@capybara-hq/kernel 0.1.0-rc.3 → 0.1.0-rc.5

package/dist/runtime/run-queue.d.ts

@@ -1,6 +1,8 @@
 import type { Logger } from "@capybara-hq/plugin-sdk";
+import type { Queue } from "bullmq";
 import { EventLog } from "../conversations/event-log.js";
 import type { ConversationKind } from "../db/kernel-schema.js";
+import { KernelConfigService } from "../infra/config/kernel-config.service.js";
 import { KernelDb } from "../infra/database/kernel-db.js";
 import { TransactionHost } from "../infra/database/transaction-host.js";
 import { KernelEvents } from "../infra/events/kernel-events.js";
@@ -29,11 +31,16 @@ export declare class RunQueue {
     private readonly events;
     private readonly log;
     readonly txHost: TransactionHost;
+    private readonly config;
+    /** The producer side of the delayed `bg-wake` job (`MaintenanceProcessor` consumes it). */
+    private readonly maintenance;
     constructor(kdb: KernelDb, eventLog: EventLog, lock: Pick<ConversationLock, "holder">, jobs: RunJobs, finisher: Pick<RunFinish, "finishIn">, 
     /** 运行以孤儿结束（标记失败 / 直接中止）后删除它的 checkpoint thread（thread_id = runId）。 */
     checkpoints: Pick<KyselyCheckpointSaver, "deleteThread">, 
     /** The run's open `Agent` call results, in the orphan repair's transaction (sub-agents spec §12). */
-    pipeline: Pick<ToolPipeline, "stopOpenAgentCalls">, events: KernelEvents, log: Pick<Logger, "error" | "warn">, txHost: TransactionHost);
+    pipeline: Pick<ToolPipeline, "stopOpenAgentCalls">, events: KernelEvents, log: Pick<Logger, "error" | "warn">, txHost: TransactionHost, config: KernelConfigService, 
+    /** The producer side of the delayed `bg-wake` job (`MaintenanceProcessor` consumes it). */
+    maintenance: Pick<Queue, "add">);
     private get db();
     /**
      * The `beforeTerminal` of an orphan repair: writes the run's open `Agent` call results in the
@@ -44,6 +51,13 @@ export declare class RunQueue {
      * 有未消费的用户消息（最新 user_message 的 seq > `consumed_seq`）、会话锁无人持有、
      * 且没有 queued / running / waiting 的运行时：插入 `runs`（queued）并投递任务，返回 runId；否则返回 null。
      *
+     * The boundary (`trigger_seq`) is that newest message; with none unconsumed, the queue's head
+     * once it is sent; with no head either, a wake-eligible finished background task or sub-agent
+     * delivery (background Bash spec §3, agent map spec §5.3, {@link wakeDecisionIn}): the run is
+     * then attributed to the event's trigger account and joins the tree's budget window, and an
+     * event inside the wake window schedules the delayed `bg-wake` job instead. A sub-agent's
+     * conversation is only woken this way ({@link childWakeIn}).
+     *
      * 孤儿清理（R29，按需进行，没有定时器）：已有的 queued / running 运行若 BullMQ 任务已不存在、
      * 会话锁也无人持有（queued 的还要超过宽限期），说明没有任何 worker 会再执行它——写
      * `run_failed { code: "run.orphaned" }` 把它标记为失败（consumed_seq 推进到它的触发消息），
@@ -52,9 +66,52 @@ export declare class RunQueue {
      *
      * agent 已软删除的会话不起新运行（孤儿清理照常进行）。
      */
-    ensureRun(conversationId: string): Promise<string | null>;
+    ensureRun(conversationId: string, 
+    /** From the `bg-wake` job: it fired `wakeRetry - 1` times before its window ended. */
+    opts?: {
+        wakeRetry?: number;
+    }): Promise<string | null>;
     /** {@link ensureRun}'s transaction: what it decided goes back through `st`. */
     private ensureRunIn;
+    /**
+     * {@link ensureRunIn} for a sub-agent's conversation (agent map spec §5.3), under its row lock.
+     * Its runs from a task, or from a message to an idle sub-agent, are started by the dispatch
+     * (`Agent`, `SendMessage`), which also records the dispatch. Here it is started from a parent's
+     * message that reached its run after that run's last model call (agent map spec §10: a message
+     * racing the run's end starts a new run, never lost); and it is woken by its own background
+     * task's finish or a delivery from its own sub-agent, under the same rules as a chat
+     * conversation. Neither happens once its last run was stopped (`aborted`): a stopped sub-agent is
+     * started again only by its parent. Both read the queue and the trigger account's access on the root
+     * of its tree; a message needs no wake setting and has no wake window. The run counts against
+     * the root's budget window and runs on the sub-agent queue under `runJobId(child, boundary
+     * seq)`. An active run is left to the orphan repair after the commit (`repairRunId`), never
+     * started past.
+     */
+    private childWakeIn;
+    /**
+     * The newest unconsumed message of an idle sub-agent (a `SendMessage` its last run did not absorb)
+     * as a run boundary, or null: none; the root's queue is paused or the root deleted (the
+     * message waits for the unpause, {@link wakeHeldTree}); or its author can no longer send to the
+     * root (it waits for a later run). The caller already refused a stopped sub-agent (last run
+     * `aborted`): a message that reached it while running and then stopped is parked, and the next
+     * run its parent starts absorbs it. Called under the sub-agent's row lock.
+     */
+    private pendingMessageIn;
+    /**
+     * Re-derives the wakes a paused root held for its tree (agent map spec §5.3: a paused root queue
+     * holds the wakes of the whole tree): after the root's queue is unpaused, {@link ensureRun} on
+     * every sub-agent conversation below it that has an unconsumed wake event or message and no
+     * active run (the root itself is the caller's own `ensureRun`). Errors are logged; the next
+     * event, run end or page opening of that conversation re-derives its wake.
+     */
+    wakeHeldTree(rootId: string): Promise<void>;
+    /**
+     * Schedules the delayed `bg-wake` job at the end of the conversation's wake window (after the
+     * commit): one job per conversation and window end, so finishes inside one window share it. It
+     * calls {@link ensureRun}, which wakes the conversation if a finish is still unconsumed. A failed
+     * enqueue is logged: the next message, run end or page opening re-derives the wake.
+     */
+    private scheduleWake;
     /**
      * Enqueues a run job under the caller's job id (resumes, dispatches, repairs), with the same
      * options as {@link ensureRun} (see {@link RunJobs.enqueue}).

package/dist/runtime/run-queue.js

@@ -8,11 +8,15 @@ var __param = (this && this.__param) || function (paramIndex, decorator) {
     return function (target, key) { decorator(target, key, paramIndex); }
 };
 import { randomUUID } from "node:crypto";
+import { getQueueToken } from "@nestjs/bullmq";
 import { Inject, Injectable } from "@nestjs/common";
 import { sql } from "kysely";
 import { visibleTo } from "../agents/agent-lookup.js";
+import { BG_WAKE_JOB, bgWakeJobId, wakeDecisionIn, } from "../background/wake-rules.js";
 import { EventLog } from "../conversations/event-log.js";
 import { dispatchQueuedIn, pauseQueueWith } from "../conversations/queue.js";
+import { conversationTree } from "../conversations/tree.js";
+import { KernelConfigService } from "../infra/config/kernel-config.service.js";
 import { KernelDb } from "../infra/database/kernel-db.js";
 import { TransactionHost } from "../infra/database/transaction-host.js";
 import { Transactional } from "../infra/database/transactional.decorator.js";
@@ -21,14 +25,23 @@ import { KERNEL_LOGGER } from "../infra/tokens.js";
 import { ToolPipeline } from "../tools/pipeline.js";
 import { KyselyCheckpointSaver } from "./checkpointer.js";
 import { ConversationLock } from "./conversation-lock.js";
+import { MAINTENANCE_QUEUE_NAME } from "./maintenance-queue.js";
 import { notifyFinished, RunFinish } from "./run-finish.js";
 import { ACTIVE_RUN_STATUSES, RunJobs, runJobId, runQueueNameFor, } from "./run-jobs.js";
+import { rootConversationOf } from "./subagents/child.js";
+import { budgetWindowOf } from "./subagents/limits.js";
 export { ACTIVE_RUN_STATUSES, RUN_QUEUE_NAME, runJobId, runQueueNameFor, SUBAGENT_RUN_QUEUE_NAME, } from "./run-jobs.js";
 /**
  * queued 运行的宽限期（秒）：`ensureRun` / 续跑先提交 runs 行、再投递任务，两步之间任务暂时不存在；
  * 只有进入 queued 超过这么久（`COALESCE(queued_at, created_at)`）仍没有任务的 queued 行才视为孤儿。
  */
 const ORPHAN_QUEUED_GRACE_SECONDS = 30;
+/**
+ * Margin added to the delayed wake job: its window is read on the database's clock, the delay runs
+ * on this process's and Redis's, so the job does not usually fire just before the window ends (one
+ * that does reschedules itself under its next retry id, `bgWakeJobId`).
+ */
+const WAKE_JOB_MARGIN_MS = 1000;
 /**
  * 运行队列（生产方）：`ensureRun` 在会话有未消费的用户消息、且没有进行中的运行时插入 `runs` 行并投递任务。
  * 「是否已有运行」以 PG 为准（会话行锁内检查 `runs` 中的 queued / running），并参考会话锁的持有者；
@@ -45,11 +58,15 @@ let RunQueue = class RunQueue {
     events;
     log;
     txHost;
+    config;
+    maintenance;
     constructor(kdb, eventLog, lock, jobs, finisher, 
     /** 运行以孤儿结束（标记失败 / 直接中止）后删除它的 checkpoint thread（thread_id = runId）。 */
     checkpoints, 
     /** The run's open `Agent` call results, in the orphan repair's transaction (sub-agents spec §12). */
-    pipeline, events, log, txHost) {
+    pipeline, events, log, txHost, config, 
+    /** The producer side of the delayed `bg-wake` job (`MaintenanceProcessor` consumes it). */
+    maintenance) {
         this.kdb = kdb;
         this.eventLog = eventLog;
         this.lock = lock;
@@ -60,6 +77,8 @@ let RunQueue = class RunQueue {
         this.events = events;
         this.log = log;
         this.txHost = txHost;
+        this.config = config;
+        this.maintenance = maintenance;
     }
     get db() {
         return this.kdb.db;
@@ -81,6 +100,13 @@ let RunQueue = class RunQueue {
      * 有未消费的用户消息（最新 user_message 的 seq > `consumed_seq`）、会话锁无人持有、
      * 且没有 queued / running / waiting 的运行时：插入 `runs`（queued）并投递任务，返回 runId；否则返回 null。
      *
+     * The boundary (`trigger_seq`) is that newest message; with none unconsumed, the queue's head
+     * once it is sent; with no head either, a wake-eligible finished background task or sub-agent
+     * delivery (background Bash spec §3, agent map spec §5.3, {@link wakeDecisionIn}): the run is
+     * then attributed to the event's trigger account and joins the tree's budget window, and an
+     * event inside the wake window schedules the delayed `bg-wake` job instead. A sub-agent's
+     * conversation is only woken this way ({@link childWakeIn}).
+     *
      * 孤儿清理（R29，按需进行，没有定时器）：已有的 queued / running 运行若 BullMQ 任务已不存在、
      * 会话锁也无人持有（queued 的还要超过宽限期），说明没有任何 worker 会再执行它——写
      * `run_failed { code: "run.orphaned" }` 把它标记为失败（consumed_seq 推进到它的触发消息），
@@ -89,7 +115,9 @@ let RunQueue = class RunQueue {
      *
      * agent 已软删除的会话不起新运行（孤儿清理照常进行）。
      */
-    async ensureRun(conversationId) {
+    async ensureRun(conversationId, 
+    /** From the `bg-wake` job: it fired `wakeRetry - 1` times before its window ended. */
+    opts = {}) {
         if ((await this.lock.holder(conversationId)) !== null)
             return null;
         const st = {
@@ -97,11 +125,23 @@ let RunQueue = class RunQueue {
             orphanRunId: null,
             dispatchedSeq: null,
             dispatchedSwitch: false,
+            dispatchedUnarchived: false,
             pausedNow: false,
             orphanStop: undefined,
+            wakeLater: null,
+            repairRunId: null,
         };
         const created = await this.ensureRunIn(conversationId, st);
+        if (st.repairRunId !== null) {
+            // a sub-agent's run nobody runs any more: requeued when queued, failed when running (its
+            // finish then reaches the parent through the wait sweep's undelivered repair)
+            await this.repairOrphan(conversationId, st.repairRunId);
+        }
+        if (st.wakeLater !== null) {
+            await this.scheduleWake(conversationId, st.wakeLater, opts.wakeRetry ?? 0);
+        }
         const { orphanSeq, orphanRunId, dispatchedSeq, dispatchedSwitch, pausedNow, orphanStop } = st;
+        const { dispatchedUnarchived } = st;
         if (orphanSeq !== null) {
             await notifyFinished(this.eventLog, conversationId, orphanSeq, "failed");
             orphanStop?.fire();
@@ -114,6 +154,8 @@ let RunQueue = class RunQueue {
             await this.eventLog.publishChanged(conversationId, "queue");
             if (dispatchedSwitch)
                 await this.eventLog.publishChanged(conversationId, "selection");
+            if (dispatchedUnarchived)
+                await this.eventLog.publishChanged(conversationId, "archived");
         }
         if (orphanRunId !== null) {
             await this.dropCheckpoints(orphanRunId);
@@ -124,10 +166,10 @@ let RunQueue = class RunQueue {
         }
         if (!created)
             return null;
-        const { runId, jobId } = created;
+        const { runId, jobId, kind } = created;
         try {
-            const queue = this.jobs.queue();
-            await this.jobs.add({ runId, conversationId }, jobId);
+            const queue = this.jobs.queue(runQueueNameFor(kind));
+            await this.jobs.enqueue({ runId, conversationId }, jobId, kind);
             // 从 Redis 读回核对：检查与 add 之间出现了同 jobId 的任务时 add 被静默忽略，撤销本次插入，
             // 不留下没有任务的 queued 行。读不到说明任务已被领取并移除（只可能是本次投递的），不撤销。
             const stored = await queue.getJob(jobId);
@@ -153,14 +195,27 @@ let RunQueue = class RunQueue {
         const t = this.db;
         const conv = await t
             .selectFrom("conversations")
-            .select(["consumed_seq", "agent_id", "kind", "queue_paused", "account_id"])
+            .select([
+            "consumed_seq",
+            "org_id",
+            "agent_id",
+            "kind",
+            "queue_paused",
+            "account_id",
+            "deleted_at",
+        ])
             .where("id", "=", conversationId)
             .forUpdate()
             .executeTakeFirst();
         // 评测回放的隐藏对话：运行由评测 worker 在里面直接驱动（工具只回放记录），绝不为它起真实运行
-        if (conv?.kind !== "chat")
+        if (conv?.kind !== "chat" && conv?.kind !== "subagent")
+            return null;
+        // A deleted conversation starts no run, whatever is still unconsumed or queued
+        if (conv.deleted_at !== null)
             return null;
-        let last = await t
+        if (conv.kind === "subagent")
+            return this.childWakeIn(conversationId, conv, st);
+        const newest = await t
             .selectFrom("conversation_events")
             .select(["seq", "payload"])
             .where("conversation_id", "=", conversationId)
@@ -168,11 +223,27 @@ let RunQueue = class RunQueue {
             .orderBy("seq", "desc")
             .limit(1)
             .executeTakeFirst();
-        if (!last || Number(last.seq) <= Number(conv.consumed_seq)) {
+        /** The boundary is a wake event: the run joins the tree's budget window (agent map §6). */
+        let isWake = false;
+        /** The run's boundary: its `trigger_seq` and the account it runs as. */
+        let last = newest && {
+            seq: newest.seq,
+            accountId: newest.payload.authorAccountId,
+        };
+        const idleActive = !last || Number(last.seq) <= Number(conv.consumed_seq)
+            ? await this.activeRun(t, conversationId)
+            : undefined;
+        if (idleActive && Number(idleActive.trigger_seq) > Number(conv.consumed_seq)) {
+            // No unconsumed message, but a run whose boundary is not consumed yet (a wake run, whose
+            // boundary is a finished background task or a sub-agent's delivery): it goes through the
+            // orphan check below with that boundary, as a message's run does
+            last = { seq: idleActive.trigger_seq, accountId: idleActive.trigger_account_id };
+        }
+        else if (!last || Number(last.seq) <= Number(conv.consumed_seq)) {
             // Nothing unconsumed: the queue's head goes out when the conversation is idle (no active
             // run, which includes runs waiting for approval or an answer) and the queue is not paused.
             // Interjections are always unconsumed messages, so the queue never jumps ahead of them.
-            if (conv.queue_paused || (await this.activeRun(t, conversationId)))
+            if (conv.queue_paused || idleActive)
                 return null;
             const owner = await t
                 .selectFrom("agents")
@@ -187,32 +258,47 @@ let RunQueue = class RunQueue {
                 .where("conversation_id", "=", conversationId)
                 .limit(1)
                 .executeTakeFirst();
-            if (!head)
-                return null;
-            // A queued message may wait for hours: it goes out only while its author can still use
-            // the conversation (active, a member, the agent visible to them; the isAccessible rule).
-            // Otherwise the queue pauses instead of posting as someone who lost access.
-            if (!(await authorCanSend(t, conversationId, conv.account_id))) {
-                await pauseQueueWith(t, conversationId, "author_cannot_send");
-                st.pausedNow = true;
-                this.log.warn("queued message not sent: its author can no longer use the conversation", {
-                    conversationId,
-                });
-                return null;
+            if (head) {
+                // A queued message may wait for hours: it goes out only while its author can still use
+                // the conversation (active, a member, the agent visible to them; the isAccessible rule).
+                // Otherwise the queue pauses instead of posting as someone who lost access.
+                if (!(await authorCanSend(t, conversationId, conv.account_id))) {
+                    await pauseQueueWith(t, conversationId, "author_cannot_send");
+                    st.pausedNow = true;
+                    this.log.warn("queued message not sent: its author can no longer use the conversation", {
+                        conversationId,
+                    });
+                    return null;
+                }
+                const sent = await dispatchQueuedIn(t, this.eventLog, conversationId);
+                if (!sent)
+                    return null;
+                if ("paused" in sent) {
+                    st.pausedNow = true;
+                    this.log.warn("queued skill invocation not sent: the skill can no longer be invoked", {
+                        conversationId,
+                    });
+                    return null;
+                }
+                st.dispatchedSeq = sent.event.seq;
+                st.dispatchedSwitch = sent.switched;
+                st.dispatchedUnarchived = sent.unarchived;
+                last = {
+                    seq: String(sent.event.seq),
+                    accountId: sent.event.payload.authorAccountId,
+                };
             }
-            const sent = await dispatchQueuedIn(t, this.eventLog, conversationId);
-            if (!sent)
-                return null;
-            if ("paused" in sent) {
-                st.pausedNow = true;
-                this.log.warn("queued skill invocation not sent: the skill can no longer be invoked", {
-                    conversationId,
-                });
-                return null;
+            else {
+                // Nothing queued either: a finished background task may wake the idle conversation
+                // (background Bash spec §3; a person's message or the queue always goes first)
+                const wake = await wakeDecisionIn(t, { id: conversationId, orgId: conv.org_id, rootId: conversationId }, Number(conv.consumed_seq), this.config.get().runtime.wakeOnBackgroundTask, (accountId) => authorCanSend(t, conversationId, accountId));
+                if (wake?.kind === "later")
+                    st.wakeLater = wake;
+                if (wake?.kind !== "wake")
+                    return null;
+                last = { seq: String(wake.seq), accountId: wake.triggerAccountId };
+                isWake = true;
             }
-            st.dispatchedSeq = sent.event.seq;
-            st.dispatchedSwitch = sent.switched;
-            last = { seq: String(sent.event.seq), payload: sent.event.payload };
         }
         const active = await this.activeRun(t, conversationId);
         if (active) {
@@ -254,7 +340,9 @@ let RunQueue = class RunQueue {
             .orderBy("seq")
             .limit(1)
             .executeTakeFirst();
-        const trigger = fork ?? last;
+        const trigger = fork
+            ? { seq: fork.seq, accountId: fork.payload.authorAccountId }
+            : last;
         // BullMQ 对重复的 jobId 静默忽略（add 返回的 Job 由调用方数据构造，看不出是否真的入队）：
         // 先显式查同一触发点的任务；仍存在（例如属于刚结束、尚未从队列移除的运行）就不插入 runs 行。
         const jobId = runJobId(conversationId, Number(trigger.seq));
@@ -283,11 +371,193 @@ let RunQueue = class RunQueue {
             conversation_id: conversationId,
             status: "queued",
             agent_version: agent.current_version,
-            trigger_account_id: trigger.payload.authorAccountId,
+            trigger_account_id: trigger.accountId,
             trigger_seq: Number(trigger.seq),
+            // a chat conversation always has its person-triggered run before any wake (a wake event
+            // comes from a run of it or of its tree); null would only make the wake run its own window
+            budget_run_id: isWake ? await budgetWindowOf(t, conversationId) : null,
+        })
+            .execute();
+        return { runId, jobId, kind: "chat" };
+    }
+    /**
+     * {@link ensureRunIn} for a sub-agent's conversation (agent map spec §5.3), under its row lock.
+     * Its runs from a task, or from a message to an idle sub-agent, are started by the dispatch
+     * (`Agent`, `SendMessage`), which also records the dispatch. Here it is started from a parent's
+     * message that reached its run after that run's last model call (agent map spec §10: a message
+     * racing the run's end starts a new run, never lost); and it is woken by its own background
+     * task's finish or a delivery from its own sub-agent, under the same rules as a chat
+     * conversation. Neither happens once its last run was stopped (`aborted`): a stopped sub-agent is
+     * started again only by its parent. Both read the queue and the trigger account's access on the root
+     * of its tree; a message needs no wake setting and has no wake window. The run counts against
+     * the root's budget window and runs on the sub-agent queue under `runJobId(child, boundary
+     * seq)`. An active run is left to the orphan repair after the commit (`repairRunId`), never
+     * started past.
+     */
+    async childWakeIn(conversationId, conv, st) {
+        const t = this.db;
+        const active = await this.activeRun(t, conversationId);
+        if (active) {
+            if (active.status !== "waiting")
+                st.repairRunId = active.id;
+            return null;
+        }
+        // a stopped sub-agent (its last run `aborted`: a person's or a tree stop) stays stopped: no
+        // message, task finish or delivery that reached it before or during the stop wakes it (a
+        // descendant's report that landed just before the stop would otherwise restart the stopped
+        // branch); they are parked for its next run, which only its parent starts (`Agent`,
+        // `SendMessage`)
+        const last = await t
+            .selectFrom("runs")
+            .select("status")
+            .where("conversation_id", "=", conversationId)
+            .orderBy("created_at", "desc")
+            .orderBy("id", "desc")
+            .limit(1)
+            .executeTakeFirst();
+        if (last?.status === "aborted")
+            return null;
+        const rootId = await rootConversationOf(t, conversationId);
+        const wake = (await this.pendingMessageIn(conversationId, Number(conv.consumed_seq), rootId)) ??
+            (await wakeDecisionIn(t, { id: conversationId, orgId: conv.org_id, rootId }, Number(conv.consumed_seq), this.config.get().runtime.wakeOnBackgroundTask, (accountId) => authorCanSend(t, rootId, accountId)));
+        if (wake?.kind === "later")
+            st.wakeLater = wake;
+        if (wake?.kind !== "wake")
+            return null;
+        const jobId = runJobId(conversationId, wake.seq);
+        const existing = await this.jobs.getJob(jobId, runQueueNameFor("subagent"));
+        if (existing) {
+            this.log.warn("run job for this trigger already exists; not enqueuing", {
+                conversationId,
+                jobId,
+                existingRunId: existing.data?.runId,
+            });
+            return null;
+        }
+        const agent = await t
+            .selectFrom("agents")
+            .select(["current_version", "deleted_at"])
+            .where("id", "=", conv.agent_id)
+            .executeTakeFirstOrThrow();
+        if (agent.deleted_at !== null)
+            return null;
+        // a sub-agent runs its definition frozen at dispatch: its runs keep the version recorded then
+        const previous = await t
+            .selectFrom("runs")
+            .select("agent_version")
+            .where("conversation_id", "=", conversationId)
+            .orderBy("created_at", "desc")
+            .limit(1)
+            .executeTakeFirst();
+        const runId = randomUUID();
+        await t
+            .insertInto("runs")
+            .values({
+            id: runId,
+            conversation_id: conversationId,
+            status: "queued",
+            agent_version: previous?.agent_version ?? agent.current_version,
+            trigger_account_id: wake.triggerAccountId,
+            trigger_seq: wake.seq,
+            budget_run_id: await budgetWindowOf(t, rootId),
         })
             .execute();
-        return { runId, jobId };
+        return { runId, jobId, kind: "subagent" };
+    }
+    /**
+     * The newest unconsumed message of an idle sub-agent (a `SendMessage` its last run did not absorb)
+     * as a run boundary, or null: none; the root's queue is paused or the root deleted (the
+     * message waits for the unpause, {@link wakeHeldTree}); or its author can no longer send to the
+     * root (it waits for a later run). The caller already refused a stopped sub-agent (last run
+     * `aborted`): a message that reached it while running and then stopped is parked, and the next
+     * run its parent starts absorbs it. Called under the sub-agent's row lock.
+     */
+    async pendingMessageIn(conversationId, consumed, rootId) {
+        const t = this.db;
+        const message = await t
+            .selectFrom("conversation_events")
+            .select(["seq", "payload"])
+            .where("conversation_id", "=", conversationId)
+            .where("type", "=", "user_message")
+            .where("seq", ">", String(consumed))
+            .orderBy("seq", "desc")
+            .limit(1)
+            .executeTakeFirst();
+        if (!message)
+            return null;
+        const root = await t
+            .selectFrom("conversations")
+            .select(["queue_paused", "deleted_at"])
+            .where("id", "=", rootId)
+            .executeTakeFirst();
+        if (!root || root.queue_paused || root.deleted_at !== null)
+            return null;
+        const accountId = message.payload.authorAccountId;
+        if (!(await authorCanSend(t, rootId, accountId)))
+            return null;
+        return { kind: "wake", seq: Number(message.seq), triggerAccountId: accountId };
+    }
+    /**
+     * Re-derives the wakes a paused root held for its tree (agent map spec §5.3: a paused root queue
+     * holds the wakes of the whole tree): after the root's queue is unpaused, {@link ensureRun} on
+     * every sub-agent conversation below it that has an unconsumed wake event or message and no
+     * active run (the root itself is the caller's own `ensureRun`). Errors are logged; the next
+     * event, run end or page opening of that conversation re-derives its wake.
+     */
+    async wakeHeldTree(rootId) {
+        const ids = (await conversationTree(this.db, rootId))
+            .filter((c) => c.level > 0)
+            .map((c) => c.id);
+        if (ids.length === 0)
+            return;
+        const held = await this.db
+            .selectFrom("conversations as c")
+            .select("c.id")
+            .where("c.id", "in", ids)
+            .where("c.deleted_at", "is", null)
+            .where(({ exists, selectFrom }) => exists(selectFrom("conversation_events as e")
+            .select(sql `1`.as("one"))
+            .whereRef("e.conversation_id", "=", "c.id")
+            .where("e.type", "in", [
+            "background_task_finished",
+            "subagent_delivery",
+            "user_message",
+        ])
+            .whereRef("e.seq", ">", "c.consumed_seq")))
+            .where(({ not, exists, selectFrom }) => not(exists(selectFrom("runs as r")
+            .select(sql `1`.as("one"))
+            .whereRef("r.conversation_id", "=", "c.id")
+            .where("r.status", "in", ACTIVE_RUN_STATUSES))))
+            .execute();
+        for (const c of held) {
+            try {
+                await this.ensureRun(c.id);
+            }
+            catch (err) {
+                this.log.warn("cannot wake a held sub-agent", { conversationId: c.id, err });
+            }
+        }
+    }
+    /**
+     * Schedules the delayed `bg-wake` job at the end of the conversation's wake window (after the
+     * commit): one job per conversation and window end, so finishes inside one window share it. It
+     * calls {@link ensureRun}, which wakes the conversation if a finish is still unconsumed. A failed
+     * enqueue is logged: the next message, run end or page opening re-derives the wake.
+     */
+    async scheduleWake(conversationId, w, retry) {
+        const data = { conversationId, ...(retry > 0 && { retry }) };
+        try {
+            await this.maintenance.add(BG_WAKE_JOB, data, {
+                // a job that fired early reschedules under a new id (its own is still taken)
+                jobId: bgWakeJobId(conversationId, w.windowEnd, retry),
+                delay: w.delayMs + WAKE_JOB_MARGIN_MS,
+                removeOnComplete: true,
+                removeOnFail: true,
+            });
+        }
+        catch (err) {
+            this.log.error("cannot schedule the background task wake", { conversationId, err });
+        }
     }
     /**
      * Enqueues a run job under the caller's job id (resumes, dispatches, repairs), with the same
@@ -407,6 +677,7 @@ let RunQueue = class RunQueue {
             "id",
             "status",
             "trigger_seq",
+            "trigger_account_id",
             "job_id",
             sql `COALESCE(queued_at, created_at) < now() - make_interval(secs => ${ORPHAN_QUEUED_GRACE_SECONDS})`.as("past_grace"),
         ])
@@ -462,7 +733,9 @@ RunQueue = __decorate([
     __param(6, Inject(ToolPipeline)),
     __param(7, Inject(KernelEvents)),
     __param(8, Inject(KERNEL_LOGGER)),
-    __param(9, Inject(TransactionHost))
+    __param(9, Inject(TransactionHost)),
+    __param(10, Inject(KernelConfigService)),
+    __param(11, Inject(getQueueToken(MAINTENANCE_QUEUE_NAME)))
 ], RunQueue);
 export { RunQueue };
 /** The conversation's owner is active and can still see its agent (`isAccessible`'s rule). */

package/dist/runtime/run-worker.d.ts

@@ -14,6 +14,7 @@ import { KyselyCheckpointSaver } from "./checkpointer.js";
 import { ConversationLock } from "./conversation-lock.js";
 import { ProviderRegistry } from "./providers.js";
 import { type OrgQuota } from "./quota.js";
+import { RunDrafts } from "./run-drafts.js";
 import { RunFinish } from "./run-finish.js";
 import type { RunJobData } from "./run-jobs.js";
 import { RunQueue } from "./run-queue.js";
@@ -70,6 +71,8 @@ export declare class RunWorker {
     private readonly pipeline;
     private readonly finisher;
     readonly txHost: TransactionHost;
+    /** The steps' live drafts this process writes for the runs it executes. */
+    private readonly drafts;
     private readonly hooks?;
     /** The two run queues' BullMQ workers (spec §7.4), once {@link start}ed on the worker role. */
     private workers;
@@ -84,7 +87,7 @@ export declare class RunWorker {
     /** 运行到达终态后删除该运行（thread_id = runId）的 checkpoint thread（只用于续跑，结束后无用）。 */
     checkpoints: Pick<KyselyCheckpointSaver, "deleteThread">, 
     /** loop 返回 `suspended` 后把运行置为 waiting（释放会话锁与配额之后）; the child lifecycle. */
-    waits: Pick<Waits, "suspend" | "childFinished" | "stopChildren">, 
+    waits: Pick<Waits, "suspend" | "childFinished" | "stopChildren" | "stopDescendantsOfStopped">, 
     /** `run.started` (the channel outbound enqueues the conversation's IM delivery). */
     events: KernelEvents, 
     /** A completed run enqueues the conversation's title job. */
@@ -98,7 +101,9 @@ export declare class RunWorker {
      * (`stopOpenAgentCalls`; a no-op without dispatches); `fire` runs their observe hooks after the
      * commit (sub-agents spec §12, review m1: a failed or stalled parent too).
      */
-    pipeline: Pick<ToolPipeline, "stopOpenAgentCalls">, finisher: Pick<RunFinish, "finish">, txHost: TransactionHost, hooks?: RunHooks | undefined);
+    pipeline: Pick<ToolPipeline, "stopOpenAgentCalls">, finisher: Pick<RunFinish, "finish">, txHost: TransactionHost, 
+    /** The steps' live drafts this process writes for the runs it executes. */
+    drafts: RunDrafts, hooks?: RunHooks | undefined);
     private get db();
     /**
      * Starts consuming (worker role only): `workers` are the `@Processor` workers of both run queues
@@ -119,13 +124,14 @@ export declare class RunWorker {
     /** A job failed (the worker's `failed` event). */
     onFailed(job: Job<RunJobData> | undefined, err: Error): void;
     /**
-     * 流式增量发布到会话频道（与 `{ kind: "events" }` 共用 `conv.<id>`），任意副本上的订阅方都能收到。
-     * 发布失败只记日志：增量只是草稿，最终内容以日志中的 assistant_message 为准。
+     * The run's stream sink: deltas on the conversation's topic (any replica's subscribers get them)
+     * and the steps' live drafts in Redis ({@link streamSink}, spec 2026-11-07 §3.2).
      */
     private sinkFor;
     /**
      * After any terminal commit of a run that is not a sub-agent's (sub-agents spec §12): stops its
-     * active children on its behalf (no account). The wait sweep is the backstop.
+     * active fork invocation child on its behalf (no account); background `Agent` children keep
+     * running (agent map spec §5.4). The wait sweep is the backstop.
      */
     private stopChildren;
     /**