@capybara-hq/kernel 0.1.0-rc.2 → 0.1.0-rc.21

package/dist/errors.js

@@ -14,7 +14,9 @@ export const ERROR_CODES = {
     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,
@@ -25,6 +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,
@@ -39,10 +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,
@@ -60,6 +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,
@@ -72,15 +90,106 @@ 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,
+    /** A theme of an imported plugin fails its file or asset checks (themes spec §4, §5). */
+    "theme.invalid": 400,
+    /** A theme of an imported plugin fails the readability checks of its merged schemes (§5). */
+    "theme.unreadable": 400,
+    /**
+     * A settings write names a theme that is not available where it is set (§7, §9): an org default
+     * off in the org, a platform default that is not a platform theme with scope `all`, or a
+     * person's pick on in none of their orgs; details `[{ theme }]`.
+     */
+    "theme.not_available": 400,
+    /**
+     * The change would take the platform default theme (`uiTheme`) away from an org: narrowing,
+     * disabling, removing it or its plugin (§7); details `[{ theme, setting: "uiTheme" }]`.
+     */
+    "theme.platform_default": 409,
+    /**
+     * Another theme with the same id is already on in an org the change would turn this one on in
+     * (§6, rule 10); details `[{ theme, orgId, themes, plugins }]`.
+     */
+    "theme.id_taken": 409,
+    /**
+     * A theme delivery URL names no stored theme version, or a file that version does not name (§8);
+     * a removed theme's versions are still served.
+     */
+    "theme.not_found": 404,
+    /**
+     * An `extension.client` section whose extension manifest lacks `permissions.client: true`
+     * (client extensions spec §5.4: client code is always declared and reviewed).
+     */
+    "client.undeclared": 400,
+    /**
+     * A client file's sha384 is not the integrity its plugin.json records (spec §9.1); details
+     * `[{ file }]` (the manifest key).
+     */
+    "client.integrity_mismatch": 400,
+    /**
+     * The client section or a client file fails the import checks (spec §4.1, §9.1: a listed file
+     * missing, a file not listed, a key that cannot be looked up exactly, a size limit, a stylesheet
+     * not wrapped in `@layer extensions` or with `@import` / an absolute `url()`); details
+     * `[{ file, reason }]`.
+     */
+    "client.invalid": 400,
+    /**
+     * The client section's `apiVersion` is newer than the deployment's `CLIENT_API_VERSION` (spec
+     * §4.2): the web host could not run the bundle; details `[{ apiVersion, supported }]`.
+     */
+    "client.api_unsupported": 400,
+    /**
+     * A client file URL (`/api/v1/client/<plugin>/<build>/<file>`, spec §9.3) the caller may not
+     * fetch: an unknown or old build, a file the build does not list (by its exact key), an extension
+     * not running or not approved, not enabled in any live org the caller belongs to, or stopped by
+     * the kill switch. One code for all, so nothing tells them apart.
+     */
+    "client.not_found": 404,
     "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,
+    /** The check is not in the setup guide rendered for that mode and audience (setup guides v2 §10.3). */
+    "catalog.check_not_found": 404,
     "marketplace.not_found": 404,
     "marketplace.entry_not_found": 404,
     "marketplace.fetch_failed": 502,
@@ -89,6 +198,7 @@ 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,
@@ -97,6 +207,10 @@ 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,
@@ -108,9 +222,15 @@ export const ERROR_CODES = {
     "plugin.upgrade_failed": 502,
     "setup.already_completed": 409,
     "setup.token_invalid": 403,
+    /** The admin overview's seven-day usage aggregate ran past its statement timeout (spec §5.5). */
+    "admin.usage_unavailable": 503,
     "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`).
@@ -118,6 +238,27 @@ 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,
@@ -130,14 +271,26 @@ 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,
+    /** The connector template's setup guide moved on since the person was shown it (guide-ack). */
+    "connector.guide_changed": 409,
+    "connector.kind_unavailable": 409,
     "connector.connect_failed": 502,
     "connector.tool_not_found": 404,
     "connector.tool_not_pending": 409,
     "connector.resources_disabled": 409,
+    /** Per-member tool policies on a connector with one org credential (decision 2026-11-09). */
+    "connector.not_personal": 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,
@@ -151,6 +304,7 @@ 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,
@@ -164,17 +318,72 @@ export const ERROR_CODES = {
     "attachment.too_many_pending": 429,
     "sandbox.unavailable": 503,
     "sandbox.busy": 409,
-    "sandbox.not_found": 404,
+    /** A sandbox file path the kernel reads is not absolute or not under the sandbox roots. */
+    "sandbox.file_invalid_path": 422,
+    /** No file at the sandbox path the kernel reads. */
+    "sandbox.file_not_found": 422,
+    /** The sandbox path the kernel reads is a directory, not one file. */
+    "sandbox.file_is_directory": 422,
+    /** The sandbox file the kernel reads is over the caller's size limit. */
+    "sandbox.file_too_large": 422,
+    /** The device is missing, deleted, or not the caller's (never 403, so ids cannot be probed). */
+    "device.not_found": 404,
+    /** The cloud device's file system is gone at the provider; reset its files to start empty. */
+    "device.files_lost": 409,
+    /** The caller may see the device but not use it (shared devices, later phases). */
+    "device.forbidden": 403,
+    /** The device's provider is not the deployment's current one, or there is none. */
+    "device.unavailable": 503,
+    /** The person already has a live device with this name (case-insensitive). */
+    "device.name_taken": 409,
+    /** The person already has `devices.maxPersonal` live personal devices. */
+    "device.limit_reached": 409,
+    /**
+     * A local device's files are on its machine: they cannot be reset from the platform; run
+     * `capybara-device logout`, then `capybara-device login` into the device, on the machine
+     * (ruling R24).
+     */
+    "device.files_on_device": 409,
+    /** The local device's machine is not connected now (no replica holds its lease). */
+    "device.offline": 503,
+    /** The local device's machine runs a `capybara-device` too old for this kernel; update it. */
+    "device.program_outdated": 426,
+    /** Device sign-in (RFC 8628 style): the person has not approved the code yet; keep polling. */
+    "device_sign_in.authorization_pending": 400,
+    /** Device sign-in: the program polls faster than the interval; add 5 seconds and keep polling. */
+    "device_sign_in.slow_down": 400,
+    /** Device sign-in: the code expired before it was approved; start over. */
+    "device_sign_in.expired": 400,
+    /** Device sign-in: the person denied the code. */
+    "device_sign_in.denied": 400,
+    /** Device sign-in: no pending code matches the one given. */
+    "device_sign_in.code_invalid": 404,
+    /** Device sign-in: the confirmation could not hand the credential to the program; start over. */
+    "device_sign_in.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,
+    /** An agent already has `AGENT_SKILLS_MAX` own skills. */
+    "skill.agent_cap": 409,
+    /** Sharing this agent skill with the org is already waiting for an org admin. */
+    "skill.share_pending": 409,
+    /** Another conversation's proposed version of this shared skill is waiting for an org admin. */
+    "skill.proposal_pending_elsewhere": 409,
+    /** A sub-agent conversation tried to register a skill on its agent (only the root registers). */
+    "skill.subagent_register": 403,
     /**
      * A person invoked a skill (`/name`, phase D) that is not one of the agent's effective skills
      * (`details: [{ reason: "not_found" }]`) or is hidden from the menu (`not_user_invocable`).
      */
     "skill.not_invocable": 400,
+    /**
+     * Uploading or sharing a skill: an org-level skill with this name already exists in the org
+     * (P3-24, F18).
+     */
+    "skill.name_exists": 409,
     "skill_change.not_found": 404,
-    "skill_change.not_owner": 403,
     "skill_change.already_decided": 409,
     "trajectory.not_found": 404,
     "eval.run_not_finished": 409,
@@ -192,6 +401,9 @@ 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/evals.module.js

@@ -27,6 +27,8 @@ import { EvalRunsService } from "./runs.service.js";
 import { EvalSetsService } from "./sets.service.js";
 import { TrajectoryService } from "./trajectory.service.js";
 import { EVAL_MAX_STALLED_COUNT, EvalWorker } from "./worker.js";
+/** The eval queue's registration (re-exported: the admin health reads its counts). */
+const EVAL_QUEUE_MODULE = QueueModule.registerQueue(EVAL_QUEUE_NAME);
 /**
  * Evals: the trajectory viewer, eval sets and cases, eval runs (the producer of
  * `cap-kernel-evals`) and the run bookkeeping ({@link EvalBudgetModule}, re-exported).
@@ -42,10 +44,17 @@ EvalsModule = __decorate([
             OrgLockModule,
             ProvidersModule,
             EvalBudgetModule,
-            QueueModule.registerQueue(EVAL_QUEUE_NAME),
+            EVAL_QUEUE_MODULE,
         ],
         providers: [TrajectoryService, EvalSetsService, EvalRunsService, EvalJobs],
-        exports: [TrajectoryService, EvalSetsService, EvalRunsService, EvalJobs, EvalBudgetModule],
+        exports: [
+            TrajectoryService,
+            EvalSetsService,
+            EvalRunsService,
+            EvalJobs,
+            EvalBudgetModule,
+            EVAL_QUEUE_MODULE,
+        ],
     })
 ], EvalsModule);
 export { EvalsModule };

package/dist/evals/evals.schemas.d.ts

@@ -0,0 +1,110 @@
+import { z } from "zod";
+import type { RouteSchemas } from "../platform/route-schemas.js";
+export declare const EvalSetCreateBody: z.ZodObject<{
+    agentId: z.ZodString;
+    name: z.ZodString;
+    description: z.ZodOptional<z.ZodString>;
+    rubric: z.ZodOptional<z.ZodString>;
+}, z.core.$strict>;
+export declare const EvalSetPatchBody: z.ZodObject<{
+    name: z.ZodOptional<z.ZodString>;
+    description: z.ZodOptional<z.ZodString>;
+    rubric: z.ZodOptional<z.ZodNullable<z.ZodString>>;
+}, z.core.$strict>;
+export declare const EvalAddCaseBody: z.ZodObject<{
+    conversationId: z.ZodString;
+    runId: z.ZodString;
+}, z.core.$strict>;
+export declare const EvalSampleBody: z.ZodObject<{
+    from: z.ZodOptional<z.ZodISODateTime>;
+    to: z.ZodOptional<z.ZodISODateTime>;
+    count: z.ZodNumber;
+    status: z.ZodOptional<z.ZodEnum<{
+        any: "any";
+        completed: "completed";
+        failed: "failed";
+    }>>;
+    minToolCalls: z.ZodOptional<z.ZodNumber>;
+}, z.core.$strict>;
+export declare const EvalCandidateSchema: z.ZodObject<{
+    agentVersion: z.ZodOptional<z.ZodNumber>;
+    model: z.ZodOptional<z.ZodObject<{
+        provider: z.ZodString;
+        model: z.ZodString;
+    }, z.core.$strict>>;
+    modelParams: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
+    instructions: z.ZodOptional<z.ZodString>;
+}, z.core.$strict>;
+/** `tokenCap` is only type-checked here; the service judges the range (`eval.token_cap_invalid`). */
+export declare const EvalRunEstimateBody: z.ZodObject<{
+    mode: z.ZodEnum<{
+        run: "run";
+        step: "step";
+    }>;
+    candidate: z.ZodObject<{
+        agentVersion: z.ZodOptional<z.ZodNumber>;
+        model: z.ZodOptional<z.ZodObject<{
+            provider: z.ZodString;
+            model: z.ZodString;
+        }, z.core.$strict>>;
+        modelParams: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
+        instructions: z.ZodOptional<z.ZodString>;
+    }, z.core.$strict>;
+    judge: z.ZodDefault<z.ZodBoolean>;
+    tokenCap: z.ZodOptional<z.ZodNumber>;
+}, z.core.$strict>;
+export declare const EvalRunStartBody: z.ZodObject<{
+    mode: z.ZodEnum<{
+        run: "run";
+        step: "step";
+    }>;
+    candidate: z.ZodObject<{
+        agentVersion: z.ZodOptional<z.ZodNumber>;
+        model: z.ZodOptional<z.ZodObject<{
+            provider: z.ZodString;
+            model: z.ZodString;
+        }, z.core.$strict>>;
+        modelParams: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
+        instructions: z.ZodOptional<z.ZodString>;
+    }, z.core.$strict>;
+    judge: z.ZodDefault<z.ZodBoolean>;
+    tokenCap: z.ZodOptional<z.ZodNumber>;
+}, z.core.$strict>;
+/**
+ * `view=light`: no content blocks (result tables and polling; the response carries `asOf`);
+ * default `full`. `updatedSince` (ISO time) takes only what changed after it.
+ */
+export declare const EvalRunResultsQuery: z.ZodObject<{
+    after: z.ZodOptional<z.ZodString>;
+    limit: z.ZodDefault<z.ZodCoercedNumber<unknown>>;
+    view: z.ZodDefault<z.ZodEnum<{
+        full: "full";
+        light: "light";
+    }>>;
+    updatedSince: z.ZodOptional<z.ZodString>;
+}, z.core.$strict>;
+export declare const EvalLabelBody: z.ZodObject<{
+    label: z.ZodNullable<z.ZodEnum<{
+        fail: "fail";
+        pass: "pass";
+    }>>;
+    note: z.ZodOptional<z.ZodString>;
+}, z.core.$strict>;
+export declare const EvalSetParams: z.ZodObject<{
+    setId: z.ZodString;
+}, z.core.$strip>;
+export declare const EvalCaseParams: z.ZodObject<{
+    setId: z.ZodString;
+    caseId: z.ZodString;
+}, z.core.$strip>;
+export declare const EvalRunParams: z.ZodObject<{
+    runId: z.ZodString;
+}, z.core.$strip>;
+export declare const EvalResultParams: z.ZodObject<{
+    runId: z.ZodString;
+    resultId: z.ZodString;
+}, z.core.$strip>;
+export declare const EvalLabelParams: z.ZodObject<{
+    resultId: z.ZodString;
+}, z.core.$strip>;
+export declare const EVAL_ROUTES: readonly RouteSchemas[];

package/dist/evals/evals.schemas.js

@@ -0,0 +1,107 @@
+import { z } from "zod";
+import { ModelRef } from "../agents/definition.js";
+const NAME = z.string().trim().min(1).max(200);
+const DESCRIPTION = z.string().max(2000);
+const RUBRIC = z.string().min(1).max(20_000);
+export const EvalSetCreateBody = z
+    .object({
+    agentId: z.string().min(1).max(100),
+    name: NAME,
+    description: DESCRIPTION.optional(),
+    rubric: RUBRIC.optional(),
+})
+    .strict();
+export const EvalSetPatchBody = z
+    .object({
+    name: NAME.optional(),
+    description: DESCRIPTION.optional(),
+    rubric: RUBRIC.nullable().optional(),
+})
+    .strict();
+export const EvalAddCaseBody = z
+    .object({ conversationId: z.string().min(1).max(100), runId: z.string().min(1).max(100) })
+    .strict();
+export const EvalSampleBody = z
+    .object({
+    from: z.iso.datetime({ offset: true }).optional(),
+    to: z.iso.datetime({ offset: true }).optional(),
+    count: z.number().int().min(1).max(50),
+    status: z.enum(["completed", "failed", "any"]).optional(),
+    minToolCalls: z.number().int().min(0).max(1000).optional(),
+})
+    .strict();
+export const EvalCandidateSchema = z
+    .object({
+    agentVersion: z.number().int().min(1).optional(),
+    model: ModelRef.strict().optional(),
+    modelParams: z.record(z.string(), z.union([z.string(), z.number(), z.boolean()])).optional(),
+    instructions: z.string().max(100_000).optional(),
+})
+    .strict();
+/** `tokenCap` is only type-checked here; the service judges the range (`eval.token_cap_invalid`). */
+export const EvalRunEstimateBody = z
+    .object({
+    mode: z.enum(["step", "run"]),
+    candidate: EvalCandidateSchema,
+    judge: z.boolean().default(false),
+    tokenCap: z.number().int().optional(),
+})
+    .strict();
+export const EvalRunStartBody = EvalRunEstimateBody;
+/**
+ * `view=light`: no content blocks (result tables and polling; the response carries `asOf`);
+ * default `full`. `updatedSince` (ISO time) takes only what changed after it.
+ */
+export const EvalRunResultsQuery = z
+    .object({
+    after: z.string().max(100).optional(),
+    limit: z.coerce.number().int().min(1).max(200).default(50),
+    view: z.enum(["full", "light"]).default("full"),
+    updatedSince: z
+        .string()
+        .max(64)
+        .refine((v) => !Number.isNaN(Date.parse(v)), { message: "must be an ISO timestamp" })
+        .optional(),
+})
+    .strict();
+export const EvalLabelBody = z
+    .object({
+    label: z.enum(["pass", "fail"]).nullable(),
+    note: z.string().max(2000).optional(),
+})
+    .strict();
+/** The path params the eval routes read (`@Param`), bounded as the operations take them. */
+const Id = z.string().min(1).max(100);
+export const EvalSetParams = z.object({ setId: Id });
+export const EvalCaseParams = z.object({ setId: Id, caseId: Id });
+export const EvalRunParams = z.object({ runId: Id });
+export const EvalResultParams = z.object({ runId: Id, resultId: Id });
+export const EvalLabelParams = z.object({ resultId: Id });
+const O = "/api/v1/orgs/:orgId";
+export const EVAL_ROUTES = [
+    { route: `POST ${O}/eval-sets`, body: EvalSetCreateBody },
+    { route: `GET ${O}/eval-sets` },
+    { route: `GET ${O}/eval-sets/:setId`, params: EvalSetParams },
+    { route: `PATCH ${O}/eval-sets/:setId`, params: EvalSetParams, body: EvalSetPatchBody },
+    { route: `DELETE ${O}/eval-sets/:setId`, params: EvalSetParams },
+    { route: `POST ${O}/eval-sets/:setId/cases`, params: EvalSetParams, body: EvalAddCaseBody },
+    {
+        route: `POST ${O}/eval-sets/:setId/cases/sample`,
+        params: EvalSetParams,
+        body: EvalSampleBody,
+    },
+    { route: `GET ${O}/eval-sets/:setId/cases/:caseId`, params: EvalCaseParams },
+    { route: `DELETE ${O}/eval-sets/:setId/cases/:caseId`, params: EvalCaseParams },
+    {
+        route: `POST ${O}/eval-sets/:setId/runs/estimate`,
+        params: EvalSetParams,
+        body: EvalRunEstimateBody,
+    },
+    { route: `POST ${O}/eval-sets/:setId/runs`, params: EvalSetParams, body: EvalRunStartBody },
+    { route: `GET ${O}/eval-sets/:setId/runs`, params: EvalSetParams },
+    { route: `GET ${O}/eval-runs/:runId`, params: EvalRunParams },
+    { route: `GET ${O}/eval-runs/:runId/results`, params: EvalRunParams, query: EvalRunResultsQuery },
+    { route: `GET ${O}/eval-runs/:runId/results/:resultId`, params: EvalResultParams },
+    { route: `POST ${O}/eval-runs/:runId/cancel`, params: EvalRunParams },
+    { route: `PUT ${O}/eval-results/:resultId/label`, params: EvalLabelParams, body: EvalLabelBody },
+];

package/dist/evals/reconstruct.d.ts

@@ -63,10 +63,7 @@ export interface ReconstructOptions {
      * eval worker reads the current one). Another version than the logged one is a candidate change
      * (`subagent_changed`).
      */
-    subagentTemplate?: {
-        id: string;
-        version: number;
-    };
+    subagentTemplate?: SubagentTemplateRef;
     /**
      * A child case under another agent version (`agentVersion` differs): that version's own
      * definition, compared with the logged sub-agent for `subagent_changed` (the candidate
@@ -87,6 +84,16 @@ export declare function reconstructRequest(s: CaseSnapshot, modelCallSeq: number
 type SubagentStarted = EventPayload<"subagent_started">;
 /** The `subagent_started` of a sub-agent's (child) log, or undefined for any other conversation. */
 export declare function subagentStartedOf(events: readonly ConversationEvent[]): SubagentStarted | undefined;
+/**
+ * A plugin template child's template as a new dispatch would use it: the selected id, the version
+ * and the platform template whose content it is (absent: the org's own template; plugin scope
+ * phase 2, a linked copy shares its id with both sources, whose version numbers may collide).
+ */
+export interface SubagentTemplateRef {
+    id: string;
+    version: number;
+    platformTemplateId?: string;
+}
 /**
  * Does a candidate change what a sub-agent (child) case would run (`subagent_changed`, spec
  * §15.1)? Pure, from the snapshot alone, comparing the model-visible parts of the effective
@@ -101,7 +108,12 @@ export declare function subagentStartedOf(events: readonly ConversationEvent[]):
  *   parent version (its own version is {@link ReconstructOptions.subagentTemplate}). Template
  *   narrowing that needs the database (inline and plugin tool allowlists, skill names) is not
  *   recomputed: those compare instructions only.
- * Models and params are not part of it: they are candidate axes of their own.
+ * Models and params are not part of it: they are candidate axes of their own. Neither is the
+ * candidate's `subagents` block as the child's own roster (allowlist, templates, limits; only a
+ * template child's entry in its parent's roster counts, `parentSide`) nor whether it offers `Agent`:
+ * whether a child has `Agent` follows its depth (`keepAgent` comes only from the logged child), and
+ * what a nesting child dispatches is recorded as its `Agent` results, which replay reuses (agent
+ * map spec §4.1, §8). A roster-only candidate therefore reports no `subagent_changed`.
  */
 export declare function subagentChanged(started: SubagentStarted, logged: AgentDefinition, candidate: AgentDefinition, sameVersion: boolean, 
 /** The parent's definition at the case's version (the snapshot's), for template children. */
@@ -115,10 +127,7 @@ export declare function subagentCaseChanged(s: CaseSnapshot, c: {
     changed: boolean;
     definition: AgentDefinition;
     agentVersion: number;
-    subagentTemplate?: {
-        id: string;
-        version: number;
-    };
+    subagentTemplate?: SubagentTemplateRef;
     subagentVersion?: AgentDefinition;
 }): boolean;
 /** 还原该请求要查的附件 id（该模型用得到的引用，去重、按出现顺序；见 `mediaRefs`）。 */