@capybara-hq/kernel 0.1.0-rc.23 → 0.1.0-rc.24

package/dist/runtime/maintenance.d.ts

@@ -16,7 +16,9 @@ export declare const MAINTENANCE_CONCURRENCY = 5;
  * The one consumer of `cap-kernel-maintenance` (decision R11 of the Nest kernel spec): dispatches
  * each one-off job by name to its domain (`wait-timeout` → {@link Waits}, `sandbox-destroy-for` and
  * `sandbox-reset-files` → {@link SandboxService}, `drift-connector` → {@link ToolDrift}, `directory-sync` →
- * {@link DirectorySync}, `bg-wake` → {@link RunQueue.ensureRun}, `owned-plugin-cleanup` →
+ * {@link DirectorySync}, `bg-wake` and `ensure-run` → {@link RunQueue.ensureRun}, `stop-tree` →
+ * {@link processStopTree} and `deliver-to-parent` → {@link processDeliverToParent} (both in
+ * runtime/subagents/tree-jobs.ts, over {@link Waits}), `owned-plugin-cleanup` →
  * {@link OwnedPluginCleanup}, `sched-fire` → {@link ScheduledTaskFirer}). The name of a periodic job (a `@LockedCron` / `@LockedInterval` job,
  * such as `sandbox-suspend-idle`) runs that job's body once ({@link ScheduledRuns.runNow}), as the
  * maintenance handlers did when those jobs were schedulers on this queue: an operator's or a
@@ -34,7 +36,7 @@ export declare class MaintenanceProcessor extends WorkerHost<Worker> {
     private readonly scheduled;
     private readonly log;
     private readonly queue;
-    constructor(waits: Pick<Waits, "processTimeout">, sandbox: Pick<SandboxService, "destroyDevice" | "resetFileSystem">, drift: Pick<ToolDrift, "check">, directory: Pick<DirectorySync, "runSyncJob">, runs: Pick<RunQueue, "ensureRun">, ownedCleanup: Pick<OwnedPluginCleanup, "run">, scheduledTasks: Pick<ScheduledTaskFirer, "fire">, scheduled: Pick<ScheduledRuns, "runNow">, log: Pick<KernelLogger, "error" | "warn">, queue: Pick<Queue, "removeJobScheduler">);
+    constructor(waits: Pick<Waits, "processTimeout" | "childFinished" | "stopChildren" | "stopDescendantsOfStopped" | "stopTree">, sandbox: Pick<SandboxService, "destroyDevice" | "resetFileSystem">, drift: Pick<ToolDrift, "check">, directory: Pick<DirectorySync, "runSyncJob">, runs: Pick<RunQueue, "ensureRun">, ownedCleanup: Pick<OwnedPluginCleanup, "run">, scheduledTasks: Pick<ScheduledTaskFirer, "fire">, scheduled: Pick<ScheduledRuns, "runNow">, log: Pick<KernelLogger, "error" | "warn">, queue: Pick<Queue, "removeJobScheduler">);
     /**
      * Removes the retired job schedulers (`RETIRED_SCHEDULER_IDS`, runtime/retired-schedulers.ts): every fixed-period
      * maintenance job is a `@LockedCron` / `@LockedInterval` now. Idempotent; called once when a

package/dist/runtime/maintenance.js

@@ -31,8 +31,9 @@ import { DRIFT_CONNECTOR_JOB } from "../tools/job-ids.js";
 import { ToolsModule } from "../tools/tools.module.js";
 import { MAINTENANCE_QUEUE_NAME } from "./maintenance-queue.js";
 import { RETIRED_SCHEDULER_IDS } from "./retired-schedulers.js";
-import { RunQueue } from "./run-queue.js";
+import { ENSURE_RUN_JOB, RunQueue } from "./run-queue.js";
 import { RuntimeModule } from "./runtime.module.js";
+import { DELIVER_TO_PARENT_JOB, processDeliverToParent, processStopTree, STOP_TREE_JOB, } from "./subagents/tree-jobs.js";
 import { WAIT_TIMEOUT_JOB, Waits } from "./waits.js";
 export { MAINTENANCE_QUEUE_NAME } from "./maintenance-queue.js";
 /** The maintenance worker's concurrency (as the hand-made worker had). */
@@ -41,7 +42,9 @@ export const MAINTENANCE_CONCURRENCY = 5;
  * The one consumer of `cap-kernel-maintenance` (decision R11 of the Nest kernel spec): dispatches
  * each one-off job by name to its domain (`wait-timeout` → {@link Waits}, `sandbox-destroy-for` and
  * `sandbox-reset-files` → {@link SandboxService}, `drift-connector` → {@link ToolDrift}, `directory-sync` →
- * {@link DirectorySync}, `bg-wake` → {@link RunQueue.ensureRun}, `owned-plugin-cleanup` →
+ * {@link DirectorySync}, `bg-wake` and `ensure-run` → {@link RunQueue.ensureRun}, `stop-tree` →
+ * {@link processStopTree} and `deliver-to-parent` → {@link processDeliverToParent} (both in
+ * runtime/subagents/tree-jobs.ts, over {@link Waits}), `owned-plugin-cleanup` →
  * {@link OwnedPluginCleanup}, `sched-fire` → {@link ScheduledTaskFirer}). The name of a periodic job (a `@LockedCron` / `@LockedInterval` job,
  * such as `sandbox-suspend-idle`) runs that job's body once ({@link ScheduledRuns.runNow}), as the
  * maintenance handlers did when those jobs were schedulers on this queue: an operator's or a
@@ -113,6 +116,18 @@ let MaintenanceProcessor = class MaintenanceProcessor extends WorkerHost {
                 await this.runs.ensureRun(conversationId, { wakeRetry: (retry ?? 0) + 1 });
                 return;
             }
+            // the ensure-run backstop of a message, wake or run end (transactional outbox spec §6)
+            case ENSURE_RUN_JOB:
+                await this.runs.ensureRun(job.data.conversationId);
+                return;
+            // the stops a run's end or a root's delete owes its sub-agent tree (outbox spec §6)
+            case STOP_TREE_JOB:
+                await processStopTree(this.waits, job.data);
+                return;
+            // a finished sub-agent run's delivery to its parent (transactional outbox spec §6)
+            case DELIVER_TO_PARENT_JOB:
+                await processDeliverToParent(this.waits, job.data);
+                return;
             // a removed org-owned plugin's extension (plugin scope phase 3, P3-20)
             case OWNED_PLUGIN_CLEANUP_JOB:
                 await this.ownedCleanup.run(job.data.itemId);

package/dist/runtime/model-call.d.ts

@@ -11,7 +11,6 @@ import type { SubagentParent } from "./loop.js";
 import { MediaCache } from "./media-cache.js";
 import { type BoundProvider, type ProviderRegistry } from "./providers.js";
 import type { StepLog } from "./step-log.js";
-import type { UsageReporter } from "./usage-report.js";
 /** 单轮收尾的尾部提醒（模型可见，英文；spec 6 轮次额度收尾）。 */
 export declare const WRAP_UP_REMINDER = "Step limit reached. Do not call tools. Summarize what you found and what remains for the user.";
 /**
@@ -48,8 +47,11 @@ export type StreamSink = (delta: {
 export interface ModelStepDeps {
     db: KernelSchemaDb;
     eventLog: EventLog;
-    /** The step's transactional writes (a model call with its tool specs, a denial with its audit row, a sub-agent's counted usage). */
-    stepLog: Pick<StepLog, "appendCounted" | "appendAudited" | "appendWithSpecs">;
+    /**
+     * The step's transactional writes (a model call with its tool specs, a denial with its audit row,
+     * usage events with their owed `usage.report` deliveries, a sub-agent's counted usage).
+     */
+    stepLog: Pick<StepLog, "appendCounted" | "appendReported" | "appendAudited" | "appendWithSpecs">;
     registry: Pick<ProviderRegistry, "resolve">;
     /** 工具注册表：每步开始时解析 agent 引用的工具（状态可能在步骤之间变化）。 */
     tools: Pick<ToolRegistry, "forAgent">;
@@ -67,11 +69,6 @@ export interface ModelStepDeps {
      * `conversationKind: "chat"` on the step input.
      */
     hooks?: HookRunner;
-    /**
-     * `usage.report` delivery (P-17): each written `usage` event is enqueued (chat runs and full-run eval
-     * replays, which both log usage). `enqueue` never throws and returns at once without handlers.
-     */
-    usageReport?: Pick<UsageReporter, "enqueue">;
 }
 /** 构建请求用到的附件来源。 */
 export interface MediaSourceDeps {
@@ -121,7 +118,8 @@ export interface ModelStepInput {
     parentRunId?: string;
     /**
      * The run's step limit (`LoopContext.limits.maxSteps`) for the turn-limit reminder; a child's
-     * may be below `runtime.maxSteps` (final review m4). Defaults to `runtime.maxSteps`.
+     * may be below `runtime.maxSteps` (final review m4). Absent: the run has no step limit, and no
+     * reminder is sent.
      */
     maxSteps?: number;
 }
@@ -176,15 +174,15 @@ export declare function runModelStep(deps: ModelStepDeps, input: ModelStepInput)
 export declare function runFinalWrapUp(deps: ModelStepDeps, input: ModelStepInput, reason?: WrapUpReason): Promise<ModelStepResult>;
 /**
  * Appends a model call's events (its reply or compaction outcome with its `usage`; also the usage
- * of a `WebFetch` extraction call, `tools/builtins/web/extract.ts`). For a
- * sub-agent's run (`input.parentRunId`) in one transaction that also adds the tokens of the usage
- * events it newly wrote to the budget run's `subagent_tokens_used` (sub-agents spec §8.3): a usage
- * event already in the log (an idempotent hit after a crash) adds nothing, so each call counts
- * exactly once.
+ * of a `WebFetch` extraction call, `tools/builtins/web/extract.ts`) in one transaction that also
+ * owes the `usage.report` delivery of the usage events it newly wrote (transactional outbox spec
+ * §8.1). For a sub-agent's run (`input.parentRunId`) that transaction also adds their tokens to the
+ * budget run's `subagent_tokens_used` (sub-agents spec §8.3): a usage event already in the log (an
+ * idempotent hit after a crash) adds nothing, so each call counts and is reported exactly once.
  */
 export declare function appendCounted(deps: {
-    eventLog: Pick<EventLog, "append" | "notify">;
-    stepLog: Pick<StepLog, "appendCounted">;
+    eventLog: Pick<EventLog, "notify">;
+    stepLog: Pick<StepLog, "appendCounted" | "appendReported">;
 }, input: {
     conversationId: string;
     parentRunId?: string | undefined;

package/dist/runtime/model-call.js

@@ -123,12 +123,6 @@ wrapUpReason) {
     const { conversationId, runId, step, signal, sink } = input;
     const existing = await findExisting(deps.eventLog, conversationId, runId, step);
     if (existing) {
-        // resumed after a crash between writing usage and enqueueing it: enqueue again (job id dedupes)
-        if (deps.usageReport && existing.usage !== null) {
-            const u = await deps.eventLog.findByIdem(conversationId, idem.usage(runId, step));
-            if (u)
-                void deps.usageReport.enqueue(conversationId, u.seq);
-        }
         // a crash between the assistant_message and clearing its draft leaves one behind
         sink({ kind: "step_end", runId, step });
         return existing;
@@ -230,8 +224,9 @@ wrapUpReason) {
             }
         }
         const { outcome, ref } = await ladder(deps, input, events, runtime, tools, wrapUpReason, streamed);
-        // assistant_message 与 usage 同一次 append = 同一事务。
-        const appended = await appendCounted(deps, input, [
+        // assistant_message and usage in one append = one transaction, which also owes the usage's
+        // usage.report delivery (transactional outbox spec §8.1)
+        await appendCounted(deps, input, [
             {
                 type: "assistant_message",
                 runId,
@@ -245,10 +240,6 @@ wrapUpReason) {
                 payload: { step, provider: ref.provider, model: ref.model, ...outcome.usage },
             },
         ]);
-        // usage.report metering (never awaited on the run's path; enqueue logs and swallows failures)
-        const usageEvent = appended.find((e) => e.type === "usage");
-        if (deps.usageReport && usageEvent)
-            void deps.usageReport.enqueue(conversationId, usageEvent.seq);
         sink({ kind: "step_end", runId, step });
         return {
             content: outcome.content,
@@ -277,18 +268,18 @@ wrapUpReason) {
 }
 /**
  * Appends a model call's events (its reply or compaction outcome with its `usage`; also the usage
- * of a `WebFetch` extraction call, `tools/builtins/web/extract.ts`). For a
- * sub-agent's run (`input.parentRunId`) in one transaction that also adds the tokens of the usage
- * events it newly wrote to the budget run's `subagent_tokens_used` (sub-agents spec §8.3): a usage
- * event already in the log (an idempotent hit after a crash) adds nothing, so each call counts
- * exactly once.
+ * of a `WebFetch` extraction call, `tools/builtins/web/extract.ts`) in one transaction that also
+ * owes the `usage.report` delivery of the usage events it newly wrote (transactional outbox spec
+ * §8.1). For a sub-agent's run (`input.parentRunId`) that transaction also adds their tokens to the
+ * budget run's `subagent_tokens_used` (sub-agents spec §8.3): a usage event already in the log (an
+ * idempotent hit after a crash) adds nothing, so each call counts and is reported exactly once.
  */
 export async function appendCounted(deps, input, events) {
     const { conversationId, parentRunId } = input;
-    if (parentRunId === undefined)
-        return deps.eventLog.append(conversationId, events);
     const budgetRunId = input.budgetRunId ?? input.parent?.budgetRunId;
-    const out = await deps.stepLog.appendCounted(conversationId, { parentRunId, ...(budgetRunId !== undefined && { budgetRunId }) }, events);
+    const out = parentRunId === undefined
+        ? await deps.stepLog.appendReported(conversationId, events)
+        : await deps.stepLog.appendCounted(conversationId, { parentRunId, ...(budgetRunId !== undefined && { budgetRunId }) }, events);
     if (out.toSeq !== null)
         await deps.eventLog.notify(conversationId, out.toSeq);
     return out.written;
@@ -691,9 +682,6 @@ async function maybeCompact(deps, input, events, runtime, tools) {
         });
     }
     const written = await appendCounted(deps, input, batch);
-    const usageEvent = written.find((e) => e.type === "usage");
-    if (deps.usageReport && usageEvent)
-        void deps.usageReport.enqueue(conversationId, usageEvent.seq);
     return mergeLocal(events, written);
 }
 /** Merges written events into the local list (idempotent hits already in it are not added twice). */
@@ -719,7 +707,9 @@ async function ladder(deps, input, events, runtime, tools, wrapUpReason, streame
         ? BUDGET_WRAP_UP_REMINDER
         : wrapUp
             ? WRAP_UP_REMINDER
-            : turnLimitReminder(step, input.maxSteps ?? runtime.maxSteps);
+            : input.maxSteps === undefined
+                ? null
+                : turnLimitReminder(step, input.maxSteps);
     let evs = events;
     const rungs = [
         { ref: agent.model, fallback: false },

package/dist/runtime/queue-lost-jobs.d.ts

@@ -0,0 +1,24 @@
+import type { Redis } from "ioredis";
+/** The Redis hash of the lost-jobs canary: `count` and `lastAt` (ISO). */
+export declare const QUEUE_LOST_JOBS_KEY = "capk:queue-lost-jobs";
+/** How long the canary stays after its last report (then the overview drops it). */
+export declare const QUEUE_LOST_JOBS_TTL_SECONDS: number;
+/** What the admin overview shows: how many run jobs the queue lost, and when it last did. */
+export interface QueueLostJobsSummary {
+    count: number;
+    lastAt: string;
+}
+/**
+ * The "queue lost jobs" canary (transactional outbox spec §8.3): the `run-orphan-check` found
+ * queued runs whose job is neither in BullMQ nor owed by the outbox, which the outbox makes
+ * impossible unless Redis lost data (its durability broke). Each report adds to a count in Redis
+ * that the admin overview shows as an attention item until a week after the last report.
+ */
+export declare class QueueLostJobs {
+    private readonly redis;
+    constructor(redis: Redis);
+    /** Adds `n` lost jobs to the canary (no-op for 0). */
+    report(n: number): Promise<void>;
+    /** The canary, or null when no job was lost in the last week. */
+    summary(): Promise<QueueLostJobsSummary | null>;
+}

package/dist/runtime/queue-lost-jobs.js

@@ -0,0 +1,51 @@
+var __decorate = (this && this.__decorate) || function (decorators, target, key, desc) {
+    var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
+    if (typeof Reflect === "object" && typeof Reflect.decorate === "function") r = Reflect.decorate(decorators, target, key, desc);
+    else for (var i = decorators.length - 1; i >= 0; i--) if (d = decorators[i]) r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
+    return c > 3 && r && Object.defineProperty(target, key, r), r;
+};
+var __param = (this && this.__param) || function (paramIndex, decorator) {
+    return function (target, key) { decorator(target, key, paramIndex); }
+};
+import { Inject, Injectable } from "@nestjs/common";
+import { REDIS } from "../infra/tokens.js";
+/** The Redis hash of the lost-jobs canary: `count` and `lastAt` (ISO). */
+export const QUEUE_LOST_JOBS_KEY = "capk:queue-lost-jobs";
+/** How long the canary stays after its last report (then the overview drops it). */
+export const QUEUE_LOST_JOBS_TTL_SECONDS = 7 * 24 * 3600;
+/**
+ * The "queue lost jobs" canary (transactional outbox spec §8.3): the `run-orphan-check` found
+ * queued runs whose job is neither in BullMQ nor owed by the outbox, which the outbox makes
+ * impossible unless Redis lost data (its durability broke). Each report adds to a count in Redis
+ * that the admin overview shows as an attention item until a week after the last report.
+ */
+let QueueLostJobs = class QueueLostJobs {
+    redis;
+    constructor(redis) {
+        this.redis = redis;
+    }
+    /** Adds `n` lost jobs to the canary (no-op for 0). */
+    async report(n) {
+        if (n <= 0)
+            return;
+        await this.redis
+            .multi()
+            .hincrby(QUEUE_LOST_JOBS_KEY, "count", n)
+            .hset(QUEUE_LOST_JOBS_KEY, "lastAt", new Date().toISOString())
+            .expire(QUEUE_LOST_JOBS_KEY, QUEUE_LOST_JOBS_TTL_SECONDS)
+            .exec();
+    }
+    /** The canary, or null when no job was lost in the last week. */
+    async summary() {
+        const h = await this.redis.hgetall(QUEUE_LOST_JOBS_KEY);
+        const count = Number(h.count ?? 0);
+        if (!(count > 0) || h.lastAt === undefined)
+            return null;
+        return { count, lastAt: h.lastAt };
+    }
+};
+QueueLostJobs = __decorate([
+    Injectable(),
+    __param(0, Inject(REDIS))
+], QueueLostJobs);
+export { QueueLostJobs };

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

@@ -1,6 +1,7 @@
-import { KernelEvents } from "../infra/events/kernel-events.js";
+import { OutboxService } from "../infra/outbox/outbox.service.js";
 import type { KernelLogger } from "../log.js";
 import { RunAbort } from "./abort.js";
+import { type ResumeOutcome, RunResume } from "./run-jobs.js";
 import { RunQueue } from "./run-queue.js";
 import { Waits } from "./waits.js";
 /**
@@ -11,13 +12,49 @@ export declare class RunControl {
     private readonly queue;
     private readonly waits;
     private readonly abort;
-    private readonly events;
+    /** The resume of a waiting run in a decision's transaction (a reply's skips). */
+    private readonly resumes;
+    /** Owes a deleted root's tree stop in the delete's transaction. */
+    private readonly outbox;
     private readonly log;
-    constructor(queue: Pick<RunQueue, "ensureRun" | "abortIfOrphaned" | "wakeHeldTree">, waits: Pick<Waits, "abortWaiting" | "childFinished" | "stopChildren" | "stopTree">, abort: Pick<RunAbort, "request">, events: KernelEvents, log: Pick<KernelLogger, "error">);
+    constructor(queue: Pick<RunQueue, "ensureRun" | "oweEnsureRun" | "abortIfOrphaned" | "wakeHeldTree" | "oweHeldTreeIn">, waits: Pick<Waits, "abortWaiting" | "childFinished" | "stopChildren" | "stopTree">, abort: Pick<RunAbort, "request">, 
+    /** The resume of a waiting run in a decision's transaction (a reply's skips). */
+    resumes: Pick<RunResume, "resumeIn">, 
+    /** Owes a deleted root's tree stop in the delete's transaction. */
+    outbox: Pick<OutboxService, "add">, log: Pick<KernelLogger, "error">);
     /** 有未消费的消息且没有进行中的运行时投递一轮运行；返回 runId 或 null。 */
     ensureRun(conversationId: string): Promise<string | null>;
+    /**
+     * The direct `ensureRun` right after a commit that owes the ensure-run backstop (a message, a
+     * queue change): a failure (Redis unreachable) is logged and never thrown, since the committed
+     * write is not undone and the backstop starts the run later (transactional outbox spec §6).
+     */
+    ensureRunAfterCommit(conversationId: string): Promise<void>;
+    /**
+     * Owes the ensure-run backstop in the caller's transaction (transactional outbox spec §6): `seq`
+     * is the event it wrote (or the conversation's newest), `orgId` the conversation's.
+     */
+    oweEnsureRun(conversationId: string, seq: number | string, orgId: string): Promise<boolean>;
     /** After a root's queue was unpaused: the wakes it held for its tree ({@link RunQueue.wakeHeldTree}). */
     wakeHeldTree(rootId: string): Promise<void>;
+    /**
+     * Owes, in the transaction that may release a root's hold (an unpause, a "send now"), the
+     * ensure-run backstop of every sub-agent the hold kept waiting ({@link RunQueue.oweHeldTreeIn});
+     * the caller's {@link wakeHeldTree} after the commit is the direct call.
+     */
+    oweHeldTreeIn(rootId: string): Promise<void>;
+    /**
+     * In the caller's transaction, after it locked the decided rows (lock order: conversations,
+     * decided rows, runs): resumes the waiting run `runId` when it waits on nothing else
+     * ({@link RunResume.resumeIn}).
+     */
+    resumeIn(runId: string, waitId: string): Promise<ResumeOutcome>;
+    /**
+     * Owes, in a root's delete transaction, the stop of its whole sub-agent tree (`stop-tree`,
+     * transactional outbox spec §6); the caller's {@link stopTree} after the commit is the direct
+     * call.
+     */
+    oweStopDeletedTree(conversationId: string, byAccountId: string | null, orgId: string): Promise<boolean>;
     /**
      * 运行是孤儿（会话锁空闲、任务已不存在，没有 worker 会收到中止广播）时直接把它写成 aborted，返回 true；
      * 否则返回 false。A stopped sub-agent then reports to its parent at once, and a stopped parent
@@ -26,8 +63,8 @@ export declare class RunControl {
     abortIfOrphaned(conversationId: string, runId: string, byAccountId: string | null): Promise<boolean>;
     /**
      * 运行在等人处理（waiting）时直接中止：取消等待、写 `run_aborted`、删除 checkpoint，返回 true；
-     * 运行已不是 waiting（例如刚被续跑）→ false. Either way its approval / question cards may have
-     * changed: `run.wait-stopped` lets the channel outbound update them.
+     * 运行已不是 waiting（例如刚被续跑）→ false. The stop's transaction owes the IM delivery that
+     * updates its approval / question cards (its appends, transactional outbox spec §8.1).
      */
     abortWaiting(conversationId: string, runId: string, byAccountId: string | null): Promise<boolean>;
     /**

package/dist/runtime/run-control.js

@@ -8,10 +8,12 @@ var __param = (this && this.__param) || function (paramIndex, decorator) {
     return function (target, key) { decorator(target, key, paramIndex); }
 };
 import { Inject, Injectable } from "@nestjs/common";
-import { KernelEvents } from "../infra/events/kernel-events.js";
+import { OutboxService } from "../infra/outbox/outbox.service.js";
 import { KERNEL_LOGGER } from "../infra/tokens.js";
 import { RunAbort } from "./abort.js";
+import { RunResume } from "./run-jobs.js";
 import { RunQueue } from "./run-queue.js";
+import { oweStopDeletedTree } from "./subagents/tree-jobs.js";
 import { Waits } from "./waits.js";
 /**
  * Starting and stopping runs from outside the run worker: the conversations API (a message, a
@@ -21,23 +23,76 @@ let RunControl = class RunControl {
     queue;
     waits;
     abort;
-    events;
+    resumes;
+    outbox;
     log;
-    constructor(queue, waits, abort, events, log) {
+    constructor(queue, waits, abort, 
+    /** The resume of a waiting run in a decision's transaction (a reply's skips). */
+    resumes, 
+    /** Owes a deleted root's tree stop in the delete's transaction. */
+    outbox, log) {
         this.queue = queue;
         this.waits = waits;
         this.abort = abort;
-        this.events = events;
+        this.resumes = resumes;
+        this.outbox = outbox;
         this.log = log;
     }
     /** 有未消费的消息且没有进行中的运行时投递一轮运行；返回 runId 或 null。 */
     ensureRun(conversationId) {
         return this.queue.ensureRun(conversationId);
     }
+    /**
+     * The direct `ensureRun` right after a commit that owes the ensure-run backstop (a message, a
+     * queue change): a failure (Redis unreachable) is logged and never thrown, since the committed
+     * write is not undone and the backstop starts the run later (transactional outbox spec §6).
+     */
+    async ensureRunAfterCommit(conversationId) {
+        try {
+            await this.queue.ensureRun(conversationId);
+        }
+        catch (err) {
+            this.log.error("cannot start the conversation's run now; the ensure-run backstop will", {
+                conversationId,
+                err,
+            });
+        }
+    }
+    /**
+     * Owes the ensure-run backstop in the caller's transaction (transactional outbox spec §6): `seq`
+     * is the event it wrote (or the conversation's newest), `orgId` the conversation's.
+     */
+    oweEnsureRun(conversationId, seq, orgId) {
+        return this.queue.oweEnsureRun(conversationId, seq, orgId);
+    }
     /** After a root's queue was unpaused: the wakes it held for its tree ({@link RunQueue.wakeHeldTree}). */
     wakeHeldTree(rootId) {
         return this.queue.wakeHeldTree(rootId);
     }
+    /**
+     * Owes, in the transaction that may release a root's hold (an unpause, a "send now"), the
+     * ensure-run backstop of every sub-agent the hold kept waiting ({@link RunQueue.oweHeldTreeIn});
+     * the caller's {@link wakeHeldTree} after the commit is the direct call.
+     */
+    oweHeldTreeIn(rootId) {
+        return this.queue.oweHeldTreeIn(rootId);
+    }
+    /**
+     * In the caller's transaction, after it locked the decided rows (lock order: conversations,
+     * decided rows, runs): resumes the waiting run `runId` when it waits on nothing else
+     * ({@link RunResume.resumeIn}).
+     */
+    resumeIn(runId, waitId) {
+        return this.resumes.resumeIn(runId, waitId);
+    }
+    /**
+     * Owes, in a root's delete transaction, the stop of its whole sub-agent tree (`stop-tree`,
+     * transactional outbox spec §6); the caller's {@link stopTree} after the commit is the direct
+     * call.
+     */
+    oweStopDeletedTree(conversationId, byAccountId, orgId) {
+        return oweStopDeletedTree(this.outbox, conversationId, byAccountId, orgId);
+    }
     /**
      * 运行是孤儿（会话锁空闲、任务已不存在，没有 worker 会收到中止广播）时直接把它写成 aborted，返回 true；
      * 否则返回 false。A stopped sub-agent then reports to its parent at once, and a stopped parent
@@ -57,13 +112,11 @@ let RunControl = class RunControl {
     }
     /**
      * 运行在等人处理（waiting）时直接中止：取消等待、写 `run_aborted`、删除 checkpoint，返回 true；
-     * 运行已不是 waiting（例如刚被续跑）→ false. Either way its approval / question cards may have
-     * changed: `run.wait-stopped` lets the channel outbound update them.
+     * 运行已不是 waiting（例如刚被续跑）→ false. The stop's transaction owes the IM delivery that
+     * updates its approval / question cards (its appends, transactional outbox spec §8.1).
      */
-    async abortWaiting(conversationId, runId, byAccountId) {
-        const r = await this.waits.abortWaiting(conversationId, runId, byAccountId);
-        await this.events.emit("run.wait-stopped", { conversationId, runId });
-        return r;
+    abortWaiting(conversationId, runId, byAccountId) {
+        return this.waits.abortWaiting(conversationId, runId, byAccountId);
     }
     /**
      * Stops a sub-agent with its descendants, or with `includeSelf: false` only the sub-agent
@@ -83,7 +136,8 @@ RunControl = __decorate([
     __param(0, Inject(RunQueue)),
     __param(1, Inject(Waits)),
     __param(2, Inject(RunAbort)),
-    __param(3, Inject(KernelEvents)),
-    __param(4, Inject(KERNEL_LOGGER))
+    __param(3, Inject(RunResume)),
+    __param(4, Inject(OutboxService)),
+    __param(5, Inject(KERNEL_LOGGER))
 ], RunControl);
 export { RunControl };

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

@@ -1,10 +1,11 @@
 import { type AppendEvent, EventLog } from "../conversations/event-log.js";
 import { KernelDb } from "../infra/database/kernel-db.js";
 import { TransactionHost } from "../infra/database/transaction-host.js";
+import { OutboxService } from "../infra/outbox/outbox.service.js";
 import { ScheduledTaskRunTracker } from "../schedules/run-tracker.js";
-/** 运行的终态。 */
-export type TerminalStatus = "completed" | "failed" | "aborted";
-export declare const TERMINAL_STATUSES: ReadonlySet<string>;
+import { ApprovalsService } from "../tools/approvals.service.js";
+import { QuestionsService } from "../tools/questions.service.js";
+import { type TerminalStatus } from "./run-jobs.js";
 export interface FinishRunArgs {
     conversationId: string;
     runId: string;
@@ -18,15 +19,10 @@ export interface FinishRunArgs {
      * 超时（max_duration）传 false：loop 没吸收的消息没有被回复，必须留给下一轮（R34）。
      */
     floorToTrigger?: boolean;
-    /**
-     * Runs in the transaction once the run is known not to be terminal, before the terminal event:
-     * a parent's open `Agent` calls get their results here (sub-agents spec §12, review m1).
-     */
-    beforeTerminal?: () => Promise<void>;
 }
 /**
- * Writes runs' terminal states (one instance; the run worker, the run queue's orphan repairs, the
- * waits and the eval replays finish runs through it).
+ * Writes runs' terminal states (one instance; the run worker, the run queue's orphan abort, the
+ * `run-orphan-check`, the waits and the eval replays finish runs through it).
  */
 export declare class RunFinish {
     private readonly kdb;
@@ -34,20 +30,51 @@ export declare class RunFinish {
     readonly txHost: TransactionHost;
     /** A scheduled task's run follows its kernel run's end in the same transaction (spec §6). */
     private readonly taskRuns;
+    /** The run's end owes the conversation's ensure-run backstop (transactional outbox spec §6). */
+    private readonly outbox;
+    /** A finished run's pending approvals are cancelled in its terminal transaction. */
+    private readonly approvals;
+    /** ... and its pending input requests. */
+    private readonly questions;
     constructor(kdb: KernelDb, eventLog: EventLog, txHost: TransactionHost, 
     /** A scheduled task's run follows its kernel run's end in the same transaction (spec §6). */
-    taskRuns: Pick<ScheduledTaskRunTracker, "onRunFinished">);
+    taskRuns: Pick<ScheduledTaskRunTracker, "onRunFinished">, 
+    /** The run's end owes the conversation's ensure-run backstop (transactional outbox spec §6). */
+    outbox: Pick<OutboxService, "add">, 
+    /** A finished run's pending approvals are cancelled in its terminal transaction. */
+    approvals: Pick<ApprovalsService, "cancelPendingIn">, 
+    /** ... and its pending input requests. */
+    questions: Pick<QuestionsService, "cancelPendingIn">);
     /**
      * 在调用方的事务内写运行的终态：会话行锁 → 运行仍未结束才写终态事件、推进 `consumed_seq`、更新 `runs.status`。
+     * It also owes the conversation's ensure-run backstop under the terminal event's seq.
+     *
+     * A finished run waits on nothing: its pending approvals and input requests (one asked just
+     * before the run failed, while it was still `running`) are cancelled before its terminal event
+     * (`approval_decided` / `question_answered` with `cancelled`). Lock order, as every path that
+     * writes them: the conversation row, the run's approval and input request rows, the run row.
+     *
+     * Sub-agents (transactional outbox spec §6): before the terminal event the run's open fork
+     * invocation gets its report (`pairStoppedForks`; an `Agent` call has its result from its
+     * dispatch), so no terminal path leaves an open call. After it, the stops and the delivery the
+     * run's end causes are owed here (runtime/subagents/tree-jobs.ts), the caller still doing them
+     * right after its commit: `stop-tree` when an active fork child remains or a sub-agent run a
+     * person stopped has sub-agents below it, `deliver-to-parent` for a sub-agent's run. The
+     * delivery is a job, not written here: this transaction holds the child's conversation and run
+     * rows, and the parent's conversation row comes first in the dispatch's lock order (parent
+     * conversation, parent run, budget run, child conversation).
      * 返回终态事件的 seq；运行已是终态（例如已被其他副本标记失败）→ 返回 null，什么都不写。
      *
      * `consumed_seq = GREATEST(consumed_seq, consumed, runs.trigger_seq)`：不回退，且失败 / 中止 / 完成都至少
      * 消费到触发该运行的消息——否则结束时的 ensureRun 会对同一条消息无限重试（失败）或卡住会话。
      * 超时（`floorToTrigger: false`）只取 `GREATEST(consumed_seq, consumed)`：没吸收的消息留给下一轮（R34）。
-     * 加锁顺序与 `RunQueue.ensureRun` 一致（先会话行、后 runs 行），不会互相死锁。
+     * Lock order as `RunQueue.ensureRun` and the decisions take it: the conversation row, the run's
+     * approval and input request rows, the run row; they never deadlock with each other.
      * 调用方在提交后对非 null 的返回值调用 {@link notifyFinished}。
      */
     finishIn(a: FinishRunArgs): Promise<number | null>;
+    /** {@link finishIn}'s owed stops and delivery of a run's sub-agent tree. */
+    private oweTreeWork;
     /** 自开事务写终态并在提交后发布会话通知；返回值同 {@link finishIn}。 */
     finish(a: FinishRunArgs): Promise<number | null>;
     private finishOwn;