@lingxiao-office/sdk 1.0.0

package/dist/agents/LeaderProgressInvariant.d.ts

@@ -0,0 +1,174 @@
+/**
+ * LeaderProgressInvariant — V3
+ *
+ * Manages progress stagnation detection, watchdog timer, and delegates
+ * eternal idle patrol to EternalLoop (unified autonomous control loop).
+ *
+ * Key changes from V2:
+ *   - Eternal patrol no longer has a hard 10-round cap.
+ *     It uses EternalLoop's exponential backoff (30s → 1m → 2m → ... → 16m max).
+ *   - Blocked state escalation flows through AlertManager (no longer just displayed).
+ *   - Token budget enforcement via EternalLoop.
+ *   - Circuit breaker on API failures.
+ *
+ * Public API unchanged — LeaderAgent main loop calls checkProgressStagnation()
+ * and maybeEternalIdlePatrol() as before.
+ */
+import type { EventEmitter } from '../core/EventEmitter.js';
+import type { DatabaseManager } from '../core/Database.js';
+import type { TaskBoard } from '../core/TaskBoard.js';
+import type { AgentPool } from './AgentPoolRuntime.js';
+import type { ChatMessage } from '../llm/types.js';
+import { type EternalRuntimeSnapshot, type TokenUsage } from '../core/EternalLoop.js';
+export interface LeaderProgressInvariantDeps {
+    sessionId: string;
+    db: DatabaseManager;
+    emitter: EventEmitter;
+    board: TaskBoard;
+    pool: AgentPool;
+    /** Returns true if the leader session is finished */
+    isFinished: () => boolean;
+    /** Returns current waitingForUser state */
+    isWaitingForUser: () => boolean;
+    /** Returns current pendingReview state */
+    isPendingReview: () => boolean;
+    /** Returns current eternalMode state */
+    isEternalMode: () => boolean;
+    /** Returns true if Leader is currently running (in the main loop, not idle) */
+    isLeaderRunning: () => boolean;
+    /** Get a snapshot of the current conversation array */
+    getConversation: () => ChatMessage[];
+    /** Get conversation length (cheaper than full array) */
+    getConversationLength: () => number;
+    /** Add and persist a message */
+    addAndPersistMessage: (msg: ChatMessage) => Promise<void>;
+    /** Call leaderThinkAndAct */
+    leaderThinkAndAct: () => Promise<void>;
+    /** Set waitingForUser */
+    setWaitingForUser: (waiting: boolean) => Promise<void>;
+    /** Record token usage for budget tracking */
+    recordTokenUsage: (usage: TokenUsage) => void;
+    /** Dispatch ready tasks in parallel (called by EternalLoop when agents are running) */
+    dispatchReadyTasks?: () => Promise<number>;
+    /** Get count of ready-to-dispatch tasks (used by EternalLoop to notify Leader) */
+    getReadyTaskCount?: () => number;
+    /** 黑板节点/边计数；返回 null 表示黑板未就绪。喂给 EternalLoop fingerprint。 */
+    getBlackboardCounts?: () => {
+        nodes: number;
+        edges: number;
+    } | null;
+    /** scratchpad review digest，喂给 EternalLoop fingerprint。 */
+    getScratchpadDigest?: () => string | null;
+    /** 最近 conversation 摘要，喂给 EternalPatrolJudge */
+    getRecentConversationDigest?: () => string;
+    /** Judge LLM client 工厂；EternalPatrolJudge 用 */
+    getEternalJudgeLlm?: () => import('../llm/ContentGenerator.js').ContentGenerator | null;
+    /** Judge model name */
+    getEternalJudgeModel?: () => string | null;
+    /** judge 决定 yield_user 时调用，让 leader 切回等待状态 */
+    yieldEternalToUser?: (reason: string) => Promise<void>;
+}
+export declare class LeaderProgressInvariant {
+    /** Progress stagnation state */
+    private lastProgressHash;
+    private stagnantRoundCount;
+    private static readonly STAGNANT_ROUNDS_THRESHOLD;
+    /** Watchdog state */
+    lastProgressAtMs: number;
+    private watchdogTimer;
+    private static readonly WATCHDOG_TIMEOUT_MS;
+    private static readonly WATCHDOG_POLL_MS;
+    /** Eternal loop (v3 — unified autonomous control) */
+    private eternalLoop;
+    private sessionId;
+    private db;
+    private emitter;
+    private board;
+    private pool;
+    private isFinished;
+    private isWaitingForUser;
+    private isPendingReview;
+    private isEternalMode;
+    private isLeaderRunning;
+    private getConversation;
+    private getConversationLength;
+    private addAndPersistMessage;
+    private leaderThinkAndAct;
+    private setWaitingForUser;
+    private recordTokenUsage;
+    private dispatchReadyTasks?;
+    private getReadyTaskCount?;
+    private getBlackboardCounts?;
+    private getScratchpadDigest?;
+    private getRecentConversationDigest?;
+    private getEternalJudgeLlm?;
+    private getEternalJudgeModel?;
+    private yieldEternalToUser?;
+    /** EventEmitter unsubscribe 句柄；dispose 时逐个调用 */
+    private eternalEventBindings;
+    constructor(deps: LeaderProgressInvariantDeps);
+    private computeProgressHash;
+    checkProgressStagnation(): void;
+    private injectStagnationBreaker;
+    markProgress(): void;
+    startWatchdog(): void;
+    stopWatchdog(): void;
+    runWatchdogCheck(): void;
+    /**
+     * Initialize the EternalLoop on first call.
+     * Separate from constructor because it depends on being in eternal mode.
+     */
+    private ensureEternalLoop;
+    /**
+     * 把外部状态变化事件转发到 EternalLoop.invalidateSilenceLock，
+     * 让 silence 期间真有进展时立刻解锁。
+     */
+    private bindEternalListeners;
+    /**
+     * Session 切换 / Leader 终止 / 切回 manual 模式时调用。
+     * 解绑 7 类事件监听 + 重置 silence lock，避免下次切回 eternal 时复活旧 fingerprint state。
+     */
+    disposeEternalListeners(): void;
+    /**
+     * 切到 eternal 时调用：若 EternalLoop 已存在但 listener 因之前切回 manual 已被解绑，
+     * 这里补上重新绑定；若 loop 还没建则等 maybeEternalIdlePatrol 懒创建。
+     */
+    rebindEternalListenersIfActive(): void;
+    /**
+     * 由 LeaderAgent 主循环 / 用户消息处理点调用，主动唤醒 silence lock。
+     */
+    invalidateEternalSilenceLock(reason?: string): void;
+    /**
+     * 主循环入口 — 替代旧 maybeEternalIdlePatrol()。
+     * 委托 EternalLoop.tick() 执行统一控制回路。
+     * 无硬上限：EternalLoop 的指数退避保证不会 spin。
+     */
+    maybeEternalIdlePatrol(): Promise<boolean>;
+    /**
+     * 执行一次 patrol（执行 prompt 注入 + LLM 调用）。
+     * 返回 true 表示 LLM 执行了工具调用。
+     */
+    private doEternalPatrol;
+    /**
+     * 从 DB 恢复 EternalLoop 状态（跨 session 重启）
+     */
+    hydrateEternalState(): Promise<void>;
+    /**
+     * 记录 token 使用（每次 LLM 调用后）
+     */
+    recordTokens(usage: TokenUsage): void;
+    /**
+     * 记录 API 成功/失败（供 circuit breaker）
+     */
+    recordApiResult(success: boolean): void;
+    /**
+     * 获取 EternalLoop 状态（供 /doctor 诊断）
+     */
+    getEternalLoopState(): import("../core/EternalLoop.js").EternalLoopState | null;
+    /**
+     * 获取 EternalMode 运行时快照。不会为了观测而创建 EternalLoop，
+     * 避免 runtime_state 查询产生 listener 绑定等副作用。
+     */
+    getEternalRuntimeSnapshot(): EternalRuntimeSnapshot;
+}
+//# sourceMappingURL=LeaderProgressInvariant.d.ts.map

package/dist/agents/LeaderProgressInvariant.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"LeaderProgressInvariant.d.ts","sourceRoot":"","sources":["../../src/agents/LeaderProgressInvariant.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,yBAAyB,CAAC;AAC5D,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAC3D,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,iBAAiB,CAAC;AAInD,OAAO,EAIL,KAAK,sBAAsB,EAC3B,KAAK,UAAU,EAGhB,MAAM,wBAAwB,CAAC;AAKhC,MAAM,WAAW,2BAA2B;IAC1C,SAAS,EAAE,MAAM,CAAC;IAClB,EAAE,EAAE,eAAe,CAAC;IACpB,OAAO,EAAE,YAAY,CAAC;IACtB,KAAK,EAAE,SAAS,CAAC;IACjB,IAAI,EAAE,SAAS,CAAC;IAChB,qDAAqD;IACrD,UAAU,EAAE,MAAM,OAAO,CAAC;IAC1B,2CAA2C;IAC3C,gBAAgB,EAAE,MAAM,OAAO,CAAC;IAChC,0CAA0C;IAC1C,eAAe,EAAE,MAAM,OAAO,CAAC;IAC/B,wCAAwC;IACxC,aAAa,EAAE,MAAM,OAAO,CAAC;IAC7B,+EAA+E;IAC/E,eAAe,EAAE,MAAM,OAAO,CAAC;IAC/B,uDAAuD;IACvD,eAAe,EAAE,MAAM,WAAW,EAAE,CAAC;IACrC,wDAAwD;IACxD,qBAAqB,EAAE,MAAM,MAAM,CAAC;IACpC,gCAAgC;IAChC,oBAAoB,EAAE,CAAC,GAAG,EAAE,WAAW,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1D,6BAA6B;IAC7B,iBAAiB,EAAE,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC;IACvC,yBAAyB;IACzB,iBAAiB,EAAE,CAAC,OAAO,EAAE,OAAO,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;IACvD,6CAA6C;IAC7C,gBAAgB,EAAE,CAAC,KAAK,EAAE,UAAU,KAAK,IAAI,CAAC;IAC9C,uFAAuF;IACvF,kBAAkB,CAAC,EAAE,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;IAC3C,kFAAkF;IAClF,iBAAiB,CAAC,EAAE,MAAM,MAAM,CAAC;IACjC,2DAA2D;IAC3D,mBAAmB,CAAC,EAAE,MAAM;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI,CAAC;IACpE,2DAA2D;IAC3D,mBAAmB,CAAC,EAAE,MAAM,MAAM,GAAG,IAAI,CAAC;IAC1C,+CAA+C;IAC/C,2BAA2B,CAAC,EAAE,MAAM,MAAM,CAAC;IAC3C,+CAA+C;IAC/C,kBAAkB,CAAC,EAAE,MAAM,OAAO,4BAA4B,EAAE,gBAAgB,GAAG,IAAI,CAAC;IACxF,uBAAuB;IACvB,oBAAoB,CAAC,EAAE,MAAM,MAAM,GAAG,IAAI,CAAC;IAC3C,8CAA8C;IAC9C,kBAAkB,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;CACxD;AAED,qBAAa,uBAAuB;IAClC,gCAAgC;IAChC,OAAO,CAAC,gBAAgB,CAAM;IAC9B,OAAO,CAAC,kBAAkB,CAAK;IAC/B,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,yBAAyB,CAAK;IAEtD,qBAAqB;IACrB,gBAAgB,SAAc;IAC9B,OAAO,CAAC,aAAa,CAA+C;IACpE,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,mBAAmB,CAAiB;IAC5D,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,gBAAgB,CAAa;IAErD,qDAAqD;IACrD,OAAO,CAAC,WAAW,CAA4B;IAE/C,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,EAAE,CAAkB;IAC5B,OAAO,CAAC,OAAO,CAAe;IAC9B,OAAO,CAAC,KAAK,CAAY;IACzB,OAAO,CAAC,IAAI,CAAY;IACxB,OAAO,CAAC,UAAU,CAAgB;IAClC,OAAO,CAAC,gBAAgB,CAAgB;IACxC,OAAO,CAAC,eAAe,CAAgB;IACvC,OAAO,CAAC,aAAa,CAAgB;IACrC,OAAO,CAAC,eAAe,CAAgB;IACvC,OAAO,CAAC,eAAe,CAAsB;IAC7C,OAAO,CAAC,qBAAqB,CAAe;IAC5C,OAAO,CAAC,oBAAoB,CAAsC;IAClE,OAAO,CAAC,iBAAiB,CAAsB;IAC/C,OAAO,CAAC,iBAAiB,CAAsC;IAC/D,OAAO,CAAC,gBAAgB,CAA8B;IACtD,OAAO,CAAC,kBAAkB,CAAC,CAAwB;IACnD,OAAO,CAAC,iBAAiB,CAAC,CAAe;IACzC,OAAO,CAAC,mBAAmB,CAAC,CAAgD;IAC5E,OAAO,CAAC,mBAAmB,CAAC,CAAsB;IAClD,OAAO,CAAC,2BAA2B,CAAC,CAAe;IACnD,OAAO,CAAC,kBAAkB,CAAC,CAAqE;IAChG,OAAO,CAAC,oBAAoB,CAAC,CAAsB;IACnD,OAAO,CAAC,kBAAkB,CAAC,CAAoC;IAE/D,gDAAgD;IAChD,OAAO,CAAC,oBAAoB,CAAyB;gBAEzC,IAAI,EAAE,2BAA2B;IA+B7C,OAAO,CAAC,mBAAmB;IAY3B,uBAAuB,IAAI,IAAI;IAqC/B,OAAO,CAAC,uBAAuB;IAqB/B,YAAY,IAAI,IAAI;IAIpB,aAAa,IAAI,IAAI;IAUrB,YAAY,IAAI,IAAI;IAQpB,gBAAgB,IAAI,IAAI;IAoExB;;;OAGG;IACH,OAAO,CAAC,iBAAiB;IAsCzB;;;OAGG;IACH,OAAO,CAAC,oBAAoB;IAe5B;;;OAGG;IACH,uBAAuB,IAAI,IAAI;IAU/B;;;OAGG;IACH,8BAA8B,IAAI,IAAI;IAMtC;;OAEG;IACH,4BAA4B,CAAC,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI;IAInD;;;;OAIG;IACG,sBAAsB,IAAI,OAAO,CAAC,OAAO,CAAC;IAMhD;;;OAGG;YACW,eAAe;IAmH7B;;OAEG;IACG,mBAAmB,IAAI,OAAO,CAAC,IAAI,CAAC;IAM1C;;OAEG;IACH,YAAY,CAAC,KAAK,EAAE,UAAU,GAAG,IAAI;IAKrC;;OAEG;IACH,eAAe,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAIvC;;OAEG;IACH,mBAAmB;IAInB;;;OAGG;IACH,yBAAyB,IAAI,sBAAsB;CAKpD"}

package/dist/agents/LeaderProgressInvariant.js

@@ -0,0 +1,479 @@
+/**
+ * LeaderProgressInvariant — V3
+ *
+ * Manages progress stagnation detection, watchdog timer, and delegates
+ * eternal idle patrol to EternalLoop (unified autonomous control loop).
+ *
+ * Key changes from V2:
+ *   - Eternal patrol no longer has a hard 10-round cap.
+ *     It uses EternalLoop's exponential backoff (30s → 1m → 2m → ... → 16m max).
+ *   - Blocked state escalation flows through AlertManager (no longer just displayed).
+ *   - Token budget enforcement via EternalLoop.
+ *   - Circuit breaker on API failures.
+ *
+ * Public API unchanged — LeaderAgent main loop calls checkProgressStagnation()
+ * and maybeEternalIdlePatrol() as before.
+ */
+import { leaderLogger } from '../core/Log.js';
+import { readPersistedEternalGoal } from '../core/EternalGoal.js';
+import { EternalLoop, createInitialEternalRuntimeSnapshot, } from '../core/EternalLoop.js';
+// EternalPatrolJudge (LLM-based) removed — eternal mode uses deterministic runtime policy only.
+import { decideEternalActionFromRuntimeState } from '../contracts/adapters/EternalPatrolPolicy.js';
+import { getPromptCatalog } from './prompts/i18n/catalog.js';
+export class LeaderProgressInvariant {
+    /** Progress stagnation state */
+    lastProgressHash = '';
+    stagnantRoundCount = 0;
+    static STAGNANT_ROUNDS_THRESHOLD = 5;
+    /** Watchdog state */
+    lastProgressAtMs = Date.now();
+    watchdogTimer = null;
+    static WATCHDOG_TIMEOUT_MS = 5 * 60 * 1000; // 5 minutes
+    static WATCHDOG_POLL_MS = 60 * 1000; // 1 minute
+    /** Eternal loop (v3 — unified autonomous control) */
+    eternalLoop = null;
+    sessionId;
+    db;
+    emitter;
+    board;
+    pool;
+    isFinished;
+    isWaitingForUser;
+    isPendingReview;
+    isEternalMode;
+    isLeaderRunning;
+    getConversation;
+    getConversationLength;
+    addAndPersistMessage;
+    leaderThinkAndAct;
+    setWaitingForUser;
+    recordTokenUsage;
+    dispatchReadyTasks;
+    getReadyTaskCount;
+    getBlackboardCounts;
+    getScratchpadDigest;
+    getRecentConversationDigest;
+    getEternalJudgeLlm;
+    getEternalJudgeModel;
+    yieldEternalToUser;
+    /** EventEmitter unsubscribe 句柄；dispose 时逐个调用 */
+    eternalEventBindings = [];
+    constructor(deps) {
+        this.sessionId = deps.sessionId;
+        this.db = deps.db;
+        this.emitter = deps.emitter;
+        this.board = deps.board;
+        this.pool = deps.pool;
+        this.isFinished = deps.isFinished;
+        this.isWaitingForUser = deps.isWaitingForUser;
+        this.isPendingReview = deps.isPendingReview;
+        this.isEternalMode = deps.isEternalMode;
+        this.isLeaderRunning = deps.isLeaderRunning ?? (() => true);
+        this.getConversation = deps.getConversation;
+        this.getConversationLength = deps.getConversationLength;
+        this.addAndPersistMessage = deps.addAndPersistMessage;
+        this.leaderThinkAndAct = deps.leaderThinkAndAct;
+        this.setWaitingForUser = deps.setWaitingForUser;
+        this.recordTokenUsage = deps.recordTokenUsage;
+        this.dispatchReadyTasks = deps.dispatchReadyTasks;
+        this.getReadyTaskCount = deps.getReadyTaskCount;
+        this.getBlackboardCounts = deps.getBlackboardCounts;
+        this.getScratchpadDigest = deps.getScratchpadDigest;
+        this.getRecentConversationDigest = deps.getRecentConversationDigest;
+        this.getEternalJudgeLlm = deps.getEternalJudgeLlm;
+        this.getEternalJudgeModel = deps.getEternalJudgeModel;
+        this.yieldEternalToUser = deps.yieldEternalToUser;
+    }
+    // ─────────────────────────────────────────────────────────────────────
+    // Progress Invariant (unchanged from V2)
+    // ─────────────────────────────────────────────────────────────────────
+    computeProgressHash() {
+        const tasks = this.board.getAllTasks();
+        const running = this.pool.getRunning().map(a => a.name).sort().join(',');
+        const stateStr = tasks.map(t => `${t.id}:${t.status}`).join('|') + '|running:' + running;
+        let hash = 5381;
+        for (let i = 0; i < stateStr.length; i++) {
+            hash = ((hash << 5) + hash) ^ stateStr.charCodeAt(i);
+            hash = hash >>> 0;
+        }
+        return hash.toString(16);
+    }
+    checkProgressStagnation() {
+        // Leader 自身正在运行（如 LLM 调用、工具执行），不算停滞
+        if (this.isLeaderRunning()) {
+            this.stagnantRoundCount = 0;
+            this.lastProgressAtMs = Date.now();
+            return;
+        }
+        if (this.pool.getRunning().length > 0) {
+            this.stagnantRoundCount = 0;
+            return;
+        }
+        if (this.isWaitingForUser() || this.isPendingReview()) {
+            this.stagnantRoundCount = 0;
+            return;
+        }
+        const hash = this.computeProgressHash();
+        if (hash === this.lastProgressHash) {
+            this.stagnantRoundCount++;
+            if (this.stagnantRoundCount >= LeaderProgressInvariant.STAGNANT_ROUNDS_THRESHOLD) {
+                leaderLogger.warn(`[ProgressInvariant] 检测到进度停滞: 连续 ${this.stagnantRoundCount} 轮哈希不变 (${hash})`);
+                this.emitter.emit('leader:progress_stagnant', {
+                    sessionId: this.sessionId,
+                    consecutiveStagnantRounds: this.stagnantRoundCount,
+                    progressHash: hash,
+                });
+                this.injectStagnationBreaker(this.stagnantRoundCount);
+                this.stagnantRoundCount = 0;
+            }
+        }
+        else {
+            this.lastProgressHash = hash;
+            this.stagnantRoundCount = 0;
+        }
+    }
+    injectStagnationBreaker(rounds) {
+        const breakerMsg = [
+            `⚠️ [系统检测] 你已连续 ${rounds} 轮没有推进任何任务。`,
+            '请立即执行以下自检：',
+            '1. 检查当前 task board 状态，是否有被阻塞的任务',
+            '2. 如果所有任务已完成，请调用 finish_session 结束会话',
+            '3. 如果存在未解决的依赖，请打破依赖链或将问题升级为新任务',
+            '4. 如果你不确定下一步，请用 ask_user 向用户确认',
+            '请立即采取行动，并使用新的证据、参数或工具路径。',
+        ].join('\n');
+        void this.addAndPersistMessage({ role: 'system', content: breakerMsg }).catch((err) => {
+            leaderLogger.warn(`[ProgressInvariant] 注入 breaker 失败: ${err instanceof Error ? err.message : String(err)}`);
+        });
+        leaderLogger.info('[ProgressInvariant] 已注入破局提示，重置停滞计数');
+    }
+    // ─────────────────────────────────────────────────────────────────────
+    // Watchdog (unchanged from V2)
+    // ─────────────────────────────────────────────────────────────────────
+    markProgress() {
+        this.lastProgressAtMs = Date.now();
+    }
+    startWatchdog() {
+        if (this.watchdogTimer)
+            return;
+        this.lastProgressAtMs = Date.now();
+        this.watchdogTimer = setInterval(() => {
+            this.runWatchdogCheck();
+        }, LeaderProgressInvariant.WATCHDOG_POLL_MS);
+        this.watchdogTimer.unref?.();
+        leaderLogger.debug('[Watchdog] 看门狗已启动');
+    }
+    stopWatchdog() {
+        if (this.watchdogTimer) {
+            clearInterval(this.watchdogTimer);
+            this.watchdogTimer = null;
+            leaderLogger.debug('[Watchdog] 看门狗已停止');
+        }
+    }
+    runWatchdogCheck() {
+        if (this.isFinished() || this.isWaitingForUser() || this.isPendingReview())
+            return;
+        // Leader 自身正在运行（LLM 调用/工具执行），重置时间戳
+        if (this.isLeaderRunning()) {
+            this.lastProgressAtMs = Date.now();
+            return;
+        }
+        // 智能 Agent watchdog：检查运行中的 Agent 是否有停滞
+        const runningAgents = this.pool.getRunning();
+        if (runningAgents.length > 0) {
+            const stagnant = this.pool.getStagnantAgents();
+            if (stagnant.length > 0) {
+                // 有 Agent 停滞，注入诊断消息给 Leader
+                const diagnostics = stagnant.map(s => `  - @${s.handle.name} (task:${s.handle.taskId}): ${s.reason} (${Math.round(s.elapsedMs / 1000)}s)${s.handle.currentToolName ? ` [tool: ${s.handle.currentToolName}]` : ''}`).join('\n');
+                const msg = [
+                    `⚠️ [Agent Watchdog] 检测到 ${stagnant.length} 个 Agent 可能停滞：`,
+                    diagnostics,
+                    '建议操作：',
+                    '- tool_stalled: 先使用 check_agent_progress 确认当前工具/日志，再优先 nudge_agent 或 retry_agent_llm',
+                    '- idle_no_progress: 使用 check_agent_progress 确认状态；不能只看任务板或时间判断',
+                    '- heartbeat_lost: 使用 check_agent_progress 确认心跳/最近进展；如确认崩溃或不可恢复，可 terminate_agent 并重新派发',
+                ].join('\n');
+                void this.addAndPersistMessage({ role: 'system', content: msg });
+                this.emitter.emit('leader:watchdog_alert', {
+                    sessionId: this.sessionId,
+                    elapsedMs: stagnant[0].elapsedMs,
+                    thresholdMs: LeaderProgressInvariant.WATCHDOG_TIMEOUT_MS,
+                    intervention: `Agent stagnation: ${stagnant.map(s => `${s.handle.name}:${s.reason}`).join(', ')}`,
+                });
+            }
+            // Agents are running (even if some stagnant), reset leader progress
+            this.lastProgressAtMs = Date.now();
+            return;
+        }
+        const elapsedMs = Date.now() - this.lastProgressAtMs;
+        if (elapsedMs < LeaderProgressInvariant.WATCHDOG_TIMEOUT_MS)
+            return;
+        const intervention = `[Watchdog] Leader 已 ${Math.round(elapsedMs / 60000)} 分钟无进度推进，自动触发自检`;
+        leaderLogger.warn(intervention);
+        this.emitter.emit('leader:watchdog_alert', {
+            sessionId: this.sessionId,
+            elapsedMs,
+            thresholdMs: LeaderProgressInvariant.WATCHDOG_TIMEOUT_MS,
+            intervention,
+        });
+        const watchdogMsg = [
+            `⚠️ [Watchdog] 系统检测到你已超过 ${Math.round(elapsedMs / 60000)} 分钟没有推进任何工作。`,
+            '请立即：',
+            '1. 检查 task board 状态（任务板状态摘要每轮已自动注入）',
+            '2. 检查是否有 Agent 需要你回复',
+            '3. 如果工作已完成，调用 finish_session 结束会话',
+            '4. 如果遇到阻塞，向用户寻求帮助',
+        ].join('\n');
+        void this.addAndPersistMessage({ role: 'system', content: watchdogMsg });
+        this.lastProgressAtMs = Date.now();
+    }
+    // ─────────────────────────────────────────────────────────────────────
+    // Eternal Patrol (V3 — delegates to EternalLoop, no hard cap)
+    // ─────────────────────────────────────────────────────────────────────
+    /**
+     * Initialize the EternalLoop on first call.
+     * Separate from constructor because it depends on being in eternal mode.
+     */
+    ensureEternalLoop() {
+        if (!this.eternalLoop) {
+            const deps = {
+                sessionId: this.sessionId,
+                db: this.db,
+                emitter: this.emitter,
+                board: this.board,
+                isEternalMode: this.isEternalMode,
+                isFinished: this.isFinished,
+                isWaitingForUser: this.isWaitingForUser,
+                isPendingReview: this.isPendingReview,
+                hasRunningAgents: () => this.pool.getRunning().length > 0,
+                getRunningAgentCount: () => this.pool.getRunning().length,
+                getMaxConcurrent: () => {
+                    const status = this.pool.getStatus();
+                    return Math.max(status.total, 5);
+                },
+                getConversationLength: this.getConversationLength,
+                getEternalGoal: () => readPersistedEternalGoal(this.db, this.sessionId),
+                recordTokenUsage: this.recordTokenUsage,
+                executeEternalPatrol: async (patrolNumber, totalPatrols, candidates) => {
+                    return await this.doEternalPatrol(patrolNumber, totalPatrols, candidates);
+                },
+                dispatchReadyTasks: this.dispatchReadyTasks ?? (async () => 0),
+                getReadyTaskCount: this.getReadyTaskCount,
+                getBlackboardCounts: this.getBlackboardCounts,
+                getScratchpadDigest: this.getScratchpadDigest,
+                getRecentConversationDigest: this.getRecentConversationDigest,
+                judgeFn: (input) => Promise.resolve(decideEternalActionFromRuntimeState(input)),
+                yieldToUser: this.yieldEternalToUser,
+            };
+            this.eternalLoop = new EternalLoop({}, deps);
+            this.bindEternalListeners(this.eternalLoop);
+        }
+        return this.eternalLoop;
+    }
+    /**
+     * 把外部状态变化事件转发到 EternalLoop.invalidateSilenceLock，
+     * 让 silence 期间真有进展时立刻解锁。
+     */
+    bindEternalListeners(loop) {
+        const invalidate = (reason) => () => loop.invalidateSilenceLock(reason);
+        const noteWorker = () => loop.noteWorkerCompletion();
+        this.eternalEventBindings = [
+            this.emitter.subscribe('task:completed', invalidate('task:completed')),
+            this.emitter.subscribe('agent:completed', invalidate('agent:completed')),
+            this.emitter.subscribe('agent:crashed', invalidate('agent:crashed')),
+            this.emitter.subscribe('worker:complete', noteWorker),
+            this.emitter.subscribe('worker:failed', noteWorker),
+            this.emitter.subscribe('blackboard:delta', invalidate('blackboard:delta')),
+            this.emitter.subscribe('team:message_sent', invalidate('team:message_sent')),
+        ];
+    }
+    /**
+     * Session 切换 / Leader 终止 / 切回 manual 模式时调用。
+     * 解绑 7 类事件监听 + 重置 silence lock，避免下次切回 eternal 时复活旧 fingerprint state。
+     */
+    disposeEternalListeners() {
+        for (const off of this.eternalEventBindings) {
+            try {
+                off();
+            }
+            catch { /* ignore */ }
+        }
+        this.eternalEventBindings = [];
+        // 彻底重置 EternalLoop 跨控制模式的状态：silence lock + idle streak + interval + outcome。
+        // 不能只 invalidateSilenceLock，否则旧 idle backoff 会在下次切回 eternal 后继续生效。
+        this.eternalLoop?.resetForControlModeSwitch();
+    }
+    /**
+     * 切到 eternal 时调用：若 EternalLoop 已存在但 listener 因之前切回 manual 已被解绑，
+     * 这里补上重新绑定；若 loop 还没建则等 maybeEternalIdlePatrol 懒创建。
+     */
+    rebindEternalListenersIfActive() {
+        if (!this.eternalLoop)
+            return; // 还没创建过 → 下一次 tick 时 ensureEternalLoop 会顺带绑定
+        if (this.eternalEventBindings.length > 0)
+            return; // 已绑过
+        this.bindEternalListeners(this.eternalLoop);
+    }
+    /**
+     * 由 LeaderAgent 主循环 / 用户消息处理点调用，主动唤醒 silence lock。
+     */
+    invalidateEternalSilenceLock(reason) {
+        this.eternalLoop?.invalidateSilenceLock(reason);
+    }
+    /**
+     * 主循环入口 — 替代旧 maybeEternalIdlePatrol()。
+     * 委托 EternalLoop.tick() 执行统一控制回路。
+     * 无硬上限：EternalLoop 的指数退避保证不会 spin。
+     */
+    async maybeEternalIdlePatrol() {
+        if (!this.isEternalMode())
+            return false;
+        const loop = this.ensureEternalLoop();
+        return await loop.tick();
+    }
+    /**
+     * 执行一次 patrol（执行 prompt 注入 + LLM 调用）。
+     * 返回 true 表示 LLM 执行了工具调用。
+     */
+    async doEternalPatrol(patrolNumber, consecutiveIdle, candidates = []) {
+        const convLenBefore = this.getConversationLength();
+        const goal = readPersistedEternalGoal(this.db, this.sessionId);
+        const hasGoal = Boolean(goal?.description?.trim());
+        const goalLines = goal
+            ? [
+                '**Eternal 目标（最高优先级，不是背景偏好）**：',
+                goal.description,
+                '',
+                '你当前处于 `/eternal <目标>` 的目标模式：职责是持续拆解、执行、验证并交付这个目标，直到目标明确完成、暂停或删除。',
+                '不要把“当前没有现成任务”理解成目标结束；没有任务就创建下一步可闭环任务，或在安全边界内直接推进。',
+                '所有自主研发、修 bug、补测试、写文档和派发 Agent 都必须优先服务这个目标。发现偏离时，先把方向拉回目标。',
+                '',
+            ]
+            : [];
+        const isEarlyPatrol = consecutiveIdle < 3;
+        const candidatePrompt = candidates.length > 0
+            ? [
+                '**结构化候选任务（按 deterministic priority score 排序）**：',
+                hasGoal
+                    ? '从以下候选任务中选择最能推进 Eternal 目标的一项执行。返回 JSON: {"chosen_task_id":"...","reason":"..."}，随后立即 dispatch 或直接执行。'
+                    : '从以下候选任务中选择一个执行。返回 JSON: {"chosen_task_id":"...","reason":"..."}，随后立即 dispatch 或直接执行。',
+                `候选: ${JSON.stringify(candidates.map((candidate) => ({
+                    id: candidate.id,
+                    title: candidate.title,
+                    score: candidate.score,
+                    type: candidate.type,
+                    reason: candidate.reason,
+                })))}`,
+                '',
+            ]
+            : [];
+        const patrolPrompt = isEarlyPatrol
+            ? [
+                hasGoal ? '[Eternal Mode — 目标完成轮次]' : '[Eternal Mode — 自主研发轮次]',
+                hasGoal
+                    ? '用户通过 /eternal 给出了要一直完成的目标。你的职责是像目标模式一样保持承诺：围绕目标找出下一步、执行、验证、收尾。'
+                    : '所有已知任务已完成，系统进入自主研发状态。你的职责是持续推进项目，做真正有价值的工作。',
+                '',
+                ...goalLines,
+                hasGoal ? '**目标推进决策框架（按优先级）**：' : '**自主研发决策框架（按优先级）**：',
+                hasGoal
+                    ? '1. **目标差距优先**：先判断目标还缺什么证据、实现、测试或文档；缺口就是下一步任务'
+                    : '1. **Bug 修复优先**：扫描最近 git log、error 日志、测试失败，发现 bug 立即创建修复任务',
+                hasGoal
+                    ? '2. **可交付闭环**：把目标拆成能创建、派发、验收的小任务；每个任务都写清验收标准'
+                    : '2. **功能完善**：对比已有功能与最佳实践，识别明显缺失的功能并实现',
+                hasGoal
+                    ? '3. **验证优先**：目标相关改动必须用测试、构建、日志或代码证据证明完成'
+                    : '3. **技术债清理**：TODO/FIXME/HACK、类型不安全、重复代码、性能瓶颈',
+                hasGoal
+                    ? '4. **阻塞处理**：如果目标被风险或缺口卡住，先做可验证侦察、最小修复或提出 grounded 决策'
+                    : '4. **测试补全**：核心逻辑覆盖不足时补充测试',
+                hasGoal
+                    ? getPromptCatalog().leader.eternalGoal.patrolDeliveryStep
+                    : '5. **文档更新**：API 变更后文档未同步、缺失使用示例',
+                '',
+                '**决策原则**：',
+                hasGoal
+                    ? '- 优先选择最能推进 Eternal 目标的工作，而不是泛化项目美化'
+                    : '- 优先选择影响面大、用户价值高的工作，而非简单的格式修复',
+                hasGoal
+                    ? '- 每轮至少产出一个推进目标的实质动作；除非目标已有完成证据'
+                    : '- 每轮至少产出一个有实质内容的任务（除非项目已无改进空间）',
+                '- 创建任务后立即 dispatch，以行动闭环替代纯分析',
+                hasGoal
+                    ? '- 不重复上一轮无产出的尝试；换一个更接近目标完成的切入点'
+                    : '- 不重复上一轮刚完成的工作类型',
+                '',
+                ...candidatePrompt,
+                hasGoal
+                    ? '发现能推进目标的工作就立即 create_task + dispatch_agent。'
+                    : '发现工作就立即 create_task + dispatch_agent。',
+                '你也可以直接动手改代码、跑测试、修 bug，不必每次都创建任务。',
+                hasGoal
+                    ? getPromptCatalog().leader.eternalGoal.patrolCompletionGuidanceEarly
+                    : '如果深度分析后确实没有高价值工作，简要说明即可。',
+                '',
+                `（Patrol #${patrolNumber}，已连续空闲 ${consecutiveIdle} 轮）`,
+            ].join('\n')
+            : [
+                hasGoal ? '[Eternal Mode — 目标完成轮次]' : '[Eternal Mode — 自主研发轮次]',
+                hasGoal
+                    ? `持续完成 Eternal 目标（Patrol #${patrolNumber}，已连续空闲 ${consecutiveIdle} 轮）。`
+                    : `持续推进项目（Patrol #${patrolNumber}，已连续空闲 ${consecutiveIdle} 轮）。`,
+                ...goalLines,
+                ...candidatePrompt,
+                hasGoal
+                    ? '按目标推进框架寻找下一步，发现即 create_task + dispatch_agent，也可以直接动手。'
+                    : '按之前定义的决策框架寻找新工作，发现即 create_task + dispatch_agent，也可以直接动手。',
+                hasGoal
+                    ? getPromptCatalog().leader.eternalGoal.patrolCompletionGuidanceLate
+                    : '确无高价值工作时无需执行工具，简要说明即可。',
+            ].join('\n');
+        // GLM 等模型拒绝只有 system 消息、没有 user 消息的请求（返回 400）。
+        // patrol prompt 改用 role: 'user' 注入，确保请求始终包含 user 消息。
+        await this.addAndPersistMessage({ role: 'user', content: patrolPrompt });
+        await this.leaderThinkAndAct();
+        // Detect tool calls from new messages
+        const newMessages = this.getConversation()
+            .slice(-(this.getConversationLength() - convLenBefore));
+        const hadToolCalls = newMessages.some((msg) => msg.role === 'assistant' && Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0);
+        return hadToolCalls;
+    }
+    /**
+     * 从 DB 恢复 EternalLoop 状态（跨 session 重启）
+     */
+    async hydrateEternalState() {
+        if (this.eternalLoop) {
+            await this.eternalLoop.hydrate();
+        }
+    }
+    /**
+     * 记录 token 使用（每次 LLM 调用后）
+     */
+    recordTokens(usage) {
+        this.recordTokenUsage(usage);
+        this.eternalLoop?.recordTokens(usage);
+    }
+    /**
+     * 记录 API 成功/失败（供 circuit breaker）
+     */
+    recordApiResult(success) {
+        this.eternalLoop?.recordApiResult(success);
+    }
+    /**
+     * 获取 EternalLoop 状态（供 /doctor 诊断）
+     */
+    getEternalLoopState() {
+        return this.eternalLoop?.state ?? null;
+    }
+    /**
+     * 获取 EternalMode 运行时快照。不会为了观测而创建 EternalLoop，
+     * 避免 runtime_state 查询产生 listener 绑定等副作用。
+     */
+    getEternalRuntimeSnapshot() {
+        const goal = readPersistedEternalGoal(this.db, this.sessionId);
+        return this.eternalLoop?.toRuntimeSnapshot(this.isEternalMode())
+            ?? createInitialEternalRuntimeSnapshot(this.isEternalMode(), {}, Date.now(), goal);
+    }
+}
+//# sourceMappingURL=LeaderProgressInvariant.js.map