@outputai/cli 0.1.12 → 0.1.13-dev.2f0a972.0

package/dist/api/generated/api.js

@@ -1,5 +1,5 @@
 /**
- * Generated by orval v8.6.2 🍺
+ * Generated by orval v8.8.0 🍺
  * Do not edit manually.
  * Output.ai API
  * API for managing and executing Output.ai workflows
@@ -24,11 +24,7 @@ export const WorkflowRunInfoStatus = {
     timed_out: 'timed_out',
     continued: 'continued',
 };
-export const PostWorkflowRun200Status = {
-    completed: 'completed',
-    failed: 'failed',
-};
-export const GetWorkflowIdStatus200Status = {
+export const WorkflowStatusResponseStatus = {
     canceled: 'canceled',
     completed: 'completed',
     continued_as_new: 'continued_as_new',
@@ -38,7 +34,7 @@ export const GetWorkflowIdStatus200Status = {
     timed_out: 'timed_out',
     unspecified: 'unspecified',
 };
-export const GetWorkflowIdResult200Status = {
+export const WorkflowResultResponseStatus = {
     completed: 'completed',
     failed: 'failed',
     canceled: 'canceled',
@@ -46,6 +42,10 @@ export const GetWorkflowIdResult200Status = {
     timed_out: 'timed_out',
     continued: 'continued',
 };
+export const PostWorkflowRun200Status = {
+    completed: 'completed',
+    failed: 'failed',
+};
 ;
 export const getGetHealthUrl = () => {
     return `/health`;
@@ -87,6 +87,24 @@ export const getWorkflowIdStatus = async (id, options) => {
         method: 'GET'
     });
 };
+export const getGetWorkflowIdRunsRidStatusUrl = (id, rid) => {
+    return `/workflow/${id}/runs/${rid}/status`;
+};
+export const getWorkflowIdRunsRidStatus = async (id, rid, options) => {
+    return customFetchInstance(getGetWorkflowIdRunsRidStatusUrl(id, rid), {
+        ...options,
+        method: 'GET'
+    });
+};
+export const getPatchWorkflowIdRunsRidStopUrl = (id, rid) => {
+    return `/workflow/${id}/runs/${rid}/stop`;
+};
+export const patchWorkflowIdRunsRidStop = async (id, rid, options) => {
+    return customFetchInstance(getPatchWorkflowIdRunsRidStopUrl(id, rid), {
+        ...options,
+        method: 'PATCH'
+    });
+};
 export const getPatchWorkflowIdStopUrl = (id) => {
     return `/workflow/${id}/stop`;
 };
@@ -96,6 +114,17 @@ export const patchWorkflowIdStop = async (id, options) => {
         method: 'PATCH'
     });
 };
+export const getPostWorkflowIdRunsRidTerminateUrl = (id, rid) => {
+    return `/workflow/${id}/runs/${rid}/terminate`;
+};
+export const postWorkflowIdRunsRidTerminate = async (id, rid, postWorkflowIdRunsRidTerminateBody, options) => {
+    return customFetchInstance(getPostWorkflowIdRunsRidTerminateUrl(id, rid), {
+        ...options,
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json', ...options?.headers },
+        body: JSON.stringify(postWorkflowIdRunsRidTerminateBody)
+    });
+};
 export const getPostWorkflowIdTerminateUrl = (id) => {
     return `/workflow/${id}/terminate`;
 };
@@ -107,15 +136,26 @@ export const postWorkflowIdTerminate = async (id, postWorkflowIdTerminateBody, o
         body: JSON.stringify(postWorkflowIdTerminateBody)
     });
 };
+export const getPostWorkflowIdRunsRidResetUrl = (id, rid) => {
+    return `/workflow/${id}/runs/${rid}/reset`;
+};
+export const postWorkflowIdRunsRidReset = async (id, rid, resetWorkflowRequest, options) => {
+    return customFetchInstance(getPostWorkflowIdRunsRidResetUrl(id, rid), {
+        ...options,
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json', ...options?.headers },
+        body: JSON.stringify(resetWorkflowRequest)
+    });
+};
 export const getPostWorkflowIdResetUrl = (id) => {
     return `/workflow/${id}/reset`;
 };
-export const postWorkflowIdReset = async (id, postWorkflowIdResetBody, options) => {
+export const postWorkflowIdReset = async (id, resetWorkflowRequest, options) => {
     return customFetchInstance(getPostWorkflowIdResetUrl(id), {
         ...options,
         method: 'POST',
         headers: { 'Content-Type': 'application/json', ...options?.headers },
-        body: JSON.stringify(postWorkflowIdResetBody)
+        body: JSON.stringify(resetWorkflowRequest)
     });
 };
 export const getGetWorkflowIdResultUrl = (id) => {
@@ -127,6 +167,15 @@ export const getWorkflowIdResult = async (id, options) => {
         method: 'GET'
     });
 };
+export const getGetWorkflowIdRunsRidResultUrl = (id, rid) => {
+    return `/workflow/${id}/runs/${rid}/result`;
+};
+export const getWorkflowIdRunsRidResult = async (id, rid, options) => {
+    return customFetchInstance(getGetWorkflowIdRunsRidResultUrl(id, rid), {
+        ...options,
+        method: 'GET'
+    });
+};
 export const getGetWorkflowIdTraceLogUrl = (id) => {
     return `/workflow/${id}/trace-log`;
 };
@@ -136,6 +185,15 @@ export const getWorkflowIdTraceLog = async (id, options) => {
         method: 'GET'
     });
 };
+export const getGetWorkflowIdRunsRidTraceLogUrl = (id, rid) => {
+    return `/workflow/${id}/runs/${rid}/trace-log`;
+};
+export const getWorkflowIdRunsRidTraceLog = async (id, rid, options) => {
+    return customFetchInstance(getGetWorkflowIdRunsRidTraceLogUrl(id, rid), {
+        ...options,
+        method: 'GET'
+    });
+};
 export const getGetWorkflowCatalogIdUrl = (id) => {
     return `/workflow/catalog/${id}`;
 };

package/dist/assets/docker/docker-compose-dev.yml

@@ -81,7 +81,7 @@ services:
         condition: service_healthy
       worker:
         condition: service_healthy
-    image: outputai/api:${OUTPUT_API_VERSION:-0.1.12}
+    image: outputai/api:${OUTPUT_API_VERSION:-0.1.13-dev.2f0a972.0}
     init: true
     networks:
       - main

package/dist/commands/workflow/generate.js

@@ -4,6 +4,7 @@ import { buildWorkflow, buildWorkflowInteractiveLoop } from '#services/workflow_
 import { ensureOutputAISystem } from '#services/coding_agents.js';
 import { getWorkflowGenerateSuccessMessage } from '#services/messages.js';
 import { DEFAULT_OUTPUT_DIRS } from '#utils/paths.js';
+import { parseWorkflowDir } from '#utils/workflow_dir_parser.js';
 import path from 'node:path';
 import * as fsSync from 'node:fs';
 import { getErrorMessage } from '#utils/error_utils.js';
@@ -83,7 +84,8 @@ export default class Generate extends Command {
         this.displaySuccess(result);
     }
     displaySuccess(result) {
-        const message = getWorkflowGenerateSuccessMessage(result.workflowName, result.targetDir, result.filesCreated);
+        const dirInfo = parseWorkflowDir(result.targetDir);
+        const message = getWorkflowGenerateSuccessMessage(result.workflowName, dirInfo.workflowId ?? result.workflowName, dirInfo.scenarioNames[0], result.targetDir, result.filesCreated);
         this.log(message);
     }
 }

package/dist/commands/workflow/generate.spec.js

@@ -2,10 +2,13 @@
 import { describe, it, expect, beforeEach, vi } from 'vitest';
 import Generate from './generate.js';
 import { generateWorkflow } from '#services/workflow_generator.js';
+import { parseWorkflowDir } from '#utils/workflow_dir_parser.js';
 import { InvalidNameError, WorkflowExistsError } from '#types/errors.js';
 vi.mock('../../services/workflow_generator.js');
+vi.mock('../../utils/workflow_dir_parser.js');
 describe('Generate Command', () => {
     let mockGenerateWorkflow;
+    let mockParseWorkflowDir;
     let logSpy;
     const createCommand = () => {
         const cmd = new Generate([], {});
@@ -20,6 +23,7 @@ describe('Generate Command', () => {
     beforeEach(() => {
         vi.clearAllMocks();
         mockGenerateWorkflow = vi.mocked(generateWorkflow);
+        mockParseWorkflowDir = vi.mocked(parseWorkflowDir);
     });
     describe('successful workflow generation', () => {
         it('should generate workflow with skeleton flag', async () => {
@@ -38,6 +42,10 @@ describe('Generate Command', () => {
                 targetDir: '/tmp/test-workflow',
                 filesCreated: ['index.ts', 'steps.ts', 'types.ts']
             });
+            mockParseWorkflowDir.mockReturnValue({
+                workflowId: 'testWorkflow',
+                scenarioNames: ['test_input']
+            });
             await cmd.run();
             expect(mockGenerateWorkflow).toHaveBeenCalledWith({
                 name: 'test-workflow',
@@ -105,6 +113,10 @@ describe('Generate Command', () => {
                 targetDir: '/custom/path/my-workflow',
                 filesCreated: ['index.ts', 'steps.ts', 'types.ts']
             });
+            mockParseWorkflowDir.mockReturnValue({
+                workflowId: 'myWorkflow',
+                scenarioNames: ['test_input']
+            });
             await cmd.run();
             expect(logSpy).toHaveBeenCalledWith(expect.stringContaining('SUCCESS!'));
             expect(logSpy).toHaveBeenCalledWith(expect.stringContaining('my-workflow'));

package/dist/commands/workflow/list.d.ts

@@ -6,6 +6,7 @@ interface WorkflowDisplay {
     inputs: string;
     outputs: string;
     scenarios: string;
+    aliases: string;
 }
 export declare function parseWorkflowForDisplay(workflow: Workflow): WorkflowDisplay;
 export default class WorkflowList extends Command {

package/dist/commands/workflow/list.js

@@ -17,7 +17,8 @@ export function parseWorkflowForDisplay(workflow) {
         description: parsed.description || 'No description',
         inputs: formatParameters(parsed.inputs),
         outputs: formatParameters(parsed.outputs),
-        scenarios: scenarioNames.length > 0 ? scenarioNames.join(', ') : 'none'
+        scenarios: scenarioNames.length > 0 ? scenarioNames.join(', ') : 'none',
+        aliases: workflow.aliases?.length ? workflow.aliases.join(', ') : 'none'
     };
 }
 function caseInsensitiveIncludes(str, filter) {
@@ -26,7 +27,10 @@ function caseInsensitiveIncludes(str, filter) {
 function matchName(filterString) {
     return workflow => {
         const name = workflow.name || '';
-        return caseInsensitiveIncludes(name, filterString);
+        if (caseInsensitiveIncludes(name, filterString)) {
+            return true;
+        }
+        return (workflow.aliases ?? []).some(alias => caseInsensitiveIncludes(alias, filterString));
     };
 }
 function sortWorkflowsByName(workflows) {
@@ -38,8 +42,8 @@ function sortWorkflowsByName(workflows) {
 }
 function createWorkflowTable(workflows, detailed) {
     const table = new Table({
-        head: ['Name', 'Description', 'Inputs', 'Outputs', 'Scenarios'],
-        colWidths: detailed ? [30, 42, 42, 42, 60] : [24, 30, 24, 24, 48],
+        head: ['Name', 'Description', 'Aliases', 'Inputs', 'Outputs', 'Scenarios'],
+        colWidths: detailed ? [28, 36, 36, 36, 36, 48] : [22, 26, 26, 22, 22, 36],
         wordWrap: true,
         style: {
             head: ['cyan']
@@ -49,13 +53,14 @@ function createWorkflowTable(workflows, detailed) {
     sortedWorkflows.forEach(workflow => {
         const display = parseWorkflowForDisplay(workflow);
         if (detailed) {
+            const aliases = workflow.aliases?.length ? workflow.aliases.join('\n') : 'none';
             const inputs = display.inputs.split(', ').join('\n');
             const outputs = display.outputs.split(', ').join('\n');
             const scenarios = display.scenarios.split(', ').join('\n');
-            table.push([display.name, display.description, inputs, outputs, scenarios]);
+            table.push([display.name, display.description, aliases, inputs, outputs, scenarios]);
         }
         else {
-            table.push([display.name, display.description, display.inputs, display.outputs, display.scenarios]);
+            table.push([display.name, display.description, display.aliases, display.inputs, display.outputs, display.scenarios]);
         }
     });
     return table.toString();
@@ -72,6 +77,7 @@ function formatWorkflowsAsJson(workflows) {
             return {
                 name: display.name,
                 description: display.description,
+                aliases: w.aliases ?? [],
                 inputs: display.inputs.split(', '),
                 outputs: display.outputs.split(', '),
                 scenarios: display.scenarios === 'none' ? [] : display.scenarios.split(', '),

package/dist/commands/workflow/list.spec.js

@@ -64,6 +64,27 @@ describe('workflow list parsing', () => {
         expect(parsed.inputs).toBe('none');
         expect(parsed.outputs).toBe('none');
         expect(parsed.scenarios).toBe('none');
+        expect(parsed.aliases).toBe('none');
+    });
+    it('should include aliases when present', async () => {
+        const { parseWorkflowForDisplay } = await import('./list.js');
+        const mockWorkflow = {
+            name: 'aliased-workflow',
+            description: 'Has aliases',
+            aliases: ['old_name', 'legacy_name']
+        };
+        const parsed = parseWorkflowForDisplay(mockWorkflow);
+        expect(parsed.aliases).toBe('old_name, legacy_name');
+    });
+    it('should show none when aliases array is empty', async () => {
+        const { parseWorkflowForDisplay } = await import('./list.js');
+        const mockWorkflow = {
+            name: 'no-aliases',
+            description: 'Empty aliases',
+            aliases: []
+        };
+        const parsed = parseWorkflowForDisplay(mockWorkflow);
+        expect(parsed.aliases).toBe('none');
     });
     it('should include scenario names when scenarios exist', async () => {
         mockListScenarios.mockReturnValueOnce(['basic', 'advanced', 'stress_test']);

package/dist/commands/workflow/run.js

@@ -55,7 +55,8 @@ export default class WorkflowRun extends Command {
         }),
         'task-queue': Flags.string({
             char: 'q',
-            description: 'Task queue name for workflow execution'
+            description: 'Task queue name for workflow execution (defaults to OUTPUT_CATALOG_ID)',
+            env: 'OUTPUT_CATALOG_ID'
         }),
         format: Flags.string({
             char: 'f',

package/dist/commands/workflow/run.spec.js

@@ -13,6 +13,7 @@ vi.mock('#utils/sleep.js', () => ({
 describe('workflow run command', () => {
     beforeEach(async () => {
         vi.clearAllMocks();
+        delete process.env.OUTPUT_CATALOG_ID;
         const { resolveInput } = await import('#utils/resolve_input.js');
         const { sleep } = await import('#utils/sleep.js');
         vi.mocked(resolveInput).mockResolvedValue({});

package/dist/commands/workflow/start.js

@@ -28,7 +28,8 @@ export default class WorkflowStart extends Command {
         }),
         'task-queue': Flags.string({
             char: 'q',
-            description: 'Task queue name for workflow execution'
+            description: 'Task queue name for workflow execution (defaults to OUTPUT_CATALOG_ID)',
+            env: 'OUTPUT_CATALOG_ID'
         })
     };
     async run() {

package/dist/commands/workflow/start.spec.js

@@ -5,6 +5,7 @@ vi.mock('../../api/generated/api.js', () => ({
 describe('workflow start command', () => {
     beforeEach(() => {
         vi.clearAllMocks();
+        delete process.env.OUTPUT_CATALOG_ID;
     });
     describe('command definition', () => {
         it('should export a valid OCLIF command', async () => {

package/dist/commands/workflow/test_eval.js

@@ -114,7 +114,7 @@ export default class WorkflowTest extends Command {
                 }
             };
             if (save) {
-                const filePath = join(dir, `${dataset.name}.yml`);
+                const filePath = dataset._source ?? join(dir, `${dataset.name}.yml`);
                 await writeDataset(updated, filePath);
                 this.log(`    Saved output to ${filePath}`);
             }
@@ -137,7 +137,7 @@ export default class WorkflowTest extends Command {
                     date: now
                 }
             };
-            const filePath = join(dir, `${dataset.name}.yml`);
+            const filePath = dataset._source ?? join(dir, `${dataset.name}.yml`);
             await writeDataset(updated, filePath);
             this.log(`  Saved eval result to ${filePath}`);
         }

package/dist/components/command_footer.d.ts

@@ -0,0 +1,8 @@
+import React from 'react';
+export interface CommandHint {
+    key: string;
+    label: string;
+}
+export declare const CommandFooter: React.FC<{
+    hints: CommandHint[];
+}>;

package/dist/components/command_footer.js

@@ -0,0 +1,4 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+import React from 'react';
+import { Box, Text } from 'ink';
+export const CommandFooter = ({ hints }) => (_jsx(Box, { marginTop: 1, children: hints.map((hint, i) => (_jsxs(React.Fragment, { children: [i > 0 && _jsx(Text, { dimColor: true, children: ' | ' }), _jsx(Text, { dimColor: true, children: '(' }), _jsx(Text, { dimColor: true, bold: true, children: hint.key }), _jsx(Text, { dimColor: true, children: ')' }), _jsx(Text, { dimColor: true, children: ` ${hint.label}` })] }, hint.key))) }));

package/dist/components/status_icon.d.ts

@@ -0,0 +1,11 @@
+import React from 'react';
+interface StatusDisplay {
+    icon: string;
+    color: string;
+}
+export declare const resolveStatus: (status: string) => StatusDisplay;
+export declare const statusColor: (status: string) => string;
+export declare const StatusIcon: React.FC<{
+    status: string;
+}>;
+export {};

package/dist/components/status_icon.js

@@ -0,0 +1,25 @@
+import { jsx as _jsx } from "react/jsx-runtime";
+import { Text } from 'ink';
+const STATUS_MAP = {
+    // Docker service health
+    healthy: { icon: '●', color: 'green' },
+    unhealthy: { icon: '○', color: 'red' },
+    starting: { icon: '◐', color: 'yellow' },
+    none: { icon: '●', color: 'blue' },
+    exited: { icon: '✗', color: 'red' },
+    // Workflow run status
+    running: { icon: '●', color: 'blue' },
+    completed: { icon: '●', color: 'green' },
+    failed: { icon: '✗', color: 'red' },
+    canceled: { icon: '○', color: 'gray' },
+    terminated: { icon: '✗', color: 'red' },
+    timed_out: { icon: '✗', color: 'red' },
+    continued: { icon: '↻', color: 'blue' }
+};
+const DEFAULT_DISPLAY = { icon: '?', color: 'white' };
+export const resolveStatus = (status) => STATUS_MAP[status] ?? DEFAULT_DISPLAY;
+export const statusColor = (status) => resolveStatus(status).color;
+export const StatusIcon = ({ status }) => {
+    const { icon, color } = resolveStatus(status);
+    return _jsx(Text, { color: color, children: icon });
+};

package/dist/components/workflow_summary.d.ts

@@ -0,0 +1,10 @@
+import React from 'react';
+export interface WorkflowSummary {
+    running: number;
+    completed: number;
+    failed: number;
+    total: number;
+}
+export declare const WorkflowSummarySection: React.FC<{
+    summary: WorkflowSummary;
+}>;

package/dist/components/workflow_summary.js

@@ -0,0 +1,4 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+import { Box, Text } from 'ink';
+import { statusColor } from '#components/status_icon.js';
+export const WorkflowSummarySection = ({ summary }) => (_jsxs(Box, { flexDirection: "column", marginTop: 1, children: [_jsx(Text, { bold: true, children: "\uD83D\uDCCB Workflows" }), _jsxs(Box, { marginTop: 1, children: [_jsxs(Text, { color: statusColor('running'), children: [summary.running, " running"] }), _jsx(Text, { children: ", " }), _jsxs(Text, { color: statusColor('failed'), children: [summary.failed, " failed"] }), _jsx(Text, { children: ", " }), _jsxs(Text, { color: statusColor('completed'), children: [summary.completed, " complete"] })] })] }));

package/dist/generated/framework_version.json

@@ -1,3 +1,3 @@
 {
-    "framework": "0.1.12"
+    "framework": "0.1.13-dev.2f0a972.0"
 }

package/dist/services/claude_client.js

@@ -105,7 +105,10 @@ function getTodoWriteMessage(message) {
     if (message.type !== 'assistant') {
         return null;
     }
-    const todoWriteMessage = message.message.content.find((c) => c?.type === 'tool_use' && c.name === 'TodoWrite');
+    const todoWriteMessage = message.message.content.find((c) => {
+        const block = c;
+        return block.type === 'tool_use' && block.name === 'TodoWrite';
+    });
     return todoWriteMessage ?? null;
 }
 function applyInstructions(message, instructions) {

package/dist/services/datasets.d.ts

@@ -8,7 +8,7 @@ export interface DatasetInfo {
 }
 export declare function resolveDatasetsDir(workflowName: string, basePath?: string): string | null;
 export declare function resolveDefaultDatasetsDir(workflowName: string, basePath?: string): string;
-export declare function readDataset(filePath: string): Promise<Dataset>;
+export declare function readDatasetFile(filePath: string): Promise<Dataset[]>;
 export declare function readAllDatasets(workflowName: string, filterNames?: string[], basePath?: string): Promise<{
     datasets: Dataset[];
     dir: string;

package/dist/services/datasets.js

@@ -24,10 +24,19 @@ export function resolveDefaultDatasetsDir(workflowName, basePath = process.cwd()
     // Default to first workflows path
     return resolve(basePath, WORKFLOWS_PATHS[0], workflowName, DATASETS_DIR);
 }
-export async function readDataset(filePath) {
-    const content = await readFile(filePath, 'utf-8');
-    const raw = yaml.load(content);
-    return DatasetSchema.parse(raw);
+export async function readDatasetFile(filePath) {
+    const raw = yaml.load(await readFile(filePath, 'utf-8'));
+    if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
+        throw new Error(`Invalid dataset file: ${filePath}`);
+    }
+    return Object.entries(raw).map(([name, body]) => {
+        if (!body || typeof body !== 'object' || !('input' in body)) {
+            throw new Error(`Dataset case "${name}" in ${filePath} is missing required "input" field`);
+        }
+        const dataset = DatasetSchema.parse({ name, ...body });
+        dataset._source = filePath;
+        return dataset;
+    });
 }
 export async function readAllDatasets(workflowName, filterNames, basePath) {
     const dir = resolveDatasetsDir(workflowName, basePath);
@@ -36,42 +45,35 @@ export async function readAllDatasets(workflowName, filterNames, basePath) {
     }
     const files = await readdir(dir);
     const ymlFiles = files.filter(f => f.endsWith('.yml') || f.endsWith('.yaml'));
+    const seen = new Set();
     const datasets = [];
     for (const file of ymlFiles) {
-        const dataset = await readDataset(join(dir, file));
-        if (filterNames && !filterNames.includes(dataset.name)) {
-            continue;
+        const cases = await readDatasetFile(join(dir, file));
+        for (const dataset of cases) {
+            if (seen.has(dataset.name)) {
+                throw new Error(`Duplicate dataset case name "${dataset.name}" found in ${file}`);
+            }
+            seen.add(dataset.name);
+            if (filterNames && !filterNames.includes(dataset.name)) {
+                continue;
+            }
+            datasets.push(dataset);
         }
-        datasets.push(dataset);
     }
     return { datasets, dir };
 }
-async function mergeWithExisting(dataset, filePath) {
-    if (!existsSync(filePath)) {
-        return dataset;
-    }
-    try {
-        const existing = await readDataset(filePath);
-        return {
-            ...existing,
-            ...dataset,
-            ground_truth: dataset.ground_truth ?? existing.ground_truth,
-            last_output: dataset.last_output ?? existing.last_output,
-            last_eval: dataset.last_eval ?? existing.last_eval
-        };
-    }
-    catch {
-        return dataset;
-    }
-}
 export async function writeDataset(dataset, filePath) {
-    const merged = await mergeWithExisting(dataset, filePath);
     const dir = resolve(filePath, '..');
     if (!existsSync(dir)) {
         await mkdir(dir, { recursive: true });
     }
-    const content = yaml.dump(merged, { lineWidth: 120, noRefs: true, sortKeys: false });
-    await writeFile(filePath, content, 'utf-8');
+    const loaded = existsSync(filePath) ? yaml.load(await readFile(filePath, 'utf-8')) : null;
+    const fileObj = (loaded && typeof loaded === 'object' && !Array.isArray(loaded)) ?
+        loaded :
+        {};
+    const { name, _source, ...caseBody } = dataset;
+    fileObj[name] = { ...fileObj[name], ...caseBody };
+    await writeFile(filePath, yaml.dump(fileObj, { lineWidth: 120, noRefs: true, sortKeys: false }), 'utf-8');
 }
 export async function listDatasets(workflowName, basePath) {
     const dir = resolveDatasetsDir(workflowName, basePath);
@@ -84,14 +86,16 @@ export async function listDatasets(workflowName, basePath) {
     for (const file of ymlFiles) {
         const filePath = join(dir, file);
         try {
-            const dataset = await readDataset(filePath);
-            results.push({
-                name: dataset.name,
-                path: filePath,
-                hasLastOutput: dataset.last_output?.output !== undefined,
-                lastOutputDate: dataset.last_output?.date,
-                lastEvalDate: dataset.last_eval?.date
-            });
+            const cases = await readDatasetFile(filePath);
+            for (const dataset of cases) {
+                results.push({
+                    name: dataset.name,
+                    path: filePath,
+                    hasLastOutput: dataset.last_output?.output !== undefined,
+                    lastOutputDate: dataset.last_output?.date,
+                    lastEvalDate: dataset.last_eval?.date
+                });
+            }
         }
         catch {
             results.push({

package/dist/services/datasets.test.d.ts

@@ -0,0 +1 @@
+export {};