@capybara-hq/kernel 0.1.0-rc.1 → 0.1.0-rc.3

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.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 {

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。 */
@@ -25,7 +40,7 @@ const notFound = () => new ApiError("trajectory.not_found", "Trajectory not foun
  * 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）
@@ -180,7 +195,7 @@ 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 +237,7 @@ 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,
@@ -320,4 +335,12 @@ 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 +1,20 @@
-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 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 +33,34 @@ 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 +92,13 @@ 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。
@@ -95,7 +112,7 @@ 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.
  */
-export declare function currentTemplate(db: KernelDb, s: CaseSnapshot): Promise<{
+export declare function currentTemplate(db: KernelSchemaDb, s: CaseSnapshot): Promise<{
     id: string;
     version: number;
 } | undefined>;