@capybara-hq/kernel 0.1.0-rc.1 → 0.1.0-rc.10

package/dist/evals/snapshot.js

@@ -97,7 +97,7 @@ export async function takeSnapshot(db, orgId, conversationId, runId, maxBytes) {
         .orderBy("version")
         .execute();
     for (const r of defRows) {
-        definitions[String(r.version)] = normalizeStoredDefinition(r.definition);
+        definitions[String(r.version)] = snapshotDefinition(r.definition);
     }
     const toolSpecs = {};
     if (hashes.size > 0) {
@@ -151,14 +151,27 @@ export async function takeSnapshot(db, orgId, conversationId, runId, maxBytes) {
     return snapshot;
 }
 /**
- * 读出库里存的快照（`eval_cases.snapshot`）：把各版本定义里改名前（D17）的 `{ kind: "app" }` 工具引用规范化为
- * connector 形（C27）。整轮回放会把定义交给实时注册表，不规范化的话连接器工具全部失效。其余字段原样。
+ * A definition as a snapshot holds it: normalized (`normalizeStoredDefinition`), and one from
+ * before tools were on by default keeps its enabled list next to it: rc tool keys are read by its
+ * extension refs (`legacyScopeOf`); the runtime ignores it (tools default-on spec §2).
+ */
+function snapshotDefinition(d) {
+    const enabled = d?.tools;
+    return {
+        ...normalizeStoredDefinition(d),
+        ...(Array.isArray(enabled) && { tools: enabled }),
+    };
+}
+/**
+ * Reads a stored snapshot (`eval_cases.snapshot`): each version's definition goes through
+ * `normalizeStoredDefinition` (rc tool keys, and an enabled tool list read as nothing turned off).
+ * Other fields as they are.
  */
 export function readSnapshot(raw) {
     const s = raw;
     const definitions = {};
     for (const [v, d] of Object.entries(s.definitions ?? {})) {
-        definitions[v] = normalizeStoredDefinition(d);
+        definitions[v] = snapshotDefinition(d);
     }
     return {
         ...s,

package/dist/evals/step-replay.d.ts

@@ -1,10 +1,10 @@
 import { type ContentBlock, type ModelUsage, type ParamValue, type StopReason } from "@capybara-hq/plugin-sdk";
 import type { AgentDefinition, ModelRef } from "../agents/definition.js";
-import type { KernelConfig } from "../config/schema.js";
 import { type Selection } from "../conversations/selection.js";
+import type { KernelConfigService } from "../infra/config/kernel-config.service.js";
 import { type MediaSourceDeps } from "../runtime/model-call.js";
 import type { ProviderRegistry } from "../runtime/providers.js";
-import { type DriftReason } from "./reconstruct.js";
+import { type DriftReason, type SubagentTemplateRef } from "./reconstruct.js";
 import type { CaseSnapshot } from "./snapshot.js";
 /**
  * 单步回放（D2 / C5）：由用例快照还原某一步的模型请求，换成候选（agent 版本 / 模型 / 参数 / 指令）重发，
@@ -33,10 +33,7 @@ export interface ResolvedCandidate {
      * A plugin template's child case: the template's current version (set by the worker), compared
      * with the logged one for `subagent_changed` when the candidate changed.
      */
-    subagentTemplate?: {
-        id: string;
-        version: number;
-    };
+    subagentTemplate?: SubagentTemplateRef;
     /**
      * A child case under another agent version: that version's own definition (with the candidate's
      * instructions), compared with the logged sub-agent for `subagent_changed`. Never rendered.
@@ -56,7 +53,7 @@ export interface ReplayOutput {
 }
 export interface StepReplayDeps {
     registry: Pick<ProviderRegistry, "resolve">;
-    config: () => KernelConfig;
+    config: Pick<KernelConfigService, "get">;
     /** 附件来源（与运行时同一个，按组织 + 对话取）；不给时媒体一律按「不再可用」占位。 */
     media?: MediaSourceDeps;
 }
@@ -95,7 +92,8 @@ export declare class ReplayAttemptError extends Error {
  * 否则带候选的定义、版本与参数——参数按候选模型的声明宽松处理，与运行时一致），用 `consumeModelStream` 收完。
  * 请求发出之前的错误原样抛出（`agent.model_unavailable`、取附件暂时失败为 `MediaUnavailableError`）；请求发出之后的
  * 错误包成 {@link ReplayAttemptError}（带估算用量，调用方计入预算后按 `cause` 分类重试）。
- * 模型按部署范围解析（与运行时一致；组织级模型准入若在第二段落地，须同样覆盖评测候选）。
+ * The candidate model resolves in the snapshot's org like the runtime (its effective provider
+ * config and disabled lists: a model disabled there is `agent.model_unavailable`, spec B §5.3).
  * 不另设单次请求的墙钟时限：与运行时一样只有流空闲看门狗（`streamIdleTimeoutMs` / `runtime.streamIdleMs`）。
  */
 export declare function replayStep(d: StepReplayDeps, s: CaseSnapshot, call: {

package/dist/evals/step-replay.js

@@ -2,7 +2,7 @@ import { sanitizeParams, } from "@capybara-hq/plugin-sdk";
 import { effectiveDefinition, NO_SELECTION, sameModel, } from "../conversations/selection.js";
 import { estimateTokens } from "../runtime/context/tokens.js";
 import { consumeModelStream, MediaUnavailableError, } from "../runtime/model-call.js";
-import { loadSnapshotMedia, reconstructRequest } from "./reconstruct.js";
+import { loadSnapshotMedia, reconstructRequest, } from "./reconstruct.js";
 /** 键排序后的 JSON（比较参数是否相同）。 */
 const canonical = (v) => JSON.stringify(Object.keys(v)
     .sort()
@@ -54,10 +54,9 @@ subagent) {
                 ? { ...(from.modelParams ?? {}) }
                 : {});
     const definition = {
-        ...from,
+        ...withInstructions(from, c.instructions),
         model,
         modelParams: params,
-        ...(c.instructions !== undefined && { instructions: c.instructions }),
     };
     const changed = agentVersion !== base.agentVersion ||
         !sameModel(model, base.definition.model) ||
@@ -73,13 +72,20 @@ subagent) {
         changed,
         overridesSwitch,
         ...(otherVersion && {
-            subagentVersion: {
-                ...versionDef,
-                ...(c.instructions !== undefined && { instructions: c.instructions }),
-            },
+            subagentVersion: withInstructions(versionDef, c.instructions),
         }),
     };
 }
+/**
+ * A candidate's `instructions` override: the prompt files it was joined from no longer describe
+ * it, so they are dropped (never model-visible either way).
+ */
+function withInstructions(def, instructions) {
+    if (instructions === undefined)
+        return def;
+    const { instructionFiles: _files, ...rest } = def;
+    return { ...rest, instructions };
+}
 /**
  * 候选请求已经发出、但没有正常收完（出错、超时、中止）：`tokens` 是这次尝试应计入预算的估算用量，
  * `cause` 是原始错误（调用方按它分类重试）。请求发出之前的失败（解析模型、取附件）原样抛出，不计费。
@@ -120,12 +126,13 @@ function baselineOf(s, call) {
  * 否则带候选的定义、版本与参数——参数按候选模型的声明宽松处理，与运行时一致），用 `consumeModelStream` 收完。
  * 请求发出之前的错误原样抛出（`agent.model_unavailable`、取附件暂时失败为 `MediaUnavailableError`）；请求发出之后的
  * 错误包成 {@link ReplayAttemptError}（带估算用量，调用方计入预算后按 `cause` 分类重试）。
- * 模型按部署范围解析（与运行时一致；组织级模型准入若在第二段落地，须同样覆盖评测候选）。
+ * The candidate model resolves in the snapshot's org like the runtime (its effective provider
+ * config and disabled lists: a model disabled there is `agent.model_unavailable`, spec B §5.3).
  * 不另设单次请求的墙钟时限：与运行时一样只有流空闲看门狗（`streamIdleTimeoutMs` / `runtime.streamIdleMs`）。
  */
 export async function replayStep(d, s, call, candidate, signal) {
     const baseline = baselineOf(s, call);
-    const { provider, info } = await d.registry.resolve(candidate.model);
+    const { provider, info } = await d.registry.resolve(s.orgId, candidate.model);
     let media;
     if (d.media) {
         try {
@@ -151,7 +158,7 @@ export async function replayStep(d, s, call, candidate, signal) {
         }),
         ...(media !== undefined && { mediaInfo: media.mediaInfo, media: media.media }),
     });
-    const idleMs = info.streamIdleTimeoutMs ?? d.config().runtime.streamIdleMs;
+    const idleMs = info.streamIdleTimeoutMs ?? d.config.get().runtime.streamIdleMs;
     const started = Date.now();
     let streamedChars = 0;
     let out;

package/dist/evals/trajectory.controller.d.ts

@@ -1,7 +1,7 @@
 import { z } from "zod";
 import { type Principal } from "../identity/principal.js";
 import { type OrgContext } from "../orgs/org.guard.js";
-import { type TrajectoryDetail, type TrajectoryListItem, type TrajectoryService } from "./trajectory.service.js";
+import { type TrajectoryDetail, type TrajectoryListItem, TrajectoryService } from "./trajectory.service.js";
 declare const ListQuery: z.ZodObject<{
     agentId: z.ZodOptional<z.ZodString>;
     kind: z.ZodOptional<z.ZodEnum<{

package/dist/evals/trajectory.controller.js

@@ -7,13 +7,13 @@ var __decorate = (this && this.__decorate) || function (decorators, target, key,
 var __param = (this && this.__param) || function (paramIndex, decorator) {
     return function (target, key) { decorator(target, key, paramIndex); }
 };
-import { Controller, Get, Inject, Param, Query } from "@nestjs/common";
+import { Controller, Get, Inject, Param } from "@nestjs/common";
 import { z } from "zod";
-import { TRAJECTORIES } from "../http/tokens.js";
-import { ZodPipe } from "../http/zod.pipe.js";
 import { CurrentPrincipal, patTokenId, requireUser, } from "../identity/principal.js";
+import { createZodDto } from "../infra/validation/create-zod-dto.js";
+import { ZQuery } from "../infra/validation/zod-params.js";
 import { CurrentOrg, OrgAction } from "../orgs/org.guard.js";
-import { RUN_STATUSES, } from "./trajectory.service.js";
+import { RUN_STATUSES, TrajectoryService, } from "./trajectory.service.js";
 const ListQuery = z.object({
     agentId: z.string().min(1).max(100).optional(),
     kind: z.enum(["chat", "eval"]).optional(),
@@ -34,6 +34,10 @@ const DetailQuery = z
 });
 /** 只给 `before` 时的页大小。 */
 const DEFAULT_RUN_PAGE = 20;
+class TrajectoryListQueryDto extends createZodDto(ListQuery, "TrajectoryListQuery") {
+}
+class TrajectoryDetailQueryDto extends createZodDto(DetailQuery, "TrajectoryDetailQuery") {
+}
 /**
  * 轨迹查看器（`/api/v1/orgs/:orgId/trajectories`）：`trajectories.read`（member 起、只给人），
  * 服务里再判定「组织管理员或 agent 负责人」；无权一律 404 `trajectory.not_found`。
@@ -71,7 +75,7 @@ __decorate([
     OrgAction("trajectories.read"),
     __param(0, CurrentPrincipal()),
     __param(1, CurrentOrg()),
-    __param(2, Query(new ZodPipe(ListQuery)))
+    __param(2, ZQuery(TrajectoryListQueryDto))
 ], TrajectoryController.prototype, "list", null);
 __decorate([
     Get(":conversationId"),
@@ -79,10 +83,10 @@ __decorate([
     __param(0, CurrentPrincipal()),
     __param(1, CurrentOrg()),
     __param(2, Param("conversationId")),
-    __param(3, Query(new ZodPipe(DetailQuery)))
+    __param(3, ZQuery(TrajectoryDetailQueryDto))
 ], TrajectoryController.prototype, "get", null);
 TrajectoryController = __decorate([
     Controller("api/v1/orgs/:orgId/trajectories"),
-    __param(0, Inject(TRAJECTORIES))
+    __param(0, Inject(TrajectoryService))
 ], TrajectoryController);
 export { TrajectoryController };

package/dist/evals/trajectory.d.ts

@@ -54,7 +54,7 @@ export interface TrajectoryToolCall {
         byAccountId: string | null;
         decidedAt: string | null;
     } | null;
-    /** 最近一次输入请求（`ask_user`、插件 `input_required`、MCP elicitation）。 */
+    /** 最近一次输入请求（`AskUserQuestion`、插件 `input_required`、MCP elicitation）。 */
     question: {
         questionId: string;
         askedAt: string;
@@ -64,7 +64,7 @@ export interface TrajectoryToolCall {
     /** A `tool.preExecute` policy hook denied the call (it never started). */
     hookDenied?: TrajectoryHookDenial;
     /**
-     * An `agent` call that dispatched (or resumed) a sub-agent (phase C, spec §15.4): the child
+     * An `Agent` call that dispatched (or resumed) a sub-agent (phase C, spec §15.4): the child
      * conversation (`subagentId`, whose trajectory the viewer links to), the run it started and that
      * run's end (`running` until its `subagent_finished`).
      */

package/dist/evals/trajectory.js

@@ -130,7 +130,7 @@ function buildRun({ started, events }, prices, forks) {
                 assistants.set(e.payload.step, [...(assistants.get(e.payload.step) ?? []), e]);
                 break;
             case "usage":
-                // a compaction or title call's usage is not a step call's: counted in the totals only (I-1)
+                // a compaction, title or web_fetch call's usage is not a step call's: counted in the totals only (I-1)
                 if (e.payload.purpose !== undefined) {
                     compactions.push(e);
                     break;

package/dist/evals/trajectory.service.d.ts

@@ -1,7 +1,8 @@
-import type { EventLog } from "../conversations/event-log.js";
-import type { ConversationKind, KernelDb } from "../db/kernel-schema.js";
-import type { SettingsService } from "../identity/settings.js";
-import type { PricesService } from "../models/prices.js";
+import { EventLog } from "../conversations/event-log.js";
+import type { ConversationKind } from "../db/kernel-schema.js";
+import { SettingsService } from "../identity/settings.js";
+import { KernelDb } from "../infra/database/kernel-db.js";
+import { PricesService } from "../models/prices.js";
 import { type TrajectoryRun } from "./trajectory.js";
 /** 轨迹列表的一项（不含对话内容，列表不写审计）。 */
 export interface TrajectoryListItem {
@@ -57,7 +58,7 @@ export type RunStatus = (typeof RUN_STATUSES)[number];
 /**
  * 轨迹查看器（D6、C1）：只给组织管理员（admin / owner）与 agent 负责人。负责人与角色每次实时读（R5）；
  * 无权查看与不存在一律 `trajectory.not_found`，不泄露对话是否存在。每次查看详情写审计 `trajectory.viewed`。
- * 企业管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
+ * 超级管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
  *
  * 列表按 `conversations.kind` 过滤，不传时只列 `chat`（评测回放的隐藏对话用 `kind: "eval"` 列出）。
  * 评测对话（`kind = 'eval'`）只给组织管理员（Ruling S1-c）：里面是管理员的候选指令，非管理员的 agent 负责人
@@ -97,8 +98,13 @@ export declare class TrajectoryService {
     get(orgId: string, accountId: string, conversationId: string, actorTokenId?: string, window?: RunWindow): Promise<TrajectoryDetail>;
     /** The parent call that first dispatched a sub-agent conversation. */
     private firstDispatch;
-    /** The agent of a sub-agent conversation's parent conversation, or null. */
-    private parentAgentId;
+    /**
+     * The agent of the root conversation of a sub-agent conversation's tree, from its parent `parentId`
+     * up (the same walk as `conversations/access.ts`: at most 16 levels, every conversation between
+     * a sub-agent conversation of the org, ending at one without a parent), or null when the chain
+     * leaves the org, passes another kind or does not end.
+     */
+    private rootAgentId;
     /**
      * 窗口内已开始（`started_at` 非空，与轨迹只列有 `run_started` 的运行一致）的运行，按 (created_at, id) 倒序取
      * `limit` 次；`before` 是上一页的 `next`（运行 id）。`next`：还有更早的运行时为本页最早一次的 id。

package/dist/evals/trajectory.service.js

@@ -1,7 +1,22 @@
+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); }
+};
+var TrajectoryService_1;
+import { Inject, Injectable } from "@nestjs/common";
 import { sql } from "kysely";
-import { isAgentVisibleTo } from "../agents/agents.service.js";
+import { isAgentVisibleTo } from "../agents/agent-lookup.js";
+import { EventLog } from "../conversations/event-log.js";
 import { ApiError } from "../errors.js";
 import { audit } from "../identity/audit.js";
+import { SettingsService } from "../identity/settings.js";
+import { KernelDb } from "../infra/database/kernel-db.js";
+import { PricesService } from "../models/prices.js";
 import { roleAtLeast } from "../orgs/roles.js";
 import { buildTrajectory } from "./trajectory.js";
 /** 运行的状态（`runs.status`），列表按最近一次运行的状态过滤用。 */
@@ -19,13 +34,13 @@ const notFound = () => new ApiError("trajectory.not_found", "Trajectory not foun
 /**
  * 轨迹查看器（D6、C1）：只给组织管理员（admin / owner）与 agent 负责人。负责人与角色每次实时读（R5）；
  * 无权查看与不存在一律 `trajectory.not_found`，不泄露对话是否存在。每次查看详情写审计 `trajectory.viewed`。
- * 企业管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
+ * 超级管理员不因身份而能看（不是组织成员时 OrgGuard 已 404）。
  *
  * 列表按 `conversations.kind` 过滤，不传时只列 `chat`（评测回放的隐藏对话用 `kind: "eval"` 列出）。
  * 评测对话（`kind = 'eval'`）只给组织管理员（Ruling S1-c）：里面是管理员的候选指令，非管理员的 agent 负责人
  * 列不出、读详情 404。
  */
-export class TrajectoryService {
+let TrajectoryService = TrajectoryService_1 = class TrajectoryService {
     db;
     eventLog;
     prices;
@@ -55,7 +70,7 @@ export class TrajectoryService {
     /** 组织 admin / owner，或该 agent 当前的负责人（且至少是 member）。agent 已删除或不在本组织 → false。 */
     async canView(orgId, accountId, agentId) {
         const role = await this.role(orgId, accountId);
-        if (!TrajectoryService.isMember(role))
+        if (!TrajectoryService_1.isMember(role))
             return false;
         const agent = await this.db
             .selectFrom("agents")
@@ -74,7 +89,7 @@ export class TrajectoryService {
      */
     async list(orgId, accountId, q) {
         const role = await this.role(orgId, accountId);
-        if (!TrajectoryService.isMember(role))
+        if (!TrajectoryService_1.isMember(role))
             return { items: [], next: null };
         const admin = roleAtLeast(role, "admin");
         // 评测对话只给组织管理员（Ruling S1-c）
@@ -165,11 +180,13 @@ export class TrajectoryService {
             .where("id", "=", conversationId)
             .where("org_id", "=", orgId)
             .executeTakeFirst();
-        // A sub-agent's trajectory is part of its parent conversation's: authorized by the parent
-        // conversation's agent, never by the dispatched agent (an org agent's owner must not read a
-        // child drawn from someone else's conversation; batch 9 review I1)
+        // A sub-agent's trajectory is part of its tree's: authorized by the ROOT conversation's agent
+        // (the rule `parentReachable` applies to the conversation itself), never by the dispatched
+        // agent nor by the direct parent's, which below an org agent's child is that org agent (an org
+        // agent's owner must not read work drawn from someone else's conversation; batch 9 review I1,
+        // agent map final review: depth 2 and below)
         const authAgentId = conv?.kind === "subagent"
-            ? await this.parentAgentId(orgId, conv.parent_conversation_id)
+            ? await this.rootAgentId(orgId, conv.parent_conversation_id)
             : conv?.agent_id;
         if (!conv || authAgentId === null || authAgentId === undefined)
             throw notFound();
@@ -180,7 +197,7 @@ export class TrajectoryService {
         if (conv.kind === "subagent" && conv.agent_id !== authAgentId) {
             const role = await this.role(orgId, accountId);
             if ((role === null || !roleAtLeast(role, "admin")) &&
-                !(await isAgentVisibleTo(this.db, conv.agent_id, accountId))) {
+                !(await isAgentVisibleTo(this.db.db, conv.agent_id, accountId))) {
                 throw notFound();
             }
         }
@@ -222,7 +239,7 @@ export class TrajectoryService {
             this.settings.get(),
             conv.kind === "subagent" ? this.firstDispatch(conv.id) : undefined,
         ]);
-        await audit(this.db, {
+        await audit(this.db.db, {
             actorAccountId: accountId,
             ...(actorTokenId !== undefined && { actorTokenId }),
             orgId,
@@ -261,17 +278,27 @@ export class TrajectoryService {
                 callId: row.parent_call_id,
             };
     }
-    /** The agent of a sub-agent conversation's parent conversation, or null. */
-    async parentAgentId(orgId, parentId) {
+    /**
+     * The agent of the root conversation of a sub-agent conversation's tree, from its parent `parentId`
+     * up (the same walk as `conversations/access.ts`: at most 16 levels, every conversation between
+     * a sub-agent conversation of the org, ending at one without a parent), or null when the chain
+     * leaves the org, passes another kind or does not end.
+     */
+    async rootAgentId(orgId, parentId) {
         if (parentId === null)
             return null;
-        const p = await this.db
-            .selectFrom("conversations")
-            .select("agent_id")
-            .where("id", "=", parentId)
-            .where("org_id", "=", orgId)
-            .executeTakeFirst();
-        return p?.agent_id ?? null;
+        const row = await sql `
+      WITH RECURSIVE up(id, parent, kind, agent_id, n) AS (
+        SELECT c.id, c.parent_conversation_id, c.kind, c.agent_id, 0
+        FROM capybara.conversations c
+        WHERE c.id = ${parentId} AND c.org_id = ${orgId}
+        UNION ALL
+        SELECT c.id, c.parent_conversation_id, c.kind, c.agent_id, up.n + 1
+        FROM capybara.conversations c JOIN up ON c.id = up.parent
+        WHERE up.kind = 'subagent' AND up.n < 16 AND c.org_id = ${orgId}
+      )
+      SELECT agent_id FROM up WHERE parent IS NULL LIMIT 1`.execute(this.db.db);
+        return row.rows[0]?.agent_id ?? null;
     }
     /**
      * 窗口内已开始（`started_at` 非空，与轨迹只列有 `run_started` 的运行一致）的运行，按 (created_at, id) 倒序取
@@ -320,4 +347,12 @@ export class TrajectoryService {
             next: rows.length > limit ? (page.at(-1)?.id ?? null) : null,
         };
     }
-}
+};
+TrajectoryService = TrajectoryService_1 = __decorate([
+    Injectable(),
+    __param(0, Inject(KernelDb)),
+    __param(1, Inject(EventLog)),
+    __param(2, Inject(PricesService)),
+    __param(3, Inject(SettingsService))
+], TrajectoryService);
+export { TrajectoryService };

package/dist/evals/worker.d.ts

@@ -1,11 +1,21 @@
-import { type Logger } from "@capybara-hq/plugin-sdk";
-import type { KernelConfig } from "../config/schema.js";
-import type { KernelDb } from "../db/kernel-schema.js";
-import { type ModelPurposesService } from "../models/purposes.js";
+import { type Job, type Worker } from "bullmq";
+import { EventLog } from "../conversations/event-log.js";
+import type { KernelSchemaDb } 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 type { KernelLogger } from "../log.js";
+import { ModelPurposesService } from "../models/purposes.js";
+import { KyselyCheckpointSaver } from "../runtime/checkpointer.js";
+import { ProviderRegistry } from "../runtime/providers.js";
 import { type OrgQuota } from "../runtime/quota.js";
-import { type RunReplayDeps } from "./run-replay.js";
+import { RunFinish } from "../runtime/run-finish.js";
+import { AgentLoops, RunSteps } from "../runtime/run-steps.js";
+import { EvalBudget } from "./budget.js";
+import { type EvalJobData } from "./jobs.js";
+import { type SubagentTemplateRef } from "./reconstruct.js";
+import { type EvalReplayHooks, EvalReplaySeeder } from "./run-replay.js";
 import { type CaseSnapshot } from "./snapshot.js";
-import { type StepReplayDeps } from "./step-replay.js";
 /**
  * 评测 worker（C7，只在含 `worker` 角色的进程启动）：领取 `cap-kernel-evals` 的任务，每个任务回放一个用例。
  * - 运行不活跃（取消 / 达到上限 / 熔断 / 已结束）→ 该用例剩余结果标 `skipped`；
@@ -24,33 +34,34 @@ import { type StepReplayDeps } from "./step-replay.js";
 export declare const EVAL_QUOTA_RETRY_MS = 5000;
 /** 可重试错误的最多重试次数。 */
 export declare const EVAL_MAX_RETRIES = 3;
-export interface EvalWorkerDeps {
-    db: KernelDb;
-    redisUrl: string;
-    /** 评测配额（`new OrgQuota(redis, EVAL_QUOTA_PREFIX)`）。 */
-    quota: OrgQuota;
-    config: () => KernelConfig;
-    log: Logger;
-    replay: StepReplayDeps;
-    /** 整轮回放（mode = run）的依赖；认领相关的两项由 worker 按每次执行补上。 */
-    runReplay: Omit<RunReplayDeps, "stillClaimed" | "onSeeded">;
-    /** 评委（用途 `eval`，D3）：开了评委的运行每条 done 的结果发一次评委请求。 */
-    purposes: Pick<ModelPurposesService, "chat">;
-    /** 测试用：覆盖拿不到配额时的延后时长。 */
-    quotaRetryMs?: number;
-}
+/** The eval worker's BullMQ stalled-job limit (the processor's `maxStalledCount`). */
+export declare const EVAL_MAX_STALLED_COUNT = 3;
+/**
+ * Test hooks of the full-run replay and the judge (`EvalReplayHooks`): a provider override in tests
+ * (`overrides: [{ provide: EVAL_REPLAY_HOOKS, useValue }]`), absent in production.
+ */
+export declare const EVAL_REPLAY_HOOKS: unique symbol;
 export declare class EvalWorker {
-    private readonly d;
+    readonly txHost: TransactionHost;
     private worker;
-    private conn;
     /** 本进程在执行的任务：停机时中止，任务回到队列。 */
     private readonly active;
     private closing;
-    constructor(d: EvalWorkerDeps);
-    start(): void;
+    private readonly d;
+    constructor(kdb: KernelDb, quota: OrgQuota, config: KernelConfigService, log: KernelLogger, purposes: ModelPurposesService, registry: ProviderRegistry, steps: RunSteps, loops: AgentLoops, eventLog: EventLog, runFinish: RunFinish, checkpoints: KyselyCheckpointSaver, seeder: EvalReplaySeeder, budget: EvalBudget, txHost: TransactionHost, hooks?: EvalReplayHooks);
+    /**
+     * Starts consuming with the `cap-kernel-evals` processor's worker (worker role, after the
+     * extensions loaded; built with `autorun: false`).
+     */
+    start(worker: Worker<EvalJobData>): void;
+    /** The worker's `error` event. */
+    onError(err: Error): void;
+    /** The worker's `failed` event. */
+    onFailed(job: Job<EvalJobData> | undefined, err: Error): void;
     /** 停止领取；中止本进程在执行的任务（回到队列由别的 worker 接着做），在 budgetMs 内等它们退出。 */
     close(budgetMs: number): Promise<void>;
-    private process;
+    /** One job: replays one case (the processor runs it in `runJob`). */
+    process(job: Job<EvalJobData>, token: string | undefined): Promise<void>;
     /** 单步回放每次请求前：运行仍在执行（{@link runActive}）且结果仍由本次执行认领。 */
     private mayRequest;
     /**
@@ -82,6 +93,13 @@ export declare class EvalWorker {
      * - 回放出错 → 结果 error 并计入熔断；停机 → 任务回到队列，接手的执行从头重跑。
      */
     private runWhole;
+    /**
+     * 写一条单步回放的结果：只在仍由本次执行认领时写结果（stalled 后被接手的旧执行不覆盖结果）。用量照样计入（请求确实
+     * 发出且已付费，与失败尝试一样不以认领为条件），走同一个原子 consumeTokens。
+     */
+    private writeStepResultIn;
+    /** 写一条整轮回放的结果（仍由本次执行认领时）；false = 已失去认领。 */
+    private writeRunResultIn;
     /**
      * 该用例剩下的结果标 skipped。整轮回放留下的评测对话里若还有没结束的运行（上一次执行崩溃在回放中途），删掉它，
      * 不留永远「执行中」的评测对话；已正常收尾（`run_aborted` / `run_finished`）的保留，轨迹里能看到停在哪里。
@@ -93,9 +111,8 @@ export declare class EvalWorker {
 }
 /**
  * A plugin template's child case: the template's current version, which a new dispatch would use
- * (a template removed since gets version 0, so it always differs). Undefined for other cases.
+ * (a template removed since gets version 0, so it always differs), with the platform template
+ * whose content that is (absent: the org's own; a link or pin change under a copy's id changes
+ * it). Undefined for other cases.
  */
-export declare function currentTemplate(db: KernelDb, s: CaseSnapshot): Promise<{
-    id: string;
-    version: number;
-} | undefined>;
+export declare function currentTemplate(db: KernelSchemaDb, s: CaseSnapshot): Promise<SubagentTemplateRef | undefined>;