@datalayer/core 1.2.27 → 1.2.29

package/lib/api/evals/__tests__/external-definition.golden.json

@@ -0,0 +1,110 @@
+{
+  "external_definition": {
+    "adapter_version": "1",
+    "capabilities": [
+      {
+        "code": "",
+        "component": "tasks.py@arithmetic:generate",
+        "kind": "solver",
+        "reason": "",
+        "status": "framework-run"
+      }
+    ],
+    "configuration": {
+      "aggregation": {},
+      "missing": [],
+      "models": ["bedrock:us.anthropic.claude-haiku-4-5"],
+      "scorers": ["includes"],
+      "task_args": {},
+      "tasks": [
+        {
+          "entry_point": "tasks.py@arithmetic",
+          "scorers": [
+            {
+              "metrics": ["inspect_ai/accuracy"],
+              "name": "includes"
+            }
+          ],
+          "solvers": [
+            {
+              "name": "generate",
+              "params": {}
+            }
+          ]
+        }
+      ],
+      "tools": ["add"]
+    },
+    "content_digest": "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb",
+    "entry_point": "tasks.py@arithmetic",
+    "framework": "inspect-ai",
+    "framework_version": "0.3.286",
+    "source": {
+      "filename": "fixture.tar.gz",
+      "kind": "archive",
+      "reference": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+    }
+  },
+  "external_plan": {
+    "adapter_version": "1",
+    "capabilities": [
+      {
+        "code": "",
+        "component": "tasks.py@arithmetic:generate",
+        "kind": "solver",
+        "reason": "",
+        "status": "framework-run"
+      },
+      {
+        "code": "",
+        "component": "tasks.py@arithmetic:includes",
+        "kind": "scorer",
+        "reason": "",
+        "status": "framework-run"
+      },
+      {
+        "code": "",
+        "component": "tasks.py@arithmetic:includes:inspect_ai/accuracy",
+        "kind": "metric",
+        "reason": "",
+        "status": "framework-run"
+      },
+      {
+        "code": "",
+        "component": "tool:add",
+        "kind": "tool",
+        "reason": "",
+        "status": "framework-run"
+      },
+      {
+        "code": "",
+        "component": "model:bedrock:us.anthropic.claude-haiku-4-5",
+        "kind": "model",
+        "reason": "",
+        "status": "framework-run"
+      }
+    ],
+    "environment": "inspect-ai-env",
+    "framework": "inspect-ai",
+    "framework_version": "0.3.286",
+    "policy": {
+      "cpu_seconds": 1800,
+      "max_bytes": 1073741824,
+      "max_files": 20000,
+      "memory_bytes": 8589934592,
+      "network": "inference",
+      "root_dirs": [],
+      "secrets": [],
+      "wall_seconds": 1800
+    },
+    "reproduction": "pip install inspect-ai==0.3.286\nDATALAYER_BASE_URL=<ai-inference>/api/ai-inference/v1 DATALAYER_API_KEY=<your Datalayer token> inspect eval tasks.py@arithmetic --model openai-api/datalayer/bedrock:us.anthropic.claude-haiku-4-5 --log-format json",
+    "resolved_configuration": {
+      "aggregation": {},
+      "approved_tools": ["add"],
+      "environment_only": false,
+      "max_samples": 1,
+      "task_args": {},
+      "task_environments": {}
+    }
+  }
+}

package/lib/api/evals/client.d.ts

@@ -9,7 +9,7 @@
  * @module api/evals/client
  */
 import { type EvalsClientOptions } from './request';
-import type { CaseListResponse, CreateReportRequest, MoveReportRequest, RegenerateReportResponse, ReportListResponse, ReportResponse, ReportsQuery, DecisionListResponse, DecisionRequest, DecisionResponse, DecisionSubject, SharedWithMeResponse, ContinueReportImportResponse, ReportImportListResponse, ReportImportRequest, ReportImportResponse, EvalsPermissionsResponse, EvalsSharingResponse, EvalsSharingUpdate, SharedEvalsRecord, EvalEvaluatorRef, EvalsetVersionListResponse, EvalsetVersionResponse, InvestigationListResponse, InvestigationResponse, InvestigationScope, InvestigationStatus, LexicalReportResponse, ReportDocumentResponse, ResumeSandboxRequest, ResumeSandboxResponse, TaskInvestigationResponse, UpdateInvestigationRequest, EvalsetCategory, ImportEvalsetRequest, ImportEvalsetResponse, CaseResultListResponse, CaseResultResponse, RunNetworkResponse, RunNetworkMessageResponse, LaunchNetworkResponse, CreateLaunchRequest, LaunchCancelResponse, LaunchListResponse, ClaimTrialResponse, LaunchPlanResponse, LaunchResponse, LiveAlertListResponse, SubjectsResponse, ReviewCaseRequest, RunCancelResponse, CaseRequest, CaseResponse, CreateEvalsetRequest, CreateExperimentRequest, CreateLiveEventRequest, CreateRunRequest, EvalKind, EvalRunEnvironment, EvalsetDeleteResponse, EvalsetListResponse, EvalsetResponse, EvalsetPackagePreviewResponse, EvalsetComparisonResponse, SubjectUsageResponse, SandboxProvenanceResponse, BenchmarkComputeResponse, EvalsSearchResponse, EvalsetLeaderboardResponse, EvalsetPublicationResponse, EvalsetPublicationListResponse, ExperimentDeleteResponse, ExperimentListResponse, ExperimentResponse, LiveEventListResponse, LiveEventResponse, LiveTargetDeleteResponse, LiveTargetListResponse, PublicEvalsetDetailsResponse, RunListResponse, RunResponse, SuccessResponse, UpdateEvalsetRequest, UpdateExperimentRequest } from './types';
+import type { AdapterInspectionResponse, EvalAdapterListResponse, InspectExternalSourceRequest, CaseListResponse, CreateReportRequest, MoveReportRequest, RegenerateReportResponse, ReportListResponse, ReportResponse, ReportsQuery, DecisionListResponse, DecisionRequest, DecisionResponse, DecisionSubject, SharedWithMeResponse, ContinueReportImportResponse, ReportImportListResponse, ReportImportRequest, ReportImportResponse, EvalsPermissionsResponse, EvalsSharingResponse, EvalsSharingUpdate, SharedEvalsRecord, EvalEvaluatorRef, EvalsetVersionListResponse, EvalsetVersionResponse, InvestigationListResponse, InvestigationResponse, InvestigationScope, InvestigationStatus, LexicalReportResponse, ReportDocumentResponse, ResumeSandboxRequest, ResumeSandboxResponse, TaskInvestigationResponse, UpdateInvestigationRequest, EvalsetCategory, ImportEvalsetRequest, ImportEvalsetResponse, CaseResultListResponse, CaseResultResponse, RunNetworkResponse, RunNetworkMessageResponse, LaunchNetworkResponse, CreateLaunchRequest, LaunchCancelResponse, LaunchListResponse, ClaimTrialResponse, LaunchPlanResponse, LaunchResponse, LiveAlertListResponse, SubjectsResponse, ReviewCaseRequest, RunCancelResponse, CaseRequest, CaseResponse, CreateEvalsetRequest, CreateExperimentRequest, CreateLiveEventRequest, CreateRunRequest, EvalKind, EvalRunEnvironment, EvalsetDeleteResponse, EvalsetListResponse, EvalsetResponse, EvalsetPackagePreviewResponse, EvalsetComparisonResponse, SubjectUsageResponse, SandboxProvenanceResponse, BenchmarkComputeResponse, EvalsSearchResponse, EvalsetLeaderboardResponse, EvalsetPublicationResponse, EvalsetPublicationListResponse, ExperimentDeleteResponse, ExperimentListResponse, ExperimentResponse, LiveEventListResponse, LiveEventResponse, LiveTargetDeleteResponse, LiveTargetListResponse, PublicEvalsetDetailsResponse, RunListResponse, RunResponse, SuccessResponse, UpdateEvalsetRequest, UpdateExperimentRequest } from './types';
 export type ListEvalsetsQuery = {
     kind?: EvalKind;
     run_environment?: EvalRunEnvironment;
@@ -19,6 +19,8 @@ export type ListEvalsetsQuery = {
     dataset?: string;
     /** …and pinned to this revision of it: one revision is not evidence about another. */
     revision?: string;
+    /** Only the benchmarks imported from this framework. */
+    framework?: string;
     limit?: number;
     offset?: number;
 };
@@ -299,6 +301,16 @@ export declare const listLaunches: (options: EvalsClientOptions, query?: ListLau
 export declare const getLaunch: (options: EvalsClientOptions, launchId: string) => Promise<LaunchResponse>;
 export declare const cancelLaunch: (options: EvalsClientOptions, launchId: string) => Promise<LaunchCancelResponse>;
 export declare const archiveLaunch: (options: EvalsClientOptions, launchId: string) => Promise<SuccessResponse>;
+/** The external frameworks this deployment offers; none hides them. */
+export declare const listEvalAdapters: (options: EvalsClientOptions) => Promise<EvalAdapterListResponse>;
+/** Keep a framework's archive by its digest and inspect it in a sandbox. */
+export declare const inspectExternalSource: (options: EvalsClientOptions, framework: string, body: InspectExternalSourceRequest) => Promise<AdapterInspectionResponse>;
+/** An inspection under way, or what it found. */
+export declare const getAdapterInspection: (options: EvalsClientOptions, inspectionId: string) => Promise<AdapterInspectionResponse>;
+/** Where an imported benchmark's original source is downloaded from. */
+export declare const externalSourceUrl: (options: EvalsClientOptions, evalsetId: string, version?: number) => string;
+/** Where a task's raw evidence is downloaded from, by the digest it names. */
+export declare const caseArtifactUrl: (options: EvalsClientOptions, runId: string, caseId: string, digest: string) => string;
 /** An evalset from a spec file, the same body the CLI and the action send. */
 export declare const importEvalset: (options: EvalsClientOptions, body: ImportEvalsetRequest) => Promise<ImportEvalsetResponse>;
 export declare const listEvalsetVersions: (options: EvalsClientOptions, evalsetId: string, query?: {

package/lib/api/evals/client.js

@@ -264,6 +264,19 @@ export const getLaunch = (options, launchId) => evalsRequest(options, `/launches
 export const cancelLaunch = (options, launchId) => evalsRequest(options, `/launches/${segment(launchId)}/cancel`, { method: 'POST' });
 export const archiveLaunch = (options, launchId) => evalsRequest(options, `/launches/${segment(launchId)}/archive`, { method: 'POST' });
 // --- Spec import and definition versions (B2-08, B2-09) --------------------
+// --- External frameworks ---------------------------------------------------
+/** The external frameworks this deployment offers; none hides them. */
+export const listEvalAdapters = (options) => evalsRequest(options, '/adapters');
+/** Keep a framework's archive by its digest and inspect it in a sandbox. */
+export const inspectExternalSource = (options, framework, body) => evalsRequest(options, `/adapters/${segment(framework)}/inspect`, { method: 'POST', body });
+/** An inspection under way, or what it found. */
+export const getAdapterInspection = (options, inspectionId) => evalsRequest(options, `/adapters/inspections/${segment(inspectionId)}`);
+/** Where an imported benchmark's original source is downloaded from. */
+export const externalSourceUrl = (options, evalsetId, version) => evalsUrl(options, `/evalsets/${segment(evalsetId)}/external-source`, {
+    version,
+});
+/** Where a task's raw evidence is downloaded from, by the digest it names. */
+export const caseArtifactUrl = (options, runId, caseId, digest) => evalsUrl(options, `/runs/${segment(runId)}/cases/${segment(caseId)}/artifacts/${segment(digest)}`);
 /** An evalset from a spec file, the same body the CLI and the action send. */
 export const importEvalset = (options, body) => evalsRequest(options, '/evalsets/import', {
     method: 'POST',

package/lib/api/evals/types.d.ts

@@ -40,6 +40,152 @@ export interface EvalCase {
     metadata: Record<string, unknown>;
 }
 /** An evalset, the benchmark: versioned cases, evaluators, metadata. */
+/** How an adapter runs a framework's component. */
+export type AdapterCapabilityStatus = 'native' | 'framework-run' | 'unsupported';
+/** One component of an external definition, as its adapter classified it. */
+export interface AdapterCapability {
+    component: string;
+    kind: string;
+    status: AdapterCapabilityStatus;
+    /** Why it cannot run, for an `unsupported` one. */
+    reason: string;
+    /** A stable code to count refusals by. */
+    code: string;
+}
+/** The immutable thing a definition was read from. */
+export interface ExternalSourceRef {
+    kind: 'archive' | 'git' | 'library' | 'registry';
+    /** The sha256 of the archive, a commit, or a pinned package. */
+    reference: string;
+    filename: string;
+}
+/** What a version and a launch keep of an imported definition. */
+export interface ExternalDefinition {
+    framework: string;
+    adapter_version: string;
+    framework_version: string;
+    source: ExternalSourceRef;
+    entry_point: string;
+    configuration: Record<string, unknown>;
+    capabilities: AdapterCapability[];
+    /** A digest covering every file of the source. */
+    content_digest: string;
+}
+/** An external framework this deployment offers. */
+export interface EvalAdapter {
+    framework: string;
+    adapter_version: string;
+    framework_version: string;
+    environment: string;
+}
+export interface EvalAdapterListResponse {
+    success: boolean;
+    adapters: EvalAdapter[];
+}
+export interface InspectedCase {
+    native_id: string;
+    name: string;
+    input: unknown;
+    target: unknown;
+    entry_point: string;
+    metadata: Record<string, unknown>;
+}
+/** What a source defines, found in a sandbox with none of it run. */
+export interface AdapterInspection {
+    framework: string;
+    adapter_version: string;
+    framework_version: string;
+    archive_digest: string;
+    content_digest: string;
+    name: string;
+    description: string;
+    entry_points: string[];
+    cases: InspectedCase[];
+    subjects: Array<Record<string, unknown>>;
+    scorers: Array<Record<string, unknown>>;
+    models: string[];
+    tools: string[];
+    secrets: string[];
+    network: 'none' | 'inference' | 'boundary';
+    environment: Record<string, unknown>;
+    capabilities: AdapterCapability[];
+    configuration: Record<string, unknown>;
+}
+/** An inspection under way, or its answer. */
+export interface AdapterInspectionRecord {
+    id: string;
+    owner_uid: string;
+    framework: string;
+    digest: string;
+    filename: string;
+    status: 'pending' | 'ready' | 'failed';
+    inspection: AdapterInspection | Record<string, never>;
+    error: string;
+    workflow_uid: string;
+    created_at: string | null;
+    updated_at: string | null;
+}
+export interface AdapterInspectionResponse {
+    success: boolean;
+    inspection: AdapterInspectionRecord;
+}
+export interface InspectExternalSourceRequest {
+    filename: string;
+    content_base64: string;
+    options?: Record<string, unknown>;
+}
+/** What was chosen from an inspection's preview. */
+export interface ExternalSelection {
+    entry_points?: string[];
+    limit?: number;
+    arguments?: Record<string, unknown>;
+    /** Catalogue models of AI Inference, one experiment each. */
+    models?: string[];
+    /** The headline score: `{scorer}` or `{kind: 'mean' | 'all'}`, with a `threshold`. */
+    aggregation?: Record<string, unknown>;
+    name?: string;
+}
+/** How a launch of an imported definition runs. */
+export interface ExternalPlan {
+    framework: string;
+    adapter_version: string;
+    framework_version: string;
+    environment: string;
+    policy: Record<string, unknown>;
+    capabilities: AdapterCapability[];
+    resolved_configuration: Record<string, unknown>;
+    /** The pinned framework command, with no secret in it. */
+    reproduction: string;
+}
+/** What a framework said of one task, beside the normalized fields. */
+export interface ExternalCaseResult {
+    framework?: string;
+    adapter_version?: string;
+    framework_version?: string;
+    native_id?: string;
+    attempt?: number;
+    scores?: Array<{
+        name: string;
+        value: unknown;
+        answer: string;
+        explanation: string;
+        metadata: Record<string, unknown>;
+    }>;
+    usage?: Record<string, unknown>;
+    details?: Record<string, unknown>;
+    /** The task's own environment, when it ran in one: what ran, by digest. */
+    environment?: {
+        uid?: string;
+        image_digest?: string;
+        base_image?: string;
+    };
+    artifacts?: Array<{
+        name: string;
+        media_type: string;
+        digest: string;
+        size: number;
+    }>;
+}
 export interface Evalset {
     id: string;
     owner_uid: string;
@@ -49,6 +195,8 @@ export interface Evalset {
     kind: EvalKind;
     category: EvalsetCategory | '';
     dataset_ref: DatasetRef | null;
+    /** Where an imported definition came from; empty for one made here. */
+    external_definition?: ExternalDefinition | Record<string, never>;
     /** Moves on every change to cases, evaluators, schema or category. */
     version: number;
     /** How many launches this benchmark has had; the next is number + 1. */
@@ -187,6 +335,8 @@ export interface EvalRun {
 }
 /** A task result: one case of one run, its own document. */
 export interface EvalTaskResult {
+    /** What a framework said of this task, beside the normalized fields. */
+    external?: ExternalCaseResult;
     id: string;
     run_id: string;
     launch_id: string;
@@ -229,6 +379,9 @@ export interface EvalTaskResult {
 }
 /** A launch: one submission of a benchmark across its experiments. */
 export interface EvalLaunch {
+    /** An imported definition's launch: what it ran from, and how. */
+    external_definition?: ExternalDefinition | Record<string, never>;
+    external_plan?: ExternalPlan | Record<string, never>;
     id: string;
     owner_uid: string;
     account_uid: string;
@@ -274,6 +427,9 @@ export interface LaunchPlanProblem {
     severity: 'error' | 'warning';
     message: string;
     experiment_id?: string;
+    /** The component an imported definition's plan refused. */
+    component?: string;
+    capability_code?: string;
 }
 /** The plan of a launch before it is made (B2-07). */
 export interface LaunchPlanResponse {
@@ -288,6 +444,10 @@ export interface LaunchPlanResponse {
      * submission.
      */
     run_mode_constraint: string;
+    /** How an imported definition would run; null for one made here. */
+    external_plan?: ExternalPlan | null;
+    /** Whom an imported definition's model calls are metered under, in a sentence. */
+    metering?: string;
     problems: LaunchPlanProblem[];
     estimate: {
         cases: number;
@@ -1032,12 +1192,19 @@ export interface CaseResultResponse {
 }
 export interface ImportEvalsetRequest {
     /** An `*.evalset.json` spec, as the CLI and the action read it. */
-    spec: Record<string, unknown>;
+    spec?: Record<string, unknown>;
+    /** Or an inspected external source. One of the two. */
+    external?: {
+        inspection_id: string;
+        selection?: ExternalSelection;
+    };
     run_environment?: EvalRunEnvironment;
 }
 export interface ImportEvalsetResponse extends EvalsetResponse {
     /** Evaluators the platform cannot run, left out of the evalset. */
     unsupported_evaluators: string[];
+    /** An external import's experiments, one per model. */
+    experiments?: EvalExperiment[];
 }
 export interface EvalsetVersionListResponse {
     success: boolean;

package/lib/api/mcp/agentAccess.d.ts

@@ -0,0 +1,10 @@
+import type { McpAgentReach, McpAgentToolsets } from '../../models/McpAgentAccess';
+export type { McpAgentNamed, McpAgentReach, McpAgentTool, McpAgentToolset, McpAgentToolsets, McpReachRow, McpSourceApprovalPolicy, McpToolsetHealth, McpToolsetHealthState, McpToolsetSelected, } from '../../models/McpAgentAccess';
+export interface McpAgentAccessFilters {
+    /** One of the caller's connected agents, by its uid. */
+    grant?: string;
+}
+/** The toolsets an agent may execute: their tools, approval and health. */
+export declare const getAgentToolsets: (token: string, filters?: McpAgentAccessFilters, baseUrl?: string) => Promise<McpAgentToolsets>;
+/** The notebooks and sources an agent can reach, and how far. */
+export declare const getAgentReach: (token: string, filters?: McpAgentAccessFilters, baseUrl?: string) => Promise<McpAgentReach>;

package/lib/api/mcp/agentAccess.js

@@ -0,0 +1,28 @@
+/*
+ * Copyright (c) 2023-2026 Datalayer, Inc.
+ * Distributed under the terms of the Modified BSD License.
+ */
+/**
+ * What one agent may execute and reach.
+ *
+ * Both routes take `grant`, the uid of one of the caller's connected agents;
+ * without it the caller is asked about itself, which is how an agent holds
+ * these answers against its own token.
+ *
+ * @module api/mcp/agentAccess
+ */
+import { DEFAULT_SERVICE_URLS } from '../constants.js';
+import { requestDatalayerAPI } from '../DatalayerApi.js';
+import { fromWire, mcpUrl } from './gateway.js';
+/** The toolsets an agent may execute: their tools, approval and health. */
+export const getAgentToolsets = async (token, filters = {}, baseUrl = DEFAULT_SERVICE_URLS.JUPYTER_MCP_SERVER) => fromWire(await requestDatalayerAPI({
+    url: mcpUrl(baseUrl, '/toolsets', { ...filters }),
+    method: 'GET',
+    token,
+}));
+/** The notebooks and sources an agent can reach, and how far. */
+export const getAgentReach = async (token, filters = {}, baseUrl = DEFAULT_SERVICE_URLS.JUPYTER_MCP_SERVER) => fromWire(await requestDatalayerAPI({
+    url: mcpUrl(baseUrl, '/reach', { ...filters }),
+    method: 'GET',
+    token,
+}));

package/lib/api/mcp/generated.js

@@ -71,6 +71,7 @@ export const MCP_GATEWAY_ROUTES = [
             { name: 'team', in: 'query' },
             { name: 'user', in: 'query' },
             { name: 'client', in: 'query' },
+            { name: 'credential', in: 'query' },
             { name: 'agent', in: 'query' },
             { name: 'tool', in: 'query' },
             { name: 'method', in: 'query' },
@@ -79,6 +80,7 @@ export const MCP_GATEWAY_ROUTES = [
             { name: 'notebook', in: 'query' },
             { name: 'source', in: 'query' },
             { name: 'task', in: 'query' },
+            { name: 'trace_id', in: 'query' },
             { name: 'since', in: 'query' },
             { name: 'until', in: 'query' },
             { name: 'cursor', in: 'query' },
@@ -156,6 +158,12 @@ export const MCP_GATEWAY_ROUTES = [
             { name: 'team', in: 'query' },
         ],
     },
+    {
+        method: 'POST',
+        path: '/api/mcp/v1/pap/checks',
+        operationId: 'pap_signed_checks_endpoint_api_mcp_v1_pap_checks_post',
+        parameters: [],
+    },
     {
         method: 'GET',
         path: '/api/mcp/v1/policy',
@@ -166,6 +174,12 @@ export const MCP_GATEWAY_ROUTES = [
             { name: 'team', in: 'query' },
         ],
     },
+    {
+        method: 'GET',
+        path: '/api/mcp/v1/reach',
+        operationId: 'agent_reach_endpoint_api_mcp_v1_reach_get',
+        parameters: [{ name: 'grant', in: 'query' }],
+    },
     {
         method: 'GET',
         path: '/api/mcp/v1/tasks',
@@ -204,6 +218,12 @@ export const MCP_GATEWAY_ROUTES = [
         operationId: 'answer_task_endpoint_api_mcp_v1_tasks__task_uid__input_post',
         parameters: [{ name: 'task_uid', in: 'path' }],
     },
+    {
+        method: 'GET',
+        path: '/api/mcp/v1/toolsets',
+        operationId: 'agent_toolsets_endpoint_api_mcp_v1_toolsets_get',
+        parameters: [{ name: 'grant', in: 'query' }],
+    },
     {
         method: 'GET',
         path: '/api/mcp/version',

package/lib/api/mcp/index.d.ts

@@ -9,6 +9,7 @@ export * from './generated';
 export * from './gateway';
 export * from './tasks';
 export * from './bindings';
+export * from './agentAccess';
 export * from './audit';
 export * from './activity';
 export * from './organizations';

package/lib/api/mcp/index.js

@@ -17,6 +17,7 @@ export * from './generated.js';
 export * from './gateway.js';
 export * from './tasks.js';
 export * from './bindings.js';
+export * from './agentAccess.js';
 export * from './audit.js';
 export * from './activity.js';
 export * from './organizations.js';

package/lib/api/mcp/observability.d.ts

@@ -39,7 +39,7 @@ export declare const MCP_DURABLE_SERVICE_NAME = "datalayer-durable";
  */
 export declare const serviceNameFor: (metric: McpMetricName) => string;
 /** The metric catalog `telemetry.py` owns; no ad-hoc counters. */
-export declare const MCP_METRIC_CATALOG: readonly ["mcp.calls", "mcp.call.duration", "mcp.refusals", "mcp.limit.unenforced", "mcp.forwards", "mcp.workers", "mcp.worker_start_seconds", "mcp.bindings", "mcp.sandbox.lost", "mcp.tasks", "mcp.task.duration", "durable.step.duration", "durable.queue.wait", "durable.recoveries", "sandbox.launch_seconds", "mcp.audit.writes", "mcp.audit.write_failures", "mcp.dependency.duration", "mcp.dependency.timeouts", "mcp.dependency.ready", "mcp.readiness.failures", "mcp.audit.deleted", "mcp.jobs", "mcp.job.duration"];
+export declare const MCP_METRIC_CATALOG: readonly ["mcp.calls", "mcp.call.duration", "mcp.refusals", "mcp.limit.unenforced", "mcp.forwards", "mcp.workers", "mcp.process.fds", "mcp.worker_start_seconds", "mcp.bindings", "mcp.sandbox.lost", "mcp.tasks", "mcp.task.duration", "durable.step.duration", "durable.queue.wait", "durable.recoveries", "sandbox.launch_seconds", "mcp.audit.writes", "mcp.audit.write_failures", "mcp.dependency.duration", "mcp.dependency.timeouts", "mcp.dependency.ready", "mcp.readiness.failures", "mcp.audit.deleted", "mcp.jobs", "mcp.job.duration"];
 export type McpMetricName = (typeof MCP_METRIC_CATALOG)[number];
 /** The four SLIs, and the catalog metric each is read from. */
 export declare const MCP_SLI_METRICS: {

package/lib/api/mcp/observability.js

@@ -79,6 +79,8 @@ export const MCP_METRIC_CATALOG = [
     'mcp.limit.unenforced',
     'mcp.forwards',
     'mcp.workers',
+    // A gateway replica: file descriptors open, and the limit (6 November 2026).
+    'mcp.process.fds',
     'mcp.worker_start_seconds',
     'mcp.bindings',
     'mcp.sandbox.lost',

package/lib/api/mcp/operations.d.ts

@@ -27,11 +27,17 @@ export interface McpWorkerList {
     /** The replica that answered. */
     replica?: string;
 }
-export interface McpWorkflowQueue {
-    name: string;
-    backlog: number;
-    /** Seconds the oldest waiting workflow has waited. */
-    oldestWaitSeconds?: number | null;
+/** One of the newest runs the durable engine reports. */
+export interface McpWorkflowRun {
+    uid: string;
+    workflow: string;
+    queue: string;
+    /** `running` (DBOS's pending and enqueued alike), `blocked`, `completed`, `failed` or `cancelled`. */
+    status: string;
+    taskUid?: string;
+    /** Unix seconds. */
+    startedAt?: number;
+    updatedAt?: number;
 }
 /** One periodic job on the replica that answered. */
 export interface McpJob {
@@ -57,14 +63,26 @@ export interface McpJobSchedule {
     holder: string;
     jobs: McpJob[];
 }
-/** Engine health, queue backlog and worker versions. */
+/**
+ * What the durable engine says about itself (`/operations/workflows`, the
+ * engine's `operations()` as the gateway forwards it): whether it is taking
+ * work, the queues it polls, the queues it was told to run that nobody serves,
+ * and its newest runs.
+ */
 export interface McpWorkflowsHealth {
     engine: 'temporal' | 'dbos' | string;
-    healthy: boolean;
+    /** Whether it takes work: launched, and its run listing readable. */
+    durable: boolean;
     detail?: string | null;
-    queues: McpWorkflowQueue[];
-    workerVersions: Record<string, number>;
-    checkedAt: string;
+    /** The queues declared on the engine, by name. */
+    queues: string[];
+    /** Queues this deployment was told to run that the catalogue does not have: work sent there is never done. */
+    queuesNobodyServes: string[];
+    workflows: string[];
+    /** Its newest runs (twenty), newest first. */
+    runs: McpWorkflowRun[];
+    databaseConfigured?: boolean;
+    applicationVersion?: string;
 }
 export declare const getGatewayVersion: (baseUrl?: string) => Promise<McpGatewayVersion>;
 export declare const getGatewayHealth: (baseUrl?: string) => Promise<McpGatewayHealth>;

package/lib/api/mcp/tasks.d.ts

@@ -16,11 +16,19 @@ export interface McpTaskListFilters {
 export declare const listTasks: (token: string, filters?: McpTaskListFilters, baseUrl?: string) => Promise<McpTaskList>;
 export declare const getTask: (token: string, taskUid: string, baseUrl?: string) => Promise<McpTask>;
 /** Stop a task that is still going. Idempotent: a finished task is answered as it is. */
-export declare const cancelTask: (token: string, taskUid: string, baseUrl?: string) => Promise<McpTask>;
+export declare const cancelTask: (token: string, taskUid: string, baseUrl?: string) => Promise<McpTaskCancelled>;
 /** What the gateway says of an answer: that it was passed on, not the task. */
 export type McpTaskAnswered = {
     uid: string;
     answered: boolean;
+    /** Set when the answer closed a run its engine no longer carried. */
+    status?: string;
+};
+/** What the gateway says of a cancel: whether it stopped, and the status. */
+export type McpTaskCancelled = {
+    uid: string;
+    cancelled: boolean;
+    status?: string;
 };
 /**
  * Answer a task that is `input_required`.