model-test-bench 1.0.2
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/LICENSE +201 -0
- package/README.md +389 -0
- package/dist/bin/mtb.d.ts +3 -0
- package/dist/bin/mtb.d.ts.map +1 -0
- package/dist/bin/mtb.js +148 -0
- package/dist/bin/mtb.js.map +1 -0
- package/dist/server/index.d.ts +13 -0
- package/dist/server/index.d.ts.map +1 -0
- package/dist/server/index.js +72 -0
- package/dist/server/index.js.map +1 -0
- package/dist/server/interfaces/evaluator.d.ts +15 -0
- package/dist/server/interfaces/evaluator.d.ts.map +1 -0
- package/dist/server/interfaces/evaluator.js +2 -0
- package/dist/server/interfaces/evaluator.js.map +1 -0
- package/dist/server/interfaces/logger.d.ts +9 -0
- package/dist/server/interfaces/logger.d.ts.map +1 -0
- package/dist/server/interfaces/logger.js +2 -0
- package/dist/server/interfaces/logger.js.map +1 -0
- package/dist/server/interfaces/runner.d.ts +9 -0
- package/dist/server/interfaces/runner.d.ts.map +1 -0
- package/dist/server/interfaces/runner.js +2 -0
- package/dist/server/interfaces/runner.js.map +1 -0
- package/dist/server/interfaces/storage.d.ts +36 -0
- package/dist/server/interfaces/storage.d.ts.map +1 -0
- package/dist/server/interfaces/storage.js +2 -0
- package/dist/server/interfaces/storage.js.map +1 -0
- package/dist/server/routes/eval-queue.d.ts +23 -0
- package/dist/server/routes/eval-queue.d.ts.map +1 -0
- package/dist/server/routes/eval-queue.js +45 -0
- package/dist/server/routes/eval-queue.js.map +1 -0
- package/dist/server/routes/evaluations.d.ts +8 -0
- package/dist/server/routes/evaluations.d.ts.map +1 -0
- package/dist/server/routes/evaluations.js +221 -0
- package/dist/server/routes/evaluations.js.map +1 -0
- package/dist/server/routes/providers.d.ts +5 -0
- package/dist/server/routes/providers.d.ts.map +1 -0
- package/dist/server/routes/providers.js +179 -0
- package/dist/server/routes/providers.js.map +1 -0
- package/dist/server/routes/run-queue.d.ts +17 -0
- package/dist/server/routes/run-queue.d.ts.map +1 -0
- package/dist/server/routes/run-queue.js +34 -0
- package/dist/server/routes/run-queue.js.map +1 -0
- package/dist/server/routes/run-sse.d.ts +18 -0
- package/dist/server/routes/run-sse.d.ts.map +1 -0
- package/dist/server/routes/run-sse.js +57 -0
- package/dist/server/routes/run-sse.js.map +1 -0
- package/dist/server/routes/runs.d.ts +9 -0
- package/dist/server/routes/runs.d.ts.map +1 -0
- package/dist/server/routes/runs.js +380 -0
- package/dist/server/routes/runs.js.map +1 -0
- package/dist/server/routes/scenarios.d.ts +5 -0
- package/dist/server/routes/scenarios.d.ts.map +1 -0
- package/dist/server/routes/scenarios.js +181 -0
- package/dist/server/routes/scenarios.js.map +1 -0
- package/dist/server/services/eval-helpers.d.ts +22 -0
- package/dist/server/services/eval-helpers.d.ts.map +1 -0
- package/dist/server/services/eval-helpers.js +75 -0
- package/dist/server/services/eval-helpers.js.map +1 -0
- package/dist/server/services/eval-parsers-debate-impl.d.ts +11 -0
- package/dist/server/services/eval-parsers-debate-impl.d.ts.map +1 -0
- package/dist/server/services/eval-parsers-debate-impl.js +133 -0
- package/dist/server/services/eval-parsers-debate-impl.js.map +1 -0
- package/dist/server/services/eval-parsers.d.ts +24 -0
- package/dist/server/services/eval-parsers.d.ts.map +1 -0
- package/dist/server/services/eval-parsers.js +153 -0
- package/dist/server/services/eval-parsers.js.map +1 -0
- package/dist/server/services/eval-prompts.d.ts +9 -0
- package/dist/server/services/eval-prompts.d.ts.map +1 -0
- package/dist/server/services/eval-prompts.js +164 -0
- package/dist/server/services/eval-prompts.js.map +1 -0
- package/dist/server/services/evaluator.d.ts +10 -0
- package/dist/server/services/evaluator.d.ts.map +1 -0
- package/dist/server/services/evaluator.js +143 -0
- package/dist/server/services/evaluator.js.map +1 -0
- package/dist/server/services/fs-adapter.d.ts +20 -0
- package/dist/server/services/fs-adapter.d.ts.map +1 -0
- package/dist/server/services/fs-adapter.js +13 -0
- package/dist/server/services/fs-adapter.js.map +1 -0
- package/dist/server/services/instruction-parser.d.ts +26 -0
- package/dist/server/services/instruction-parser.d.ts.map +1 -0
- package/dist/server/services/instruction-parser.js +121 -0
- package/dist/server/services/instruction-parser.js.map +1 -0
- package/dist/server/services/log-rotator.d.ts +20 -0
- package/dist/server/services/log-rotator.d.ts.map +1 -0
- package/dist/server/services/log-rotator.js +60 -0
- package/dist/server/services/log-rotator.js.map +1 -0
- package/dist/server/services/logger.d.ts +15 -0
- package/dist/server/services/logger.d.ts.map +1 -0
- package/dist/server/services/logger.js +69 -0
- package/dist/server/services/logger.js.map +1 -0
- package/dist/server/services/model-factory.d.ts +10 -0
- package/dist/server/services/model-factory.d.ts.map +1 -0
- package/dist/server/services/model-factory.js +33 -0
- package/dist/server/services/model-factory.js.map +1 -0
- package/dist/server/services/runner.d.ts +9 -0
- package/dist/server/services/runner.d.ts.map +1 -0
- package/dist/server/services/runner.js +99 -0
- package/dist/server/services/runner.js.map +1 -0
- package/dist/server/services/seeder.d.ts +5 -0
- package/dist/server/services/seeder.d.ts.map +1 -0
- package/dist/server/services/seeder.js +79 -0
- package/dist/server/services/seeder.js.map +1 -0
- package/dist/server/services/storage-test-helpers.d.ts +15 -0
- package/dist/server/services/storage-test-helpers.d.ts.map +1 -0
- package/dist/server/services/storage-test-helpers.js +151 -0
- package/dist/server/services/storage-test-helpers.js.map +1 -0
- package/dist/server/services/storage.d.ts +35 -0
- package/dist/server/services/storage.d.ts.map +1 -0
- package/dist/server/services/storage.js +219 -0
- package/dist/server/services/storage.js.map +1 -0
- package/dist/server/services/tools.d.ts +6 -0
- package/dist/server/services/tools.d.ts.map +1 -0
- package/dist/server/services/tools.js +94 -0
- package/dist/server/services/tools.js.map +1 -0
- package/dist/server/services/transcript-formatter.d.ts +18 -0
- package/dist/server/services/transcript-formatter.d.ts.map +1 -0
- package/dist/server/services/transcript-formatter.js +227 -0
- package/dist/server/services/transcript-formatter.js.map +1 -0
- package/dist/server/services/update-checker.d.ts +3 -0
- package/dist/server/services/update-checker.d.ts.map +1 -0
- package/dist/server/services/update-checker.js +34 -0
- package/dist/server/services/update-checker.js.map +1 -0
- package/dist/server/types/evaluation.d.ts +94 -0
- package/dist/server/types/evaluation.d.ts.map +1 -0
- package/dist/server/types/evaluation.js +5 -0
- package/dist/server/types/evaluation.js.map +1 -0
- package/dist/server/types/index.d.ts +5 -0
- package/dist/server/types/index.d.ts.map +1 -0
- package/dist/server/types/index.js +5 -0
- package/dist/server/types/index.js.map +1 -0
- package/dist/server/types/provider.d.ts +23 -0
- package/dist/server/types/provider.d.ts.map +1 -0
- package/dist/server/types/provider.js +5 -0
- package/dist/server/types/provider.js.map +1 -0
- package/dist/server/types/run.d.ts +31 -0
- package/dist/server/types/run.d.ts.map +1 -0
- package/dist/server/types/run.js +5 -0
- package/dist/server/types/run.js.map +1 -0
- package/dist/server/types/scenario.d.ts +19 -0
- package/dist/server/types/scenario.d.ts.map +1 -0
- package/dist/server/types/scenario.js +5 -0
- package/dist/server/types/scenario.js.map +1 -0
- package/dist/src/server/index.d.ts +13 -0
- package/dist/src/server/index.d.ts.map +1 -0
- package/dist/src/server/index.js +72 -0
- package/dist/src/server/index.js.map +1 -0
- package/dist/src/server/interfaces/evaluator.d.ts +15 -0
- package/dist/src/server/interfaces/evaluator.d.ts.map +1 -0
- package/dist/src/server/interfaces/evaluator.js +2 -0
- package/dist/src/server/interfaces/evaluator.js.map +1 -0
- package/dist/src/server/interfaces/logger.d.ts +9 -0
- package/dist/src/server/interfaces/logger.d.ts.map +1 -0
- package/dist/src/server/interfaces/logger.js +2 -0
- package/dist/src/server/interfaces/logger.js.map +1 -0
- package/dist/src/server/interfaces/runner.d.ts +9 -0
- package/dist/src/server/interfaces/runner.d.ts.map +1 -0
- package/dist/src/server/interfaces/runner.js +2 -0
- package/dist/src/server/interfaces/runner.js.map +1 -0
- package/dist/src/server/interfaces/storage.d.ts +36 -0
- package/dist/src/server/interfaces/storage.d.ts.map +1 -0
- package/dist/src/server/interfaces/storage.js +2 -0
- package/dist/src/server/interfaces/storage.js.map +1 -0
- package/dist/src/server/routes/eval-queue.d.ts +23 -0
- package/dist/src/server/routes/eval-queue.d.ts.map +1 -0
- package/dist/src/server/routes/eval-queue.js +45 -0
- package/dist/src/server/routes/eval-queue.js.map +1 -0
- package/dist/src/server/routes/evaluations.d.ts +8 -0
- package/dist/src/server/routes/evaluations.d.ts.map +1 -0
- package/dist/src/server/routes/evaluations.js +221 -0
- package/dist/src/server/routes/evaluations.js.map +1 -0
- package/dist/src/server/routes/providers.d.ts +5 -0
- package/dist/src/server/routes/providers.d.ts.map +1 -0
- package/dist/src/server/routes/providers.js +179 -0
- package/dist/src/server/routes/providers.js.map +1 -0
- package/dist/src/server/routes/run-queue.d.ts +17 -0
- package/dist/src/server/routes/run-queue.d.ts.map +1 -0
- package/dist/src/server/routes/run-queue.js +34 -0
- package/dist/src/server/routes/run-queue.js.map +1 -0
- package/dist/src/server/routes/run-sse.d.ts +18 -0
- package/dist/src/server/routes/run-sse.d.ts.map +1 -0
- package/dist/src/server/routes/run-sse.js +57 -0
- package/dist/src/server/routes/run-sse.js.map +1 -0
- package/dist/src/server/routes/runs.d.ts +9 -0
- package/dist/src/server/routes/runs.d.ts.map +1 -0
- package/dist/src/server/routes/runs.js +380 -0
- package/dist/src/server/routes/runs.js.map +1 -0
- package/dist/src/server/routes/scenarios.d.ts +5 -0
- package/dist/src/server/routes/scenarios.d.ts.map +1 -0
- package/dist/src/server/routes/scenarios.js +181 -0
- package/dist/src/server/routes/scenarios.js.map +1 -0
- package/dist/src/server/services/eval-helpers.d.ts +22 -0
- package/dist/src/server/services/eval-helpers.d.ts.map +1 -0
- package/dist/src/server/services/eval-helpers.js +75 -0
- package/dist/src/server/services/eval-helpers.js.map +1 -0
- package/dist/src/server/services/eval-parsers-debate-impl.d.ts +11 -0
- package/dist/src/server/services/eval-parsers-debate-impl.d.ts.map +1 -0
- package/dist/src/server/services/eval-parsers-debate-impl.js +133 -0
- package/dist/src/server/services/eval-parsers-debate-impl.js.map +1 -0
- package/dist/src/server/services/eval-parsers.d.ts +24 -0
- package/dist/src/server/services/eval-parsers.d.ts.map +1 -0
- package/dist/src/server/services/eval-parsers.js +153 -0
- package/dist/src/server/services/eval-parsers.js.map +1 -0
- package/dist/src/server/services/eval-prompts.d.ts +9 -0
- package/dist/src/server/services/eval-prompts.d.ts.map +1 -0
- package/dist/src/server/services/eval-prompts.js +164 -0
- package/dist/src/server/services/eval-prompts.js.map +1 -0
- package/dist/src/server/services/evaluator.d.ts +10 -0
- package/dist/src/server/services/evaluator.d.ts.map +1 -0
- package/dist/src/server/services/evaluator.js +143 -0
- package/dist/src/server/services/evaluator.js.map +1 -0
- package/dist/src/server/services/fs-adapter.d.ts +20 -0
- package/dist/src/server/services/fs-adapter.d.ts.map +1 -0
- package/dist/src/server/services/fs-adapter.js +13 -0
- package/dist/src/server/services/fs-adapter.js.map +1 -0
- package/dist/src/server/services/instruction-parser.d.ts +26 -0
- package/dist/src/server/services/instruction-parser.d.ts.map +1 -0
- package/dist/src/server/services/instruction-parser.js +121 -0
- package/dist/src/server/services/instruction-parser.js.map +1 -0
- package/dist/src/server/services/log-rotator.d.ts +20 -0
- package/dist/src/server/services/log-rotator.d.ts.map +1 -0
- package/dist/src/server/services/log-rotator.js +60 -0
- package/dist/src/server/services/log-rotator.js.map +1 -0
- package/dist/src/server/services/logger.d.ts +15 -0
- package/dist/src/server/services/logger.d.ts.map +1 -0
- package/dist/src/server/services/logger.js +69 -0
- package/dist/src/server/services/logger.js.map +1 -0
- package/dist/src/server/services/model-factory.d.ts +10 -0
- package/dist/src/server/services/model-factory.d.ts.map +1 -0
- package/dist/src/server/services/model-factory.js +33 -0
- package/dist/src/server/services/model-factory.js.map +1 -0
- package/dist/src/server/services/runner.d.ts +9 -0
- package/dist/src/server/services/runner.d.ts.map +1 -0
- package/dist/src/server/services/runner.js +99 -0
- package/dist/src/server/services/runner.js.map +1 -0
- package/dist/src/server/services/seeder.d.ts +5 -0
- package/dist/src/server/services/seeder.d.ts.map +1 -0
- package/dist/src/server/services/seeder.js +79 -0
- package/dist/src/server/services/seeder.js.map +1 -0
- package/dist/src/server/services/storage.d.ts +35 -0
- package/dist/src/server/services/storage.d.ts.map +1 -0
- package/dist/src/server/services/storage.js +219 -0
- package/dist/src/server/services/storage.js.map +1 -0
- package/dist/src/server/services/tools.d.ts +6 -0
- package/dist/src/server/services/tools.d.ts.map +1 -0
- package/dist/src/server/services/tools.js +94 -0
- package/dist/src/server/services/tools.js.map +1 -0
- package/dist/src/server/services/transcript-formatter.d.ts +18 -0
- package/dist/src/server/services/transcript-formatter.d.ts.map +1 -0
- package/dist/src/server/services/transcript-formatter.js +227 -0
- package/dist/src/server/services/transcript-formatter.js.map +1 -0
- package/dist/src/server/services/update-checker.d.ts +3 -0
- package/dist/src/server/services/update-checker.d.ts.map +1 -0
- package/dist/src/server/services/update-checker.js +34 -0
- package/dist/src/server/services/update-checker.js.map +1 -0
- package/dist/src/server/types/evaluation.d.ts +94 -0
- package/dist/src/server/types/evaluation.d.ts.map +1 -0
- package/dist/src/server/types/evaluation.js +5 -0
- package/dist/src/server/types/evaluation.js.map +1 -0
- package/dist/src/server/types/index.d.ts +5 -0
- package/dist/src/server/types/index.d.ts.map +1 -0
- package/dist/src/server/types/index.js +5 -0
- package/dist/src/server/types/index.js.map +1 -0
- package/dist/src/server/types/provider.d.ts +23 -0
- package/dist/src/server/types/provider.d.ts.map +1 -0
- package/dist/src/server/types/provider.js +5 -0
- package/dist/src/server/types/provider.js.map +1 -0
- package/dist/src/server/types/run.d.ts +31 -0
- package/dist/src/server/types/run.d.ts.map +1 -0
- package/dist/src/server/types/run.js +5 -0
- package/dist/src/server/types/run.js.map +1 -0
- package/dist/src/server/types/scenario.d.ts +19 -0
- package/dist/src/server/types/scenario.d.ts.map +1 -0
- package/dist/src/server/types/scenario.js +5 -0
- package/dist/src/server/types/scenario.js.map +1 -0
- package/dist/web/assets/index-AJu1Yn5F.js +70 -0
- package/dist/web/assets/index-C_ioEISr.css +1 -0
- package/dist/web/index.html +15 -0
- package/docs/schemas/provider-api.example.json +12 -0
- package/docs/schemas/provider-openai.example.json +11 -0
- package/docs/schemas/scenario-baseline.example.json +24 -0
- package/docs/schemas/scenario-carwash-baseline.example.json +22 -0
- package/docs/schemas/scenario-carwash-with-system-prompt.example.json +24 -0
- package/docs/schemas/scenario-golden-rules-baseline.example.json +24 -0
- package/docs/schemas/scenario-golden-rules-with-system-prompt.example.json +28 -0
- package/docs/schemas/scenario-negative-analysis-baseline.example.json +23 -0
- package/docs/schemas/scenario-negative-analysis-with-system-prompt.example.json +25 -0
- package/docs/schemas/scenario-with-system-prompt.example.json +25 -0
- package/package.json +97 -0
|
@@ -0,0 +1,380 @@
|
|
|
1
|
+
// ---------------------------------------------------------------------------
|
|
2
|
+
// Run routes — start, list, get, delete, stream runs
|
|
3
|
+
// ---------------------------------------------------------------------------
|
|
4
|
+
import { Router } from 'express';
|
|
5
|
+
import { v4 as uuidv4 } from 'uuid';
|
|
6
|
+
import { handleSSEConnection, broadcastSSE, closeSSE } from './run-sse.js';
|
|
7
|
+
import { EvalQueue } from './eval-queue.js';
|
|
8
|
+
// ---------------------------------------------------------------------------
|
|
9
|
+
// Helpers
|
|
10
|
+
// ---------------------------------------------------------------------------
|
|
11
|
+
function paramId(req) {
|
|
12
|
+
const raw = req.params.id;
|
|
13
|
+
return Array.isArray(raw) ? raw[0] : raw;
|
|
14
|
+
}
|
|
15
|
+
/** Strip messages from a run for the summary endpoint. */
|
|
16
|
+
function runSummary(run) {
|
|
17
|
+
const { messages: _messages, ...rest } = run;
|
|
18
|
+
return rest;
|
|
19
|
+
}
|
|
20
|
+
/** Build EvaluatorConfig[] from reviewer provider snapshots. N-1 as Evaluators, last as Synthesizer. */
|
|
21
|
+
function buildEvaluatorConfigs(reviewerSnapshots) {
|
|
22
|
+
return reviewerSnapshots.map((p, idx) => ({
|
|
23
|
+
providerName: p.providerName,
|
|
24
|
+
model: p.model,
|
|
25
|
+
apiKey: p.apiKey,
|
|
26
|
+
baseUrl: p.baseUrl,
|
|
27
|
+
role: idx < reviewerSnapshots.length - 1
|
|
28
|
+
? (reviewerSnapshots.length > 2 ? `Evaluator ${idx + 1}` : 'Evaluator')
|
|
29
|
+
: 'Synthesizer',
|
|
30
|
+
}));
|
|
31
|
+
}
|
|
32
|
+
// Import and re-export RunQueue from extracted module
|
|
33
|
+
import { RunQueue } from './run-queue.js';
|
|
34
|
+
export { RunQueue } from './run-queue.js';
|
|
35
|
+
// ---------------------------------------------------------------------------
|
|
36
|
+
// Abort controller registry (for cancelling running runs)
|
|
37
|
+
// ---------------------------------------------------------------------------
|
|
38
|
+
const abortControllers = new Map();
|
|
39
|
+
// ---------------------------------------------------------------------------
|
|
40
|
+
// Router factory
|
|
41
|
+
// ---------------------------------------------------------------------------
|
|
42
|
+
export function createRunRoutes(storage, runner, logger, evaluator, queue) {
|
|
43
|
+
const router = Router();
|
|
44
|
+
const runQueue = queue ?? new RunQueue(1);
|
|
45
|
+
const sseSubscribers = new Map();
|
|
46
|
+
const evalQueue = new EvalQueue(1);
|
|
47
|
+
// POST / — start a new run
|
|
48
|
+
router.post('/', async (req, res) => {
|
|
49
|
+
try {
|
|
50
|
+
const body = req.body;
|
|
51
|
+
const providerId = body.providerId;
|
|
52
|
+
const scenarioId = body.scenarioId;
|
|
53
|
+
if (!providerId || !scenarioId) {
|
|
54
|
+
res.status(400).json({ error: 'providerId and scenarioId are required' });
|
|
55
|
+
return;
|
|
56
|
+
}
|
|
57
|
+
const provider = await storage.getProvider(providerId);
|
|
58
|
+
if (!provider) {
|
|
59
|
+
res.status(404).json({ error: 'Provider not found' });
|
|
60
|
+
return;
|
|
61
|
+
}
|
|
62
|
+
const scenario = await storage.getScenario(scenarioId);
|
|
63
|
+
if (!scenario) {
|
|
64
|
+
res.status(404).json({ error: 'Scenario not found' });
|
|
65
|
+
return;
|
|
66
|
+
}
|
|
67
|
+
// Optional reviewer providers for auto-evaluation
|
|
68
|
+
const rawReviewerIds = body.reviewerProviderIds;
|
|
69
|
+
let reviewerProviderIds;
|
|
70
|
+
let reviewerProviderSnapshots;
|
|
71
|
+
if (Array.isArray(rawReviewerIds) && rawReviewerIds.length > 0) {
|
|
72
|
+
reviewerProviderIds = [];
|
|
73
|
+
reviewerProviderSnapshots = [];
|
|
74
|
+
for (const rid of rawReviewerIds) {
|
|
75
|
+
if (typeof rid !== 'string') {
|
|
76
|
+
res.status(400).json({ error: 'reviewerProviderIds must be an array of strings' });
|
|
77
|
+
return;
|
|
78
|
+
}
|
|
79
|
+
const reviewerProvider = await storage.getProvider(rid);
|
|
80
|
+
if (!reviewerProvider) {
|
|
81
|
+
res.status(404).json({ error: `Reviewer provider not found: ${rid}` });
|
|
82
|
+
return;
|
|
83
|
+
}
|
|
84
|
+
reviewerProviderIds.push(rid);
|
|
85
|
+
reviewerProviderSnapshots.push(reviewerProvider);
|
|
86
|
+
}
|
|
87
|
+
}
|
|
88
|
+
const maxEvalRounds = typeof body.maxEvalRounds === 'number'
|
|
89
|
+
? Math.min(5, Math.max(1, body.maxEvalRounds))
|
|
90
|
+
: undefined;
|
|
91
|
+
const now = new Date().toISOString();
|
|
92
|
+
const run = {
|
|
93
|
+
id: uuidv4(),
|
|
94
|
+
providerId,
|
|
95
|
+
scenarioId,
|
|
96
|
+
status: 'pending',
|
|
97
|
+
providerSnapshot: provider,
|
|
98
|
+
scenarioSnapshot: scenario,
|
|
99
|
+
messages: [],
|
|
100
|
+
resultText: '',
|
|
101
|
+
totalCostUsd: 0,
|
|
102
|
+
durationMs: 0,
|
|
103
|
+
numTurns: 0,
|
|
104
|
+
reviewerProviderIds,
|
|
105
|
+
reviewerProviderSnapshots,
|
|
106
|
+
maxEvalRounds,
|
|
107
|
+
createdAt: now,
|
|
108
|
+
updatedAt: now,
|
|
109
|
+
};
|
|
110
|
+
await storage.saveRun(run);
|
|
111
|
+
// Enqueue async execution
|
|
112
|
+
const abortController = new AbortController();
|
|
113
|
+
abortControllers.set(run.id, abortController);
|
|
114
|
+
runQueue.enqueue({
|
|
115
|
+
run,
|
|
116
|
+
execute: async () => {
|
|
117
|
+
const callbacks = {
|
|
118
|
+
onMessage(message) {
|
|
119
|
+
broadcastSSE(run.id, 'message', { type: 'sdkMessage', ...message }, sseSubscribers);
|
|
120
|
+
},
|
|
121
|
+
onStatusChange(status) {
|
|
122
|
+
const updatedRun = {
|
|
123
|
+
...run,
|
|
124
|
+
status,
|
|
125
|
+
updatedAt: new Date().toISOString(),
|
|
126
|
+
};
|
|
127
|
+
broadcastSSE(run.id, 'message', { type: 'status', status }, sseSubscribers);
|
|
128
|
+
storage.saveRun(updatedRun).catch((saveErr) => {
|
|
129
|
+
logger.error('Failed to persist status change', {
|
|
130
|
+
runId: run.id,
|
|
131
|
+
status,
|
|
132
|
+
error: String(saveErr),
|
|
133
|
+
});
|
|
134
|
+
});
|
|
135
|
+
},
|
|
136
|
+
};
|
|
137
|
+
try {
|
|
138
|
+
const result = await runner.executeRun(provider, scenario, run, callbacks, abortController);
|
|
139
|
+
await storage.saveRun(result);
|
|
140
|
+
// Broadcast completed/failed run to SSE clients
|
|
141
|
+
broadcastSSE(run.id, 'message', { type: 'runComplete', run: runSummary(result) }, sseSubscribers);
|
|
142
|
+
// Auto-trigger evaluation if reviewer providers were configured
|
|
143
|
+
if (result.status === 'completed' && evaluator && reviewerProviderSnapshots && reviewerProviderSnapshots.length > 0) {
|
|
144
|
+
try {
|
|
145
|
+
const evaluatorConfigs = buildEvaluatorConfigs(reviewerProviderSnapshots);
|
|
146
|
+
const evalRequest = {
|
|
147
|
+
runId: result.id,
|
|
148
|
+
evaluators: evaluatorConfigs,
|
|
149
|
+
maxRounds: maxEvalRounds ?? 1,
|
|
150
|
+
};
|
|
151
|
+
const evalNow = new Date().toISOString();
|
|
152
|
+
const evaluation = {
|
|
153
|
+
id: uuidv4(),
|
|
154
|
+
runId: result.id,
|
|
155
|
+
status: 'pending',
|
|
156
|
+
evaluators: evaluatorConfigs,
|
|
157
|
+
rounds: [],
|
|
158
|
+
answerComparison: { matches: false, explanation: '', similarity: 0 },
|
|
159
|
+
criticalResults: [],
|
|
160
|
+
setupCompliance: {
|
|
161
|
+
instructionCompliance: { followed: [], violated: [], notApplicable: [], overallCompliance: 0 },
|
|
162
|
+
},
|
|
163
|
+
synthesis: { dimensionScores: {}, weightedTotal: 0, confidence: 0, dissenting: [] },
|
|
164
|
+
ledger: [],
|
|
165
|
+
totalCostUsd: 0,
|
|
166
|
+
createdAt: evalNow,
|
|
167
|
+
updatedAt: evalNow,
|
|
168
|
+
};
|
|
169
|
+
await storage.saveEvaluation(evaluation);
|
|
170
|
+
// Link evaluation to the run
|
|
171
|
+
const linkedRun = {
|
|
172
|
+
...result,
|
|
173
|
+
evaluationId: evaluation.id,
|
|
174
|
+
updatedAt: new Date().toISOString(),
|
|
175
|
+
};
|
|
176
|
+
await storage.saveRun(linkedRun);
|
|
177
|
+
// Broadcast the evaluation ID to SSE subscribers
|
|
178
|
+
broadcastSSE(run.id, 'message', { type: 'evaluation', evaluationId: evaluation.id }, sseSubscribers);
|
|
179
|
+
// Enqueue evaluation execution
|
|
180
|
+
evalQueue.enqueue({
|
|
181
|
+
evaluation,
|
|
182
|
+
execute: async () => {
|
|
183
|
+
const evalCallbacks = {
|
|
184
|
+
onStatusChange(evalStatus) {
|
|
185
|
+
broadcastSSE(run.id, 'message', { type: 'evalStatus', status: evalStatus }, sseSubscribers);
|
|
186
|
+
const updatedEval = {
|
|
187
|
+
...evaluation,
|
|
188
|
+
status: evalStatus,
|
|
189
|
+
updatedAt: new Date().toISOString(),
|
|
190
|
+
};
|
|
191
|
+
storage.saveEvaluation(updatedEval).catch((saveErr) => {
|
|
192
|
+
logger.error('Failed to persist eval status', {
|
|
193
|
+
evalId: evaluation.id,
|
|
194
|
+
error: String(saveErr),
|
|
195
|
+
});
|
|
196
|
+
});
|
|
197
|
+
},
|
|
198
|
+
onProgress() { },
|
|
199
|
+
onMessage() { },
|
|
200
|
+
};
|
|
201
|
+
try {
|
|
202
|
+
const evalResult = await evaluator.evaluateRun(linkedRun, scenario, provider, evalRequest, evalCallbacks);
|
|
203
|
+
const finalEval = {
|
|
204
|
+
...evalResult,
|
|
205
|
+
id: evaluation.id,
|
|
206
|
+
updatedAt: new Date().toISOString(),
|
|
207
|
+
};
|
|
208
|
+
await storage.saveEvaluation(finalEval);
|
|
209
|
+
broadcastSSE(run.id, 'message', { type: 'evalComplete', evaluationId: evaluation.id }, sseSubscribers);
|
|
210
|
+
closeSSE(run.id, sseSubscribers);
|
|
211
|
+
}
|
|
212
|
+
catch (evalErr) {
|
|
213
|
+
logger.error('Auto-evaluation failed', {
|
|
214
|
+
runId: run.id,
|
|
215
|
+
evalId: evaluation.id,
|
|
216
|
+
error: String(evalErr),
|
|
217
|
+
});
|
|
218
|
+
const failedEval = {
|
|
219
|
+
...evaluation,
|
|
220
|
+
status: 'failed',
|
|
221
|
+
updatedAt: new Date().toISOString(),
|
|
222
|
+
};
|
|
223
|
+
await storage.saveEvaluation(failedEval).catch(() => { });
|
|
224
|
+
closeSSE(run.id, sseSubscribers);
|
|
225
|
+
}
|
|
226
|
+
},
|
|
227
|
+
});
|
|
228
|
+
}
|
|
229
|
+
catch (evalSetupErr) {
|
|
230
|
+
logger.error('Failed to set up auto-evaluation', {
|
|
231
|
+
runId: run.id,
|
|
232
|
+
error: String(evalSetupErr),
|
|
233
|
+
});
|
|
234
|
+
// Run itself succeeded — don't fail it because eval setup failed
|
|
235
|
+
}
|
|
236
|
+
}
|
|
237
|
+
}
|
|
238
|
+
catch (err) {
|
|
239
|
+
logger.error('Run execution failed', {
|
|
240
|
+
runId: run.id,
|
|
241
|
+
error: String(err),
|
|
242
|
+
});
|
|
243
|
+
const failedRun = {
|
|
244
|
+
...run,
|
|
245
|
+
status: 'failed',
|
|
246
|
+
error: err instanceof Error ? err.message : String(err),
|
|
247
|
+
updatedAt: new Date().toISOString(),
|
|
248
|
+
};
|
|
249
|
+
await storage.saveRun(failedRun).catch((saveErr) => {
|
|
250
|
+
logger.error('Failed to save failed run', {
|
|
251
|
+
runId: run.id,
|
|
252
|
+
error: String(saveErr),
|
|
253
|
+
});
|
|
254
|
+
});
|
|
255
|
+
}
|
|
256
|
+
finally {
|
|
257
|
+
abortControllers.delete(run.id);
|
|
258
|
+
// Don't close SSE yet if evaluation is running — it will close when eval completes
|
|
259
|
+
if (!reviewerProviderSnapshots || reviewerProviderSnapshots.length === 0) {
|
|
260
|
+
closeSSE(run.id, sseSubscribers);
|
|
261
|
+
}
|
|
262
|
+
}
|
|
263
|
+
},
|
|
264
|
+
});
|
|
265
|
+
res.status(202).json({ id: run.id, status: 'pending' });
|
|
266
|
+
}
|
|
267
|
+
catch (err) {
|
|
268
|
+
logger.error('Failed to start run', { error: String(err) });
|
|
269
|
+
res.status(500).json({ error: 'Failed to start run' });
|
|
270
|
+
}
|
|
271
|
+
});
|
|
272
|
+
// GET / — list runs
|
|
273
|
+
router.get('/', async (req, res) => {
|
|
274
|
+
try {
|
|
275
|
+
const filter = {};
|
|
276
|
+
if (typeof req.query.providerId === 'string')
|
|
277
|
+
filter.providerId = req.query.providerId;
|
|
278
|
+
if (typeof req.query.scenarioId === 'string')
|
|
279
|
+
filter.scenarioId = req.query.scenarioId;
|
|
280
|
+
const runs = await storage.listRuns(filter);
|
|
281
|
+
res.json(runs.map(runSummary));
|
|
282
|
+
}
|
|
283
|
+
catch (err) {
|
|
284
|
+
logger.error('Failed to list runs', { error: String(err) });
|
|
285
|
+
res.status(500).json({ error: 'Failed to list runs' });
|
|
286
|
+
}
|
|
287
|
+
});
|
|
288
|
+
// GET /:id — get full run
|
|
289
|
+
router.get('/:id', async (req, res) => {
|
|
290
|
+
try {
|
|
291
|
+
const id = paramId(req);
|
|
292
|
+
const run = await storage.getRun(id);
|
|
293
|
+
if (!run) {
|
|
294
|
+
res.status(404).json({ error: 'Run not found' });
|
|
295
|
+
return;
|
|
296
|
+
}
|
|
297
|
+
res.json(run);
|
|
298
|
+
}
|
|
299
|
+
catch (err) {
|
|
300
|
+
logger.error('Failed to get run', { id: paramId(req), error: String(err) });
|
|
301
|
+
res.status(500).json({ error: 'Failed to get run' });
|
|
302
|
+
}
|
|
303
|
+
});
|
|
304
|
+
// GET /:id/summary — get run without messages
|
|
305
|
+
router.get('/:id/summary', async (req, res) => {
|
|
306
|
+
try {
|
|
307
|
+
const id = paramId(req);
|
|
308
|
+
const run = await storage.getRun(id);
|
|
309
|
+
if (!run) {
|
|
310
|
+
res.status(404).json({ error: 'Run not found' });
|
|
311
|
+
return;
|
|
312
|
+
}
|
|
313
|
+
res.json(runSummary(run));
|
|
314
|
+
}
|
|
315
|
+
catch (err) {
|
|
316
|
+
logger.error('Failed to get run summary', { id: paramId(req), error: String(err) });
|
|
317
|
+
res.status(500).json({ error: 'Failed to get run summary' });
|
|
318
|
+
}
|
|
319
|
+
});
|
|
320
|
+
// DELETE /:id — delete run (or abort if running)
|
|
321
|
+
router.delete('/:id', async (req, res) => {
|
|
322
|
+
try {
|
|
323
|
+
const id = paramId(req);
|
|
324
|
+
const run = await storage.getRun(id);
|
|
325
|
+
if (!run) {
|
|
326
|
+
res.status(404).json({ error: 'Run not found' });
|
|
327
|
+
return;
|
|
328
|
+
}
|
|
329
|
+
// Abort if running
|
|
330
|
+
const controller = abortControllers.get(id);
|
|
331
|
+
if (controller) {
|
|
332
|
+
controller.abort();
|
|
333
|
+
abortControllers.delete(id);
|
|
334
|
+
}
|
|
335
|
+
const deleted = await storage.deleteRun(id);
|
|
336
|
+
if (!deleted) {
|
|
337
|
+
res.status(404).json({ error: 'Run not found' });
|
|
338
|
+
return;
|
|
339
|
+
}
|
|
340
|
+
res.status(204).send();
|
|
341
|
+
}
|
|
342
|
+
catch (err) {
|
|
343
|
+
logger.error('Failed to delete run', { id: paramId(req), error: String(err) });
|
|
344
|
+
res.status(500).json({ error: 'Failed to delete run' });
|
|
345
|
+
}
|
|
346
|
+
});
|
|
347
|
+
// GET /:id/stream — SSE endpoint
|
|
348
|
+
router.get('/:id/stream', async (req, res) => {
|
|
349
|
+
try {
|
|
350
|
+
const id = paramId(req);
|
|
351
|
+
const run = await storage.getRun(id);
|
|
352
|
+
if (!run) {
|
|
353
|
+
res.status(404).json({ error: 'Run not found' });
|
|
354
|
+
return;
|
|
355
|
+
}
|
|
356
|
+
// If run is already terminal, send final status and close immediately
|
|
357
|
+
const terminalStatuses = ['completed', 'failed', 'cancelled'];
|
|
358
|
+
if (terminalStatuses.includes(run.status)) {
|
|
359
|
+
res.writeHead(200, {
|
|
360
|
+
'Content-Type': 'text/event-stream',
|
|
361
|
+
'Cache-Control': 'no-cache',
|
|
362
|
+
Connection: 'keep-alive',
|
|
363
|
+
'Access-Control-Allow-Origin': '*',
|
|
364
|
+
});
|
|
365
|
+
res.flushHeaders();
|
|
366
|
+
res.write(`event: status\ndata: ${JSON.stringify(run.status)}\n\n`);
|
|
367
|
+
res.write('event: done\ndata: {}\n\n');
|
|
368
|
+
res.end();
|
|
369
|
+
return;
|
|
370
|
+
}
|
|
371
|
+
handleSSEConnection(req, res, id, sseSubscribers);
|
|
372
|
+
}
|
|
373
|
+
catch (err) {
|
|
374
|
+
logger.error('Failed to start SSE stream', { id: paramId(req), error: String(err) });
|
|
375
|
+
res.status(500).json({ error: 'Failed to start stream' });
|
|
376
|
+
}
|
|
377
|
+
});
|
|
378
|
+
return router;
|
|
379
|
+
}
|
|
380
|
+
//# sourceMappingURL=runs.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"runs.js","sourceRoot":"","sources":["../../../../src/server/routes/runs.ts"],"names":[],"mappings":"AAAA,8EAA8E;AAC9E,qDAAqD;AACrD,8EAA8E;AAE9E,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,EAAE,IAAI,MAAM,EAAE,MAAM,MAAM,CAAC;AAOpC,OAAO,EAAE,mBAAmB,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AAE3E,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAE5C,8EAA8E;AAC9E,UAAU;AACV,8EAA8E;AAE9E,SAAS,OAAO,CAAC,GAAY;IAC3B,MAAM,GAAG,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;IAC1B,OAAO,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC;AAC3C,CAAC;AAED,0DAA0D;AAC1D,SAAS,UAAU,CAAC,GAAQ;IAC1B,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,GAAG,IAAI,EAAE,GAAG,GAAG,CAAC;IAC7C,OAAO,IAAI,CAAC;AACd,CAAC;AAED,wGAAwG;AACxG,SAAS,qBAAqB,CAAC,iBAAsC;IACnE,OAAO,iBAAiB,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,EAAE,CAAC,CAAC;QACxC,YAAY,EAAE,CAAC,CAAC,YAAY;QAC5B,KAAK,EAAE,CAAC,CAAC,KAAK;QACd,MAAM,EAAE,CAAC,CAAC,MAAM;QAChB,OAAO,EAAE,CAAC,CAAC,OAAO;QAClB,IAAI,EAAE,GAAG,GAAG,iBAAiB,CAAC,MAAM,GAAG,CAAC;YACtC,CAAC,CAAC,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,aAAa,GAAG,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,WAAW,CAAC;YACvE,CAAC,CAAC,aAAa;KAClB,CAAC,CAAC,CAAC;AACN,CAAC;AAED,sDAAsD;AACtD,OAAO,EAAE,QAAQ,EAAE,MAAM,gBAAgB,CAAC;AAC1C,OAAO,EAAE,QAAQ,EAAE,MAAM,gBAAgB,CAAC;AAE1C,8EAA8E;AAC9E,0DAA0D;AAC1D,8EAA8E;AAE9E,MAAM,gBAAgB,GAAG,IAAI,GAAG,EAA2B,CAAC;AAE5D,8EAA8E;AAC9E,iBAAiB;AACjB,8EAA8E;AAE9E,MAAM,UAAU,eAAe,CAC7B,OAAiB,EACjB,MAAe,EACf,MAAe,EACf,SAAsB,EACtB,KAAgB;IAEhB,MAAM,MAAM,GAAG,MAAM,EAAE,CAAC;IACxB,MAAM,QAAQ,GAAG,KAAK,IAAI,IAAI,QAAQ,CAAC,CAAC,CAAC,CAAC;IAC1C,MAAM,cAAc,GAAqB,IAAI,GAAG,EAAE,CAAC;IAEnD,MAAM,SAAS,GAAG,IAAI,SAAS,CAAC,CAAC,CAAC,CAAC;IAEnC,2BAA2B;IAC3B,MAAM,CAAC,IAAI,CAAC,GAAG,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QACrD,IAAI,CAAC;YACH,MAAM,IAAI,GAAG,GAAG,CAAC,IAA+B,CAAC;YACjD,MAAM,UAAU,GAAG,IAAI,CAAC,UAAgC,CAAC;YACzD,MAAM,UAAU,GAAG,IAAI,CAAC,UAAgC,CAAC;YAEzD,IAAI,CAAC,UAAU,IAAI,CAAC,UAAU,EAAE,CAAC;gBAC/B,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,wCAAwC,EAAE,CAAC,CAAC;gBAC1E,OAAO;YACT,CAAC;YAED,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,WAAW,CAAC,UAAU,CAAC,CAAC;YACvD,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACd,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;gBACtD,OAAO;YACT,CAAC;YAED,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,WAAW,CAAC,UAAU,CAAC,CAAC;YACvD,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACd,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;gBACtD,OAAO;YACT,CAAC;YAED,kDAAkD;YAClD,MAAM,cAAc,GAAG,IAAI,CAAC,mBAAmB,CAAC;YAChD,IAAI,mBAAyC,CAAC;YAC9C,IAAI,yBAAiD,CAAC;YACtD,IAAI,KAAK,CAAC,OAAO,CAAC,cAAc,CAAC,IAAI,cAAc,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC/D,mBAAmB,GAAG,EAAE,CAAC;gBACzB,yBAAyB,GAAG,EAAE,CAAC;gBAC/B,KAAK,MAAM,GAAG,IAAI,cAAc,EAAE,CAAC;oBACjC,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE,CAAC;wBAC5B,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,iDAAiD,EAAE,CAAC,CAAC;wBACnF,OAAO;oBACT,CAAC;oBACD,MAAM,gBAAgB,GAAG,MAAM,OAAO,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;oBACxD,IAAI,CAAC,gBAAgB,EAAE,CAAC;wBACtB,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,gCAAgC,GAAG,EAAE,EAAE,CAAC,CAAC;wBACvE,OAAO;oBACT,CAAC;oBACD,mBAAmB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;oBAC9B,yBAAyB,CAAC,IAAI,CAAC,gBAAgB,CAAC,CAAC;gBACnD,CAAC;YACH,CAAC;YAED,MAAM,aAAa,GAAG,OAAO,IAAI,CAAC,aAAa,KAAK,QAAQ;gBAC1D,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,aAAa,CAAC,CAAC;gBAC9C,CAAC,CAAC,SAAS,CAAC;YAEd,MAAM,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;YACrC,MAAM,GAAG,GAAQ;gBACf,EAAE,EAAE,MAAM,EAAE;gBACZ,UAAU;gBACV,UAAU;gBACV,MAAM,EAAE,SAAS;gBACjB,gBAAgB,EAAE,QAAQ;gBAC1B,gBAAgB,EAAE,QAAQ;gBAC1B,QAAQ,EAAE,EAAE;gBACZ,UAAU,EAAE,EAAE;gBACd,YAAY,EAAE,CAAC;gBACf,UAAU,EAAE,CAAC;gBACb,QAAQ,EAAE,CAAC;gBACX,mBAAmB;gBACnB,yBAAyB;gBACzB,aAAa;gBACb,SAAS,EAAE,GAAG;gBACd,SAAS,EAAE,GAAG;aACf,CAAC;YAEF,MAAM,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;YAE3B,0BAA0B;YAC1B,MAAM,eAAe,GAAG,IAAI,eAAe,EAAE,CAAC;YAC9C,gBAAgB,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,EAAE,eAAe,CAAC,CAAC;YAE9C,QAAQ,CAAC,OAAO,CAAC;gBACf,GAAG;gBACH,OAAO,EAAE,KAAK,IAAI,EAAE;oBAClB,MAAM,SAAS,GAAiB;wBAC9B,SAAS,CAAC,OAAO;4BACf,YAAY,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,YAAY,EAAE,GAAG,OAAO,EAAE,EAAE,cAAc,CAAC,CAAC;wBACtF,CAAC;wBACD,cAAc,CAAC,MAAiB;4BAC9B,MAAM,UAAU,GAAQ;gCACtB,GAAG,GAAG;gCACN,MAAM;gCACN,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;6BACpC,CAAC;4BACF,YAAY,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,EAAE,EAAE,cAAc,CAAC,CAAC;4BAC5E,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,KAAK,CAAC,CAAC,OAAO,EAAE,EAAE;gCAC5C,MAAM,CAAC,KAAK,CAAC,iCAAiC,EAAE;oCAC9C,KAAK,EAAE,GAAG,CAAC,EAAE;oCACb,MAAM;oCACN,KAAK,EAAE,MAAM,CAAC,OAAO,CAAC;iCACvB,CAAC,CAAC;4BACL,CAAC,CAAC,CAAC;wBACL,CAAC;qBACF,CAAC;oBAEF,IAAI,CAAC;wBACH,MAAM,MAAM,GAAG,MAAM,MAAM,CAAC,UAAU,CAAC,QAAQ,EAAE,QAAQ,EAAE,GAAG,EAAE,SAAS,EAAE,eAAe,CAAC,CAAC;wBAC5F,MAAM,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;wBAE9B,gDAAgD;wBAChD,YAAY,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,aAAa,EAAE,GAAG,EAAE,UAAU,CAAC,MAAM,CAAC,EAAE,EAAE,cAAc,CAAC,CAAC;wBAElG,gEAAgE;wBAChE,IAAI,MAAM,CAAC,MAAM,KAAK,WAAW,IAAI,SAAS,IAAI,yBAAyB,IAAI,yBAAyB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;4BACpH,IAAI,CAAC;gCACH,MAAM,gBAAgB,GAAG,qBAAqB,CAAC,yBAAyB,CAAC,CAAC;gCAC1E,MAAM,WAAW,GAAsB;oCACrC,KAAK,EAAE,MAAM,CAAC,EAAE;oCAChB,UAAU,EAAE,gBAAgB;oCAC5B,SAAS,EAAE,aAAa,IAAI,CAAC;iCAC9B,CAAC;gCAEF,MAAM,OAAO,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;gCACzC,MAAM,UAAU,GAAe;oCAC7B,EAAE,EAAE,MAAM,EAAE;oCACZ,KAAK,EAAE,MAAM,CAAC,EAAE;oCAChB,MAAM,EAAE,SAAS;oCACjB,UAAU,EAAE,gBAAgB;oCAC5B,MAAM,EAAE,EAAE;oCACV,gBAAgB,EAAE,EAAE,OAAO,EAAE,KAAK,EAAE,WAAW,EAAE,EAAE,EAAE,UAAU,EAAE,CAAC,EAAE;oCACpE,eAAe,EAAE,EAAE;oCACnB,eAAe,EAAE;wCACf,qBAAqB,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,aAAa,EAAE,EAAE,EAAE,iBAAiB,EAAE,CAAC,EAAE;qCAC/F;oCACD,SAAS,EAAE,EAAE,eAAe,EAAE,EAAE,EAAE,aAAa,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,UAAU,EAAE,EAAE,EAAE;oCACnF,MAAM,EAAE,EAAE;oCACV,YAAY,EAAE,CAAC;oCACf,SAAS,EAAE,OAAO;oCAClB,SAAS,EAAE,OAAO;iCACnB,CAAC;gCAEF,MAAM,OAAO,CAAC,cAAc,CAAC,UAAU,CAAC,CAAC;gCAEzC,6BAA6B;gCAC7B,MAAM,SAAS,GAAQ;oCACrB,GAAG,MAAM;oCACT,YAAY,EAAE,UAAU,CAAC,EAAE;oCAC3B,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;iCACpC,CAAC;gCACF,MAAM,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC;gCAEjC,iDAAiD;gCACjD,YAAY,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,YAAY,EAAE,YAAY,EAAE,UAAU,CAAC,EAAE,EAAE,EAAE,cAAc,CAAC,CAAC;gCAErG,+BAA+B;gCAC/B,SAAS,CAAC,OAAO,CAAC;oCAChB,UAAU;oCACV,OAAO,EAAE,KAAK,IAAI,EAAE;wCAClB,MAAM,aAAa,GAAwB;4CACzC,cAAc,CAAC,UAA4B;gDACzC,YAAY,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,YAAY,EAAE,MAAM,EAAE,UAAU,EAAE,EAAE,cAAc,CAAC,CAAC;gDAC5F,MAAM,WAAW,GAAe;oDAC9B,GAAG,UAAU;oDACb,MAAM,EAAE,UAAU;oDAClB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;iDACpC,CAAC;gDACF,OAAO,CAAC,cAAc,CAAC,WAAW,CAAC,CAAC,KAAK,CAAC,CAAC,OAAO,EAAE,EAAE;oDACpD,MAAM,CAAC,KAAK,CAAC,+BAA+B,EAAE;wDAC5C,MAAM,EAAE,UAAU,CAAC,EAAE;wDACrB,KAAK,EAAE,MAAM,CAAC,OAAO,CAAC;qDACvB,CAAC,CAAC;gDACL,CAAC,CAAC,CAAC;4CACL,CAAC;4CACD,UAAU,KAA4C,CAAC;4CACvD,SAAS,KAAgD,CAAC;yCAC3D,CAAC;wCAEF,IAAI,CAAC;4CACH,MAAM,UAAU,GAAG,MAAM,SAAS,CAAC,WAAW,CAC5C,SAAS,EAAE,QAAQ,EAAE,QAAQ,EAAE,WAAW,EAAE,aAAa,CAC1D,CAAC;4CACF,MAAM,SAAS,GAAe;gDAC5B,GAAG,UAAU;gDACb,EAAE,EAAE,UAAU,CAAC,EAAE;gDACjB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;6CACpC,CAAC;4CACF,MAAM,OAAO,CAAC,cAAc,CAAC,SAAS,CAAC,CAAC;4CACxC,YAAY,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,cAAc,EAAE,YAAY,EAAE,UAAU,CAAC,EAAE,EAAE,EAAE,cAAc,CAAC,CAAC;4CACvG,QAAQ,CAAC,GAAG,CAAC,EAAE,EAAE,cAAc,CAAC,CAAC;wCACnC,CAAC;wCAAC,OAAO,OAAO,EAAE,CAAC;4CACjB,MAAM,CAAC,KAAK,CAAC,wBAAwB,EAAE;gDACrC,KAAK,EAAE,GAAG,CAAC,EAAE;gDACb,MAAM,EAAE,UAAU,CAAC,EAAE;gDACrB,KAAK,EAAE,MAAM,CAAC,OAAO,CAAC;6CACvB,CAAC,CAAC;4CACH,MAAM,UAAU,GAAe;gDAC7B,GAAG,UAAU;gDACb,MAAM,EAAE,QAAQ;gDAChB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;6CACpC,CAAC;4CACF,MAAM,OAAO,CAAC,cAAc,CAAC,UAAU,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;4CACzD,QAAQ,CAAC,GAAG,CAAC,EAAE,EAAE,cAAc,CAAC,CAAC;wCACnC,CAAC;oCACH,CAAC;iCACF,CAAC,CAAC;4BACL,CAAC;4BAAC,OAAO,YAAY,EAAE,CAAC;gCACtB,MAAM,CAAC,KAAK,CAAC,kCAAkC,EAAE;oCAC/C,KAAK,EAAE,GAAG,CAAC,EAAE;oCACb,KAAK,EAAE,MAAM,CAAC,YAAY,CAAC;iCAC5B,CAAC,CAAC;gCACH,iEAAiE;4BACnE,CAAC;wBACH,CAAC;oBACH,CAAC;oBAAC,OAAO,GAAG,EAAE,CAAC;wBACb,MAAM,CAAC,KAAK,CAAC,sBAAsB,EAAE;4BACnC,KAAK,EAAE,GAAG,CAAC,EAAE;4BACb,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC;yBACnB,CAAC,CAAC;wBACH,MAAM,SAAS,GAAQ;4BACrB,GAAG,GAAG;4BACN,MAAM,EAAE,QAAQ;4BAChB,KAAK,EAAE,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC;4BACvD,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;yBACpC,CAAC;wBACF,MAAM,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,KAAK,CAAC,CAAC,OAAO,EAAE,EAAE;4BACjD,MAAM,CAAC,KAAK,CAAC,2BAA2B,EAAE;gCACxC,KAAK,EAAE,GAAG,CAAC,EAAE;gCACb,KAAK,EAAE,MAAM,CAAC,OAAO,CAAC;6BACvB,CAAC,CAAC;wBACL,CAAC,CAAC,CAAC;oBACL,CAAC;4BAAS,CAAC;wBACT,gBAAgB,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;wBAChC,mFAAmF;wBACnF,IAAI,CAAC,yBAAyB,IAAI,yBAAyB,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;4BACzE,QAAQ,CAAC,GAAG,CAAC,EAAE,EAAE,cAAc,CAAC,CAAC;wBACnC,CAAC;oBACH,CAAC;gBACH,CAAC;aACF,CAAC,CAAC;YAEH,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,MAAM,EAAE,SAAS,EAAE,CAAC,CAAC;QAC1D,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,qBAAqB,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC5D,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,qBAAqB,EAAE,CAAC,CAAC;QACzD,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,oBAAoB;IACpB,MAAM,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QACpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAuC,EAAE,CAAC;YACtD,IAAI,OAAO,GAAG,CAAC,KAAK,CAAC,UAAU,KAAK,QAAQ;gBAAE,MAAM,CAAC,UAAU,GAAG,GAAG,CAAC,KAAK,CAAC,UAAU,CAAC;YACvF,IAAI,OAAO,GAAG,CAAC,KAAK,CAAC,UAAU,KAAK,QAAQ;gBAAE,MAAM,CAAC,UAAU,GAAG,GAAG,CAAC,KAAK,CAAC,UAAU,CAAC;YAEvF,MAAM,IAAI,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;YAC5C,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC;QACjC,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,qBAAqB,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC5D,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,qBAAqB,EAAE,CAAC,CAAC;QACzD,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,0BAA0B;IAC1B,MAAM,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QACvD,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YACrC,IAAI,CAAC,GAAG,EAAE,CAAC;gBACT,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC,CAAC;gBACjD,OAAO;YACT,CAAC;YACD,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAChB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,mBAAmB,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC5E,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,mBAAmB,EAAE,CAAC,CAAC;QACvD,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,8CAA8C;IAC9C,MAAM,CAAC,GAAG,CAAC,cAAc,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QAC/D,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YACrC,IAAI,CAAC,GAAG,EAAE,CAAC;gBACT,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC,CAAC;gBACjD,OAAO;YACT,CAAC;YACD,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC;QAC5B,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,2BAA2B,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACpF,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,2BAA2B,EAAE,CAAC,CAAC;QAC/D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,iDAAiD;IACjD,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QAC1D,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YACrC,IAAI,CAAC,GAAG,EAAE,CAAC;gBACT,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC,CAAC;gBACjD,OAAO;YACT,CAAC;YAED,mBAAmB;YACnB,MAAM,UAAU,GAAG,gBAAgB,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC5C,IAAI,UAAU,EAAE,CAAC;gBACf,UAAU,CAAC,KAAK,EAAE,CAAC;gBACnB,gBAAgB,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YAC9B,CAAC;YAED,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC;YAC5C,IAAI,CAAC,OAAO,EAAE,CAAC;gBACb,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC,CAAC;gBACjD,OAAO;YACT,CAAC;YACD,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;QACzB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,sBAAsB,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC/E,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,sBAAsB,EAAE,CAAC,CAAC;QAC1D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,iCAAiC;IACjC,MAAM,CAAC,GAAG,CAAC,aAAa,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QAC9D,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YACrC,IAAI,CAAC,GAAG,EAAE,CAAC;gBACT,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC,CAAC;gBACjD,OAAO;YACT,CAAC;YAED,sEAAsE;YACtE,MAAM,gBAAgB,GAAgB,CAAC,WAAW,EAAE,QAAQ,EAAE,WAAW,CAAC,CAAC;YAC3E,IAAI,gBAAgB,CAAC,QAAQ,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;gBAC1C,GAAG,CAAC,SAAS,CAAC,GAAG,EAAE;oBACjB,cAAc,EAAE,mBAAmB;oBACnC,eAAe,EAAE,UAAU;oBAC3B,UAAU,EAAE,YAAY;oBACxB,6BAA6B,EAAE,GAAG;iBACnC,CAAC,CAAC;gBACH,GAAG,CAAC,YAAY,EAAE,CAAC;gBACnB,GAAG,CAAC,KAAK,CAAC,wBAAwB,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;gBACpE,GAAG,CAAC,KAAK,CAAC,2BAA2B,CAAC,CAAC;gBACvC,GAAG,CAAC,GAAG,EAAE,CAAC;gBACV,OAAO;YACT,CAAC;YAED,mBAAmB,CAAC,GAAG,EAAE,GAAG,EAAE,EAAE,EAAE,cAAc,CAAC,CAAC;QACpD,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,4BAA4B,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACrF,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,wBAAwB,EAAE,CAAC,CAAC;QAC5D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,OAAO,MAAM,CAAC;AAChB,CAAC"}
|
|
@@ -0,0 +1,5 @@
|
|
|
1
|
+
import { Router } from 'express';
|
|
2
|
+
import type { IStorage } from '../interfaces/storage.js';
|
|
3
|
+
import type { ILogger } from '../interfaces/logger.js';
|
|
4
|
+
export declare function createScenarioRoutes(storage: IStorage, logger: ILogger): Router;
|
|
5
|
+
//# sourceMappingURL=scenarios.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"scenarios.d.ts","sourceRoot":"","sources":["../../../../src/server/routes/scenarios.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AAGjC,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,0BAA0B,CAAC;AACzD,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,yBAAyB,CAAC;AAuEvD,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,GAAG,MAAM,CA0H/E"}
|
|
@@ -0,0 +1,181 @@
|
|
|
1
|
+
import { Router } from 'express';
|
|
2
|
+
import { v4 as uuidv4 } from 'uuid';
|
|
3
|
+
// ─── Helpers ───────────────────────────────────────────────────────────
|
|
4
|
+
function paramId(req) {
|
|
5
|
+
const raw = req.params.id;
|
|
6
|
+
return Array.isArray(raw) ? raw[0] : raw;
|
|
7
|
+
}
|
|
8
|
+
const VALID_CATEGORIES = [
|
|
9
|
+
'planning', 'instruction-following', 'reasoning', 'tool-strategy',
|
|
10
|
+
'error-handling', 'ambiguity-handling', 'scope-management', 'custom',
|
|
11
|
+
];
|
|
12
|
+
function scenarioMetadata(s) {
|
|
13
|
+
return {
|
|
14
|
+
id: s.id,
|
|
15
|
+
name: s.name,
|
|
16
|
+
category: s.category,
|
|
17
|
+
createdAt: s.createdAt,
|
|
18
|
+
};
|
|
19
|
+
}
|
|
20
|
+
function validateScoringWeights(dimensions) {
|
|
21
|
+
if (dimensions.length === 0)
|
|
22
|
+
return true;
|
|
23
|
+
const sum = dimensions.reduce((acc, d) => acc + d.weight, 0);
|
|
24
|
+
return Math.abs(sum - 1.0) < 0.001;
|
|
25
|
+
}
|
|
26
|
+
function validateScenarioBody(body) {
|
|
27
|
+
const errors = [];
|
|
28
|
+
if (typeof body !== 'object' || body === null) {
|
|
29
|
+
errors.push({ field: 'body', message: 'Request body must be a JSON object' });
|
|
30
|
+
return errors;
|
|
31
|
+
}
|
|
32
|
+
const b = body;
|
|
33
|
+
if (typeof b.name !== 'string' || b.name.trim().length === 0) {
|
|
34
|
+
errors.push({ field: 'name', message: 'name is required' });
|
|
35
|
+
}
|
|
36
|
+
if (typeof b.prompt !== 'string' || b.prompt.trim().length === 0) {
|
|
37
|
+
errors.push({ field: 'prompt', message: 'prompt is required' });
|
|
38
|
+
}
|
|
39
|
+
if (typeof b.category !== 'string' || !VALID_CATEGORIES.includes(b.category)) {
|
|
40
|
+
errors.push({
|
|
41
|
+
field: 'category',
|
|
42
|
+
message: `category must be one of: ${VALID_CATEGORIES.join(', ')}`,
|
|
43
|
+
});
|
|
44
|
+
}
|
|
45
|
+
if (b.enabledTools !== undefined && !Array.isArray(b.enabledTools)) {
|
|
46
|
+
errors.push({ field: 'enabledTools', message: 'enabledTools must be an array of strings' });
|
|
47
|
+
}
|
|
48
|
+
if (Array.isArray(b.scoringDimensions) && b.scoringDimensions.length > 0) {
|
|
49
|
+
if (!validateScoringWeights(b.scoringDimensions)) {
|
|
50
|
+
errors.push({
|
|
51
|
+
field: 'scoringDimensions',
|
|
52
|
+
message: 'scoringDimensions weights must sum to 1.0',
|
|
53
|
+
});
|
|
54
|
+
}
|
|
55
|
+
}
|
|
56
|
+
return errors;
|
|
57
|
+
}
|
|
58
|
+
// ─── Router factory ────────────────────────────────────────────────────
|
|
59
|
+
export function createScenarioRoutes(storage, logger) {
|
|
60
|
+
const router = Router();
|
|
61
|
+
// GET / — list all scenarios (metadata only)
|
|
62
|
+
router.get('/', async (_req, res) => {
|
|
63
|
+
try {
|
|
64
|
+
const scenarios = await storage.listScenarios();
|
|
65
|
+
res.json(scenarios.map(scenarioMetadata));
|
|
66
|
+
}
|
|
67
|
+
catch (err) {
|
|
68
|
+
logger.error('Failed to list scenarios', { error: String(err) });
|
|
69
|
+
res.status(500).json({ error: 'Failed to list scenarios' });
|
|
70
|
+
}
|
|
71
|
+
});
|
|
72
|
+
// GET /:id — get full scenario
|
|
73
|
+
router.get('/:id', async (req, res) => {
|
|
74
|
+
try {
|
|
75
|
+
const id = paramId(req);
|
|
76
|
+
const scenario = await storage.getScenario(id);
|
|
77
|
+
if (!scenario) {
|
|
78
|
+
res.status(404).json({ error: 'Scenario not found' });
|
|
79
|
+
return;
|
|
80
|
+
}
|
|
81
|
+
res.json(scenario);
|
|
82
|
+
}
|
|
83
|
+
catch (err) {
|
|
84
|
+
logger.error('Failed to get scenario', { id: paramId(req), error: String(err) });
|
|
85
|
+
res.status(500).json({ error: 'Failed to get scenario' });
|
|
86
|
+
}
|
|
87
|
+
});
|
|
88
|
+
// POST / — create custom scenario
|
|
89
|
+
router.post('/', async (req, res) => {
|
|
90
|
+
const errors = validateScenarioBody(req.body);
|
|
91
|
+
if (errors.length > 0) {
|
|
92
|
+
res.status(400).json({ errors });
|
|
93
|
+
return;
|
|
94
|
+
}
|
|
95
|
+
try {
|
|
96
|
+
const body = req.body;
|
|
97
|
+
const now = new Date().toISOString();
|
|
98
|
+
const scenario = {
|
|
99
|
+
id: uuidv4(),
|
|
100
|
+
name: body.name.trim(),
|
|
101
|
+
category: body.category,
|
|
102
|
+
prompt: body.prompt.trim(),
|
|
103
|
+
systemPrompt: typeof body.systemPrompt === 'string' ? body.systemPrompt : '',
|
|
104
|
+
enabledTools: Array.isArray(body.enabledTools) ? body.enabledTools : [],
|
|
105
|
+
expectedAnswer: typeof body.expectedAnswer === 'string' ? body.expectedAnswer : '',
|
|
106
|
+
criticalRequirements: Array.isArray(body.criticalRequirements) ? body.criticalRequirements : [],
|
|
107
|
+
gradingGuidelines: typeof body.gradingGuidelines === 'string' ? body.gradingGuidelines : '',
|
|
108
|
+
scoringDimensions: Array.isArray(body.scoringDimensions) ? body.scoringDimensions : [],
|
|
109
|
+
createdAt: now,
|
|
110
|
+
updatedAt: now,
|
|
111
|
+
};
|
|
112
|
+
await storage.saveScenario(scenario);
|
|
113
|
+
res.status(201).json(scenario);
|
|
114
|
+
}
|
|
115
|
+
catch (err) {
|
|
116
|
+
logger.error('Failed to create scenario', { error: String(err) });
|
|
117
|
+
res.status(500).json({ error: 'Failed to create scenario' });
|
|
118
|
+
}
|
|
119
|
+
});
|
|
120
|
+
// PUT /:id — update custom scenario
|
|
121
|
+
router.put('/:id', async (req, res) => {
|
|
122
|
+
try {
|
|
123
|
+
const id = paramId(req);
|
|
124
|
+
const existing = await storage.getScenario(id);
|
|
125
|
+
if (!existing) {
|
|
126
|
+
res.status(404).json({ error: 'Scenario not found' });
|
|
127
|
+
return;
|
|
128
|
+
}
|
|
129
|
+
const errors = validateScenarioBody(req.body);
|
|
130
|
+
if (errors.length > 0) {
|
|
131
|
+
res.status(400).json({ errors });
|
|
132
|
+
return;
|
|
133
|
+
}
|
|
134
|
+
const body = req.body;
|
|
135
|
+
const updated = {
|
|
136
|
+
...existing,
|
|
137
|
+
name: body.name.trim(),
|
|
138
|
+
category: body.category,
|
|
139
|
+
prompt: body.prompt.trim(),
|
|
140
|
+
systemPrompt: typeof body.systemPrompt === 'string' ? body.systemPrompt : existing.systemPrompt,
|
|
141
|
+
enabledTools: Array.isArray(body.enabledTools) ? body.enabledTools : existing.enabledTools,
|
|
142
|
+
expectedAnswer: typeof body.expectedAnswer === 'string' ? body.expectedAnswer : existing.expectedAnswer,
|
|
143
|
+
criticalRequirements: Array.isArray(body.criticalRequirements)
|
|
144
|
+
? body.criticalRequirements
|
|
145
|
+
: existing.criticalRequirements,
|
|
146
|
+
gradingGuidelines: typeof body.gradingGuidelines === 'string'
|
|
147
|
+
? body.gradingGuidelines
|
|
148
|
+
: existing.gradingGuidelines,
|
|
149
|
+
scoringDimensions: Array.isArray(body.scoringDimensions)
|
|
150
|
+
? body.scoringDimensions
|
|
151
|
+
: existing.scoringDimensions,
|
|
152
|
+
updatedAt: new Date().toISOString(),
|
|
153
|
+
};
|
|
154
|
+
await storage.saveScenario(updated);
|
|
155
|
+
res.json(updated);
|
|
156
|
+
}
|
|
157
|
+
catch (err) {
|
|
158
|
+
logger.error('Failed to update scenario', { id: paramId(req), error: String(err) });
|
|
159
|
+
res.status(500).json({ error: 'Failed to update scenario' });
|
|
160
|
+
}
|
|
161
|
+
});
|
|
162
|
+
// DELETE /:id — delete custom scenario
|
|
163
|
+
router.delete('/:id', async (req, res) => {
|
|
164
|
+
try {
|
|
165
|
+
const id = paramId(req);
|
|
166
|
+
const existing = await storage.getScenario(id);
|
|
167
|
+
if (!existing) {
|
|
168
|
+
res.status(404).json({ error: 'Scenario not found' });
|
|
169
|
+
return;
|
|
170
|
+
}
|
|
171
|
+
await storage.deleteScenario(id);
|
|
172
|
+
res.status(204).send();
|
|
173
|
+
}
|
|
174
|
+
catch (err) {
|
|
175
|
+
logger.error('Failed to delete scenario', { id: paramId(req), error: String(err) });
|
|
176
|
+
res.status(500).json({ error: 'Failed to delete scenario' });
|
|
177
|
+
}
|
|
178
|
+
});
|
|
179
|
+
return router;
|
|
180
|
+
}
|
|
181
|
+
//# sourceMappingURL=scenarios.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"scenarios.js","sourceRoot":"","sources":["../../../../src/server/routes/scenarios.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,EAAE,IAAI,MAAM,EAAE,MAAM,MAAM,CAAC;AAMpC,0EAA0E;AAE1E,SAAS,OAAO,CAAC,GAAY;IAC3B,MAAM,GAAG,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;IAC1B,OAAO,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC;AAC3C,CAAC;AAED,MAAM,gBAAgB,GAAgC;IACpD,UAAU,EAAE,uBAAuB,EAAE,WAAW,EAAE,eAAe;IACjE,gBAAgB,EAAE,oBAAoB,EAAE,kBAAkB,EAAE,QAAQ;CACrE,CAAC;AAOF,SAAS,gBAAgB,CAAC,CAAW;IACnC,OAAO;QACL,EAAE,EAAE,CAAC,CAAC,EAAE;QACR,IAAI,EAAE,CAAC,CAAC,IAAI;QACZ,QAAQ,EAAE,CAAC,CAAC,QAAQ;QACpB,SAAS,EAAE,CAAC,CAAC,SAAS;KACvB,CAAC;AACJ,CAAC;AAED,SAAS,sBAAsB,CAAC,UAAuC;IACrE,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACzC,MAAM,GAAG,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;IAC7D,OAAO,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG,KAAK,CAAC;AACrC,CAAC;AAED,SAAS,oBAAoB,CAAC,IAAa;IACzC,MAAM,MAAM,GAAsB,EAAE,CAAC;IACrC,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;QAC9C,MAAM,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,oCAAoC,EAAE,CAAC,CAAC;QAC9E,OAAO,MAAM,CAAC;IAChB,CAAC;IACD,MAAM,CAAC,GAAG,IAA+B,CAAC;IAE1C,IAAI,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC7D,MAAM,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,kBAAkB,EAAE,CAAC,CAAC;IAC9D,CAAC;IACD,IAAI,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACjE,MAAM,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,QAAQ,EAAE,OAAO,EAAE,oBAAoB,EAAE,CAAC,CAAC;IAClE,CAAC;IACD,IAAI,OAAO,CAAC,CAAC,QAAQ,KAAK,QAAQ,IAAI,CAAC,gBAAgB,CAAC,QAAQ,CAAC,CAAC,CAAC,QAA4B,CAAC,EAAE,CAAC;QACjG,MAAM,CAAC,IAAI,CAAC;YACV,KAAK,EAAE,UAAU;YACjB,OAAO,EAAE,4BAA4B,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE;SACnE,CAAC,CAAC;IACL,CAAC;IACD,IAAI,CAAC,CAAC,YAAY,KAAK,SAAS,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,YAAY,CAAC,EAAE,CAAC;QACnE,MAAM,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,cAAc,EAAE,OAAO,EAAE,0CAA0C,EAAE,CAAC,CAAC;IAC9F,CAAC;IACD,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,iBAAiB,CAAC,IAAI,CAAC,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACzE,IAAI,CAAC,sBAAsB,CAAC,CAAC,CAAC,iBAAuC,CAAC,EAAE,CAAC;YACvE,MAAM,CAAC,IAAI,CAAC;gBACV,KAAK,EAAE,mBAAmB;gBAC1B,OAAO,EAAE,2CAA2C;aACrD,CAAC,CAAC;QACL,CAAC;IACH,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,0EAA0E;AAE1E,MAAM,UAAU,oBAAoB,CAAC,OAAiB,EAAE,MAAe;IACrE,MAAM,MAAM,GAAG,MAAM,EAAE,CAAC;IAExB,6CAA6C;IAC7C,MAAM,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,EAAE,IAAa,EAAE,GAAa,EAAE,EAAE;QACrD,IAAI,CAAC;YACH,MAAM,SAAS,GAAG,MAAM,OAAO,CAAC,aAAa,EAAE,CAAC;YAChD,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,gBAAgB,CAAC,CAAC,CAAC;QAC5C,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,0BAA0B,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACjE,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,0BAA0B,EAAE,CAAC,CAAC;QAC9D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,+BAA+B;IAC/B,MAAM,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QACvD,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;YAC/C,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACd,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;gBACtD,OAAO;YACT,CAAC;YACD,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACrB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,wBAAwB,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACjF,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,wBAAwB,EAAE,CAAC,CAAC;QAC5D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,kCAAkC;IAClC,MAAM,CAAC,IAAI,CAAC,GAAG,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QACrD,MAAM,MAAM,GAAG,oBAAoB,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAC9C,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACtB,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC;YACjC,OAAO;QACT,CAAC;QACD,IAAI,CAAC;YACH,MAAM,IAAI,GAAG,GAAG,CAAC,IAA+B,CAAC;YACjD,MAAM,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;YACrC,MAAM,QAAQ,GAAa;gBACzB,EAAE,EAAE,MAAM,EAAE;gBACZ,IAAI,EAAG,IAAI,CAAC,IAAe,CAAC,IAAI,EAAE;gBAClC,QAAQ,EAAE,IAAI,CAAC,QAA4B;gBAC3C,MAAM,EAAG,IAAI,CAAC,MAAiB,CAAC,IAAI,EAAE;gBACtC,YAAY,EAAE,OAAO,IAAI,CAAC,YAAY,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE;gBAC5E,YAAY,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE;gBACvE,cAAc,EAAE,OAAO,IAAI,CAAC,cAAc,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE;gBAClF,oBAAoB,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAC,CAAC,EAAE;gBAC/F,iBAAiB,EAAE,OAAO,IAAI,CAAC,iBAAiB,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE;gBAC3F,iBAAiB,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,iBAAiB,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE;gBACtF,SAAS,EAAE,GAAG;gBACd,SAAS,EAAE,GAAG;aACf,CAAC;YACF,MAAM,OAAO,CAAC,YAAY,CAAC,QAAQ,CAAC,CAAC;YACrC,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACjC,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,2BAA2B,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAClE,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,2BAA2B,EAAE,CAAC,CAAC;QAC/D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,oCAAoC;IACpC,MAAM,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QACvD,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;YAC/C,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACd,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;gBACtD,OAAO;YACT,CAAC;YACD,MAAM,MAAM,GAAG,oBAAoB,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;YAC9C,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACtB,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC;gBACjC,OAAO;YACT,CAAC;YACD,MAAM,IAAI,GAAG,GAAG,CAAC,IAA+B,CAAC;YACjD,MAAM,OAAO,GAAa;gBACxB,GAAG,QAAQ;gBACX,IAAI,EAAG,IAAI,CAAC,IAAe,CAAC,IAAI,EAAE;gBAClC,QAAQ,EAAE,IAAI,CAAC,QAA4B;gBAC3C,MAAM,EAAG,IAAI,CAAC,MAAiB,CAAC,IAAI,EAAE;gBACtC,YAAY,EAAE,OAAO,IAAI,CAAC,YAAY,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,QAAQ,CAAC,YAAY;gBAC/F,YAAY,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,QAAQ,CAAC,YAAY;gBAC1F,cAAc,EAAE,OAAO,IAAI,CAAC,cAAc,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,QAAQ,CAAC,cAAc;gBACvG,oBAAoB,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,oBAAoB,CAAC;oBAC5D,CAAC,CAAC,IAAI,CAAC,oBAAoB;oBAC3B,CAAC,CAAC,QAAQ,CAAC,oBAAoB;gBACjC,iBAAiB,EAAE,OAAO,IAAI,CAAC,iBAAiB,KAAK,QAAQ;oBAC3D,CAAC,CAAC,IAAI,CAAC,iBAAiB;oBACxB,CAAC,CAAC,QAAQ,CAAC,iBAAiB;gBAC9B,iBAAiB,EAAE,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,iBAAiB,CAAC;oBACtD,CAAC,CAAC,IAAI,CAAC,iBAAiB;oBACxB,CAAC,CAAC,QAAQ,CAAC,iBAAiB;gBAC9B,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;aACpC,CAAC;YACF,MAAM,OAAO,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC;YACpC,GAAG,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;QACpB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,2BAA2B,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACpF,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,2BAA2B,EAAE,CAAC,CAAC;QAC/D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,uCAAuC;IACvC,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE,KAAK,EAAE,GAAY,EAAE,GAAa,EAAE,EAAE;QAC1D,IAAI,CAAC;YACH,MAAM,EAAE,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;YACxB,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;YAC/C,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACd,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;gBACtD,OAAO;YACT,CAAC;YACD,MAAM,OAAO,CAAC,cAAc,CAAC,EAAE,CAAC,CAAC;YACjC,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;QACzB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,KAAK,CAAC,2BAA2B,EAAE,EAAE,EAAE,EAAE,OAAO,CAAC,GAAG,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACpF,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,2BAA2B,EAAE,CAAC,CAAC;QAC/D,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,OAAO,MAAM,CAAC;AAChB,CAAC"}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
import type { Scenario, AnswerComparison, CriticalPartResult, InstructionCompliance } from '../types/index.js';
|
|
2
|
+
import type { ScoreParseResult, ComplianceParseResult } from './eval-parsers.js';
|
|
3
|
+
/** Shared state accumulated per evaluator across rounds. */
|
|
4
|
+
export interface EvaluatorAccumulator {
|
|
5
|
+
role: string;
|
|
6
|
+
costUsd: number;
|
|
7
|
+
tokensIn: number;
|
|
8
|
+
tokensOut: number;
|
|
9
|
+
rounds: number;
|
|
10
|
+
scoreResult: Partial<ScoreParseResult>;
|
|
11
|
+
complianceResult: Partial<ComplianceParseResult>;
|
|
12
|
+
assessmentText: string;
|
|
13
|
+
}
|
|
14
|
+
/** Check whether evaluator scores have converged (within 1 point per dimension). */
|
|
15
|
+
export declare function checkConsensus(accumulators: readonly EvaluatorAccumulator[]): boolean;
|
|
16
|
+
/** Aggregate answer closeness from all evaluators into a comparison. */
|
|
17
|
+
export declare function buildAnswerComparison(accumulators: readonly EvaluatorAccumulator[]): AnswerComparison;
|
|
18
|
+
/** Build critical requirement results by checking evaluator-flagged misses. */
|
|
19
|
+
export declare function buildCriticalResults(accumulators: readonly EvaluatorAccumulator[], scenario: Scenario): CriticalPartResult[];
|
|
20
|
+
/** Merge compliance results from all evaluators into a single report. */
|
|
21
|
+
export declare function mergeCompliance(accumulators: readonly EvaluatorAccumulator[]): InstructionCompliance;
|
|
22
|
+
//# sourceMappingURL=eval-helpers.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"eval-helpers.d.ts","sourceRoot":"","sources":["../../../../src/server/services/eval-helpers.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EACV,QAAQ,EACR,gBAAgB,EAChB,kBAAkB,EAClB,qBAAqB,EACtB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,gBAAgB,EAAE,qBAAqB,EAAE,MAAM,mBAAmB,CAAC;AAEjF,4DAA4D;AAC5D,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,EAAE,MAAM,CAAC;IACjB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACvC,gBAAgB,EAAE,OAAO,CAAC,qBAAqB,CAAC,CAAC;IACjD,cAAc,EAAE,MAAM,CAAC;CACxB;AAED,oFAAoF;AACpF,wBAAgB,cAAc,CAAC,YAAY,EAAE,SAAS,oBAAoB,EAAE,GAAG,OAAO,CAerF;AAED,wEAAwE;AACxE,wBAAgB,qBAAqB,CACnC,YAAY,EAAE,SAAS,oBAAoB,EAAE,GAC5C,gBAAgB,CAgBlB;AAED,+EAA+E;AAC/E,wBAAgB,oBAAoB,CAClC,YAAY,EAAE,SAAS,oBAAoB,EAAE,EAC7C,QAAQ,EAAE,QAAQ,GACjB,kBAAkB,EAAE,CAStB;AAED,yEAAyE;AACzE,wBAAgB,eAAe,CAC7B,YAAY,EAAE,SAAS,oBAAoB,EAAE,GAC5C,qBAAqB,CAwBvB"}
|