@outputai/cli 0.1.13-dev.01b8dea.0 → 0.1.13-dev.59a1b6d.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/api/generated/api.d.ts +323 -85
- package/dist/api/generated/api.js +66 -8
- package/dist/assets/docker/docker-compose-dev.yml +1 -1
- package/dist/commands/credentials/set.d.ts +14 -0
- package/dist/commands/credentials/set.js +57 -0
- package/dist/commands/credentials/set.spec.d.ts +1 -0
- package/dist/commands/credentials/set.spec.js +95 -0
- package/dist/commands/workflow/runs/list.d.ts +1 -0
- package/dist/commands/workflow/runs/list.js +6 -0
- package/dist/commands/workflow/test_eval.js +2 -2
- package/dist/generated/framework_version.json +1 -1
- package/dist/services/datasets.d.ts +1 -1
- package/dist/services/datasets.js +41 -37
- package/dist/services/datasets.test.d.ts +1 -0
- package/dist/services/datasets.test.js +202 -0
- package/dist/services/workflow_runs.d.ts +1 -0
- package/dist/services/workflow_runs.js +3 -0
- package/dist/utils/format_workflow_result.d.ts +3 -3
- package/package.json +4 -4
|
@@ -24,11 +24,7 @@ export const WorkflowRunInfoStatus = {
|
|
|
24
24
|
timed_out: 'timed_out',
|
|
25
25
|
continued: 'continued',
|
|
26
26
|
};
|
|
27
|
-
export const
|
|
28
|
-
completed: 'completed',
|
|
29
|
-
failed: 'failed',
|
|
30
|
-
};
|
|
31
|
-
export const GetWorkflowIdStatus200Status = {
|
|
27
|
+
export const WorkflowStatusResponseStatus = {
|
|
32
28
|
canceled: 'canceled',
|
|
33
29
|
completed: 'completed',
|
|
34
30
|
continued_as_new: 'continued_as_new',
|
|
@@ -38,7 +34,7 @@ export const GetWorkflowIdStatus200Status = {
|
|
|
38
34
|
timed_out: 'timed_out',
|
|
39
35
|
unspecified: 'unspecified',
|
|
40
36
|
};
|
|
41
|
-
export const
|
|
37
|
+
export const WorkflowResultResponseStatus = {
|
|
42
38
|
completed: 'completed',
|
|
43
39
|
failed: 'failed',
|
|
44
40
|
canceled: 'canceled',
|
|
@@ -46,6 +42,10 @@ export const GetWorkflowIdResult200Status = {
|
|
|
46
42
|
timed_out: 'timed_out',
|
|
47
43
|
continued: 'continued',
|
|
48
44
|
};
|
|
45
|
+
export const PostWorkflowRun200Status = {
|
|
46
|
+
completed: 'completed',
|
|
47
|
+
failed: 'failed',
|
|
48
|
+
};
|
|
49
49
|
;
|
|
50
50
|
export const getGetHealthUrl = () => {
|
|
51
51
|
return `/health`;
|
|
@@ -87,6 +87,24 @@ export const getWorkflowIdStatus = async (id, options) => {
|
|
|
87
87
|
method: 'GET'
|
|
88
88
|
});
|
|
89
89
|
};
|
|
90
|
+
export const getGetWorkflowIdRunsRidStatusUrl = (id, rid) => {
|
|
91
|
+
return `/workflow/${id}/runs/${rid}/status`;
|
|
92
|
+
};
|
|
93
|
+
export const getWorkflowIdRunsRidStatus = async (id, rid, options) => {
|
|
94
|
+
return customFetchInstance(getGetWorkflowIdRunsRidStatusUrl(id, rid), {
|
|
95
|
+
...options,
|
|
96
|
+
method: 'GET'
|
|
97
|
+
});
|
|
98
|
+
};
|
|
99
|
+
export const getPatchWorkflowIdRunsRidStopUrl = (id, rid) => {
|
|
100
|
+
return `/workflow/${id}/runs/${rid}/stop`;
|
|
101
|
+
};
|
|
102
|
+
export const patchWorkflowIdRunsRidStop = async (id, rid, options) => {
|
|
103
|
+
return customFetchInstance(getPatchWorkflowIdRunsRidStopUrl(id, rid), {
|
|
104
|
+
...options,
|
|
105
|
+
method: 'PATCH'
|
|
106
|
+
});
|
|
107
|
+
};
|
|
90
108
|
export const getPatchWorkflowIdStopUrl = (id) => {
|
|
91
109
|
return `/workflow/${id}/stop`;
|
|
92
110
|
};
|
|
@@ -96,6 +114,17 @@ export const patchWorkflowIdStop = async (id, options) => {
|
|
|
96
114
|
method: 'PATCH'
|
|
97
115
|
});
|
|
98
116
|
};
|
|
117
|
+
export const getPostWorkflowIdRunsRidTerminateUrl = (id, rid) => {
|
|
118
|
+
return `/workflow/${id}/runs/${rid}/terminate`;
|
|
119
|
+
};
|
|
120
|
+
export const postWorkflowIdRunsRidTerminate = async (id, rid, postWorkflowIdRunsRidTerminateBody, options) => {
|
|
121
|
+
return customFetchInstance(getPostWorkflowIdRunsRidTerminateUrl(id, rid), {
|
|
122
|
+
...options,
|
|
123
|
+
method: 'POST',
|
|
124
|
+
headers: { 'Content-Type': 'application/json', ...options?.headers },
|
|
125
|
+
body: JSON.stringify(postWorkflowIdRunsRidTerminateBody)
|
|
126
|
+
});
|
|
127
|
+
};
|
|
99
128
|
export const getPostWorkflowIdTerminateUrl = (id) => {
|
|
100
129
|
return `/workflow/${id}/terminate`;
|
|
101
130
|
};
|
|
@@ -107,15 +136,26 @@ export const postWorkflowIdTerminate = async (id, postWorkflowIdTerminateBody, o
|
|
|
107
136
|
body: JSON.stringify(postWorkflowIdTerminateBody)
|
|
108
137
|
});
|
|
109
138
|
};
|
|
139
|
+
export const getPostWorkflowIdRunsRidResetUrl = (id, rid) => {
|
|
140
|
+
return `/workflow/${id}/runs/${rid}/reset`;
|
|
141
|
+
};
|
|
142
|
+
export const postWorkflowIdRunsRidReset = async (id, rid, resetWorkflowRequest, options) => {
|
|
143
|
+
return customFetchInstance(getPostWorkflowIdRunsRidResetUrl(id, rid), {
|
|
144
|
+
...options,
|
|
145
|
+
method: 'POST',
|
|
146
|
+
headers: { 'Content-Type': 'application/json', ...options?.headers },
|
|
147
|
+
body: JSON.stringify(resetWorkflowRequest)
|
|
148
|
+
});
|
|
149
|
+
};
|
|
110
150
|
export const getPostWorkflowIdResetUrl = (id) => {
|
|
111
151
|
return `/workflow/${id}/reset`;
|
|
112
152
|
};
|
|
113
|
-
export const postWorkflowIdReset = async (id,
|
|
153
|
+
export const postWorkflowIdReset = async (id, resetWorkflowRequest, options) => {
|
|
114
154
|
return customFetchInstance(getPostWorkflowIdResetUrl(id), {
|
|
115
155
|
...options,
|
|
116
156
|
method: 'POST',
|
|
117
157
|
headers: { 'Content-Type': 'application/json', ...options?.headers },
|
|
118
|
-
body: JSON.stringify(
|
|
158
|
+
body: JSON.stringify(resetWorkflowRequest)
|
|
119
159
|
});
|
|
120
160
|
};
|
|
121
161
|
export const getGetWorkflowIdResultUrl = (id) => {
|
|
@@ -127,6 +167,15 @@ export const getWorkflowIdResult = async (id, options) => {
|
|
|
127
167
|
method: 'GET'
|
|
128
168
|
});
|
|
129
169
|
};
|
|
170
|
+
export const getGetWorkflowIdRunsRidResultUrl = (id, rid) => {
|
|
171
|
+
return `/workflow/${id}/runs/${rid}/result`;
|
|
172
|
+
};
|
|
173
|
+
export const getWorkflowIdRunsRidResult = async (id, rid, options) => {
|
|
174
|
+
return customFetchInstance(getGetWorkflowIdRunsRidResultUrl(id, rid), {
|
|
175
|
+
...options,
|
|
176
|
+
method: 'GET'
|
|
177
|
+
});
|
|
178
|
+
};
|
|
130
179
|
export const getGetWorkflowIdTraceLogUrl = (id) => {
|
|
131
180
|
return `/workflow/${id}/trace-log`;
|
|
132
181
|
};
|
|
@@ -136,6 +185,15 @@ export const getWorkflowIdTraceLog = async (id, options) => {
|
|
|
136
185
|
method: 'GET'
|
|
137
186
|
});
|
|
138
187
|
};
|
|
188
|
+
export const getGetWorkflowIdRunsRidTraceLogUrl = (id, rid) => {
|
|
189
|
+
return `/workflow/${id}/runs/${rid}/trace-log`;
|
|
190
|
+
};
|
|
191
|
+
export const getWorkflowIdRunsRidTraceLog = async (id, rid, options) => {
|
|
192
|
+
return customFetchInstance(getGetWorkflowIdRunsRidTraceLogUrl(id, rid), {
|
|
193
|
+
...options,
|
|
194
|
+
method: 'GET'
|
|
195
|
+
});
|
|
196
|
+
};
|
|
139
197
|
export const getGetWorkflowCatalogIdUrl = (id) => {
|
|
140
198
|
return `/workflow/catalog/${id}`;
|
|
141
199
|
};
|
|
@@ -0,0 +1,14 @@
|
|
|
1
|
+
import { Command } from '@oclif/core';
|
|
2
|
+
export default class CredentialsSet extends Command {
|
|
3
|
+
static description: string;
|
|
4
|
+
static examples: string[];
|
|
5
|
+
static args: {
|
|
6
|
+
path: import("@oclif/core/interfaces").Arg<string, Record<string, unknown>>;
|
|
7
|
+
value: import("@oclif/core/interfaces").Arg<string, Record<string, unknown>>;
|
|
8
|
+
};
|
|
9
|
+
static flags: {
|
|
10
|
+
environment: import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
|
|
11
|
+
workflow: import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
|
|
12
|
+
};
|
|
13
|
+
run(): Promise<void>;
|
|
14
|
+
}
|
|
@@ -0,0 +1,57 @@
|
|
|
1
|
+
import { Args, Command, Flags } from '@oclif/core';
|
|
2
|
+
import { load as parseYaml, dump as stringifyYaml } from 'js-yaml';
|
|
3
|
+
import { decryptCredentials, credentialsExist, writeEncrypted, resolveCredentialsPath } from '#services/credentials_service.js';
|
|
4
|
+
const setNestedValue = (obj, dotPath, value) => {
|
|
5
|
+
const parts = dotPath.split('.');
|
|
6
|
+
const parent = parts.slice(0, -1).reduce((current, key) => {
|
|
7
|
+
if (!current[key] || typeof current[key] !== 'object') {
|
|
8
|
+
current[key] = {};
|
|
9
|
+
}
|
|
10
|
+
return current[key];
|
|
11
|
+
}, obj);
|
|
12
|
+
parent[parts[parts.length - 1]] = value;
|
|
13
|
+
};
|
|
14
|
+
export default class CredentialsSet extends Command {
|
|
15
|
+
static description = 'Set a credential value by dot-notation path';
|
|
16
|
+
static examples = [
|
|
17
|
+
'<%= config.bin %> <%= command.id %> anthropic.api_key sk-ant-...',
|
|
18
|
+
'<%= config.bin %> <%= command.id %> openai.api_key sk-... --environment production',
|
|
19
|
+
'<%= config.bin %> <%= command.id %> stripe.key sk_live_... --workflow my_workflow'
|
|
20
|
+
];
|
|
21
|
+
static args = {
|
|
22
|
+
path: Args.string({
|
|
23
|
+
description: 'Dot-notation path to the credential (e.g. anthropic.api_key)',
|
|
24
|
+
required: true
|
|
25
|
+
}),
|
|
26
|
+
value: Args.string({
|
|
27
|
+
description: 'Value to set',
|
|
28
|
+
required: true
|
|
29
|
+
})
|
|
30
|
+
};
|
|
31
|
+
static flags = {
|
|
32
|
+
environment: Flags.string({
|
|
33
|
+
char: 'e',
|
|
34
|
+
description: 'Target environment (e.g. production, development)'
|
|
35
|
+
}),
|
|
36
|
+
workflow: Flags.string({
|
|
37
|
+
char: 'w',
|
|
38
|
+
description: 'Target a specific workflow directory'
|
|
39
|
+
})
|
|
40
|
+
};
|
|
41
|
+
async run() {
|
|
42
|
+
const { args, flags } = await this.parse(CredentialsSet);
|
|
43
|
+
const environment = flags.environment;
|
|
44
|
+
const workflow = flags.workflow;
|
|
45
|
+
if (environment && workflow) {
|
|
46
|
+
this.error('Cannot specify both --environment and --workflow.');
|
|
47
|
+
}
|
|
48
|
+
if (!credentialsExist(environment, workflow)) {
|
|
49
|
+
this.error(`No credentials file found at ${resolveCredentialsPath(environment, workflow)}. Run "output credentials init" first.`);
|
|
50
|
+
}
|
|
51
|
+
const plaintext = decryptCredentials(environment, workflow);
|
|
52
|
+
const data = (parseYaml(plaintext) || {});
|
|
53
|
+
setNestedValue(data, args.path, args.value);
|
|
54
|
+
writeEncrypted(environment, stringifyYaml(data), workflow);
|
|
55
|
+
this.log(`Set ${args.path}`);
|
|
56
|
+
}
|
|
57
|
+
}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
export {};
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
/* eslint-disable @typescript-eslint/no-explicit-any */
|
|
2
|
+
import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest';
|
|
3
|
+
import * as credentialsService from '#services/credentials_service.js';
|
|
4
|
+
import CredentialsSet from './set.js';
|
|
5
|
+
vi.mock('#services/credentials_service.js');
|
|
6
|
+
vi.mock('js-yaml', () => ({
|
|
7
|
+
load: vi.fn((yaml) => {
|
|
8
|
+
if (yaml.includes('sk-existing')) {
|
|
9
|
+
return { anthropic: { api_key: 'sk-existing' } };
|
|
10
|
+
}
|
|
11
|
+
return {};
|
|
12
|
+
}),
|
|
13
|
+
dump: vi.fn((obj) => JSON.stringify(obj))
|
|
14
|
+
}));
|
|
15
|
+
describe('credentials set command', () => {
|
|
16
|
+
beforeEach(() => {
|
|
17
|
+
vi.clearAllMocks();
|
|
18
|
+
vi.mocked(credentialsService.credentialsExist).mockReturnValue(true);
|
|
19
|
+
vi.mocked(credentialsService.decryptCredentials).mockReturnValue('anthropic:\n api_key: sk-existing\n');
|
|
20
|
+
vi.mocked(credentialsService.writeEncrypted).mockImplementation(() => { });
|
|
21
|
+
});
|
|
22
|
+
afterEach(() => {
|
|
23
|
+
vi.restoreAllMocks();
|
|
24
|
+
});
|
|
25
|
+
const createTestCommand = (parsedArgs = {}, flags = {}) => {
|
|
26
|
+
const cmd = new CredentialsSet([], {});
|
|
27
|
+
cmd.log = vi.fn();
|
|
28
|
+
cmd.error = vi.fn((msg) => {
|
|
29
|
+
throw new Error(msg);
|
|
30
|
+
});
|
|
31
|
+
Object.defineProperty(cmd, 'parse', {
|
|
32
|
+
value: vi.fn().mockResolvedValue({
|
|
33
|
+
args: { path: 'anthropic.api_key', value: 'sk-new-key', ...parsedArgs },
|
|
34
|
+
flags: { environment: undefined, workflow: undefined, ...flags }
|
|
35
|
+
}),
|
|
36
|
+
configurable: true
|
|
37
|
+
});
|
|
38
|
+
return cmd;
|
|
39
|
+
};
|
|
40
|
+
describe('command structure', () => {
|
|
41
|
+
it('should have correct description', () => {
|
|
42
|
+
expect(CredentialsSet.description).toContain('credential value');
|
|
43
|
+
});
|
|
44
|
+
it('should have required path and value arguments', () => {
|
|
45
|
+
expect(CredentialsSet.args.path).toBeDefined();
|
|
46
|
+
expect(CredentialsSet.args.path.required).toBe(true);
|
|
47
|
+
expect(CredentialsSet.args.value).toBeDefined();
|
|
48
|
+
expect(CredentialsSet.args.value.required).toBe(true);
|
|
49
|
+
});
|
|
50
|
+
it('should have environment and workflow flags', () => {
|
|
51
|
+
expect(CredentialsSet.flags.environment).toBeDefined();
|
|
52
|
+
expect(CredentialsSet.flags.workflow).toBeDefined();
|
|
53
|
+
});
|
|
54
|
+
});
|
|
55
|
+
describe('command execution', () => {
|
|
56
|
+
it('should decrypt, update, and re-encrypt credentials', async () => {
|
|
57
|
+
const cmd = createTestCommand();
|
|
58
|
+
await cmd.run();
|
|
59
|
+
expect(credentialsService.decryptCredentials).toHaveBeenCalledWith(undefined, undefined);
|
|
60
|
+
expect(credentialsService.writeEncrypted).toHaveBeenCalledWith(undefined, expect.any(String), undefined);
|
|
61
|
+
expect(cmd.log).toHaveBeenCalledWith('Set anthropic.api_key');
|
|
62
|
+
});
|
|
63
|
+
it('should create nested keys that do not exist', async () => {
|
|
64
|
+
vi.mocked(credentialsService.decryptCredentials).mockReturnValue('');
|
|
65
|
+
const cmd = createTestCommand({ path: 'new.nested.key', value: 'my-value' });
|
|
66
|
+
await cmd.run();
|
|
67
|
+
expect(credentialsService.writeEncrypted).toHaveBeenCalledTimes(1);
|
|
68
|
+
expect(cmd.log).toHaveBeenCalledWith('Set new.nested.key');
|
|
69
|
+
});
|
|
70
|
+
it('should pass environment flag to service functions', async () => {
|
|
71
|
+
const cmd = createTestCommand({}, { environment: 'production' });
|
|
72
|
+
await cmd.run();
|
|
73
|
+
expect(credentialsService.credentialsExist).toHaveBeenCalledWith('production', undefined);
|
|
74
|
+
expect(credentialsService.decryptCredentials).toHaveBeenCalledWith('production', undefined);
|
|
75
|
+
expect(credentialsService.writeEncrypted).toHaveBeenCalledWith('production', expect.any(String), undefined);
|
|
76
|
+
});
|
|
77
|
+
it('should pass workflow flag to service functions', async () => {
|
|
78
|
+
const cmd = createTestCommand({}, { workflow: 'my_workflow' });
|
|
79
|
+
await cmd.run();
|
|
80
|
+
expect(credentialsService.credentialsExist).toHaveBeenCalledWith(undefined, 'my_workflow');
|
|
81
|
+
expect(credentialsService.decryptCredentials).toHaveBeenCalledWith(undefined, 'my_workflow');
|
|
82
|
+
expect(credentialsService.writeEncrypted).toHaveBeenCalledWith(undefined, expect.any(String), 'my_workflow');
|
|
83
|
+
});
|
|
84
|
+
it('should error when both environment and workflow are specified', async () => {
|
|
85
|
+
const cmd = createTestCommand({}, { environment: 'production', workflow: 'my_workflow' });
|
|
86
|
+
await expect(cmd.run()).rejects.toThrow('Cannot specify both');
|
|
87
|
+
});
|
|
88
|
+
it('should error when credentials file does not exist', async () => {
|
|
89
|
+
vi.mocked(credentialsService.credentialsExist).mockReturnValue(false);
|
|
90
|
+
vi.mocked(credentialsService.resolveCredentialsPath).mockReturnValue('/project/config/credentials.yml.enc');
|
|
91
|
+
const cmd = createTestCommand();
|
|
92
|
+
await expect(cmd.run()).rejects.toThrow('No credentials file found');
|
|
93
|
+
});
|
|
94
|
+
});
|
|
95
|
+
});
|
|
@@ -6,6 +6,7 @@ export default class WorkflowRunsList extends Command {
|
|
|
6
6
|
workflowName: import("@oclif/core/interfaces").Arg<string | undefined, Record<string, unknown>>;
|
|
7
7
|
};
|
|
8
8
|
static flags: {
|
|
9
|
+
'task-queue': import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
|
|
9
10
|
limit: import("@oclif/core/interfaces").OptionFlag<number, import("@oclif/core/interfaces").CustomOptions>;
|
|
10
11
|
format: import("@oclif/core/interfaces").OptionFlag<string, import("@oclif/core/interfaces").CustomOptions>;
|
|
11
12
|
};
|
|
@@ -65,6 +65,11 @@ export default class WorkflowRunsList extends Command {
|
|
|
65
65
|
})
|
|
66
66
|
};
|
|
67
67
|
static flags = {
|
|
68
|
+
'task-queue': Flags.string({
|
|
69
|
+
char: 'q',
|
|
70
|
+
description: 'Filter runs by task queue (defaults to OUTPUT_CATALOG_ID)',
|
|
71
|
+
env: 'OUTPUT_CATALOG_ID'
|
|
72
|
+
}),
|
|
68
73
|
limit: Flags.integer({
|
|
69
74
|
char: 'l',
|
|
70
75
|
description: 'Maximum number of runs to return',
|
|
@@ -81,6 +86,7 @@ export default class WorkflowRunsList extends Command {
|
|
|
81
86
|
const { args, flags } = await this.parse(WorkflowRunsList);
|
|
82
87
|
const { runs, count } = await fetchWorkflowRuns({
|
|
83
88
|
workflowType: args.workflowName,
|
|
89
|
+
taskQueue: flags['task-queue'],
|
|
84
90
|
limit: flags.limit
|
|
85
91
|
});
|
|
86
92
|
if (runs.length === 0) {
|
|
@@ -114,7 +114,7 @@ export default class WorkflowTest extends Command {
|
|
|
114
114
|
}
|
|
115
115
|
};
|
|
116
116
|
if (save) {
|
|
117
|
-
const filePath = join(dir, `${dataset.name}.yml`);
|
|
117
|
+
const filePath = dataset._source ?? join(dir, `${dataset.name}.yml`);
|
|
118
118
|
await writeDataset(updated, filePath);
|
|
119
119
|
this.log(` Saved output to ${filePath}`);
|
|
120
120
|
}
|
|
@@ -137,7 +137,7 @@ export default class WorkflowTest extends Command {
|
|
|
137
137
|
date: now
|
|
138
138
|
}
|
|
139
139
|
};
|
|
140
|
-
const filePath = join(dir, `${dataset.name}.yml`);
|
|
140
|
+
const filePath = dataset._source ?? join(dir, `${dataset.name}.yml`);
|
|
141
141
|
await writeDataset(updated, filePath);
|
|
142
142
|
this.log(` Saved eval result to ${filePath}`);
|
|
143
143
|
}
|
|
@@ -8,7 +8,7 @@ export interface DatasetInfo {
|
|
|
8
8
|
}
|
|
9
9
|
export declare function resolveDatasetsDir(workflowName: string, basePath?: string): string | null;
|
|
10
10
|
export declare function resolveDefaultDatasetsDir(workflowName: string, basePath?: string): string;
|
|
11
|
-
export declare function
|
|
11
|
+
export declare function readDatasetFile(filePath: string): Promise<Dataset[]>;
|
|
12
12
|
export declare function readAllDatasets(workflowName: string, filterNames?: string[], basePath?: string): Promise<{
|
|
13
13
|
datasets: Dataset[];
|
|
14
14
|
dir: string;
|
|
@@ -24,10 +24,19 @@ export function resolveDefaultDatasetsDir(workflowName, basePath = process.cwd()
|
|
|
24
24
|
// Default to first workflows path
|
|
25
25
|
return resolve(basePath, WORKFLOWS_PATHS[0], workflowName, DATASETS_DIR);
|
|
26
26
|
}
|
|
27
|
-
export async function
|
|
28
|
-
const
|
|
29
|
-
|
|
30
|
-
|
|
27
|
+
export async function readDatasetFile(filePath) {
|
|
28
|
+
const raw = yaml.load(await readFile(filePath, 'utf-8'));
|
|
29
|
+
if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
|
|
30
|
+
throw new Error(`Invalid dataset file: ${filePath}`);
|
|
31
|
+
}
|
|
32
|
+
return Object.entries(raw).map(([name, body]) => {
|
|
33
|
+
if (!body || typeof body !== 'object' || !('input' in body)) {
|
|
34
|
+
throw new Error(`Dataset case "${name}" in ${filePath} is missing required "input" field`);
|
|
35
|
+
}
|
|
36
|
+
const dataset = DatasetSchema.parse({ name, ...body });
|
|
37
|
+
dataset._source = filePath;
|
|
38
|
+
return dataset;
|
|
39
|
+
});
|
|
31
40
|
}
|
|
32
41
|
export async function readAllDatasets(workflowName, filterNames, basePath) {
|
|
33
42
|
const dir = resolveDatasetsDir(workflowName, basePath);
|
|
@@ -36,42 +45,35 @@ export async function readAllDatasets(workflowName, filterNames, basePath) {
|
|
|
36
45
|
}
|
|
37
46
|
const files = await readdir(dir);
|
|
38
47
|
const ymlFiles = files.filter(f => f.endsWith('.yml') || f.endsWith('.yaml'));
|
|
48
|
+
const seen = new Set();
|
|
39
49
|
const datasets = [];
|
|
40
50
|
for (const file of ymlFiles) {
|
|
41
|
-
const
|
|
42
|
-
|
|
43
|
-
|
|
51
|
+
const cases = await readDatasetFile(join(dir, file));
|
|
52
|
+
for (const dataset of cases) {
|
|
53
|
+
if (seen.has(dataset.name)) {
|
|
54
|
+
throw new Error(`Duplicate dataset case name "${dataset.name}" found in ${file}`);
|
|
55
|
+
}
|
|
56
|
+
seen.add(dataset.name);
|
|
57
|
+
if (filterNames && !filterNames.includes(dataset.name)) {
|
|
58
|
+
continue;
|
|
59
|
+
}
|
|
60
|
+
datasets.push(dataset);
|
|
44
61
|
}
|
|
45
|
-
datasets.push(dataset);
|
|
46
62
|
}
|
|
47
63
|
return { datasets, dir };
|
|
48
64
|
}
|
|
49
|
-
async function mergeWithExisting(dataset, filePath) {
|
|
50
|
-
if (!existsSync(filePath)) {
|
|
51
|
-
return dataset;
|
|
52
|
-
}
|
|
53
|
-
try {
|
|
54
|
-
const existing = await readDataset(filePath);
|
|
55
|
-
return {
|
|
56
|
-
...existing,
|
|
57
|
-
...dataset,
|
|
58
|
-
ground_truth: dataset.ground_truth ?? existing.ground_truth,
|
|
59
|
-
last_output: dataset.last_output ?? existing.last_output,
|
|
60
|
-
last_eval: dataset.last_eval ?? existing.last_eval
|
|
61
|
-
};
|
|
62
|
-
}
|
|
63
|
-
catch {
|
|
64
|
-
return dataset;
|
|
65
|
-
}
|
|
66
|
-
}
|
|
67
65
|
export async function writeDataset(dataset, filePath) {
|
|
68
|
-
const merged = await mergeWithExisting(dataset, filePath);
|
|
69
66
|
const dir = resolve(filePath, '..');
|
|
70
67
|
if (!existsSync(dir)) {
|
|
71
68
|
await mkdir(dir, { recursive: true });
|
|
72
69
|
}
|
|
73
|
-
const
|
|
74
|
-
|
|
70
|
+
const loaded = existsSync(filePath) ? yaml.load(await readFile(filePath, 'utf-8')) : null;
|
|
71
|
+
const fileObj = (loaded && typeof loaded === 'object' && !Array.isArray(loaded)) ?
|
|
72
|
+
loaded :
|
|
73
|
+
{};
|
|
74
|
+
const { name, _source, ...caseBody } = dataset;
|
|
75
|
+
fileObj[name] = { ...fileObj[name], ...caseBody };
|
|
76
|
+
await writeFile(filePath, yaml.dump(fileObj, { lineWidth: 120, noRefs: true, sortKeys: false }), 'utf-8');
|
|
75
77
|
}
|
|
76
78
|
export async function listDatasets(workflowName, basePath) {
|
|
77
79
|
const dir = resolveDatasetsDir(workflowName, basePath);
|
|
@@ -84,14 +86,16 @@ export async function listDatasets(workflowName, basePath) {
|
|
|
84
86
|
for (const file of ymlFiles) {
|
|
85
87
|
const filePath = join(dir, file);
|
|
86
88
|
try {
|
|
87
|
-
const
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
89
|
+
const cases = await readDatasetFile(filePath);
|
|
90
|
+
for (const dataset of cases) {
|
|
91
|
+
results.push({
|
|
92
|
+
name: dataset.name,
|
|
93
|
+
path: filePath,
|
|
94
|
+
hasLastOutput: dataset.last_output?.output !== undefined,
|
|
95
|
+
lastOutputDate: dataset.last_output?.date,
|
|
96
|
+
lastEvalDate: dataset.last_eval?.date
|
|
97
|
+
});
|
|
98
|
+
}
|
|
95
99
|
}
|
|
96
100
|
catch {
|
|
97
101
|
results.push({
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
export {};
|