@capybara-hq/kernel 0.1.0-rc.4 → 0.1.0-rc.6

package/dist/runtime/run-queue.js

@@ -15,6 +15,7 @@ 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";
@@ -27,6 +28,8 @@ 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；
@@ -98,9 +101,11 @@ let RunQueue = class RunQueue {
      * 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 (background
-     * Bash spec §3, {@link wakeDecisionIn}): the run is then attributed to the task's trigger
-     * account, and a finish inside the wake window schedules the delayed `bg-wake` job instead.
+     * 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
@@ -124,8 +129,14 @@ let RunQueue = class RunQueue {
             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);
         }
@@ -155,10 +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);
@@ -197,11 +208,13 @@ let RunQueue = class RunQueue {
             .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;
+        if (conv.kind === "subagent")
+            return this.childWakeIn(conversationId, conv, st);
         const newest = await t
             .selectFrom("conversation_events")
             .select(["seq", "payload"])
@@ -210,6 +223,8 @@ let RunQueue = class RunQueue {
             .orderBy("seq", "desc")
             .limit(1)
             .executeTakeFirst();
+        /** 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,
@@ -220,8 +235,8 @@ let RunQueue = class RunQueue {
             : 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): it goes through the orphan check below with
-            // that boundary, as a message's run does
+            // 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)) {
@@ -276,12 +291,13 @@ let RunQueue = class RunQueue {
             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 }, Number(conv.consumed_seq), this.config.get().runtime.wakeOnBackgroundTask, (accountId) => authorCanSend(t, conversationId, accountId));
+                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;
             }
         }
         const active = await this.activeRun(t, conversationId);
@@ -357,9 +373,170 @@ let RunQueue = class RunQueue {
             agent_version: agent.current_version,
             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 };
+        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, 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

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;
     /**
@@ -164,7 +170,7 @@ export declare class RunWorker {
      * model removed), writes `model_switched { model: null, effort: null, reason: "unavailable" }`
      * for this run (idempotent) and clears the cached selection: the run and the next ones use the
      * agent's model. A provider that only failed to list its models is not "gone" (the run's ladder
-     * handles it).
+     * handles it). Resolved for the conversation's org (its effective provider config).
      */
     private fallBackIfUnavailable;
     /**

package/dist/runtime/run-worker.js

@@ -10,7 +10,7 @@ var __param = (this && this.__param) || function (paramIndex, decorator) {
 import { Inject, Injectable, Optional } from "@nestjs/common";
 import { DelayedError } from "bullmq";
 import { sql } from "kysely";
-import { conversationTopic, EventLog, } from "../conversations/event-log.js";
+import { EventLog } from "../conversations/event-log.js";
 import { isAbsorbable, } from "../conversations/events.js";
 import { effectiveDefinition, NO_SELECTION, readSelection, sameModel, sameSelection, } from "../conversations/selection.js";
 import { TitleQueue } from "../conversations/titles.js";
@@ -32,10 +32,11 @@ import { ConversationLock } from "./conversation-lock.js";
 import { RunError } from "./model-call.js";
 import { ProviderListError, ProviderRegistry } from "./providers.js";
 import { ORG_QUOTA_LEASE_MS, orgRunLimit, RUN_QUOTA } from "./quota.js";
+import { RunDrafts, streamSink } from "./run-drafts.js";
 import { RunFinish, TERMINAL_STATUSES } from "./run-finish.js";
 import { RunQueue } from "./run-queue.js";
 import { AgentLoops, RunSteps } from "./run-steps.js";
-import { dispatchOfChildRun } from "./subagents/child.js";
+import { dispatchOfChildRun, rootConversationOf } from "./subagents/child.js";
 import { childRunDefinition } from "./subagents/definition.js";
 import { forkBoundary } from "./subagents/skill-fork.js";
 import { Waits } from "./waits.js";
@@ -86,6 +87,7 @@ let RunWorker = class RunWorker {
     pipeline;
     finisher;
     txHost;
+    drafts;
     hooks;
     /** The two run queues' BullMQ workers (spec §7.4), once {@link start}ed on the worker role. */
     workers = [];
@@ -114,7 +116,9 @@ let RunWorker = 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, finisher, txHost, hooks) {
+    pipeline, finisher, txHost, 
+    /** The steps' live drafts this process writes for the runs it executes. */
+    drafts, hooks) {
         this.kdb = kdb;
         this.eventLog = eventLog;
         this.lock = lock;
@@ -134,6 +138,7 @@ let RunWorker = class RunWorker {
         this.pipeline = pipeline;
         this.finisher = finisher;
         this.txHost = txHost;
+        this.drafts = drafts;
         this.hooks = hooks;
         this.loop = loops.create(steps.deps);
     }
@@ -178,35 +183,16 @@ let RunWorker = class RunWorker {
         }
     }
     /**
-     * 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).
      */
     sinkFor(conversationId) {
-        // 本次运行执行中推出的文本字符数（按步，step_reset 清零）：本进程正在执行的这次运行自己的计数
-        let at = { runId: "", step: -1, chars: 0 };
-        return (d) => {
-            if (at.runId !== d.runId || at.step !== d.step || d.kind === "step_reset") {
-                at = { runId: d.runId, step: d.step, chars: 0 };
-            }
-            const offset = at.chars;
-            if (d.kind === "text")
-                at.chars += d.text?.length ?? 0;
-            const msg = {
-                kind: "delta",
-                runId: d.runId,
-                step: d.step,
-                delta: d.kind,
-                ...(d.text !== undefined && { text: d.text }),
-                ...(d.kind === "text" && { offset }),
-            };
-            this.bus
-                .publish(conversationTopic(conversationId), msg)
-                .catch((err) => this.log.warn("stream delta publish failed", { conversationId, err }));
-        };
+        return streamSink(conversationId, { bus: this.bus, drafts: this.drafts, log: this.log });
     }
     /**
      * 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.
      */
     async stopChildren(runId) {
         try {
@@ -228,6 +214,14 @@ let RunWorker = class RunWorker {
         catch (err) {
             this.log.error("cannot report finished sub-agent run to its parent", { runId, err });
         }
+        // a sub-agent a person stopped stops its descendants too, also one it dispatched after the
+        // tree stop read the tree (agent map spec §6); any other end leaves them running
+        try {
+            await this.waits.stopDescendantsOfStopped(runId);
+        }
+        catch (err) {
+            this.log.error("cannot stop the descendants of a stopped sub-agent", { runId, err });
+        }
     }
     /**
      * 任务结束后的 ensureRun，失败按 {@link ENSURE_RETRY_DELAYS_MS} 退避重试；停机开始后不再等待间隔
@@ -369,6 +363,7 @@ let RunWorker = class RunWorker {
             "trigger_seq",
             "started_at",
             "waited_ms",
+            "budget_run_id",
         ])
             .where("id", "=", runId)
             .executeTakeFirst();
@@ -386,7 +381,7 @@ let RunWorker = class RunWorker {
             return;
         const conv = await this.db
             .selectFrom("conversations")
-            .select(["org_id", "agent_id", "kind"])
+            .select(["org_id", "agent_id", "kind", "subagent_depth"])
             .where("id", "=", conversationId)
             .executeTakeFirst();
         if (!conv)
@@ -532,18 +527,36 @@ let RunWorker = class RunWorker {
             maxSteps = Math.min(started.payload.maxSteps, runtime.maxSteps);
             const row = await dispatchOfChildRun(this.db, runId);
             const otherAgent = started.payload.template.source === "agent";
+            const parentConversationId = row?.parentConversationId ?? started.payload.parent.conversationId;
+            // a sub-agent conversation is at depth 1 at least: one written before depths were recorded
+            // (default 0) was a child of a root
+            const depth = Math.max(conv.subagent_depth, 1) - 1;
+            const rootConversationId = depth === 0
+                ? parentConversationId
+                : await rootConversationOf(this.db, parentConversationId);
+            // the tree's budget window of THIS run (agent map spec §6): its own `budget_run_id`, set when
+            // it was created (a dispatch, or a wake run from the root's current window), else its
+            // dispatch's; never derived from the parent run, which for a wake run is the first dispatch's
+            const budgetRunId = run.budget_run_id ?? row?.budgetRunId;
             parent = row
                 ? {
                     conversationId: row.parentConversationId,
                     runId: row.parentRunId,
                     callId: row.parentCallId,
+                    depth,
+                    rootConversationId,
                     otherAgent,
+                    ...(budgetRunId !== undefined && { budgetRunId }),
+                    ...(row.forSkill === true && { forSkill: true }),
                 }
                 : {
                     conversationId: started.payload.parent.conversationId,
                     runId: started.payload.parent.runId,
                     callId: started.payload.parent.callId,
+                    depth,
+                    rootConversationId,
                     otherAgent,
+                    ...(budgetRunId !== undefined && { budgetRunId }),
                 };
         }
         else {
@@ -568,7 +581,7 @@ let RunWorker = class RunWorker {
         if (run.status === "queued" || startedAt === null) {
             if (!existingStart) {
                 if (saved !== undefined)
-                    await this.fallBackIfUnavailable(conversationId, runId, saved);
+                    await this.fallBackIfUnavailable(conv.org_id, conversationId, runId, saved);
                 const started = await this.appendRunStarted(conversationId, run, agent, saved);
                 agent = started.agent;
                 const toSeq = started.toSeq;
@@ -703,6 +716,10 @@ let RunWorker = class RunWorker {
                 });
                 return "finished";
             }
+            finally {
+                // the execution here ended: what this process held of the run's drafts goes
+                this.drafts.release(runId);
+            }
             if (state.abandoned)
                 return "abandoned";
             // 锁在运行中丢失：无论 loop 返回什么都不写终态，任务回到队列由持锁者续跑。
@@ -772,14 +789,14 @@ let RunWorker = class RunWorker {
      * model removed), writes `model_switched { model: null, effort: null, reason: "unavailable" }`
      * for this run (idempotent) and clears the cached selection: the run and the next ones use the
      * agent's model. A provider that only failed to list its models is not "gone" (the run's ladder
-     * handles it).
+     * handles it). Resolved for the conversation's org (its effective provider config).
      */
-    async fallBackIfUnavailable(conversationId, runId, agent) {
+    async fallBackIfUnavailable(orgId, conversationId, runId, agent) {
         const sel = await readSelection(this.db, conversationId);
         if (sel.model === null || sameModel(sel.model, agent.model))
             return;
         try {
-            await this.models.resolve(sel.model);
+            await this.models.resolve(orgId, sel.model);
             return;
         }
         catch (e) {
@@ -816,7 +833,8 @@ let RunWorker = class RunWorker {
         const agent = saved !== undefined
             ? effectiveDefinition(saved, await readSelection(t, conversationId))
             : current;
-        // a wake run's boundary is a finished background task, not a message (background Bash §3)
+        // a wake run's boundary is a finished background task (background Bash §3) or a sub-agent's
+        // delivery (agent map spec §5.3), not a message
         const boundary = await t
             .selectFrom("conversation_events")
             .select("type")
@@ -836,6 +854,10 @@ let RunWorker = class RunWorker {
                         trigger: "background_task",
                         triggerSeq: Number(run.trigger_seq),
                     }),
+                    ...(boundary?.type === "subagent_delivery" && {
+                        trigger: "subagent",
+                        triggerSeq: Number(run.trigger_seq),
+                    }),
                 },
             },
         ]);
@@ -927,7 +949,8 @@ RunWorker = __decorate([
     __param(18, Inject(ToolPipeline)),
     __param(19, Inject(RunFinish)),
     __param(20, Inject(TransactionHost)),
-    __param(21, Optional()),
-    __param(21, Inject(RUN_HOOKS))
+    __param(21, Inject(RunDrafts)),
+    __param(22, Optional()),
+    __param(22, Inject(RUN_HOOKS))
 ], RunWorker);
 export { RunWorker };

package/dist/runtime/runtime.module.js

@@ -25,6 +25,7 @@ import { MAINTENANCE_QUEUE_NAME } from "./maintenance-queue.js";
 import { ProvidersModule } from "./providers.js";
 import { EVAL_QUOTA, EVAL_QUOTA_PREFIX, OrgQuota, RUN_QUOTA, RUN_QUOTA_PREFIX } from "./quota.js";
 import { RunControl } from "./run-control.js";
+import { RunDrafts } from "./run-drafts.js";
 import { RunFinish } from "./run-finish.js";
 import { RunJobsModule } from "./run-jobs.js";
 import { RunQueue } from "./run-queue.js";
@@ -75,6 +76,7 @@ RuntimeModule = __decorate([
             Waits,
             DecisionResumer,
             RunControl,
+            RunDrafts,
             AgentLoops,
             RunSteps,
             RunWorker,
@@ -102,6 +104,7 @@ RuntimeModule = __decorate([
             RunQueue,
             Waits,
             RunControl,
+            RunDrafts,
             AgentLoops,
             RunSteps,
             RunWorker,

package/dist/runtime/step-log.d.ts

@@ -19,11 +19,16 @@ export declare class StepLog {
     constructor(kdb: KernelDb, eventLog: EventLog, txHost: TransactionHost);
     /**
      * Appends events of a sub-agent's run and adds the tokens of the usage events it newly wrote to
-     * the parent 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. Returns
-     * what was written (hits included) and the highest new seq (null: nothing new).
+     * the `subagent_tokens_used` of the tree's budget run (sub-agents spec §8.3, agent map spec §6:
+     * the whole tree counts on one run): `counter.budgetRunId`, the run's own window
+     * (`SubagentParent.budgetRunId`), else the parent run's ({@link budgetRunOf}). A usage event
+     * already in the log (an idempotent hit after a crash) adds nothing, so each call counts exactly
+     * once. Returns what was written (hits included) and the highest new seq (null: nothing new).
      */
-    appendCounted(conversationId: string, parentRunId: string, events: AppendEvent[]): Promise<{
+    appendCounted(conversationId: string, counter: {
+        parentRunId: string;
+        budgetRunId?: string | undefined;
+    }, events: AppendEvent[]): Promise<{
         written: ConversationEvent[];
         toSeq: number | null;
     }>;

package/dist/runtime/step-log.js

@@ -16,6 +16,7 @@ import { KernelDb } from "../infra/database/kernel-db.js";
 import { TransactionHost } from "../infra/database/transaction-host.js";
 import { Transactional } from "../infra/database/transactional.decorator.js";
 import { recordToolSpecs } from "../tools/specs.js";
+import { budgetRunOf } from "./subagents/limits.js";
 /**
  * The writes of a run step that go to the event log together with another row, each in a
  * transaction of its own (the caller notifies the log after the commit): a model call with the
@@ -33,11 +34,13 @@ let StepLog = class StepLog {
     }
     /**
      * Appends events of a sub-agent's run and adds the tokens of the usage events it newly wrote to
-     * the parent 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. Returns
-     * what was written (hits included) and the highest new seq (null: nothing new).
+     * the `subagent_tokens_used` of the tree's budget run (sub-agents spec §8.3, agent map spec §6:
+     * the whole tree counts on one run): `counter.budgetRunId`, the run's own window
+     * (`SubagentParent.budgetRunId`), else the parent run's ({@link budgetRunOf}). A usage event
+     * already in the log (an idempotent hit after a crash) adds nothing, so each call counts exactly
+     * once. Returns what was written (hits included) and the highest new seq (null: nothing new).
      */
-    async appendCounted(conversationId, parentRunId, events) {
+    async appendCounted(conversationId, counter, events) {
         const t = this.kdb.db;
         // the conversation row lock the append takes anyway, taken first to read the seq before it
         const before = await t
@@ -56,7 +59,7 @@ let StepLog = class StepLog {
             await t
                 .updateTable("runs")
                 .set({ subagent_tokens_used: sql `subagent_tokens_used + ${tokens}` })
-                .where("id", "=", parentRunId)
+                .where("id", "=", counter.budgetRunId ?? (await budgetRunOf(t, counter.parentRunId)))
                 .execute();
         }
         return { written, toSeq: fresh.length > 0 ? Math.max(...fresh.map((e) => e.seq)) : null };