment.isLongTask}, cont=${judgment.isContinuation}, label=${judgment.newTaskLabel ?? "none"} (LLM=${Date.now() - l15LlmStart}ms)`,
+    );
+
+    // Trace L1.5 decision
+    traceServerTaskDecision({
+      sessionId,
+      judgment: judgment as unknown as Record<string, unknown>,
+      durationMs: Date.now() - startMs,
+      logger: this.deps.logger,
+    });
+
+    // ─── Phase 2: Write phase (SHORT LOCK) ────────────────────────────────
+    // Only lock during state.json read-modify-write. Lock TTL is short (10s)
+    // because this phase only does file I/O (no LLM calls).
+    const lockKey = `offload-state:${task.instanceId}:${sessionId}`;
+    const lockOwner = task.id;
+    let locked = false;
+    for (let attempt = 0; attempt < 10; attempt++) {
+      locked = await this.deps.stateBackend.acquireLock(lockKey, lockOwner, 10_000);
+      if (locked) break;
+      await new Promise((r) => setTimeout(r, 100 + attempt * 50));
+    }
+    if (!locked) {
+      this.deps.logger.warn(`[offload-server] L1.5: failed to acquire write lock after 10 attempts (ts=${boundaryTimestamp}), skipping`);
+      return;
+    }
+
+    try {
+      // Re-read state (fresh, may have been modified by another L1.5 that finished earlier)
+      const state = await this.readState(storage, basePath);
+
+      // CAS check: find boundary by timestamp
+      const boundaryIdx = state.boundaries.findIndex((b) => b.timestamp === boundaryTimestamp);
+      if (boundaryIdx < 0) {
+        this.deps.logger.warn(`[offload-server] L1.5: boundary not found for ts=${boundaryTimestamp}, skipping`);
+        return;
+      }
+
+      // CAS check: already backfilled by another concurrent L1.5? Skip.
+      if (state.boundaries[boundaryIdx].targetMmd !== "_pending") {
+        this.deps.logger.info(`[offload-server] L1.5: boundary[ts=${boundaryTimestamp}] already backfilled, skipping`);
+        return;
+      }
+
+      // Apply task transition
+      await handleTaskTransition(state, judgment, storage, basePath);
+
+      // Backfill boundary targetMmd
+      state.boundaries[boundaryIdx].targetMmd = state.activeMmdFile;
+
+      // Update state
+      await this.writeState(storage, basePath, state);
+
+      // Check L2 trigger: scan ALL null entries grouped by targetMmd.
+      // This covers the case where L1 completed before L1.5 backfill —
+      // L1 skipped entries with _pending boundaries, so L1.5 must pick them up.
+      const entriesRaw = await storage.readFile(`${basePath}/entries.jsonl`);
+      if (entriesRaw) {
+        const allEntries = parseJsonl<OffloadEntry>(entriesRaw);
+        const l15NodeMapping = await this.readNodeMapping(storage, basePath);
+
+        // Group null entries by their resolved targetMmd (using all boundaries, not just current)
+        const nullByMmd = new Map<string, number>();
+        for (const e of allEntries) {
+          if (this.getEffectiveNodeId(e, l15NodeMapping) !== null || !e.timestamp) continue;
+          const boundary = this.findBoundaryByTimestamp(state.boundaries, e.timestamp);
+          if (!boundary) continue;
+          if (boundary.targetMmd === "_pending" || !boundary.targetMmd) continue;
+          nullByMmd.set(boundary.targetMmd, (nullByMmd.get(boundary.targetMmd) ?? 0) + 1);
+        }
+
+        for (const [mmdFile, nullCount] of nullByMmd) {
+          if (nullCount >= this.config.l2NullThreshold) {
+            await this.deps.stateBackend.setTimerIfEarlier(
+              task.instanceId,
+              `offload-l2:${task.instanceId}:${sessionId}:${mmdFile}`,
+              Date.now() + 1_000,
+            );
+            this.deps.logger.info(`[offload-server] L1.5: L2 timer set (fast, mmd=${mmdFile}, nullCount=${nullCount})`);
+          } else if (nullCount > 0) {
+            await this.deps.stateBackend.setTimerIfEarlier(
+              task.instanceId,
+              `offload-l2:${task.instanceId}:${sessionId}:${mmdFile}`,
+              Date.now() + 30_000,
+            );
+            this.deps.logger.info(`[offload-server] L1.5: L2 timer set after backfill (mmd=${mmdFile}, nullCount=${nullCount})`);
+          }
+        }
+      }
+    } finally {
+      if (locked) {
+        await this.deps.stateBackend.releaseLock(lockKey, lockOwner);
+      }
+    }
+
+    this.deps.logger.info(
+      `[offload-server] L1.5 complete: boundary[ts=${boundaryTimestamp}] (total=${Date.now() - startMs}ms)`,
+    );
+  }
+
+  // ═══════════════════════════════════════════════════════════════════════════
+  // L2: MMD generation / update + node_id backfill
+  // ═══════════════════════════════════════════════════════════════════════════
+
+  async executeOffloadL2(task: TaskPayload, _signal?: AbortSignal): Promise<void> {
+    const startMs = Date.now();
+    const sessionId = this.extractSessionId(task);
+    if (!sessionId) return;
+
+    let targetMmdFile: string | undefined;
+    const data = task.data as Record<string, unknown>;
+    targetMmdFile = data.targetMmdFile as string | undefined;
+
+    // Extract targetMmdFile from timer member if available
+    // Format: "offload-l2:{instanceId}:{sessionId}:{mmdFile}" → extract last segment ending in .mmd
+    if (!targetMmdFile && data.timerMember) {
+      const timerMember = data.timerMember as string;
+      const mmdMatch = timerMember.match(/(\d+-[^:]+\.mmd)$/);
+      if (mmdMatch) {
+        targetMmdFile = mmdMatch[1];
+      }
+    }
+
+    const storage = await this.resolveStorageOrThrow(task.instanceId);
+    const basePath = buildOffloadBasePath(sessionId);
+    const state = await this.readState(storage, basePath);
+
+    // Resolve targetMmdFile: from task data, timer member, or state (fallback)
+    if (!targetMmdFile) {
+      targetMmdFile = state.activeMmdFile ?? undefined;
+    }
+    if (!targetMmdFile) return;
+
+    // 1. Read all entries + node mapping
+    const entriesRaw = await storage.readFile(`${basePath}/entries.jsonl`);
+    if (!entriesRaw) return;
+    const allEntries = parseJsonl<OffloadEntry>(entriesRaw);
+    const nodeMapping = await this.readNodeMapping(storage, basePath);
+
+    // 2. Filter: entries belonging to targetMmdFile with null node_id (after join)
+    const relevantEntries = allEntries.filter((e) => {
+      if (this.getEffectiveNodeId(e, nodeMapping) !== null) return false;
+      if (!e.timestamp) return false;
+      const boundary = this.findBoundaryByTimestamp(state.boundaries, e.timestamp);
+      return boundary?.targetMmd === targetMmdFile;
+    });
+
+    if (relevantEntries.length === 0) return;
+
+    // Guard: if oldest null entry is older than 10 minutes, give up and assign fallback node_id
+    // to prevent infinite L2 retries when LLM consistently fails to map certain entries.
+    const L2_MAX_AGE_MS = 10 * 60 * 1000;
+    const oldestTs = relevantEntries.reduce((min, e) => {
+      const t = new Date(e.timestamp).getTime();
+      return t < min ? t : min;
+    }, Infinity);
+    if (Date.now() - oldestTs > L2_MAX_AGE_MS) {
+      this.deps.logger.warn(
+        `[offload-server] L2: ${relevantEntries.length} entries exceeded max age (10min), assigning fallback node_id`,
+      );
+      const fallbackMappings = relevantEntries.map((e) => ({
+        tool_call_id: e.tool_call_id,
+        node_id: `${targetMmdFile!.replace(/\.mmd$/, "")}-orphan`,
+      }));
+      const mappingPath = `${basePath}/node-mapping.jsonl`;
+      await storage.appendFile(mappingPath, serializeJsonl(fallbackMappings));
+      return;
+    }
+
+    // 3. Read existing MMD
+    const mmdPath = `${basePath}/mmds/${targetMmdFile}`;
+    const existingMmd = await storage.readFile(mmdPath);
+
+    // 4. Build prompt (include recent context for better MMD generation)
+    const recentHistory = await storage.readFile(`${basePath}/recent-context.txt`) ?? null;
+    const taskLabel = targetMmdFile.replace(/^\d+-/, "").replace(/\.mmd$/, "") || "task";
+    const prefixMatch = targetMmdFile.match(/^(\d+)-/);
+    const mmdPrefix = prefixMatch ? prefixMatch[1] : "000";
+    const charCount = existingMmd?.length ?? 0;
+
+    const userPrompt = buildL2UserPrompt({
+      existingMmd: existingMmd || null,
+      entries: relevantEntries,
+      recentHistory,
+      taskLabel,
+      mmdPrefix,
+      charCount,
+    });
+
+    // 5. Call LLM
+    let result: L2ParsedResponse | null;
+    let rawL2Response: string | undefined;
+    const l2LlmStart = Date.now();
+    try {
+      const response = await this.deps.llmClient.chat({
+        model: this.config.l2Model,
+        messages: [
+          { role: "system", content: L2_SYSTEM_PROMPT },
+          { role: "user", content: userPrompt },
+        ],
+        temperature: this.config.l2Temperature,
+        max_tokens: this.config.l2MaxTokens,
+        timeoutMs: this.config.l2TimeoutMs,
+      });
+      rawL2Response = response;
+      result = parseL2Response(response);
+      traceServerModelIo({
+        sessionId: sessionId!,
+        stage: "L2",
+        model: this.config.l2Model,
+        systemPrompt: L2_SYSTEM_PROMPT,
+        userPrompt,
+        responseContent: response,
+        status: "ok",
+        durationMs: Date.now() - l2LlmStart,
+        logger: this.deps.logger,
+      });
+    } catch (err) {
+      traceServerModelIo({
+        sessionId: sessionId!,
+        stage: "L2",
+        model: this.config.l2Model,
+        systemPrompt: L2_SYSTEM_PROMPT,
+        userPrompt,
+        responseContent: rawL2Response ?? "",
+        status: "error",
+        errorMessage: String(err),
+        durationMs: Date.now() - l2LlmStart,
+        logger: this.deps.logger,
+      });
+      this.deps.logger.error(`[offload-server] L2 LLM failed:`, err);
+      throw err;
+    }
+
+    if (!result) {
+      this.deps.logger.warn(`[offload-server] L2: parse failed, raw=${rawL2Response?.slice(0, 500) ?? "(empty)"}`);
+      // Schedule retry so remaining null entries are not orphaned
+      await this.deps.stateBackend.setTimerIfEarlier(
+        task.instanceId,
+        `offload-l2:${task.instanceId}:${sessionId}:${targetMmdFile}`,
+        Date.now() + 30_000,
+      );
+      return;
+    }
+
+    // 6. Apply MMD update
+    const updatedMmd = this.applyL2Result(existingMmd ?? "", result);
+    await storage.writeFile(mmdPath, updatedMmd);
+
+    // 7. Backfill: write node mappings to separate file (avoids overwriting entries.jsonl)
+    const mappingEntries = Object.entries(result.nodeMapping).map(([toolCallId, nodeId]) => ({
+      tool_call_id: toolCallId,
+      node_id: nodeId,
+    }));
+    if (mappingEntries.length > 0) {
+      const mappingPath = `${basePath}/node-mapping.jsonl`;
+      await storage.appendFile(mappingPath, serializeJsonl(mappingEntries));
+    }
+
+    // 8. Check remaining null entries for this MMD — re-read node-mapping for accurate count
+    //    (another L2 may have concurrently written mappings, or LLM may have missed some)
+    const freshNodeMapping = await this.readNodeMapping(storage, basePath);
+    const remainingNull = relevantEntries.filter(
+      (e) => !freshNodeMapping.has(e.tool_call_id) && !result!.nodeMapping[e.tool_call_id],
+    ).length;
+    if (remainingNull > 0) {
+      await this.deps.stateBackend.setTimerIfEarlier(
+        task.instanceId,
+        `offload-l2:${task.instanceId}:${sessionId}:${targetMmdFile}`,
+        Date.now() + 30_000,
+      );
+      this.deps.logger.info(`[offload-server] L2 retry timer set (mmd=${targetMmdFile}, remainingNull=${remainingNull})`);
+    }
+
+    this.deps.logger.info(
+      `[offload-server] L2 complete: ${Object.keys(result.nodeMapping).length} entries mapped, action=${result.fileAction} (${Date.now() - startMs}ms)`,
+    );
+  }
+
+  // ═══════════════════════════════════════════════════════════════════════════
+  // Helpers
+  // ═══════════════════════════════════════════════════════════════════════════
+
+  /**
+   * Extract sessionId from task data or task.sessionId.
+   */
+  private extractSessionId(task: TaskPayload): string | undefined {
+    const data = task.data as Record<string, unknown> | undefined;
+    // Prefer explicit sessionId in data
+    if (data?.sessionId && typeof data.sessionId === "string") {
+      return data.sessionId;
+    }
+    // Fallback: task-level sessionId
+    if (task.sessionId) {
+      return task.sessionId;
+    }
+    return undefined;
+  }
+
+  private async resolveStorageOrThrow(instanceId: string): Promise<StorageAdapter> {
+    const storage = await this.deps.resolveStorage(instanceId);
+    if (!storage) throw new Error(`Storage unavailable for instance ${instanceId}`);
+    return storage;
+  }
+
+  private async readState(storage: StorageAdapter, basePath: string): Promise<OffloadState> {
+    const raw = await storage.readFile(`${basePath}/state.json`);
+    if (!raw) return defaultOffloadState();
+    try {
+      return { ...defaultOffloadState(), ...JSON.parse(raw) };
+    } catch {
+      return defaultOffloadState();
+    }
+  }
+
+  private async writeState(storage: StorageAdapter, basePath: string, state: OffloadState): Promise<void> {
+    await storage.writeFile(`${basePath}/state.json`, JSON.stringify(state));
+  }
+
+  private findBoundaryByTimestamp(
+    boundaries: OffloadState["boundaries"],
+    entryTimestamp: string,
+  ): OffloadState["boundaries"][number] | null {
+    if (boundaries.length === 0) return null;
+    // Find the last boundary whose timestamp <= entryTimestamp
+    let result: OffloadState["boundaries"][number] | null = null;
+    for (const b of boundaries) {
+      if (b.timestamp <= entryTimestamp) result = b;
+      else break;
+    }
+    return result;
+  }
+
+  private applyL2Result(existingMmd: string, result: L2ParsedResponse): string {
+    if (result.fileAction === "write" && result.mmdContent) {
+      return result.mmdContent;
+    }
+
+    if (result.fileAction === "replace" && result.replaceBlocks?.length) {
+      const lines = existingMmd.split("\n");
+      // Sort blocks by startLine descending to avoid offset issues
+      const sorted = [...result.replaceBlocks]
+        .filter((block) => block.startLine >= 1 && block.endLine >= block.startLine && block.startLine <= lines.length)
+        .sort((a, b) => b.startLine - a.startLine);
+      for (const block of sorted) {
+        const start = block.startLine - 1; // 0-based
+        const end = Math.min(block.endLine, lines.length); // clamp endLine
+        const deleteCount = end - block.startLine + 1;
+        const newLines = block.content.split("\n");
+        lines.splice(start, deleteCount, ...newLines);
+      }
+      return lines.join("\n");
+    }
+
+    return existingMmd;
+  }
+
+  /**
+   * Read node-mapping.jsonl and build a tool_call_id → node_id map.
+   */
+  private async readNodeMapping(storage: StorageAdapter, basePath: string): Promise<Map<string, string>> {
+    const raw = await storage.readFile(`${basePath}/node-mapping.jsonl`);
+    const map = new Map<string, string>();
+    if (!raw) return map;
+    const lines = parseJsonl<{ tool_call_id: string; node_id: string }>(raw);
+    for (const line of lines) {
+      if (line.tool_call_id && line.node_id) {
+        map.set(line.tool_call_id, line.node_id);
+      }
+    }
+    return map;
+  }
+
+  /**
+   * Get effective node_id for an entry: check node-mapping first, then entry's own field.
+   */
+  private getEffectiveNodeId(entry: OffloadEntry, nodeMapping: Map<string, string>): string | null {
+    return nodeMapping.get(entry.tool_call_id) ?? entry.node_id;
+  }
+}

src/offload_server/opik-tracer.ts

@@ -0,0 +1,336 @@
+/**
+ * Opik observability tracer for offload server.
+ * Wraps the opik npm package with graceful degradation when not installed.
+ * Configuration is read from environment variables (no OpenClaw plugin config).
+ */
+
+// Opik client types (minimal shape to avoid hard dependency)
+interface OpikClient {
+  trace(params: Record<string, unknown>): OpikTrace;
+  flush(): Promise<void>;
+}
+interface OpikTrace {
+  update(params: Record<string, unknown>): void;
+  end(): void;
+  span(params: Record<string, unknown>): OpikSpan;
+}
+interface OpikSpan {
+  update(params: Record<string, unknown>): void;
+  end(): void;
+}
+
+// ─── Module State ────────────────────────────────────────────────────────────
+
+let client: OpikClient | null = null;
+let tracerEnabled = false;
+let tracerInitTried = false;
+
+// ─── Helpers ─────────────────────────────────────────────────────────────────
+
+function extractLayerTag(stage: string): string {
+  const match = stage.match(/^(L\d+(?:\.\d+)?)/i);
+  if (!match) return "Lx-unknown";
+  return match[1].toUpperCase();
+}
+
+function durationBucketTag(ms: number): string {
+  if (typeof ms !== "number" || ms < 0) return "duration:unknown";
+  if (ms < 1000) return "duration:<1s";
+  if (ms < 5000) return "duration:1-5s";
+  if (ms < 15000) return "duration:5-15s";
+  if (ms < 30000) return "duration:15-30s";
+  return "duration:>30s";
+}
+
+function formatDuration(ms: number): string {
+  if (typeof ms !== "number" || ms < 0) return "?";
+  if (ms < 1000) return `${Math.round(ms)}ms`;
+  return `${(ms / 1000).toFixed(2)}s`;
+}
+
+// ─── Logger Interface ────────────────────────────────────────────────────────
+
+export interface TracerLogger {
+  info: (message: string, ...args: unknown[]) => void;
+  warn: (message: string, ...args: unknown[]) => void;
+  debug?: (message: string, ...args: unknown[]) => void;
+}
+
+// ─── Init ────────────────────────────────────────────────────────────────────
+
+/**
+ * Initialize the offload server Opik tracer.
+ * Reads config from environment variables:
+ *   OPIK_ENABLED        — set to "true" to enable (default: disabled)
+ *   OPIK_URL_OVERRIDE   — Opik server URL
+ *   OPIK_API_KEY        — API key
+ *   OPIK_WORKSPACE      — workspace name (default: "default")
+ *   OPIK_PROJECT_NAME   — project name (default: "openclaw-offload-server")
+ */
+export async function initServerOpikTracer(logger: TracerLogger): Promise<void> {
+  if (tracerInitTried) return;
+  tracerInitTried = true;
+  try {
+    const enabled = process.env.OPIK_ENABLED === "true";
+    if (!enabled) {
+      logger.debug?.("[offload-server] Opik tracer disabled (OPIK_ENABLED != true)");
+      return;
+    }
+
+    const apiUrl = process.env.OPIK_URL_OVERRIDE;
+    const apiKey = process.env.OPIK_API_KEY;
+    const workspaceName = process.env.OPIK_WORKSPACE ?? "default";
+    const projectName = process.env.OPIK_PROJECT_NAME ?? "openclaw-offload-server";
+
+    // Dynamic import — graceful when opik is not installed
+    let OpikConstructor: new (params: Record<string, unknown>) => OpikClient;
+    let disableOpikLogger: (() => void) | undefined;
+    try {
+      const opikModule = await import("opik") as {
+        Opik: new (params: Record<string, unknown>) => OpikClient;
+        disableLogger?: () => void;
+        setLoggerLevel?: (level: string) => void;
+      };
+      OpikConstructor = opikModule.Opik;
+      disableOpikLogger = opikModule.disableLogger;
+    } catch {
+      logger.debug?.("[offload-server] opik package not available, tracer disabled");
+      return;
+    }
+
+    // Suppress opik internal logs (flush messages, ANSI color noise)
+    if (disableOpikLogger) {
+      disableOpikLogger();
+    }
+
+    client = new OpikConstructor({
+      ...(apiKey ? { apiKey } : {}),
+      ...(apiUrl ? { apiUrl } : {}),
+      workspaceName,
+      projectName,
+    });
+    tracerEnabled = true;
+    logger.info(
+      `[offload-server] Opik tracer enabled: project=${projectName}, workspace=${workspaceName}`,
+    );
+  } catch (err) {
+    tracerEnabled = false;
+    client = null;
+    logger.warn(`[offload-server] Opik tracer init failed: ${String(err)}`);
+  }
+}
+
+/**
+ * Check if the tracer is enabled and ready to trace.
+ */
+export function isTracerEnabled(): boolean {
+  return tracerEnabled && client !== null;
+}
+
+// ─── Trace: Model I/O (L1/L1.5/L2 LLM calls) ───────────────────────────────
+
+/**
+ * Trace LLM model I/O for offload server L1/L1.5/L2 stages.
+ */
+export function traceServerModelIo(params: {
+  sessionId: string;
+  stage: string;
+  model: string;
+  systemPrompt: string;
+  userPrompt: string;
+  responseContent: string;
+  usage?: { promptTokens?: number; completionTokens?: number; totalTokens?: number };
+  status: "ok" | "error";
+  errorMessage?: string;
+  durationMs: number;
+  logger?: TracerLogger;
+}): void {
+  if (!tracerEnabled || !client) return;
+  try {
+    const layerTag = extractLayerTag(params.stage);
+    const threadId = params.sessionId || `offload-server-${Date.now()}`;
+    const dur = params.durationMs;
+    const durStr = formatDuration(dur);
+    const durBucket = durationBucketTag(dur);
+    const skTag = `session:${params.sessionId || "unknown"}`;
+
+    const trace = client.trace({
+      name: `${params.model} · offload-server · ${params.stage} · ${durStr}`,
+      threadId,
+      metadata: {
+        plugin: "openclaw-offload-server",
+        category: "llm",
+        stage: params.stage,
+        layer: layerTag,
+        model: params.model,
+        sessionId: params.sessionId,
+        durationMs: dur,
+        duration: durStr,
+      },
+      tags: ["offload-server", "llm", layerTag, durBucket, skTag],
+    });
+
+    const span = trace.span({
+      name: `${params.model} · ${params.stage} · ${durStr}`,
+      type: "llm",
+      model: params.model,
+      input: {
+        systemPrompt: params.systemPrompt,
+        userPrompt: params.userPrompt,
+      },
+      metadata: {
+        stage: params.stage,
+        layer: layerTag,
+        sessionId: params.sessionId,
+        durationMs: dur,
+        duration: durStr,
+      },
+    });
+
+    span.update({
+      output: {
+        responseContent: params.responseContent,
+        usage: params.usage,
+        durationMs: dur,
+        duration: durStr,
+        error: params.errorMessage,
+      },
+      metadata: {
+        status: params.status,
+        durationMs: dur,
+      },
+    });
+    span.end();
+    trace.end();
+    void client.flush().catch(() => undefined);
+  } catch (err) {
+    params.logger?.warn?.(`[offload-server] Opik model I/O trace failed: ${String(err)}`);
+  }
+}
+
+// ─── Trace: Compaction Decision ──────────────────────────────────────────────
+
+/**
+ * Trace compaction (L3) decision and results.
+ */
+export function traceServerCompaction(params: {
+  sessionId: string;
+  level: string;
+  ratio: number;
+  contextWindow: number;
+  totalTokensBefore: number;
+  totalTokensAfter: number;
+  originalMsgCount: number;
+  compactedMsgCount: number;
+  report: Record<string, unknown>;
+  messages: unknown[];
+  durationMs: number;
+  logger?: TracerLogger;
+}): void {
+  if (!tracerEnabled || !client) return;
+  try {
+    const threadId = params.sessionId || `offload-server-${Date.now()}`;
+    const dur = params.durationMs;
+    const durStr = formatDuration(dur);
+    const durBucket = durationBucketTag(dur);
+    const skTag = `session:${params.sessionId || "unknown"}`;
+
+    const trace = client.trace({
+      name: `compaction · L3 · ${params.level} · ${durStr} [${params.sessionId}]`,
+      threadId,
+      input: {
+        level: params.level,
+        ratio: params.ratio,
+        contextWindow: params.contextWindow,
+        totalTokensBefore: params.totalTokensBefore,
+        originalMsgCount: params.originalMsgCount,
+      },
+      metadata: {
+        plugin: "openclaw-offload-server",
+        category: "compaction",
+        stage: "L3",
+        layer: "L3",
+        level: params.level,
+        sessionId: params.sessionId,
+        durationMs: dur,
+        duration: durStr,
+      },
+      tags: ["offload-server", "compaction", "L3", `level:${params.level}`, durBucket, skTag],
+    });
+
+    // Serialize messages for full snapshot
+    const serializedMessages = params.messages.map((msg: any, i: number) => {
+      const role = msg.role ?? "unknown";
+      const content = typeof msg.content === "string"
+        ? msg.content
+        : Array.isArray(msg.content)
+          ? msg.content.map((c: any) => {
+              if (c.type === "text") return c.text;
+              if (c.type === "tool_use") return `[tool_use: ${c.name} id=${c.id}]`;
+              if (c.type === "tool_result") return `[tool_result: id=${c.tool_use_id} content=${typeof c.content === "string" ? c.content.slice(0, 500) : JSON.stringify(c.content).slice(0, 500)}]`;
+              return `[${c.type ?? "unknown"}]`;
+            }).join("\n")
+          : "";
+      return { i, role, content, ...(msg._mmdContextMessage ? { mmdCtx: true } : {}), ...(msg._offloaded ? { offloaded: true } : {}) };
+    });
+
+    trace.update({
+      output: {
+        totalTokensAfter: params.totalTokensAfter,
+        compactedMsgCount: params.compactedMsgCount,
+        tokenReduction: params.totalTokensBefore - params.totalTokensAfter,
+        report: params.report,
+        messages: serializedMessages,
+      },
+    });
+    trace.end();
+    void client.flush().catch(() => undefined);
+  } catch (err) {
+    params.logger?.warn?.(`[offload-server] Opik compaction trace failed: ${String(err)}`);
+  }
+}
+
+// ─── Trace: Task Decision (L1.5 judgment result) ─────────────────────────────
+
+/**
+ * Trace L1.5 task judgment decision.
+ */
+export function traceServerTaskDecision(params: {
+  sessionId: string;
+  judgment: Record<string, unknown>;
+  durationMs: number;
+  logger?: TracerLogger;
+}): void {
+  if (!tracerEnabled || !client) return;
+  try {
+    const threadId = params.sessionId || `offload-server-${Date.now()}`;
+    const dur = params.durationMs;
+    const durStr = formatDuration(dur);
+    const skTag = `session:${params.sessionId || "unknown"}`;
+
+    const trace = client.trace({
+      name: `task-decision · L1.5 · ${durStr} [${params.sessionId}]`,
+      threadId,
+      input: {
+        stage: "L1.5",
+        sessionId: params.sessionId,
+      },
+      metadata: {
+        plugin: "openclaw-offload-server",
+        category: "decision",
+        stage: "L1.5",
+        layer: "L1.5",
+        sessionId: params.sessionId,
+        durationMs: dur,
+        duration: durStr,
+      },
+      tags: ["offload-server", "decision", "L1.5", skTag],
+    });
+
+    trace.update({ output: params.judgment });
+    trace.end();
+    void client.flush().catch(() => undefined);
+  } catch (err) {
+    params.logger?.warn?.(`[offload-server] Opik task decision trace failed: ${String(err)}`);
+  }
+}

src/offload_server/parsers/json-utils.ts

@@ -0,0 +1,92 @@
+/**
+ * JSON / Mermaid extraction utilities for LLM response parsing.
+ */
+
+/**
+ * Extract JSON from raw LLM output. Tolerates markdown fences, extra text.
+ */
+export function extractJson<T>(raw: string): T | null {
+  if (!raw || typeof raw !== "string") return null;
+
+  const trimmed = raw.trim();
+
+  // 1. Direct parse
+  try {
+    return JSON.parse(trimmed) as T;
+  } catch {
+    // continue
+  }
+
+  // 2. Extract from ```json ... ``` fence
+  const jsonFenceMatch = trimmed.match(/```(?:json)?\s*\n?([\s\S]*?)```/);
+  if (jsonFenceMatch) {
+    try {
+      return JSON.parse(jsonFenceMatch[1].trim()) as T;
+    } catch {
+      // continue
+    }
+  }
+
+  // 3. Extract first { ... } (greedy last })
+  const objStart = trimmed.indexOf("{");
+  const objEnd = trimmed.lastIndexOf("}");
+  if (objStart !== -1 && objEnd > objStart) {
+    try {
+      return JSON.parse(trimmed.slice(objStart, objEnd + 1)) as T;
+    } catch {
+      // continue
+    }
+  }
+
+  // 4. Extract first [ ... ] (greedy last ])
+  const arrStart = trimmed.indexOf("[");
+  const arrEnd = trimmed.lastIndexOf("]");
+  if (arrStart !== -1 && arrEnd > arrStart) {
+    try {
+      return JSON.parse(trimmed.slice(arrStart, arrEnd + 1)) as T;
+    } catch {
+      // continue
+    }
+  }
+
+  return null;
+}
+
+/**
+ * Extract mermaid content from a ```mermaid ... ``` code fence.
+ */
+export function extractMermaidFromFence(raw: string): string | null {
+  if (!raw) return null;
+  const match = raw.match(/```mermaid\s*\n?([\s\S]*?)```/);
+  return match ? match[1].trim() : null;
+}
+
+/**
+ * Parse JSONL (newline-delimited JSON) string into array.
+ * Corrupted lines are silently skipped (logged via optional callback).
+ */
+export function parseJsonl<T>(
+  content: string,
+  onBadLine?: (line: string, error: unknown) => void,
+): T[] {
+  if (!content || !content.trim()) return [];
+  const results: T[] = [];
+  for (const line of content.split("\n")) {
+    const trimmed = line.trim();
+    if (trimmed.length === 0) continue;
+    try {
+      results.push(JSON.parse(trimmed) as T);
+    } catch (err) {
+      onBadLine?.(trimmed, err);
+    }
+  }
+  return results;
+}
+
+/**
+ * Serialize array to JSONL string (trailing newline).
+ */
+export function serializeJsonl<T>(items: T[]): string {
+  if (items.length === 0) return "";
+  return items.map((item) => JSON.stringify(item)).join("\n") + "\n";
+}

src/offload_server/parsers/l1-parser.ts

@@ -0,0 +1,41 @@
+/**
+ * L1 Response Parser — extracts OffloadEntry[] from LLM output.
+ */
+import { extractJson } from "./json-utils.js";
+import type { OffloadEntry } from "../types.js";
+
+interface RawL1Entry {
+  tool_call?: string;
+  summary?: string;
+  tool_call_id?: string;
+  timestamp?: string;
+  score?: number;
+}
+
+/**
+ * Parse L1 LLM response into OffloadEntry array.
+ * Tolerant of markdown wrapping, missing fields, etc.
+ */
+export function parseL1Response(raw: string): OffloadEntry[] {
+  const parsed = extractJson<RawL1Entry[]>(raw);
+  if (!parsed || !Array.isArray(parsed)) return [];
+
+  const entries: OffloadEntry[] = [];
+  for (const item of parsed) {
+    if (!item || typeof item !== "object") continue;
+    const toolCallId = item.tool_call_id ?? "";
+    if (!toolCallId) continue;
+
+    entries.push({
+      tool_call_id: toolCallId,
+      tool_call: item.tool_call ?? "",
+      summary: item.summary ?? "",
+      timestamp: item.timestamp ?? "",
+      score: typeof item.score === "number" ? Math.min(10, Math.max(0, item.score)) : 5,
+      node_id: null,
+      seq: -1,
+    });
+  }
+
+  return entries;
+}

src/offload_server/parsers/l15-parser.ts

@@ -0,0 +1,57 @@
+/**
+ * L1.5 Response Parser — extracts TaskJudgment from LLM output.
+ */
+import { extractJson } from "./json-utils.js";
+import type { TaskJudgment } from "../types.js";
+
+interface RawL15Response {
+  taskCompleted?: boolean | null;
+  isContinuation?: boolean | null;
+  isLongTask?: boolean | null;
+  continuationMmdFile?: string | null;
+  newTaskLabel?: string | null;
+}
+
+/**
+ * Parse L1.5 LLM response into TaskJudgment.
+ * Returns null if completely unparseable or all-null (LLM unavailable).
+ */
+export function parseL15Response(raw: string): TaskJudgment | null {
+  const parsed = extractJson<RawL15Response>(raw);
+  if (!parsed || typeof parsed !== "object") return null;
+
+  if (
+    parsed.taskCompleted == null &&
+    parsed.isContinuation == null &&
+    parsed.isLongTask == null
+  ) {
+    return null;
+  }
+
+  return {
+    taskCompleted: toBool(parsed.taskCompleted),
+    isContinuation: toBool(parsed.isContinuation),
+    isLongTask: toBool(parsed.isLongTask),
+    continuationMmdFile:
+      typeof parsed.continuationMmdFile === "string" && isSafeFilename(parsed.continuationMmdFile)
+        ? parsed.continuationMmdFile
+        : undefined,
+    newTaskLabel:
+      typeof parsed.newTaskLabel === "string" ? parsed.newTaskLabel : undefined,
+  };
+}
+
+/** Safely coerce LLM value to boolean, handling string "false"/"0". */
+function toBool(value: unknown): boolean {
+  if (typeof value === "string") {
+    return value.toLowerCase() !== "false" && value !== "0" && value !== "";
+  }
+  return Boolean(value);
+}
+
+/** Validate that a filename is safe (no path traversal or special chars). */
+function isSafeFilename(name: string): boolean {
+  if (!name) return false;
+  if (name.includes("/") || name.includes("\\") || name.includes("..")) return false;
+  return /^[a-zA-Z0-9_.\-]+$/.test(name);
+}

src/offload_server/parsers/l2-parser.ts

@@ -0,0 +1,76 @@
+/**
+ * L2 Response Parser — extracts MMD generation results from LLM output.
+ */
+import { extractJson, extractMermaidFromFence } from "./json-utils.js";
+import type { L2ParsedResponse } from "../types.js";
+
+interface RawL2Response {
+  file_action?: string;
+  mmd_content?: string | null;
+  replace_blocks?: Array<{
+    start_line?: number | string;
+    end_line?: number | string;
+    content?: string;
+  }> | null;
+  node_mapping?: Record<string, string>;
+}
+
+/**
+ * Parse L2 LLM response into structured result.
+ * Returns null if parsing fails completely.
+ */
+export function parseL2Response(raw: string): L2ParsedResponse | null {
+  const parsed = extractJson<RawL2Response>(raw);
+  if (!parsed || typeof parsed !== "object") {
+    // Fallback: try extracting ```mermaid ... ``` code block
+    const mmd = extractMermaidFromFence(raw);
+    if (mmd) {
+      return { fileAction: "write", mmdContent: mmd, nodeMapping: {} };
+    }
+    return null;
+  }
+
+  const fileAction = parsed.file_action === "replace" ? "replace" : "write";
+
+  // Extract mmd_content (may be wrapped in code fence)
+  let mmdContent: string | undefined;
+  if (fileAction === "write") {
+    if (parsed.mmd_content) {
+      mmdContent =
+        extractMermaidFromFence(parsed.mmd_content) ?? parsed.mmd_content;
+    } else {
+      const fallbackMmd = extractMermaidFromFence(raw);
+      if (fallbackMmd) mmdContent = fallbackMmd;
+    }
+  }
+
+  // Parse replace_blocks
+  let replaceBlocks: L2ParsedResponse["replaceBlocks"] | undefined;
+  if (fileAction === "replace" && Array.isArray(parsed.replace_blocks)) {
+    replaceBlocks = [];
+    for (const block of parsed.replace_blocks) {
+      if (!block || typeof block !== "object") continue;
+      const startLine = Number(block.start_line);
+      const endLine = Number(block.end_line);
+      if (isNaN(startLine) || isNaN(endLine)) continue;
+
+      let content = block.content ?? "";
+      const extracted = extractMermaidFromFence(content);
+      if (extracted) content = extracted;
+
+      replaceBlocks.push({ startLine, endLine, content });
+    }
+  }
+
+  // Parse node_mapping
+  const nodeMapping: Record<string, string> = {};
+  if (parsed.node_mapping && typeof parsed.node_mapping === "object") {
+    for (const [key, value] of Object.entries(parsed.node_mapping)) {
+      if (typeof value === "string") {
+        nodeMapping[key] = value;
+      }
+    }
+  }
+
+  return { fileAction, mmdContent, replaceBlocks, nodeMapping };
+}

src/offload_server/prompts/l1-prompt.ts

@@ -0,0 +1,81 @@
+/**
+ * L1 Summarization Prompt — converts ToolPairs into OffloadEntry summaries.
+ */
+import type { ToolPair } from "../types.js";
+
+const PARAMS_MAX_LEN = 500;
+const RESULT_MAX_LEN = 2000;
+const COMPRESS_THRESHOLD = 200;
+
+export const L1_SYSTEM_PROMPT = `你是一个专为 AI 编码助手提供支持的"工具结果摘要器"。你的核心任务是深度理解当前的对话上下文，并将繁杂的工具调用与执行结果（一对toolcall和tool result整合成一条summary输出），提炼为高信息密度的 JSON 数组。
+
+在生成摘要前，请务必进行以下内部思考：
+1. 任务对齐：结合最近的对话记录，识别用户当前的核心目标和最新意图。若上下文存在冲突，始终以最新的用户意图为准。
+2. 价值过滤：忽略工具如何工作的冗余细节，直接提取"发现了什么关键线索"、"做了什么关键动作"、"修改了什么具体内容"或"遇到了什么具体报错"。
+3. 影响评估：判断该结果对当前任务的实质性影响（例如：证实了某个假设、推进了哪一步、做出了什么决策，或因为什么报错导致了阻塞）。
+
+【输出格式要求】
+你必须且只能输出一个合法的 JSON 对象数组 [{...}]，每个对象**必须**包含以下字段：
+- "tool_call": 工具调用的简洁描述。处理规则如下：
+  · 如果输入中该 tool pair 标记了 [NEEDS_COMPRESS]，你必须将工具名+关键参数压缩为一句简洁的描述（≤150字符），保留工具名、操作目标（如文件路径、命令意图），省略内联脚本/大段内容的细节。
+    示例：exec({"command":"python3 -c 'import csv; ...200行脚本...'"}) → "exec: 运行 Python （xx/xx/xx.sh，标明具体路径和文件）脚本分析 sales_channels.csv 数据质量"
+    示例：write_file({"path":"/root/app.py","content":"...5000字符..."}) → "write_file: 写入 /root/app.py (Flask 应用主文件)，大致内容是……"
+  · 如果未标记 [NEEDS_COMPRESS]，直接简述工具与参数即可（系统会用原始值覆盖）。
+- "summary": 融合上述思考的精炼总结（≤200个字符）。必须一针见血地说清楚结果的业务价值，以及它对任务的推进/阻塞作用。
+- "tool_call_id": 原始的 tool_call_id（必须原样透传）。
+- "timestamp": 原始的 ISO 8601 时间戳（必须原样透传）。
+- "score"（**必填**）: 结合信息密度和任务目的分析summary对于原文的可替代性，范围在0-10之间，越接近10表示summary越能替代原文。
+
+【严格规则】
+只允许输出纯 JSON 数组，严禁输出思考过程或其他解释性文本。`;
+
+/**
+ * Build the L1 user prompt for summarization.
+ */
+export function buildL1UserPrompt(
+  recentContext: string,
+  pairs: ToolPair[],
+): string {
+  const parts: string[] = [];
+
+  parts.push("## 最近的对话上下文（用于理解当前任务）：");
+  parts.push(recentContext || "(无可用上下文)");
+  parts.push("\n## Tool call/result pairs to summarize:");
+
+  for (let i = 0; i < pairs.length; i++) {
+    const p = pairs[i];
+    const paramsStr = truncate(stringify(p.params), PARAMS_MAX_LEN);
+    const resultStr = truncate(stringify(p.result), RESULT_MAX_LEN);
+    const canonical = `${p.toolName}(${stringify(p.params)})`;
+    const needsCompress = canonical.length > COMPRESS_THRESHOLD;
+
+    parts.push(`--- Tool Pair ${i + 1} ---`);
+    parts.push(`tool_call_id: ${p.toolCallId}`);
+    parts.push(`timestamp: ${p.timestamp}`);
+    if (needsCompress) {
+      parts.push(`Tool: ${p.toolName} [NEEDS_COMPRESS]`);
+    } else {
+      parts.push(`Tool: ${p.toolName}`);
+    }
+    parts.push(`Params: ${paramsStr}`);
+    parts.push(`Result: ${resultStr}\n`);
+  }
+
+  parts.push("Summarize each pair into the JSON array format described.");
+  return parts.join("\n");
+}
+
+function stringify(value: unknown): string {
+  if (value == null) return "";
+  if (typeof value === "string") return value;
+  try {
+    return JSON.stringify(value);
+  } catch {
+    return String(value);
+  }
+}
+
+function truncate(s: string, maxLen: number): string {
+  if (s.length <= maxLen) return s;
+  return s.slice(0, maxLen) + "...";
+}

src/offload_server/prompts/l15-prompt.ts

@@ -0,0 +1,81 @@
+/**
+ * L1.5 Task Judgment Prompt — determines task lifecycle.
+ */
+import type { MmdMeta } from "../types.js";
+
+export const L15_SYSTEM_PROMPT = `你是一个面向 AI 编码助手的"任务生命周期门神"。
+你的职责是交叉分析提供的三个输入源，精准研判任务状态，并输出纯 JSON 对象。
+
+【输入数据利用指南（必须遵循的思考链路）】
+1. 第一步 - 剖析 recentMessages（识别意图）：根据当前和历史对话，提取用户最新回复的核心诉求。判断是"继续排查"、"宣布完工（如：跑通了）"、"单轮闲聊问答"还是"开启全新需求"。
+2. 第二步 - 对齐 currentMmd（评估当前基线）：将用户的最新意图与 currentMmd 的完整 Mermaid 内容进行比对——关注 taskGoal、各节点的 status（done/doing/todo）以及 summary。如果诉求完全超出了当前图表的范畴或目标已实现（所有节点 done 且无后续），则 taskCompleted 为 true。若仍在解决图表中的子问题（包括 doing 节点或修 bug），则为 false。(如果没有currentMmd，就只根据当前对话和历史对话来判断是否继续任务)
+3. 第三步 - 检索 availableMmds（判断是否延续）：如果判定要开启新任务（isLongTask=true 且 taskCompleted=true/当前无任务），必须扫描 availableMmds 的 taskGoal 和时间信息。若新诉求与列表中某个旧任务高度重合（如回到昨天没做完的模块），则是延续（isContinuation=true）。
+
+【严格 JSON 输出格式】
+务必输出合法的纯 JSON 对象，格式如下：
+{
+  "taskCompleted": boolean, // 当前任务是否已结束（如果 currentMmd 为 none，这里必须填 true）
+  "isLongTask": boolean,    // 最新诉求是否是需要多步操作的复杂工程（普通技术问答、闲聊填 false）
+  "isContinuation": boolean, // 是否在延续 availableMmds 中的历史任务
+  "continuationMmdFile": "string|null", // 若延续旧任务，精确填入 availableMmds 中的文件名（不含路径前缀），否则为 null
+  "newTaskLabel": "string|null" // 若是全新长任务，生成简短标签（≤30字符，kebab-case，如 "refactor-api"），否则为 null
+}
+
+只输出纯 JSON 对象，绝不允许包含解释文字。`;
+
+export interface L15CurrentMmd {
+  filename: string;
+  content: string;
+}
+
+/**
+ * Build the L1.5 user prompt for task judgment.
+ */
+export function buildL15UserPrompt(
+  recentMessages: string,
+  currentMmd: L15CurrentMmd | null,
+  metas: MmdMeta[],
+): string {
+  const parts: string[] = [];
+
+  parts.push("## 1. 最近的对话上下文 (Recent messages):");
+  parts.push(recentMessages);
+  parts.push("\n## 2. 当前挂载的任务图 (Active Mermaid — 完整内容):");
+
+  if (currentMmd && currentMmd.filename) {
+    parts.push(`**File:** ${currentMmd.filename}`);
+    parts.push(`\n\`\`\`mermaid\n${currentMmd.content}\n\`\`\``);
+  } else {
+    parts.push("(none - 当前处于闲置状态，无活跃任务)");
+  }
+
+  parts.push("\n## 3. 历史可用的任务图 (Available Mermaid task files):");
+
+  if (metas.length === 0) {
+    parts.push("(none - 暂无历史长任务)");
+  } else {
+    for (const m of metas) {
+      const total = m.doneCount + m.doingCount + m.todoCount;
+      parts.push(`- **${m.filename}**`);
+      parts.push(`  taskGoal: ${m.taskGoal}`);
+      parts.push(
+        `  progress: ${m.doneCount}/${total} done, ${m.doingCount} doing, ${m.todoCount} todo`,
+      );
+      if (m.updatedTime) {
+        parts.push(`  lastUpdated: ${m.updatedTime}`);
+      }
+      if (m.nodeSummaries && m.nodeSummaries.length > 0) {
+        parts.push("  recentNodes:");
+        for (const n of m.nodeSummaries) {
+          parts.push(`    - [${n.nodeId}] (${n.status}) ${n.summary}`);
+        }
+      }
+      parts.push("");
+    }
+  }
+
+  parts.push(
+    "请严格根据系统指令的【三步思考链路】进行研判，并输出合法的 JSON 对象。",
+  );
+  return parts.join("\n");
+}

src/offload_server/prompts/l2-prompt.ts

@@ -0,0 +1,113 @@
+/**
+ * L2 MMD Generation Prompt — generates/updates Mermaid flowcharts.
+ */
+import type { OffloadEntry } from "../types.js";
+
+export const L2_SYSTEM_PROMPT = `你是一个究极实用主义的 AI 任务拓扑架构师与视觉叙事者。
+你的核心逻辑是用尽量少的字符表达尽量多的信息，让LLM模型能看懂，不是为人类服务，尽量减少无用的视觉符号。任务是将底层工具调用记录，升维映射为一张高度语义化、表现力丰富且极度克制的 Mermaid (flowchart TD) 认知状态机。你要根据当前任务和意图，归纳"过去"，要思考"未来"如何用这些已有的信息（你只需要记录已有信息，不需要写下一步规划）并标记"雷区"。保持图表的高度概括性。
+
+【高阶认知与拓扑指南（你的自主权与极简原则）】
+1. 弹性聚合：你拥有决定节点拆合的完全自主权。对于连续的、意图相同的常规动作（如连续查看多个文件以了解上下文），建议合并为一个宏观节点；，但保留关键转折点或重大发现为独立节点。图表必须保持宏观和克制，绝不事无巨细地记流水账。
+2. 认知墓碑 (防重蹈覆辙)：遇到彻底走不通的死胡同或引发严重报错的废弃方案，可以建立警示节点（status: blocked）（如果是价值不高的fail信息则不需要记录）。
+3. 结论导向的摘要：节点的 summary（注意：尽量小于150字）应聚焦于"得出了什么结论"或"发生了什么实质改变"，而非罗列琐碎的数据或参数，记得保持极简原则。
+4. 要实事求是，你的任务是记录并归纳已经发生的事情，不是规划未来的具体操作，未发生的节点不要写，记录的已发生节点要有对应的消息来源（对应标注node_id）。
+【符号即语义：高维认知字典（你的核心武器）】为了极致压缩 Token 并为你下一步推理提供"认知锚点"，请自由使用不同的mmd形状来代表不同的节点逻辑。让形状替你说话，省略冗余的文字描述。
+
+【高度自由的拓扑与极简法则】
+1. 语义浓缩：既然形状已经表达了"领域"，你的 summary 必须极其精简（≤150字），如"发现死锁"、"依赖冲突"、"已修复"。
+2. 弹性拓扑：自主使用带标签的连线（-->|测试失败|）和虚线（-.->|参考|）来构建"依赖树"和"假设验证环"。不要记流水账。
+3. 动态更新 (Token 极简)：
+   - replace (增量微调)：仅修改现有节点的状态、时间戳、短文本或追加极少节点时。
+   - write (全量重写)：逻辑大洗牌、重构图表或初始化时。
+注意：Existing Mermaid content 中每行开头都带有行号标记（如 "L1: ..."），这些行号仅供你在 replace 模式中引用，不是 MMD 内容的一部分。
+
+【严格的工程底线】
+1.节点标准格式：NodeID["阶段名: 宏观动作简述<br/>status: done|doing|paused|blocked <br/>summary: 核心结论摘要<br/>Timestamp: ISO8601"]
+2. 全员归宿映射：输入的每一个新 tool_call_id，都必须在 node_mapping 中被分配到一个 Node ID；MMD里的每一个node都应该有源头的tool_call消息来源，不能乱编，绝对不允许遗漏！（Node_id和tool_call_id是一对多的关系）
+3. 你可以通过各种整合方法，尽量把更新后mmd文件大小控制在4000字以内
+
+【严格时间戳与元数据规则】
+1. 顶部元数据（必填）：%%{ "taskGoal": "一句话总结此次任务的目标（可动态更新）", "progress（0-100）": "进度百分比（严格点，几乎确认完成再打到90+)", createdTime": "ISO时间", "updatedTime": "ISO时间" }%%（updatedTime为node中的最新时间）。
+2. 节点内时间：如果合并了多个新条目，节点内的 Timestamp 必须取其中最新的 ISO 时间。
+
+【严格 JSON 输出格式】
+务必正确转义双引号。所有 Mermaid 代码（无论是 mmd_content 还是 replace_blocks 中的 content）都必须用 \`\`\`mermaid ... \`\`\` 代码块包裹起来。必须输出如下 JSON 结构：
+{
+  "file_action": "replace 或 write",
+  "mmd_content": "完整的、带转义的 .mmd 代码，必须用 \`\`\`mermaid ... \`\`\` 包裹。（仅在 file_action 为 write 时填写，否则必须设为 null）",
+  "replace_blocks": [
+    {
+      "start_line": "需要更新范围的起始行号（整数，对应 Existing Mermaid content 中的 L 标号）",
+      "end_line": "需要更新范围的结束行号（整数，包含该行）。要在某行之前插入新内容而不删除任何行，将 start_line 设为该行号，end_line 设为 start_line - 1",
+      "content": "替换后的新内容（不需要带行号前缀），必须用 \`\`\`mermaid ... \`\`\` 包裹"
+    }
+  ],
+  "node_mapping": {
+    "tool_call_id_1": "001-N1",
+    "tool_call_id_2": "001-N1"
+  }
+}
+
+注意：node_mapping 中的 Node ID 必须是 MMD 中实际使用的完整 ID（包含 MMD prefix，如 "001-N1"），不能只写短 ID（如 "N1"）。
+仅输出纯 JSON 对象，绝不允许包含任何解释。`;
+
+/**
+ * Build the L2 user prompt for MMD generation.
+ */
+export function buildL2UserPrompt(opts: {
+  existingMmd: string | null;
+  entries: OffloadEntry[];
+  recentHistory?: string | null;
+  currentTurn?: string | null;
+  taskLabel: string;
+  mmdPrefix: string;
+  charCount: number;
+}): string {
+  const { existingMmd, entries, recentHistory, currentTurn, taskLabel, mmdPrefix, charCount } = opts;
+  const parts: string[] = [];
+
+  // History section
+  if (recentHistory) {
+    parts.push(`## 近期对话历史：\n${recentHistory}`);
+  } else {
+    parts.push("## 近期对话历史：\n(无可用历史)");
+  }
+
+  if (currentTurn) {
+    parts.push(`\n## 当前最新一轮：\n${currentTurn}`);
+  }
+
+  parts.push(`\n## MMD prefix: ${mmdPrefix}`);
+  parts.push(`（所有节点 ID 必须以此前缀开头，如 ${mmdPrefix}-N1, ${mmdPrefix}-N2...）`);
+  parts.push(`\n## Current task label: ${taskLabel}`);
+
+  // Char count warning
+  if (charCount > 2500) {
+    parts.push(`\n## Current MMD size: ${charCount} chars (budget: 4000 chars)`);
+    parts.push("⚠ 接近上限，请积极合并节点、精简 summary，优先使用 replace 模式微调而非 write 全量重写。");
+  } else if (charCount > 2000) {
+    parts.push(`\n## Current MMD size: ${charCount} chars (budget: 4000 chars)`);
+    parts.push("注意控制增长，合并同类节点。");
+  }
+
+  // Existing MMD with line numbers
+  parts.push("\n## Existing Mermaid content:");
+  if (existingMmd) {
+    const lines = existingMmd.split("\n");
+    for (let i = 0; i < lines.length; i++) {
+      parts.push(`L${i + 1}: ${lines[i]}`);
+    }
+  } else {
+    parts.push("(empty — create new)");
+  }
+
+  // New entries
+  parts.push("\n## New offload entries to incorporate:");
+  for (let i = 0; i < entries.length; i++) {
+    const e = entries[i];
+    parts.push(`${i + 1}. [${e.tool_call_id}] ${e.tool_call} → ${e.summary} (${e.timestamp})`);
+  }
+
+  parts.push("\n请根据系统指令生成/更新 Mermaid 流程图，并输出合法的 JSON 对象（含 node_mapping）。");
+  return parts.join("\n");
+}

src/offload_server/router.ts

@@ -0,0 +1,88 @@
+/**
+ * Offload V2 Router — route registration and dispatch.
+ */
+import type http from "node:http";
+import type { StorageAdapter } from "../core/storage/adapter.js";
+import type { IStateBackend } from "../core/state/types.js";
+import type { OffloadExecutorConfig } from "./types.js";
+import { defaultOffloadConfig } from "./types.js";
+import { parseV2Auth, successEnvelope, errorEnvelope, makeRequestId } from "../gateway/v2-router.js";
+import { handleIngest } from "./ingest-handler.js";
+import { handleMmdQuery } from "./mmd-handler.js";
+import { handleCompaction } from "./compact/compaction-handler.js";
+import { MmdQuerySchema } from "./schemas.js";
+
+export interface OffloadV2Deps {
+  resolveStorage?: (instanceId: string) => Promise<StorageAdapter | undefined>;
+  getStorage: () => StorageAdapter | undefined;
+  logger: { info: (...args: unknown[]) => void; warn: (...args: unknown[]) => void; error: (...args: unknown[]) => void };
+  stateBackend?: IStateBackend;
+  config?: OffloadExecutorConfig;
+}
+
+/**
+ * Handle offload V2 routes. Returns true if the request was handled.
+ */
+export async function handleOffloadV2Route(
+  req: http.IncomingMessage,
+  res: http.ServerResponse,
+  pathname: string,
+  method: string,
+  parseJsonBody: <T>(req: http.IncomingMessage) => Promise<T>,
+  sendJson: (res: http.ServerResponse, status: number, body: unknown) => void,
+  deps: OffloadV2Deps,
+): Promise<boolean> {
+  if (!pathname.startsWith("/v2/offload/")) return false;
+
+  const requestId = makeRequestId();
+
+  // Auth
+  const auth = parseV2Auth(req, res, requestId, sendJson);
+  if (!auth) return true; // 401 already sent
+
+  // Resolve storage
+  const storage =
+    (await deps.resolveStorage?.(auth.serviceId)) ?? deps.getStorage();
+  if (!storage) {
+    sendJson(res, 503, errorEnvelope(503, "Storage unavailable", requestId));
+    return true;
+  }
+
+  const config = deps.config ?? defaultOffloadConfig();
+  // Normalize trailing slash for consistent route matching
+  const normalizedPath = pathname.endsWith("/") && pathname.length > 1 ? pathname.slice(0, -1) : pathname;
+  const route = `${method} ${normalizedPath}`;
+
+  switch (route) {
+    case "POST /v2/offload/ingest":
+      await handleIngest(req, res, auth, {
+        storage,
+        stateBackend: deps.stateBackend,
+        config,
+        logger: deps.logger,
+      }, requestId, parseJsonBody, sendJson, successEnvelope, errorEnvelope);
+      return true;
+
+    case "POST /v2/offload/query-mmd": {
+      const body = await parseJsonBody<{ session_id?: string; limit?: number }>(req);
+      const parsed = MmdQuerySchema.safeParse(body);
+      if (!parsed.success) {
+        sendJson(res, 400, errorEnvelope(400, "missing or invalid session_id in body", requestId));
+        return true;
+      }
+      await handleMmdQuery(req, res, auth, storage, requestId, sendJson, successEnvelope, errorEnvelope, parsed.data.session_id, parsed.data.limit);
+      return true;
+    }
+
+    case "POST /v2/offload/compact":
+      await handleCompaction(req, res, auth, {
+        storage,
+        config,
+        logger: deps.logger,
+      }, requestId, parseJsonBody, sendJson, successEnvelope, errorEnvelope);
+      return true;
+
+    default:
+      return false;
+  }
+}

src/offload_server/schemas.ts

@@ -0,0 +1,80 @@
+/**
+ * Offload Server — Request validation schemas (Zod).
+ */
+import { z } from "zod";
+
+/** Safe session ID: alphanumeric, underscore, hyphen, dot, colon allowed. No slashes or path traversal. Max 500 chars. */
+const safeSessionId = z.string().min(1).max(500, {
+  message: "sessionId must not exceed 500 characters",
+}).regex(/^[a-zA-Z0-9_.\-:]+$/, {
+  message: "Must only contain alphanumeric, underscore, hyphen, dot, or colon characters",
+});
+
+const ToolPairSchema = z.object({
+  tool_name: z.string(),
+  tool_call_id: z.string(),
+  params: z.unknown(),
+  result: z.unknown(),
+  error: z.string().optional(),
+  timestamp: z.string(),
+  duration_ms: z.number().optional(),
+});
+
+/** Recent message item: user/assistant text only (no tool_call/tool_result). */
+const RecentMessageSchema = z.object({
+  role: z.enum(["user", "assistant"]),
+  content: z.string(),
+});
+
+export const IngestRequestSchema = z
+  .object({
+    session_id: safeSessionId,
+    tool_pairs: z.array(ToolPairSchema).default([]),
+    /** Current user prompt that triggers L1.5 task judgment. Must be non-empty (whitespace-only is rejected). */
+    prompt: z.string().trim().min(1, { message: "prompt must not be empty or whitespace-only" }).optional(),
+    /** Recent history messages (user/assistant only, no tool calls). */
+    recent_messages: z.array(RecentMessageSchema).optional(),
+  })
+  .refine(
+    (data) => data.tool_pairs.length > 0 || (data.prompt && data.prompt.length > 0),
+    { message: "Either tool_pairs must be non-empty or prompt must be provided" },
+  );
+
+export type IngestRequest = z.infer<typeof IngestRequestSchema>;
+
+/**
+ * Each message must have non-empty `role` and `content` fields.
+ * Intentionally lenient: any non-empty string role is accepted to support
+ * OpenAI / Anthropic / OpenClaw-wrapped formats without over-specifying.
+ */
+const CompactionMessageSchema = z
+  .record(z.string(), z.unknown())
+  .refine(
+    (msg) => typeof msg["role"] === "string" && (msg["role"] as string).length > 0,
+    { message: "Each message must have a non-empty 'role' field" },
+  );
+
+export const CompactionRequestSchema = z.object({
+  session_id: safeSessionId,
+  messages: z.array(CompactionMessageSchema),
+  ratio: z.number().min(0).max(2),
+});
+
+export type CompactionRequest = z.infer<typeof CompactionRequestSchema>;
+
+/** Extended compaction schema with token metadata for L3 compression. */
+export const CompactionRequestSchemaV2 = z.object({
+  session_id: safeSessionId,
+  messages: z.array(CompactionMessageSchema),
+  ratio: z.number().min(0).max(2),
+  context_window: z.number().int().min(1),
+  total_tokens: z.number().int().min(0),
+  message_tokens: z.array(z.number()).optional(),
+});
+
+export type CompactionRequestV2 = z.infer<typeof CompactionRequestSchemaV2>;
+
+export const MmdQuerySchema = z.object({
+  session_id: safeSessionId,
+  limit: z.number().int().min(1).optional(),
+});

src/offload_server/session-utils.ts

@@ -0,0 +1,29 @@
+/**
+ * Offload Session Utilities — unified sessionId sanitization.
+ *
+ * sessionId: opaque identifier from client (free format, e.g. "main:main", "coding:session-001")
+ * sanitizedSessionId: filesystem-safe version used for:
+ *   - File/directory paths (offload/{sanitizedSessionId}/)
+ *   - Distributed lock keys
+ *   - Task queue session identifiers
+ *
+ * Conversion: replace all non-alphanumeric/dot/hyphen/underscore chars with "_"
+ */
+
+/**
+ * Convert a raw sessionId to a filesystem-safe string.
+ * Replaces colons and other unsafe characters with underscores.
+ */
+export function sanitizeSessionId(sessionId: string): string {
+  return sessionId.replace(/[^a-zA-Z0-9._\-]/g, "_");
+}
+
+/**
+ * Build the offload storage base path for a session.
+ */
+export function buildOffloadBasePath(sessionId: string): string {
+  return `offload/${sanitizeSessionId(sessionId)}`;
+}
+
+// Legacy compat: keep old name as alias (will remove after all callers migrated)
+export const toSessionId = sanitizeSessionId;

src/offload_server/task-transition.ts

@@ -0,0 +1,132 @@
+/**
+ * Task Transition logic — manages MMD file creation/switching based on L1.5 judgment.
+ */
+import type { StorageAdapter } from "../core/storage/adapter.js";
+import type { OffloadState, TaskJudgment, MmdMeta } from "./types.js";
+
+/**
+ * Apply task transition based on L1.5 judgment.
+ * Mutates state in place; may create new MMD file in COS.
+ * @param basePath - Full storage base path, e.g. "offload/default/agent_main_tui-xxx"
+ */
+export async function handleTaskTransition(
+  state: OffloadState,
+  judgment: TaskJudgment,
+  storage: StorageAdapter,
+  basePath: string,
+): Promise<void> {
+  if (judgment.taskCompleted && judgment.isLongTask && !judgment.isContinuation && judgment.newTaskLabel) {
+    // CASE 1: New long task → create new MMD file
+    const filename = await generateMmdFilename(storage, basePath, judgment.newTaskLabel);
+    await storage.writeFile(`${basePath}/mmds/${filename}`, "");
+    state.activeMmdFile = filename;
+  } else if (judgment.taskCompleted && judgment.isContinuation && judgment.continuationMmdFile) {
+    // CASE 2: Continue historical task → switch to that MMD
+    state.activeMmdFile = judgment.continuationMmdFile;
+  } else if (judgment.taskCompleted && !judgment.isLongTask) {
+    // CASE 3: Short task / casual chat → clear active MMD
+    state.activeMmdFile = null;
+  } else if (!judgment.taskCompleted && judgment.isLongTask && !state.activeMmdFile) {
+    // CASE 5: Task in progress but no active MMD yet → create one
+    const label = judgment.newTaskLabel || "current-task";
+    const filename = await generateMmdFilename(storage, basePath, label);
+    await storage.writeFile(`${basePath}/mmds/${filename}`, "");
+    state.activeMmdFile = filename;
+  }
+  // CASE 4: !taskCompleted + activeMmdFile already set → keep unchanged
+}
+
+/**
+ * Generate a new MMD filename with auto-incrementing sequence number.
+ * Format: "003-refactor-api.mmd"
+ */
+async function generateMmdFilename(
+  storage: StorageAdapter,
+  basePath: string,
+  label: string,
+): Promise<string> {
+  const mmdsPrefix = `${basePath}/mmds/`;
+  const existingFiles = await storage.readdirNames(mmdsPrefix, ".mmd");
+
+  let maxSeq = 0;
+  for (const f of existingFiles) {
+    const match = f.match(/^(\d+)-/);
+    if (match) {
+      const seq = parseInt(match[1], 10);
+      if (seq > maxSeq) maxSeq = seq;
+    }
+  }
+
+  const nextSeq = String(maxSeq + 1).padStart(3, "0");
+  const safeLabel = label
+    .toLowerCase()
+    .replace(/[^a-z0-9-]/g, "-")
+    .replace(/-+/g, "-")
+    .replace(/^-|-$/g, "")
+    .slice(0, 30);
+
+  return `${nextSeq}-${safeLabel || "task"}.mmd`;
+}
+
+/**
+ * Extract MmdMeta from MMD file content (parses %%{ ... }%% header).
+ */
+export function extractMmdMeta(filename: string, content: string): MmdMeta {
+  const defaults: MmdMeta = {
+    filename,
+    taskGoal: "",
+    doneCount: 0,
+    doingCount: 0,
+    todoCount: 0,
+    updatedTime: null,
+    nodeSummaries: [],
+  };
+
+  if (!content) return defaults;
+
+  // Parse %%{ ... }%% metadata line
+  const metaMatch = content.match(/%%\{([\s\S]*?)\}%%/);
+  if (metaMatch) {
+    try {
+      const meta = JSON.parse(`{${metaMatch[1]}}`);
+      defaults.taskGoal = meta.taskGoal ?? "";
+      defaults.updatedTime = meta.updatedTime ?? null;
+    } catch {
+      // ignore parse errors
+    }
+  }
+
+  // Count node statuses and extract summaries
+  const nodeRegex = /(\w[\w-]*)\["([^"]*?)"\]/g;
+  let nodeMatch;
+  while ((nodeMatch = nodeRegex.exec(content)) !== null) {
+    const nodeId = nodeMatch[1];
+    const nodeContent = nodeMatch[2];
+    const statusM = nodeContent.match(/status:\s*(done|doing|todo|paused|blocked)/i);
+    const summaryM = nodeContent.match(/summary:\s*(.+?)(?:<br\/>|$)/i);
+    if (statusM) {
+      const s = statusM[1].toLowerCase();
+      if (s === "done") defaults.doneCount++;
+      else if (s === "doing") defaults.doingCount++;
+      else if (s === "todo") defaults.todoCount++;
+      defaults.nodeSummaries!.push({
+        nodeId,
+        status: s,
+        summary: summaryM?.[1]?.trim() ?? "",
+      });
+    }
+  }
+
+  // Fallback: if regex didn't catch statuses, try simple pattern
+  if (defaults.doneCount === 0 && defaults.doingCount === 0 && defaults.todoCount === 0) {
+    const statusMatches = content.matchAll(/status:\s*(done|doing|todo|paused|blocked)/gi);
+    for (const m of statusMatches) {
+      const s = m[1].toLowerCase();
+      if (s === "done") defaults.doneCount++;
+      else if (s === "doing") defaults.doingCount++;
+      else if (s === "todo") defaults.todoCount++;
+    }
+  }
+
+  return defaults;
+}

src/offload_server/types.ts

@@ -0,0 +1,163 @@
+/**
+ * Offload Server — Independent type definitions.
+ * This module does NOT import from src/offload/.
+ */
+
+// ─── ToolPair (ingest 写入 pending.jsonl) ────────────────────────────────────
+
+export interface ToolPair {
+  toolName: string;
+  toolCallId: string;
+  params: unknown;
+  result: unknown;
+  error?: string;
+  timestamp: string;
+  durationMs?: number;
+}
+
+// ─── OffloadEntry (L1 产出, entries.jsonl 每行) ──────────────────────────────
+
+export interface OffloadEntry {
+  tool_call_id: string;
+  tool_call: string;
+  summary: string;
+  timestamp: string;
+  score: number;
+  node_id: string | null;
+  /** Full relative path to ref file storing original tool result (e.g. "offload/{sessionId}/refs/call_214.md"), readable via tdai_read_cos */
+  result_ref?: string;
+}
+
+// ─── TaskJudgment (L1.5 产出) ────────────────────────────────────────────────
+
+export interface TaskJudgment {
+  taskCompleted: boolean;
+  isLongTask: boolean;
+  isContinuation: boolean;
+  continuationMmdFile?: string;
+  newTaskLabel?: string;
+}
+
+// ─── TaskBoundary (L1.5 写入 state.boundaries) ──────────────────────────────
+
+export interface TaskBoundary {
+  targetMmd: string | null;
+  timestamp: string;
+}
+
+// ─── MmdMeta (从 MMD 文件首行 %%{...}%% 解析) ──────────────────────────────
+
+export interface MmdMeta {
+  filename: string;
+  taskGoal: string;
+  doneCount: number;
+  doingCount: number;
+  todoCount: number;
+  updatedTime?: string | null;
+  nodeSummaries?: Array<{ nodeId: string; status: string; summary: string }>;
+}
+
+// ─── OffloadState (state.json) ───────────────────────────────────────────────
+
+export interface OffloadState {
+  activeMmdFile: string | null;
+  boundaries: TaskBoundary[];
+  lastL15CreatedAt: number;
+}
+
+// ─── CompactState (compact-state.json) ───────────────────────────────────────
+
+export interface CompactState {
+  confirmedOffloadIds: string[];
+  deletedOffloadIds: string[];
+  lastCompactedAt: string;
+}
+
+// ─── L2 Parsed Response ──────────────────────────────────────────────────────
+
+export interface L2ParsedResponse {
+  fileAction: "write" | "replace";
+  mmdContent?: string;
+  replaceBlocks?: Array<{
+    startLine: number;
+    endLine: number;
+    content: string;
+  }>;
+  nodeMapping: Record<string, string>;
+}
+
+// ─── Configuration ───────────────────────────────────────────────────────────
+
+export interface OffloadExecutorConfig {
+  forceTriggerThreshold: number;
+  pendingMaxAgeSeconds: number;
+
+  l1Model: string;
+  l1Temperature: number;
+  l1MaxTokens: number;
+  l1TimeoutMs: number;
+
+  l15Model: string;
+  l15Temperature: number;
+  l15MaxTokens: number;
+  l15TimeoutMs: number;
+
+  l2Model: string;
+  l2Temperature: number;
+  l2MaxTokens: number;
+  l2TimeoutMs: number;
+  l2NullThreshold: number;
+
+  mildOffloadRatio: number;
+  aggressiveCompressRatio: number;
+  emergencyCompressRatio: number;
+
+  maxRetries: number;
+}
+
+export function defaultOffloadConfig(): OffloadExecutorConfig {
+  return {
+    forceTriggerThreshold: 4,
+    pendingMaxAgeSeconds: 30,
+
+    l1Model: "",  // uses gateway llm.model
+    l1Temperature: 0.3,
+    l1MaxTokens: 8000,
+    l1TimeoutMs: 120_000,
+
+    l15Model: "", // uses gateway llm.model
+    l15Temperature: 0.2,
+    l15MaxTokens: 3000,
+    l15TimeoutMs: 120_000,
+
+    l2Model: "",  // uses gateway llm.model
+    l2Temperature: 0.4,
+    l2MaxTokens: 16000,
+    l2TimeoutMs: 120_000,
+    l2NullThreshold: 6,
+
+    mildOffloadRatio: 0.5,
+    aggressiveCompressRatio: 0.85,
+    emergencyCompressRatio: 0.95,
+
+    maxRetries: 3,
+  };
+}
+
+// ─── Default empty state ─────────────────────────────────────────────────────
+
+export function defaultOffloadState(): OffloadState {
+  return {
+    activeMmdFile: null,
+    boundaries: [],
+    lastL15CreatedAt: 0,
+  };
+}
+
+export function defaultCompactState(): CompactState {
+  return {
+    confirmedOffloadIds: [],
+    deletedOffloadIds: [],
+    lastCompactedAt: "",
+  };
+}

src/services/pipeline-worker.ts

@@ -42,6 +42,9 @@ export interface TaskExecutor {
   executeL2(task: TaskPayload, signal?: AbortSignal): Promise<void>;
   executeL3(task: TaskPayload, signal?: AbortSignal): Promise<void>;
   executeFlush?(task: TaskPayload, signal?: AbortSignal): Promise<void>;
+  executeOffloadL1?(task: TaskPayload, signal?: AbortSignal): Promise<void>;
+  executeOffloadL15?(task: TaskPayload, signal?: AbortSignal): Promise<void>;
+  executeOffloadL2?(task: TaskPayload, signal?: AbortSignal): Promise<void>;
 }
 
 export interface PipelineWorkerConfig {
@@ -249,11 +252,67 @@ export class PipelineWorker {
     const lockKey = this.getLockKey(task);
     const retryCount = (task.data?.retryCount as number) ?? 0;
 
+    // Lock-free path: offload-l1 doesn't need distributed lock
+    if (lockKey === null) {
+      this.runningTasks.set(task.id, task);
+      try {
+        await this.executeTask(task, undefined);
+
+        // ACK
+        const msgId = (task as any)._msgId;
+        if (msgId) await this.backend.ackTask(msgId);
+
+        this.metrics.tasksCompleted++;
+        this.logger?.debug?.(`${TAG} Task completed (lock-free): ${task.type} [${task.instanceId}/${task.sessionId}]`);
+      } catch (err) {
+        const errMsg = err instanceof Error ? err.message : String(err);
+        this.metrics.tasksFailed++;
+
+        if (retryCount < this.config.maxRetries) {
+          const delay = this.config.retryBaseDelayMs * Math.pow(3, retryCount);
+          this.logger.warn(`${TAG} Task failed (lock-free, retry ${retryCount + 1}/${this.config.maxRetries}, delay=${delay}ms): ${errMsg}`);
+          const msgId = (task as any)._msgId;
+          if (msgId) { try { await this.backend.ackTask(msgId); } catch { /* best effort */ } }
+          await this.sleep(delay);
+          await this.reEnqueue(task, retryCount + 1);
+          this.metrics.tasksRetried++;
+        } else {
+          await this.moveToDeadLetter(task, errMsg, retryCount);
+        }
+      } finally {
+        this.runningTasks.delete(task.id);
+
+        // Deferred enqueue (same as locked path)
+        const deferred = (task as any)._deferredEnqueue as TaskPayload[] | undefined;
+        if (deferred?.length) {
+          for (const dTask of deferred) {
+            try {
+              await this.backend.enqueueTask(dTask);
+              this.logger?.debug?.(`${TAG} Deferred enqueue: ${dTask.type} [${dTask.id}]`);
+            } catch (err) {
+              this.logger?.warn?.(`${TAG} Deferred enqueue failed: ${err instanceof Error ? err.message : String(err)}`);
+            }
+          }
+        }
+      }
+      return;
+    }
+
     // Step 1: 抢分布式锁
     const locked = await this.backend.acquireLock(lockKey, this.config.workerId, this.config.lockTtlMs);
     if (!locked) {
       this.metrics.lockConflicts++;
 
+      // offload-l2: skip immediately on lock conflict (idempotent timer will re-trigger)
+      if (task.type === "offload-l2") {
+        this.logger?.debug?.(`${TAG} Lock conflict [offload-l2] (task=${task.id}): ${lockKey}, skip (timer will re-trigger)`);
+        const msgId = (task as any)._msgId;
+        if (msgId) {
+          try { await this.backend.ackTask(msgId); } catch { /* best effort */ }
+        }
+        return;
+      }
+
       // Lock conflict: current coroutine waits locally (no re-enqueue to stream).
       // Exponential backoff: 200ms → 600ms → 1.8s → 5s (capped), retry until lockTtlMs exhausted.
       // 旧版本固定 sleep(5000) 在 instance 级锁下会造成排队体感差(同 instance 多 session 累积秒级延迟);
@@ -265,14 +324,14 @@ export class PipelineWorker {
       let delay = 200;
       while (Date.now() < deadline && this.running) {
         attempt++;
-        this.logger?.debug?.(`${TAG} Lock conflict: ${lockKey}, retry ${attempt} after ${delay}ms`);
+        this.logger?.debug?.(`${TAG} Lock conflict [${task.type}] (task=${task.id}): ${lockKey}, retry ${attempt} after ${delay}ms`);
         await this.sleep(delay);
         acquired = await this.backend.acquireLock(lockKey, this.config.workerId, this.config.lockTtlMs);
         if (acquired) break;
         delay = Math.min(delay * 3, 5000);
       }
       if (!acquired) {
-        this.logger?.warn?.(`${TAG} Lock conflict timeout: ${lockKey}, dropping task`);
+        this.logger?.warn?.(`${TAG} Lock conflict timeout [${task.type}] (task=${task.id}): ${lockKey}, dropping task`);
         // CR-1 fix: ACK to prevent stale recovery from re-claiming this message in an
         // infinite loop. Without it, XPENDING keeps returning this msgId every
         // pendingRecoveryIntervalMs, exhausting worker slots.
@@ -396,6 +455,20 @@ export class PipelineWorker {
       this.activeLocks.delete(lockKey);
       this.runningTasks.delete(task.id);
       try { await this.backend.releaseLock(lockKey, this.config.workerId); } catch { /* best effort */ }
+
+      // Step 7: 延迟入队 — executor 可通过 task._deferredEnqueue 暂存需要在锁释放后才入队的任务，
+      // 避免新任务立即被消费时因同 session 锁仍被持有而产生不必要的锁冲突。
+      const deferred = (task as any)._deferredEnqueue as TaskPayload[] | undefined;
+      if (deferred?.length) {
+        for (const dTask of deferred) {
+          try {
+            await this.backend.enqueueTask(dTask);
+            this.logger?.debug?.(`${TAG} Deferred enqueue: ${dTask.type} [${dTask.id}]`);
+          } catch (err) {
+            this.logger?.warn?.(`${TAG} Deferred enqueue failed: ${err instanceof Error ? err.message : String(err)}`);
+          }
+        }
+      }
     }
   }
 
@@ -405,6 +478,9 @@ export class PipelineWorker {
       case "L2": return this.executor.executeL2(task, signal);
       case "L3": return this.executor.executeL3(task, signal);
       case "flush": return this.executor.executeFlush?.(task, signal) ?? this.executor.executeL1(task, signal);
+      case "offload-l1": return this.executor.executeOffloadL1?.(task, signal);
+      case "offload-l15": return this.executor.executeOffloadL15?.(task, signal);
+      case "offload-l2": return this.executor.executeOffloadL2?.(task, signal);
       default:
         this.logger.warn(`${TAG} Unknown task type: ${task.type}`);
     }
@@ -494,7 +570,22 @@ export class PipelineWorker {
    *   升级窗口期，新旧 pod 看到的 lock key 格式不同，可能短暂破坏跨 pod
    *   互斥。缓解: 升级时停所有 worker → 等 pending 任务清空 → 启动新版。
    */
-  private getLockKey(task: TaskPayload): string {
+  private getLockKey(task: TaskPayload): string | null {
+    // offload-l1 is lock-free: rename guarantees exclusive file ownership,
+    // appendFile is atomic (O_APPEND), and state.json is read-only for L1.
+    if (task.type === "offload-l1") return null;
+
+    // offload-l2: per-MMD lock so different MMDs can be processed concurrently.
+    if (task.type === "offload-l2") {
+      const mmdFile = (task.data as any)?.targetMmdFile ?? "default";
+      return `pipeline:{${task.instanceId}}:offload-l2:${mmdFile}`;
+    }
+
+    // offload-l15: lock-free at worker level. The executor acquires a short
+    // lock only during the final write phase (state.json update), allowing
+    // multiple L1.5 LLM calls to run concurrently without blocking each other.
+    if (task.type === "offload-l15") return null;
+
     if (this.config.lockGranularity === "instance") {
       return `pipeline:{${task.instanceId}}`;
     }

src/services/timer-scanner.ts

@@ -152,7 +152,7 @@ export class TimerScanner {
         }
 
         for (const entry of expired) {
-          const { instanceId, sessionId, taskType, priority } = this.parseShardMember(entry.member);
+          const { instanceId, sessionId, taskType, priority, timerType } = this.parseShardMember(entry.member);
 
           const task: TaskPayload = {
             id: `${taskType}-${instanceId.slice(-8)}-${sessionId.slice(-8)}-${now}`,
@@ -161,7 +161,7 @@ export class TimerScanner {
             sessionId,
             priority,
             createdAt: now,
-            data: { triggeredBy: "timer_scanner", timerMember: `${sessionId}:${taskType === "L1" ? "L1_idle" : taskType === "L2" ? "L2_schedule" : "L3"}`, instanceId },
+            data: { triggeredBy: "timer_scanner", timerMember: `${sessionId}:${timerType}`, instanceId },
           };
 
           await this.backend.enqueueTask(task);
@@ -191,7 +191,7 @@ export class TimerScanner {
    * Parse shard member format: "{instanceId}\x00{sessionId}:{timerType}"
    * Example: "mem-j4wjesud\x00sess_001:L1_idle" → { instanceId: "mem-j4wjesud", sessionId: "sess_001", taskType: "L1" }
    */
-  private parseShardMember(member: string): { instanceId: string; sessionId: string; taskType: "L1" | "L2" | "L3" | "flush"; priority: number } {
+  private parseShardMember(member: string): { instanceId: string; sessionId: string; taskType: TaskPayload["type"]; priority: number; timerType: string } {
     const sep = member.indexOf("\x00");
     let instanceId: string;
     let rest: string;
@@ -206,19 +206,51 @@ export class TimerScanner {
       rest = member.slice(firstColon + 1);
     }
 
-    // rest = "sessionId:timerType" (e.g. "sess_001:L1_idle", "sess_001:L2_schedule")
+    // rest = timer member after instanceId separator
+    // New unified format: "offload-{type}:{embeddedInstanceId}:{sessionId}[:{extra}]" (prefix-based)
+    // Legacy format: "sessionId:L1_idle" or "sessionId:L2_schedule"
+
+    // Check for offload prefix format first
+    if (rest.startsWith("offload-l15:")) {
+      // Skip embedded instanceId: "offload-l15:{instanceId}:{sessionId}"
+      const afterPrefix = rest.slice("offload-l15:".length);
+      const colonIdx = afterPrefix.indexOf(":");
+      const sessionId = colonIdx > 0 ? afterPrefix.slice(colonIdx + 1) : afterPrefix;
+      return { instanceId, sessionId, taskType: "offload-l15", priority: 0, timerType: rest };
+    }
+    if (rest.startsWith("offload-l2:")) {
+      // Skip embedded instanceId: "offload-l2:{instanceId}:{sessionId}[:{mmdFile}]"
+      const afterPrefix = rest.slice("offload-l2:".length);
+      const colonIdx = afterPrefix.indexOf(":");
+      let sessionId = colonIdx > 0 ? afterPrefix.slice(colonIdx + 1) : afterPrefix;
+      // Strip trailing ":{mmdFile}" from sessionId
+      if (sessionId.endsWith(".mmd")) {
+        const lastColon = sessionId.lastIndexOf(":");
+        if (lastColon > 0) sessionId = sessionId.slice(0, lastColon);
+      }
+      return { instanceId, sessionId, taskType: "offload-l2", priority: 1, timerType: rest };
+    }
+    if (rest.startsWith("offload-l1:")) {
+      // Skip embedded instanceId: "offload-l1:{instanceId}:{sessionId}"
+      const afterPrefix = rest.slice("offload-l1:".length);
+      const colonIdx = afterPrefix.indexOf(":");
+      const sessionId = colonIdx > 0 ? afterPrefix.slice(colonIdx + 1) : afterPrefix;
+      return { instanceId, sessionId, taskType: "offload-l1", priority: 0, timerType: rest };
+    }
+
+    // For non-offload types, use lastColon (original logic)
     const lastColon = rest.lastIndexOf(":");
     if (lastColon <= 0) {
-      return { instanceId, sessionId: rest, taskType: "L1", priority: 0 };
+      return { instanceId, sessionId: rest, taskType: "L1", priority: 0, timerType: "L1_idle" };
     }
 
     const sessionId = rest.slice(0, lastColon);
     const timerType = rest.slice(lastColon + 1);
 
-    if (timerType.startsWith("L1")) return { instanceId, sessionId, taskType: "L1", priority: 0 };
-    if (timerType.startsWith("L2")) return { instanceId, sessionId, taskType: "L2", priority: 1 };
-    if (timerType.startsWith("L3")) return { instanceId, sessionId, taskType: "L3", priority: 2 };
-    return { instanceId, sessionId, taskType: "flush", priority: 0 };
+    if (timerType.startsWith("L1")) return { instanceId, sessionId, taskType: "L1", priority: 0, timerType };
+    if (timerType.startsWith("L2")) return { instanceId, sessionId, taskType: "L2", priority: 1, timerType };
+    if (timerType.startsWith("L3")) return { instanceId, sessionId, taskType: "L3", priority: 2, timerType };
+    return { instanceId, sessionId, taskType: "flush", priority: 0, timerType };
   }
 }
 

src/utils/clean-context-runner.ts

@@ -5,7 +5,7 @@
  * Guarantees:
  * 1. Blank conversation history (temporary session file)
  * 2. Independent system prompt (only the task prompt)
- * 3. No tool calls (tools restricted to minimal read-only set to avoid empty tools[] rejection by some providers)
+ * 3. No tool calls when enableTools=false (disableTools:true — no tool definitions sent to API)
  * 4. No contamination from the main agent's context
  */
 
@@ -17,6 +17,7 @@ import { fileURLToPath, pathToFileURL } from "node:url";
 import type { OpenClawPluginApi } from "openclaw/plugin-sdk/core";
 import { getEnv } from "./env.js";
 import { report } from "../core/report/reporter.js";
+import type { Logger } from "../core/types.js";
 
 /**
  * Resolve a preferred temporary directory for memory-tdai operations.
@@ -49,12 +50,7 @@ function resolveOpenClawTmpDir(): string {
 
 const TAG = "[memory-tdai] [runner]";
 
-interface RunnerLogger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+type RunnerLogger = Logger;
 
 // Dynamic import type — runEmbeddedPiAgent is an internal API
 // Prefer the public plugin runtime signature so host-injected runtimes stay assignable.
@@ -394,10 +390,11 @@ export class CleanContextRunner {
         },
         tools: {
           ...((this.options.config as Record<string, unknown>)?.tools as Record<string, unknown> | undefined),
-          // When enableTools=false we still keep one lightweight read-only tool
-          // so that the tools array sent to the API is non-empty.
-          // Some providers (e.g. qwencode) reject tools:[] with minItems:1 validation.
-          allow: this.options.enableTools ? ["read", "write", "edit"] : ["read"],
+          // When enableTools=true, restrict to the minimal set needed for
+          // scene extraction (read/write/edit).
+          // When enableTools=false, pass an empty allow list — disableTools:true
+          // will prevent tools from being sent to the API entirely.
+          allow: this.options.enableTools ? ["read", "write", "edit"] : [],
         },
         // Override the full agent system prompt with the caller's extraction-specific
         // system prompt. This replaces OpenClaw's default system prompt (identity,
@@ -455,11 +452,13 @@ export class CleanContextRunner {
         runId,
         provider: this.resolvedProvider,
         model: this.resolvedModel,
-        // Do NOT pass disableTools:true — that produces tools:[] which some
-        // providers (qwencode) reject with "[] is too short - 'tools'".
-        // Instead rely on cleanConfig.tools.allow to restrict the tool set
-        // to a minimal read-only tool (when enableTools=false).
-        disableTools: false,
+        // When enableTools=false, pass disableTools:true so that no tool
+        // definitions are sent to the API. This avoids polluting the LLM
+        // context with tool schemas and prevents the model from attempting
+        // tool calls during pure text extraction tasks.
+        // If a provider (e.g. qwencode) rejects empty tools[], users should
+        // switch to StandaloneLLMRunner via LLM configuration instead.
+        disableTools: !this.options.enableTools,
         extraSystemPrompt: effectiveSystemPrompt,
         streamParams: {
           maxTokens: params.maxTokens,

src/utils/memory-cleaner.ts

@@ -3,13 +3,7 @@ import path from "node:path";
 
 import type { IMemoryStore } from "../core/store/types.js";
 import { ManagedTimer } from "./managed-timer.js";
-
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+import type { Logger } from "../core/types.js";
 
 export interface MemoryCleanerOptions {
   baseDir: string;

src/utils/pipeline-factory.ts

@@ -37,6 +37,7 @@ import { PersonaTrigger } from "../core/persona/persona-trigger.js";
 import { PersonaGenerator } from "../core/persona/persona-generator.js";
 import { pullProfilesToLocal, syncLocalProfilesToStore } from "../core/profile/profile-sync.js";
 import type { StorageAdapter } from "../core/storage/adapter.js";
+import type { Logger } from "../core/types.js";
 
 const TAG = "[memory-tdai] [pipeline-factory]";
 
@@ -72,12 +73,8 @@ function supportsProfileSyncWrite(store?: IMemoryStore): boolean {
 // Logger interface
 // ============================
 
-export interface PipelineLogger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
+/** @deprecated Use `Logger` from `../core/types.js` directly. */
+export type PipelineLogger = Logger;
 
 // ============================
 // Factory options

src/utils/pipeline-manager.ts

@@ -81,18 +81,12 @@ import { SessionFilter } from "./session-filter.js";
 import { ManagedTimer } from "./managed-timer.js";
 import { SerialQueue } from "./serial-queue.js";
 import { report } from "../core/report/reporter.js";
+import type { Logger } from "../core/types.js";
 
 // ============================
 // Types
 // ============================
 
-interface Logger {
-  debug?: (message: string) => void;
-  info: (message: string) => void;
-  warn: (message: string) => void;
-  error: (message: string) => void;
-}
-
 /** A single captured message ready for L1 processing. */
 export interface CapturedMessage {
   role: "user" | "assistant" | "tool";