Package v0.2.12→v0.2.124: manifest, actions, version graph

startos/versions/v0.2.13.ts

@@ -0,0 +1,18 @@
+import { VersionInfo } from '@start9labs/start-sdk'
+
+export const v_0_2_13 = VersionInfo.of({
+  version: '0.2.13:2',
+  releaseNotes: {
+    en_US:
+      'Dashboard v3 — backend-agnostic performance benchmarks and per-model failure analysis.\n\n' +
+      '• New "Transcription speed by model (normalized)" table — reports wall-clock processing time per minute of audio (ms/min). Lets the operator compare e.g. Gemini Flash vs. Parakeet on a like-for-like benchmark: if Gemini takes 2.5 s/min and Parakeet takes 0.8 s/min on the same audio, the difference is now visible at a glance.\n\n' +
+      '• New "Analysis speed by model" table — reports wall-clock per 1000 input tokens, since analyze calls work on text not audio. Same like-for-like comparison story between Gemini Pro analysis vs. Gemma analysis.\n\n' +
+      '• Per-model success rate, computed against attempted calls (success + error). Refused calls are excluded — they never reached the backend, so they shouldn\'t count against the model\'s reliability.\n\n' +
+      '• New "Top failure modes by model" table — top 3 normalized error signatures per model with occurrence counts. Tells the operator at a glance whether a model is failing for one specific reason (Gemini quota, Parakeet timeout) or distributed noise. Error strings are normalized — timestamps, UUIDs, ports, hex IDs collapse to placeholders so near-identical errors bucket together.\n\n' +
+      '• Transcribe routes now record `audio_seconds` (via ffprobe) alongside every audit entry, which is the data the new normalized benchmarks consume. Audit log records grow by ~15 bytes per row.',
+  },
+  migrations: {
+    up: async ({ effects }) => {},
+    down: async ({ effects }) => {},
+  },
+})