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

package/dist/errors.js

@@ -8,12 +8,15 @@ export const ERROR_CODES = {
     validation_failed: 400,
     forbidden: 403,
     conflict: 409,
+    "lock.acquire_failed": 409,
     payload_too_large: 413,
     unsupported_media_type: 415,
     rate_limited: 429,
     "account.login_taken": 409,
     "account.not_found": 404,
-    "account.last_enterprise_admin": 409,
+    "account.last_super_admin": 409,
+    "account.initial_password_unavailable": 404,
+    "account.create_needs_approval": 409,
     "auth.account_disabled": 403,
     "auth.required": 401,
     "auth.token_expired": 401,
@@ -24,6 +27,7 @@ export const ERROR_CODES = {
     "auth.csrf": 403,
     "auth.rate_limited": 429,
     "auth.session_required": 403,
+    "auth.password_change_required": 403,
     "token.not_found": 404,
     "org.not_found": 404,
     "org.creation_disabled": 403,
@@ -38,10 +42,15 @@ export const ERROR_CODES = {
     "invitation.expired": 410,
     "invitation.used": 409,
     "extension.not_found": 404,
+    /** The extension registers no tool of that name (org tool switches, layering spec A §7). */
+    "extension.tool_not_found": 404,
     "extension.unavailable": 503,
     "extension.no_org_config": 400,
     "extension.config_invalid": 400,
     "extension.config_readonly": 409,
+    "extension.field_not_overridable": 400,
+    "extension.no_org_override": 404,
+    "extension.needs_config": 409,
     "extension.dependency_missing": 409,
     "extension.has_dependents": 409,
     "extension.not_enabled": 409,
@@ -59,6 +68,16 @@ export const ERROR_CODES = {
     "extension.name_mismatch": 409,
     "extension.downgrade_unsupported": 409,
     "extension.not_disabled": 409,
+    /**
+     * Another enabled extension provides a single-provider extension point this one provides
+     * (details `[{ extensionPoint, enabled: { name, displayName } }]`): disable it first.
+     */
+    "extension.single_provider_enabled": 409,
+    /**
+     * A required bootstrap plugin provides a single-provider extension point this one provides
+     * (details as above, `required: true`): it cannot be disabled, so it cannot be replaced.
+     */
+    "extension.single_provider_required": 409,
     "packages.npm_failed": 502,
     "bootstrap.integrity_mismatch": 500,
     "bootstrap.download_failed": 502,
@@ -71,15 +90,50 @@ export const ERROR_CODES = {
     "catalog.dependency_cycle": 409,
     "catalog.version_exists": 409,
     "catalog.item_not_found": 404,
+    /** An org-owned item its owner has not made shareable cannot be shown or associated (P3-16). */
+    "catalog.not_shareable": 409,
+    /** The owner org cannot be removed from its own item's list (P3-17). */
+    "catalog.owner_required": 409,
+    /**
+     * Only the org's own rows can be made shareable: not an agent's skill, not a copy linked to a
+     * platform item another owner has (P3-6).
+     */
+    "catalog.not_owned": 409,
+    /** The item (or the version asked for) waits for the super admin's confirmation (P3-9). */
+    "catalog.item_pending": 409,
+    /** Deleting a platform plugin whose extension is still enabled on the deployment. */
+    "catalog.extension_enabled": 409,
+    /** A super admin confirmation raced the item's removal (P3-21). */
+    "catalog.item_removed": 409,
+    /** What deleting an org's own plugin does changed since its preview (P3-19). */
+    "catalog.preview_stale": 409,
     "catalog.name_conflict": 409,
     "directory.not_configured": 409,
     "catalog.item_disabled": 403,
+    "plugin.disabled": 409,
+    "plugin.scope_locked": 409,
+    "plugin.switch_stale": 409,
+    "plugin.builtin": 409,
+    "plugin.scope_stale": 409,
+    "component.not_found": 404,
+    "component.version_stale": 409,
+    /** A platform component offered to the org but off there (one more connector from a template). */
+    "component.off": 409,
     "catalog.approval_required": 403,
     "catalog.request_not_found": 404,
+    /** No open org notice with that id for an admin of the org (inbox). */
+    "inbox.notice_not_found": 404,
     "catalog.request_decided": 409,
+    /** A pending request of the same kind for the same item or entry exists (spec §5.1 rule 5). */
+    "catalog.request_pending": 409,
+    /** The item is not in the state the request kind asks to change (already offered, on, ...). */
+    "catalog.request_invalid": 409,
+    /** The deployment setting `orgMarketplaceBrowse` is off: no marketplace requests. */
+    "catalog.marketplace_browse_disabled": 409,
     "catalog.custom_disabled": 403,
     "catalog.not_deployed": 409,
     "catalog.icon_not_found": 404,
+    "catalog.guide_not_found": 404,
     "marketplace.not_found": 404,
     "marketplace.entry_not_found": 404,
     "marketplace.fetch_failed": 502,
@@ -88,6 +142,7 @@ export const ERROR_CODES = {
     "marketplace.private_address": 422,
     "marketplace.redirect": 422,
     "marketplace.incompatible": 409,
+    "marketplace.entry_deprecated": 409,
     "marketplace.package_mismatch": 409,
     "marketplace.builtin": 409,
     "marketplace.git_failed": 502,
@@ -96,6 +151,10 @@ export const ERROR_CODES = {
     "marketplace.host_not_allowed": 403,
     "marketplace.path_unsupported": 422,
     "plugin.review_stale": 409,
+    "plugin.name_taken": 409,
+    "plugin.name_reserved": 409,
+    "plugin.dependency_unavailable": 409,
+    "plugin.extension_upload_pending": 501,
     "plugin.staging_not_found": 404,
     "plugin.install_failed": 502,
     "plugin.install_not_found": 404,
@@ -109,7 +168,11 @@ export const ERROR_CODES = {
     "setup.token_invalid": 403,
     "agent.not_found": 404,
     "agent.version_conflict": 409,
+    /** The profile was changed since the client read it (`expectedUpdatedAt` is stale). */
+    "agent.profile_conflict": 409,
     "agent.model_unavailable": 400,
+    /** A save gave `instructionFiles` and an `instructions` that is not their join. */
+    "agent.instructions_mismatch": 400,
     /**
      * `skills.pinned` of an agent definition is refused (details: `{ skill, reason, totalBytes? }`;
      * reasons `not_selected`, `not_found`, `duplicate`, `too_many`, `too_large`, `manual_only`).
@@ -117,6 +180,27 @@ export const ERROR_CODES = {
     "agent.pinned_invalid": 400,
     /** `subagents` of an agent definition is refused (details: `{ reason, path, value? }`). */
     "agent.subagents_invalid": 400,
+    /** Creating from a template with a visibility other than `private` without `agents.manage`. */
+    "agent.visibility_forbidden": 403,
+    /** Creating from a template: the template or the org changed since the preview (P4-3). */
+    "agent.template_preview_stale": 409,
+    /** The agent has no pending connector with that key (ruling P4-6). */
+    "agent.pending_connector_not_found": 404,
+    /**
+     * The connector does not fit the pending entry (ruling P4-T6-1; details `[{ reason }]`:
+     * `template` another or no connector template, `protocol` another kind, `tools` none of the
+     * entry's tools, `not_synced` no tool snapshot yet: an OAuth connector to authorize first).
+     */
+    "agent.pending_connector_mismatch": 400,
+    /** Requesting connector setup for an agent with no pending connectors (ruling P4-T6-7). */
+    "agent.no_pending_connectors": 409,
+    /**
+     * An org template's bare name is a type the org can already use (spec §5 rule 10, ruling
+     * P4-9; details `[{ owner, templateId }]`, `owner` as `plugin.name_taken`'s).
+     */
+    "agent_template.name_taken": 409,
+    /** An org template was saved since the editor read it (`expectedVersion` is stale, P4-12). */
+    "agent_template.version_conflict": 409,
     "conversation.not_found": 404,
     "conversation.output_not_found": 404,
     "conversation.artifact_not_found": 404,
@@ -129,14 +213,22 @@ export const ERROR_CODES = {
     /** A sub-agent's conversation takes no messages from people (sub-agents spec §6.1). */
     "conversation.read_only": 409,
     "run.not_active": 409,
+    /** A run that is not one of the conversation's (the live draft route, spec 2026-11-07 §3.2). */
+    "run.not_found": 404,
+    /** A background task that is not in this conversation (or its sub-agents), or does not exist. */
+    "background_task.not_found": 404,
+    /** The task's output cannot be read now (its sandbox is not running and no finish recorded it). */
+    "background_task.output_unavailable": 409,
     "connector.not_found": 404,
     "connector.slug_taken": 409,
+    "connector.kind_unavailable": 409,
     "connector.connect_failed": 502,
     "connector.tool_not_found": 404,
     "connector.tool_not_pending": 409,
     "connector.resources_disabled": 409,
     "approval.not_found": 404,
     "approval.already_decided": 409,
+    "approval.scope_not_allowed": 409,
     "question.not_found": 404,
     "question.already_answered": 409,
     "tool_output.not_found": 404,
@@ -150,6 +242,7 @@ export const ERROR_CODES = {
     "channel.webhook_not_found": 404,
     "channel.webhook_unavailable": 409,
     "channel.webhook_retry": 503,
+    "channel.guide_changed": 409,
     "ui.not_found": 404,
     "ui.render_failed": 502,
     "ui.action_failed": 502,
@@ -163,7 +256,42 @@ export const ERROR_CODES = {
     "attachment.too_many_pending": 429,
     "sandbox.unavailable": 503,
     "sandbox.busy": 409,
-    "sandbox.not_found": 404,
+    /** The environment is missing, deleted, or not the caller's (never 403, so ids cannot be probed). */
+    "environment.not_found": 404,
+    /** The environment's file system is gone at the provider; reset its files to start empty. */
+    "environment.files_lost": 409,
+    /** The caller may see the environment but not use it (project environments, later phases). */
+    "environment.forbidden": 403,
+    /** The environment's provider is not the deployment's current one, or there is none. */
+    "environment.unavailable": 503,
+    /** The person already has a live environment with this name (case-insensitive). */
+    "environment.name_taken": 409,
+    /** The person already has `environments.maxPersonal` live personal environments. */
+    "environment.limit_reached": 409,
+    /**
+     * A connected environment's files are on its machine: they cannot be reset from the platform;
+     * run `capybara-device logout`, then `capybara-device login` into the environment, on the machine
+     * (ruling R24).
+     */
+    "environment.files_on_device": 409,
+    /** The environment's connected machine is not connected now (no replica holds its lease). */
+    "environment.offline": 503,
+    /** The connected machine runs a device program too old for this kernel; update the program. */
+    "environment.program_outdated": 426,
+    /** Device sign-in (RFC 8628 style): the person has not approved the code yet; keep polling. */
+    "device.authorization_pending": 400,
+    /** Device sign-in: the program polls faster than the interval; add 5 seconds and keep polling. */
+    "device.slow_down": 400,
+    /** Device sign-in: the code expired before it was approved; start over. */
+    "device.expired": 400,
+    /** Device sign-in: the person denied the code. */
+    "device.denied": 400,
+    /** Device sign-in: no pending code matches the one given. */
+    "device.code_invalid": 404,
+    /** Device sign-in: the confirmation could not hand the credential to the program; start over. */
+    "device.unavailable": 503,
+    /** The machine credential is unknown or revoked. */
+    "device.credential_invalid": 401,
     "skill.not_found": 404,
     /** Also a skill neither the model nor a person could use (phase D: both invocation flags off). */
     "skill.invalid_package": 400,
@@ -172,6 +300,8 @@ export const ERROR_CODES = {
      * (`details: [{ reason: "not_found" }]`) or is hidden from the menu (`not_user_invocable`).
      */
     "skill.not_invocable": 400,
+    /** Creating a skill from text (P3-24, F18): an org-level skill with this name already exists in the org. */
+    "skill.name_exists": 409,
     "skill_change.not_found": 404,
     "skill_change.not_owner": 403,
     "skill_change.already_decided": 409,
@@ -191,6 +321,9 @@ export const ERROR_CODES = {
     "eval.candidate_invalid": 400,
     "model.unconfigured": 409,
     "model.purpose_unknown": 404,
+    "model.provider_not_found": 404,
+    /** An extension's `ctx.models.embed` call was denied by a `model.request` policy hook. */
+    "model.hook_denied": 403,
     "eval.result_not_found": 404,
 };
 /** 业务代码抛出的 API 错误；由 HTTP 层的错误过滤器转成错误信封。 */

package/dist/evals/budget.d.ts

@@ -1,5 +1,7 @@
 import type { ModelUsage } from "@capybara-hq/plugin-sdk";
-import type { KernelDb } from "../db/kernel-schema.js";
+import type { KernelSchemaDb } from "../db/kernel-schema.js";
+import { KernelDb } from "../infra/database/kernel-db.js";
+import { TransactionHost } from "../infra/database/transaction-host.js";
 /**
  * 评测运行的 token 预算（C7）：`eval_runs.tokens_used` 在 PG 里原子累加，达到 `token_cap` 的同一条语句把
  * 状态置为 `capped`；worker 每次请求前查 {@link runActive}，不活跃就不再发新请求。在途请求可以越过上限，
@@ -13,7 +15,7 @@ export declare const usageTokens: (u: ModelUsage) => number;
  * 累加用量：`UPDATE … RETURNING`，累加后 ≥ 上限且运行仍在执行 → 同一语句置 `capped`、记 `finished_at`。
  * 运行已不在（被删）时返回 `{ used: 0, capped: true }`，调用方同样停下。
  */
-export declare function consumeTokens(db: KernelDb, runId: string, tokens: number): Promise<{
+export declare function consumeTokens(db: KernelSchemaDb, runId: string, tokens: number): Promise<{
     used: number;
     capped: boolean;
 }>;
@@ -27,35 +29,54 @@ export declare const TOO_MANY_ERRORS: {
  * 同一语句置 `failed`、记原因与 `finished_at`；成功 → 清零。返回运行是否因此被熔断。
  * 多个任务并发时「连续」按写入顺序计（近似），足以挡住「每个请求都失败」的运行。
  */
-export declare function recordOutcome(db: KernelDb, runId: string, ok: boolean, maxConsecutive: number): Promise<{
+export declare function recordOutcome(db: KernelSchemaDb, runId: string, ok: boolean, maxConsecutive: number): Promise<{
     tripped: boolean;
 }>;
 /**
  * 运行停下（取消 / 越限 / 熔断 / 删除评测集）时：还在等评委的结果（`judge_status = 'pending'`）记 `skipped`，
  * 否则这些用例永远算不上做完。给了 `caseId` 只处理该用例。在途的评委请求做完后写不进来（条件是 pending），用量照样计入。
  */
-export declare function skipPendingJudges(db: KernelDb, runId: string, caseId?: string): Promise<void>;
-/**
- * 重算 `cases_done` = 结果全部结束、且没有等评委的结果的用例数（幂等：任务重做不会重复计数），达到 `cases_total` 且运行仍在执行 →
- * `completed`。先锁运行行再数（READ COMMITTED 下新语句看得到其他任务已提交的结果），并发完成的任务不会互相覆盖。
- * worker 每做完一个用例、取消、入队失败时都调用。在事务里调用时锁随事务释放。
- */
-export declare function recountCases(db: KernelDb, runId: string): Promise<void>;
+export declare function skipPendingJudges(db: KernelSchemaDb, runId: string, caseId?: string): Promise<void>;
 /**
  * 运行仍在执行（queued / running），且所属组织与评测集的 agent 都没有删除：取消、达到上限、已结束、运行行已删除、
  * 组织或 agent 已删除都为 false（删除组织 / agent 时会显式取消运行，这里再兜一次底，挡住删除与请求之间的竞争）。
  */
-export declare function runActive(db: KernelDb, runId: string): Promise<boolean>;
+export declare function runActive(db: KernelSchemaDb, runId: string): Promise<boolean>;
 /**
- * 删除组织或 agent 时（调用方的事务内）取消其在执行的评测运行：运行置 `cancelled`，未开始的结果 `skipped`、
- * 等评委的记 `skipped`，重算进度，每个运行写审计 `eval.run_cancelled`（detail 带原因）。在途请求做完后写不进新的请求
- * （worker 每次请求前查 {@link runActive}）。运行创建者离开组织不取消（Ruling S1-b：管理员的评测运行是组织的工作）。
+ * The eval run bookkeeping that needs a transaction: the progress recount (it locks the run row)
+ * and the cancellation of an org's or agent's runs when it is deleted. One provider, shared by the
+ * evals, the agents and the orgs.
  */
-export declare function cancelEvalRuns(trx: KernelDb, scope: {
-    orgId: string;
-    agentId?: string;
-}, actor: {
-    accountId: string;
-    tokenId?: string;
-}, reason: "org_deleted" | "agent_deleted"): Promise<string[]>;
+export declare class EvalBudget {
+    private readonly db;
+    readonly txHost: TransactionHost;
+    constructor(db: KernelDb, txHost: TransactionHost);
+    /**
+     * 重算 `cases_done` = 结果全部结束、且没有等评委的结果的用例数（幂等：任务重做不会重复计数），达到 `cases_total` 且运行仍在执行 →
+     * `completed`。先锁运行行再数（READ COMMITTED 下新语句看得到其他任务已提交的结果），并发完成的任务不会互相覆盖。
+     * worker 每做完一个用例、取消、入队失败时都调用。Joins the caller's transaction (the lock is released with it),
+     * else runs in its own.
+     */
+    recountCases(runId: string): Promise<void>;
+    /**
+     * 删除组织或 agent 时（调用方的事务内）取消其在执行的评测运行：运行置 `cancelled`，未开始的结果 `skipped`、
+     * 等评委的记 `skipped`，重算进度，每个运行写审计 `eval.run_cancelled`（detail 带原因）。在途请求做完后写不进新的请求
+     * （worker 每次请求前查 {@link runActive}）。运行创建者离开组织不取消（Ruling S1-b：管理员的评测运行是组织的工作）。
+     */
+    cancelRuns(scope: {
+        orgId: string;
+        agentId?: string;
+    }, actor: {
+        accountId: string;
+        tokenId?: string;
+    }, reason: "org_deleted" | "agent_deleted"): Promise<string[]>;
+    /**
+     * A run that stopped (cancelled, its set deleted): its results not started yet are `skipped`,
+     * the pending judges too, and the progress is recounted, in the caller's transaction.
+     */
+    skipRun(runId: string): Promise<void>;
+}
+/** {@link EvalBudget} on its own: the agents and the orgs import it without the evals. */
+export declare class EvalBudgetModule {
+}
 export declare const addUsage: (a: ModelUsage, b: ModelUsage) => ModelUsage;

package/dist/evals/budget.js

@@ -1,5 +1,18 @@
+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, Module } from "@nestjs/common";
 import { sql } from "kysely";
 import { audit } from "../identity/audit.js";
+import { KernelDb } from "../infra/database/kernel-db.js";
+import { TransactionHost } from "../infra/database/transaction-host.js";
+import { Transactional } from "../infra/database/transactional.decorator.js";
 /**
  * 评测运行的 token 预算（C7）：`eval_runs.tokens_used` 在 PG 里原子累加，达到 `token_cap` 的同一条语句把
  * 状态置为 `capped`；worker 每次请求前查 {@link runActive}，不活跃就不再发新请求。在途请求可以越过上限，
@@ -83,13 +96,44 @@ export async function skipPendingJudges(db, runId, caseId) {
     await q.execute();
 }
 /**
- * 重算 `cases_done` = 结果全部结束、且没有等评委的结果的用例数（幂等：任务重做不会重复计数），达到 `cases_total` 且运行仍在执行 →
- * `completed`。先锁运行行再数（READ COMMITTED 下新语句看得到其他任务已提交的结果），并发完成的任务不会互相覆盖。
- * worker 每做完一个用例、取消、入队失败时都调用。在事务里调用时锁随事务释放。
+ * 运行仍在执行（queued / running），且所属组织与评测集的 agent 都没有删除：取消、达到上限、已结束、运行行已删除、
+ * 组织或 agent 已删除都为 false（删除组织 / agent 时会显式取消运行，这里再兜一次底，挡住删除与请求之间的竞争）。
  */
-export async function recountCases(db, runId) {
-    const run = async (trx) => {
-        const row = await trx
+export async function runActive(db, runId) {
+    const r = await db
+        .selectFrom("eval_runs as r")
+        .innerJoin("eval_sets as s", "s.id", "r.set_id")
+        .innerJoin("agents as a", "a.id", "s.agent_id")
+        .innerJoin("orgs as o", "o.id", "r.org_id")
+        .select("r.id")
+        .where("r.id", "=", runId)
+        .where("r.status", "in", [...ACTIVE_EVAL_STATUSES])
+        .where("a.deleted_at", "is", null)
+        .where("o.deleted_at", "is", null)
+        .executeTakeFirst();
+    return r !== undefined;
+}
+/**
+ * The eval run bookkeeping that needs a transaction: the progress recount (it locks the run row)
+ * and the cancellation of an org's or agent's runs when it is deleted. One provider, shared by the
+ * evals, the agents and the orgs.
+ */
+let EvalBudget = class EvalBudget {
+    db;
+    txHost;
+    constructor(db, txHost) {
+        this.db = db;
+        this.txHost = txHost;
+    }
+    /**
+     * 重算 `cases_done` = 结果全部结束、且没有等评委的结果的用例数（幂等：任务重做不会重复计数），达到 `cases_total` 且运行仍在执行 →
+     * `completed`。先锁运行行再数（READ COMMITTED 下新语句看得到其他任务已提交的结果），并发完成的任务不会互相覆盖。
+     * worker 每做完一个用例、取消、入队失败时都调用。Joins the caller's transaction (the lock is released with it),
+     * else runs in its own.
+     */
+    async recountCases(runId) {
+        const db = this.db.db;
+        const row = await db
             .selectFrom("eval_runs")
             .select(["cases_total", "status"])
             .where("id", "=", runId)
@@ -102,10 +146,10 @@ export async function recountCases(db, runId) {
         SELECT case_id FROM capybara.eval_results WHERE run_id = ${runId}
         GROUP BY case_id
         HAVING bool_and(status IN ('done', 'error', 'skipped') AND judge_status <> 'pending')
-      ) finished`.execute(trx);
+      ) finished`.execute(db);
         const done = Number(r.rows[0]?.n ?? 0);
         const complete = done >= row.cases_total && ACTIVE_EVAL_STATUSES.includes(row.status);
-        await trx
+        await db
             .updateTable("eval_runs")
             .set({
             cases_done: done,
@@ -113,66 +157,74 @@ export async function recountCases(db, runId) {
         })
             .where("id", "=", runId)
             .execute();
-    };
-    if (db.isTransaction)
-        await run(db);
-    else
-        await db.transaction().execute((trx) => run(trx));
-}
-/**
- * 运行仍在执行（queued / running），且所属组织与评测集的 agent 都没有删除：取消、达到上限、已结束、运行行已删除、
- * 组织或 agent 已删除都为 false（删除组织 / agent 时会显式取消运行，这里再兜一次底，挡住删除与请求之间的竞争）。
- */
-export async function runActive(db, runId) {
-    const r = await db
-        .selectFrom("eval_runs as r")
-        .innerJoin("eval_sets as s", "s.id", "r.set_id")
-        .innerJoin("agents as a", "a.id", "s.agent_id")
-        .innerJoin("orgs as o", "o.id", "r.org_id")
-        .select("r.id")
-        .where("r.id", "=", runId)
-        .where("r.status", "in", [...ACTIVE_EVAL_STATUSES])
-        .where("a.deleted_at", "is", null)
-        .where("o.deleted_at", "is", null)
-        .executeTakeFirst();
-    return r !== undefined;
-}
-/**
- * 删除组织或 agent 时（调用方的事务内）取消其在执行的评测运行：运行置 `cancelled`，未开始的结果 `skipped`、
- * 等评委的记 `skipped`，重算进度，每个运行写审计 `eval.run_cancelled`（detail 带原因）。在途请求做完后写不进新的请求
- * （worker 每次请求前查 {@link runActive}）。运行创建者离开组织不取消（Ruling S1-b：管理员的评测运行是组织的工作）。
- */
-export async function cancelEvalRuns(trx, scope, actor, reason) {
-    let q = trx
-        .updateTable("eval_runs")
-        .set({ status: "cancelled", finished_at: sql `now()` })
-        .where("org_id", "=", scope.orgId)
-        .where("status", "in", [...ACTIVE_EVAL_STATUSES]);
-    if (scope.agentId !== undefined) {
-        const agentId = scope.agentId;
-        q = q.where("set_id", "in", (eb) => eb.selectFrom("eval_sets").select("id").where("agent_id", "=", agentId));
     }
-    const runs = await q.returning("id").execute();
-    for (const r of runs) {
-        await trx
+    /**
+     * 删除组织或 agent 时（调用方的事务内）取消其在执行的评测运行：运行置 `cancelled`，未开始的结果 `skipped`、
+     * 等评委的记 `skipped`，重算进度，每个运行写审计 `eval.run_cancelled`（detail 带原因）。在途请求做完后写不进新的请求
+     * （worker 每次请求前查 {@link runActive}）。运行创建者离开组织不取消（Ruling S1-b：管理员的评测运行是组织的工作）。
+     */
+    async cancelRuns(scope, actor, reason) {
+        const db = this.db.db;
+        let q = db
+            .updateTable("eval_runs")
+            .set({ status: "cancelled", finished_at: sql `now()` })
+            .where("org_id", "=", scope.orgId)
+            .where("status", "in", [...ACTIVE_EVAL_STATUSES]);
+        if (scope.agentId !== undefined) {
+            const agentId = scope.agentId;
+            q = q.where("set_id", "in", (eb) => eb.selectFrom("eval_sets").select("id").where("agent_id", "=", agentId));
+        }
+        const runs = await q.returning("id").execute();
+        for (const r of runs) {
+            await this.skipRun(r.id);
+            await audit(db, {
+                actorAccountId: actor.accountId,
+                ...(actor.tokenId !== undefined && { actorTokenId: actor.tokenId }),
+                orgId: scope.orgId,
+                action: "eval.run_cancelled",
+                target: r.id,
+                detail: { reason },
+            });
+        }
+        return runs.map((r) => r.id);
+    }
+    /**
+     * A run that stopped (cancelled, its set deleted): its results not started yet are `skipped`,
+     * the pending judges too, and the progress is recounted, in the caller's transaction.
+     */
+    async skipRun(runId) {
+        await this.db
             .updateTable("eval_results")
             .set({ status: "skipped", updated_at: sql `now()` })
-            .where("run_id", "=", r.id)
+            .where("run_id", "=", runId)
             .where("status", "=", "pending")
             .execute();
-        await skipPendingJudges(trx, r.id);
-        await recountCases(trx, r.id);
-        await audit(trx, {
-            actorAccountId: actor.accountId,
-            ...(actor.tokenId !== undefined && { actorTokenId: actor.tokenId }),
-            orgId: scope.orgId,
-            action: "eval.run_cancelled",
-            target: r.id,
-            detail: { reason },
-        });
+        await skipPendingJudges(this.db.db, runId);
+        await this.recountCases(runId);
     }
-    return runs.map((r) => r.id);
-}
+};
+__decorate([
+    Transactional()
+], EvalBudget.prototype, "recountCases", null);
+__decorate([
+    Transactional({ propagation: "mandatory" })
+], EvalBudget.prototype, "cancelRuns", null);
+__decorate([
+    Transactional({ propagation: "mandatory" })
+], EvalBudget.prototype, "skipRun", null);
+EvalBudget = __decorate([
+    Injectable(),
+    __param(0, Inject(KernelDb)),
+    __param(1, Inject(TransactionHost))
+], EvalBudget);
+export { EvalBudget };
+/** {@link EvalBudget} on its own: the agents and the orgs import it without the evals. */
+let EvalBudgetModule = class EvalBudgetModule {
+};
+EvalBudgetModule = __decorate([
+    Module({ providers: [EvalBudget], exports: [EvalBudget] })
+], EvalBudgetModule);
+export { EvalBudgetModule };
 export const addUsage = (a, b) => ({
     inputTokens: a.inputTokens + b.inputTokens,
     outputTokens: a.outputTokens + b.outputTokens,

package/dist/evals/evals.module.d.ts

@@ -0,0 +1,31 @@
+import { WorkerHost } from "@nestjs/bullmq";
+import { type DynamicModule } from "@nestjs/common";
+import type { Job, Worker } from "bullmq";
+import type { KernelConfig } from "../config/schema.js";
+import { type EvalJobData } from "./jobs.js";
+import { EvalWorker } from "./worker.js";
+/**
+ * Evals: the trajectory viewer, eval sets and cases, eval runs (the producer of
+ * `cap-kernel-evals`) and the run bookkeeping ({@link EvalBudgetModule}, re-exported).
+ */
+export declare class EvalsModule {
+}
+/** The consumer of `cap-kernel-evals` (one case per job): {@link EvalWorker}. */
+export declare class EvalsProcessor extends WorkerHost<Worker<EvalJobData>> {
+    private readonly evals;
+    constructor(evals: EvalWorker);
+    process(job: Job<EvalJobData>, token?: string): Promise<void>;
+    failed(job: Job<EvalJobData> | undefined, err: Error): void;
+    error(err: Error): void;
+}
+/**
+ * The eval worker (step and full-run replays, the judge) over the runtime's loop and model
+ * registry, and the consumer of `cap-kernel-evals`, its options from the startup config
+ * (concurrency `evals.workerConcurrency`, the run queues' job lock and stalled detection). Imported
+ * only by the worker role (`WorkerModule`), so other roles get no processor and no BullMQ worker.
+ * The worker never starts by itself: the worker role (`WorkerRole`) starts it after the extensions
+ * loaded, and drains it on stop.
+ */
+export declare class EvalWorkersModule {
+    static forRoot(config: KernelConfig): DynamicModule;
+}