feat(studio): connect Studio to Intake based eval schema, update modal

web/packages/studio/src/components/evaluation/experimentEvalConfig.ts

@@ -0,0 +1,68 @@
+// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+import { createEvaluation, filesListFilesetFiles } from '@nemo/sdk/generated/platform/api';
+import type { ExperimentResponse } from '@nemo/sdk/generated/platform/schema';
+import { buildEvalJobName } from '@studio/components/evaluation/submitEvaluationJob';
+
+/** Experiment metadata key holding the name of the fileset that stores its eval config.
+ *  Metadata values are plain strings, which is all a fileset name needs to be. */
+export const EVAL_CONFIG_FILESET_KEY = 'eval_config_fileset';
+
+/** Flat filename the reusable config is stored as inside its fileset. Every Experiment's
+ *  fileset must carry one at the root — there is no per-run file picker. */
+export const EVAL_CONFIG_FILENAME = 'eval-config.json';
+
+/** The fileset an Experiment stores its eval config in, or null when it names none. */
+export const experimentFilesetName = (experiment: ExperimentResponse): string | null =>
+  experiment.metadata?.[EVAL_CONFIG_FILESET_KEY] ?? null;
+
+/** Why an Experiment cannot be run against, or null when it can. Two ways to be invalid:
+ *  it names no fileset, or the fileset it names has no eval-config.json at the root. */
+export const experimentConfigError = async (
+  workspace: string,
+  experiment: ExperimentResponse,
+  signal?: AbortSignal
+): Promise<string | null> => {
+  const filesetName = experimentFilesetName(experiment);
+  if (!filesetName) {
+    return `Experiment "${experiment.name}" has no eval config fileset. Pick another experiment, or create one from a template.`;
+  }
+  const files = await filesListFilesetFiles(workspace, filesetName, undefined, signal).catch(
+    () => null
+  );
+  if (!files) {
+    return `Could not read fileset "${filesetName}" for experiment "${experiment.name}". Pick another experiment.`;
+  }
+  const hasConfig = (files.data ?? []).some((file) => file.path === EVAL_CONFIG_FILENAME);
+  return hasConfig
+    ? null
+    : `Fileset "${filesetName}" has no ${EVAL_CONFIG_FILENAME} at its root, so experiment "${experiment.name}" cannot be run. Pick another experiment.`;
+};
+
+/** Create the Intake Evaluation this run publishes under, returning its **name** —
+ *  which is what ``publication.intake.evaluation_id`` takes (the entity id is not it).
+ *  ``experimentId`` conversely is the Experiment's **id**, not its name. */
+export const createRunEvaluation = async (
+  workspace: string,
+  {
+    experimentId,
+    experimentName,
+    filesetName,
+    signal,
+  }: {
+    experimentId: string;
+    experimentName: string;
+    filesetName: string;
+    signal?: AbortSignal;
+  }
+): Promise<string> => {
+  // Same normalisation as the job name, so the pair reads as one run.
+  const name = buildEvalJobName(experimentName);
+  await createEvaluation(
+    workspace,
+    { name, experiment_ids: [experimentId], dataset_name: filesetName },
+    signal
+  );
+  return name;
+};

web/packages/studio/src/components/evaluation/submitEvaluationJob.test.ts

@@ -6,6 +6,8 @@ import {
   buildAgentEvalRequestBody,
   buildAgentTarget,
   buildDatasetAgentTarget,
+  buildDatasetEvalRequestBody,
+  type DatasetEvalSpec,
   buildEvalJobName,
   buildPersistedSpec,
   injectJudgeModel,
@@ -143,6 +145,63 @@ describe('buildAgentEvalRequestBody', () => {
     expect(body.spec.labels).toBeUndefined();
     expect(body.name).toBeUndefined();
   });
+
+  it('names the job after the experiment, falling back to the fileset', () => {
+    const withExperiment = buildAgentEvalRequestBody(persisted(), {
+      workspace: 'ws-a',
+      agent: 'a',
+      filesetName: 'wise-blue-data',
+      experimentName: 'wise-blue',
+    });
+    expect(withExperiment.name).toMatch(/^wise-blue-[a-z0-9]{8}$/);
+
+    const filesetOnly = buildAgentEvalRequestBody(persisted(), {
+      workspace: 'ws-a',
+      agent: 'a',
+      filesetName: 'wise-blue-data',
+    });
+    expect(filesetOnly.name).toMatch(/^wise-blue-data-[a-z0-9]{8}$/);
+  });
+
+  it('publishes to the named evaluation when one is selected', () => {
+    const body = buildAgentEvalRequestBody(persisted(), {
+      workspace: 'ws-a',
+      agent: 'a',
+      evaluationId: 'nightly-eval-a3f2',
+    });
+    // agent_name is omitted deliberately: the backend derives it from the agent target.
+    expect(body.spec.publication).toEqual({ intake: { evaluation_id: 'nightly-eval-a3f2' } });
+  });
+
+  it('omits publication entirely when no evaluation is selected', () => {
+    const body = buildAgentEvalRequestBody(persisted(), { workspace: 'ws-a', agent: 'a' });
+    expect(body.spec.publication).toBeUndefined();
+    expect('publication' in body.spec).toBe(false);
+  });
+});
+
+describe('buildDatasetEvalRequestBody', () => {
+  const datasetConfig: DatasetEvalSpec = {
+    dataset: [{ prompt: '2+2?' }],
+    metrics: [metric],
+    prompt_template: '{{item.prompt}}',
+  };
+
+  it('carries publication through the dataset path too, and omits it otherwise', () => {
+    const withPublication = buildDatasetEvalRequestBody(
+      datasetConfig,
+      { workspace: 'ws-a', agent: 'a', evaluationId: 'eval-1' },
+      null
+    );
+    expect(withPublication.spec.publication).toEqual({ intake: { evaluation_id: 'eval-1' } });
+
+    const without = buildDatasetEvalRequestBody(
+      datasetConfig,
+      { workspace: 'ws-a', agent: 'a' },
+      null
+    );
+    expect(without.spec.publication).toBeUndefined();
+  });
 });
 
 describe('buildEvalJobName', () => {

web/packages/studio/src/components/evaluation/submitEvaluationJob.ts

@@ -9,11 +9,16 @@ import { PLATFORM_BASE_URL } from '@studio/constants/environment';
 export const CREATE_NEW = '__create_new__';
 
 export const MODE_DEFAULT = 'default';
-export const MODE_FILESET = 'fileset';
+/** Re-run an existing Experiment, which owns the fileset holding its eval config. */
+export const MODE_EXPERIMENT = 'experiment';
 
-/** Suggested name for a new eval-config fileset (e.g. "wise-blue"). */
+/** Suggested name for a new experiment (e.g. "wise-blue"). */
 export const generateEvalConfigName = (): string => generateDefaultName({ length: 2 });
 
+/** The fileset that stores an experiment's eval config and data artifacts. */
+export const filesetNameForExperiment = (experimentName: string): string =>
+  `${experimentName}-data`;
+
 /** Default parallelism for a submitted eval (Studio default; the config value is a hint). */
 export const DEFAULT_MAX_CONCURRENT_TASKS = 1;
 
@@ -100,8 +105,25 @@ export interface SubmitSelections {
   agent: string;
   /** Eval-config fileset name, stored under spec.labels.eval_config_fileset for display. */
   filesetName?: string;
+  /** Experiment this run belongs to; names the job so it reads as one of that experiment's runs. */
+  experimentName?: string;
+  /** Name of an existing Intake Evaluation to publish results under. The job fails if it
+   *  names nothing — the worker never creates it. Omitted means the run publishes nowhere. */
+  evaluationId?: string;
 }
 
+/** ``spec.publication`` for a run that asked to publish, or nothing at all. ``agent_name`` is
+ *  left off deliberately: the backend derives it from the agent target. */
+const publicationSpec = (evaluationId: string | undefined) =>
+  evaluationId ? { publication: { intake: { evaluation_id: evaluationId } } } : {};
+
+/** ``{ name }`` for the job, stemmed from the experiment it belongs to and falling back to the
+ *  fileset for a submit that names no experiment. Absent when neither is known. */
+const jobName = (selections: SubmitSelections) => {
+  const stem = selections.experimentName ?? selections.filesetName;
+  return stem ? { name: buildEvalJobName(stem) } : {};
+};
+
 /** Strip an optional ``workspace/`` prefix, returning the bare model/agent name. */
 export const bareName = (value: string): string =>
   value.includes('/') ? (value.split('/').pop() ?? value) : value;
@@ -171,12 +193,13 @@ export const buildAgentEvalRequestBody = (
   spec: PersistedEvalSpec,
   selections: SubmitSelections
 ) => ({
-  ...(selections.filesetName ? { name: buildEvalJobName(selections.filesetName) } : {}),
+  ...jobName(selections),
   spec: {
     tasks: spec.tasks,
     target: buildAgentTarget(selections.workspace, selections.agent),
     max_concurrent_tasks: spec.max_concurrent_tasks ?? DEFAULT_MAX_CONCURRENT_TASKS,
     ...(selections.filesetName ? { labels: { eval_config_fileset: selections.filesetName } } : {}),
+    ...publicationSpec(selections.evaluationId),
   },
 });
 
@@ -195,7 +218,7 @@ export const buildDatasetEvalRequestBody = (
   selections: SubmitSelections,
   judgeModel: string | null
 ) => ({
-  ...(selections.filesetName ? { name: buildEvalJobName(selections.filesetName) } : {}),
+  ...jobName(selections),
   spec: {
     dataset: spec.dataset,
     metrics: judgeModel
@@ -205,6 +228,7 @@ export const buildDatasetEvalRequestBody = (
     prompt_template: spec.prompt_template,
     ...(spec.field_mapping ? { field_mapping: spec.field_mapping } : {}),
     params: AGENT_RUN_PARAMS,
+    ...publicationSpec(selections.evaluationId),
   },
 });
 

web/packages/studio/src/mocks/handlers.ts

@@ -24,6 +24,7 @@ import {
   mockEvaluationSessionsPage,
   mockEvaluationsPage,
   mockExperiment,
+  mockExperimentsPage,
 } from '@studio/mocks/intake/experiments';
 import {
   createMockAnnotation,
@@ -483,6 +484,9 @@ export const handlers = [
     const session = mockSessionById(String(params['sessionId']));
     return session ? HttpResponse.json(session) : new HttpResponse(null, { status: 404 });
   }),
+  http.get('*/apis/intake/v2/workspaces/:workspace/experiments', () =>
+    HttpResponse.json(mockExperimentsPage())
+  ),
   http.get('*/apis/intake/v2/workspaces/:workspace/experiments/:name', ({ params }) =>
     HttpResponse.json(mockExperiment(String(params['name'])))
   ),

web/packages/studio/src/mocks/intake/experiments.ts

@@ -7,6 +7,7 @@ import type {
   EvaluationSessionResponse,
   EvaluationSessionResponsesPage,
   ExperimentResponse,
+  ExperimentResponsesPage,
 } from '@nemo/sdk/generated/platform/schema';
 
 const WORKSPACE = 'default';
@@ -37,6 +38,11 @@ export const mockEvaluationsPage = (): EvaluationResponsesPage => ({
   data: MOCK_EVALUATION_NAMES.map(mockEvaluation),
 });
 
+/** The group the mock evaluations belong to, so a caller resolving experiment_ids finds a name. */
+export const mockExperimentsPage = (): ExperimentResponsesPage => ({
+  data: [{ ...mockExperiment('my-group'), id: 'grp_my-group' }],
+});
+
 const mockRun = (
   evaluationName: string,
   sessionId: string,

web/packages/studio/src/routes/agents/AgentDetailRoute/EvaluationsTab.tsx

@@ -1,93 +1,49 @@
 // SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
-import { RelativeTime } from '@nemo/common/src/components/RelativeTime';
-import { StatusBadge } from '@nemo/common/src/components/StatusBadge';
-import { Button, Flex, Stack, Text } from '@nvidia/foundations-react-core';
-import {
-  EVAL_JOB_KIND_LABEL,
-  evalJobDetailRoute,
-  type EvalJobRow,
-  hasMixedEvalKinds,
-} from '@studio/api/evaluation/utils';
+import { SegmentedControl, Stack } from '@nvidia/foundations-react-core';
+import { EvaluationsTable } from '@studio/routes/agents/AgentDetailRoute/evaluations/EvaluationsTable';
+import { ExperimentsTable } from '@studio/routes/agents/AgentDetailRoute/evaluations/ExperimentsTable';
+import { groupByExperiment } from '@studio/routes/agents/AgentDetailRoute/evaluations/groupByExperiment';
 import { DetailPanel } from '@studio/routes/agents/AgentDetailRoute/overview/DetailPanel';
-import { getAgentEvaluationsListRoute } from '@studio/routes/utils';
-import type { FC } from 'react';
-import { Link } from 'react-router';
+import type { AgentEvaluationRow } from '@studio/routes/agents/AgentDetailRoute/useAgentDetails';
+import { type FC, useMemo, useState } from 'react';
+
+const VIEW_EVALUATIONS = 'evaluations';
+const VIEW_EXPERIMENTS = 'experiments';
+
+const VIEW_ITEMS = [
+  { value: VIEW_EVALUATIONS, children: 'Evaluations' },
+  { value: VIEW_EXPERIMENTS, children: 'Experiments' },
+];
 
 interface EvaluationsTabProps {
   workspace: string;
-  evals: EvalJobRow[];
-  onRunEvaluation: () => void;
+  evals: AgentEvaluationRow[];
 }
 
-/** Recent evaluation jobs for the agent, linking through to each run. */
-export const EvaluationsTab: FC<EvaluationsTabProps> = ({ workspace, evals, onRunEvaluation }) => {
-  const showKind = hasMixedEvalKinds(evals);
+/** The agent's published evaluations, either flat or rolled up by experiment. Both views read the
+ *  same evaluations; only the grouping differs, which is why this is a toggle and not a tab. */
+export const EvaluationsTab: FC<EvaluationsTabProps> = ({ workspace, evals }) => {
+  const [view, setView] = useState<string>(VIEW_EVALUATIONS);
+  const experiments = useMemo(() => groupByExperiment(evals), [evals]);
 
   return (
-    <DetailPanel
-      title="Recent evaluations"
-      flush
-      slotAction={
-        <Button kind="secondary" size="small" onClick={onRunEvaluation}>
-          Run evaluation
-        </Button>
-      }
-    >
-      {evals.length === 0 ? (
-        <Stack gap="2" className="p-4">
-          <Text color="secondary">No evaluation jobs found for this agent.</Text>
-          <Link to={getAgentEvaluationsListRoute(workspace)} className="text-xs">
-            View all evaluations →
-          </Link>
-        </Stack>
-      ) : (
-        <Stack gap="0">
-          {evals.map((job, index) => (
-            <Link
-              key={job.id}
-              to={evalJobDetailRoute(workspace, job)}
-              className="text-inherit no-underline"
-            >
-              <Flex
-                align="center"
-                gap="2"
-                className={`px-4 py-4 hover:bg-surface-hover ${index > 0 ? 'border-t border-base' : ''}`}
-              >
-                <Stack gap="1" className="min-w-0 flex-1">
-                  <Flex align="baseline" gap="2" className="min-w-0">
-                    <Text kind="body/semibold/md" className="truncate">
-                      {job.name}
-                    </Text>
-                    {showKind && (
-                      <Text kind="body/regular/sm" color="secondary" className="shrink-0">
-                        ({EVAL_JOB_KIND_LABEL[job.kind]})
-                      </Text>
-                    )}
-                  </Flex>
-                  {job.configLabel && (
-                    <Text kind="body/regular/sm" color="secondary" className="truncate">
-                      Eval Config: {job.configLabel}
-                    </Text>
-                  )}
-                </Stack>
-                <Stack gap="1" align="end" className="shrink-0">
-                  <StatusBadge status={job.status} />
-                  <Text kind="body/regular/sm" color="secondary">
-                    {job.created_at ? <RelativeTime datetime={job.created_at} /> : '—'}
-                  </Text>
-                </Stack>
-              </Flex>
-            </Link>
-          ))}
-          <div className="border-t border-base px-4 py-3">
-            <Link to={getAgentEvaluationsListRoute(workspace)} className="text-xs">
-              View all evaluations →
-            </Link>
-          </div>
-        </Stack>
-      )}
+    <DetailPanel title="Evaluations" flush>
+      <Stack gap="density-lg" className="p-4">
+        <SegmentedControl
+          className="w-fit"
+          aria-label="Group evaluations"
+          value={view}
+          onValueChange={setView}
+          items={VIEW_ITEMS}
+        />
+        {view === VIEW_EXPERIMENTS ? (
+          <ExperimentsTable workspace={workspace} experiments={experiments} />
+        ) : (
+          <EvaluationsTable workspace={workspace} evaluations={evals} />
+        )}
+      </Stack>
     </DetailPanel>
   );
 };