d

.env.example

@@ -1,3 +1,5 @@
 DISCORD_TOKEN=
 GEMINI_API_KEY=
 GEMINI_MODEL=
+# Opcional: se usa si GEMINI_MODEL sigue saturado después de los reintentos
+GEMINI_FALLBACK_MODEL=

src/gemini.ts

@@ -1,9 +1,18 @@
-import { GoogleGenAI, FileState, createPartFromUri, createUserContent, type File } from '@google/genai';
+import {
+  ApiError,
+  FileState,
+  GoogleGenAI,
+  createPartFromUri,
+  createUserContent,
+  type Content,
+  type File,
+} from '@google/genai';
 import type { AudioTrack } from './recorder.js';
 import { buildPrompt } from './prompt.js';
 
 const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });
 export const MODEL = process.env.GEMINI_MODEL!;
+export const FALLBACK_MODEL = process.env.GEMINI_FALLBACK_MODEL;
 
 const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
 
@@ -17,6 +26,30 @@ async function upload(track: AudioTrack): Promise<File> {
   return file;
 }
 
+// 429/500/503 suelen ser picos de demanda: reintenta y, si hay GEMINI_FALLBACK_MODEL, prueba con ese
+const RETRY_DELAYS = [5_000, 15_000, 30_000];
+
+async function generateWithRetry(contents: Content) {
+  const models = [MODEL, FALLBACK_MODEL].filter((m): m is string => !!m);
+  let lastErr: unknown;
+  for (const model of models) {
+    for (let attempt = 0; attempt <= RETRY_DELAYS.length; attempt++) {
+      try {
+        return await ai.models.generateContent({ model, contents });
+      } catch (err) {
+        lastErr = err;
+        const status = err instanceof ApiError ? err.status : 0;
+        if (![429, 500, 503].includes(status)) throw err;
+        const delay = RETRY_DELAYS[attempt];
+        if (delay === undefined) break;
+        console.warn(`${model} respondió ${status}, reintento en ${delay / 1000}s`);
+        await sleep(delay);
+      }
+    }
+  }
+  throw lastErr;
+}
+
 export async function generateNotes(tracks: AudioTrack[], startedAt: Date): Promise<string> {
   const files: File[] = [];
   try {
@@ -26,10 +59,8 @@ export async function generateNotes(tracks: AudioTrack[], startedAt: Date): Prom
       `Pista ${i + 1}: ${track.name}`,
       createPartFromUri(files[i].uri!, files[i].mimeType!),
     ]);
-    const res = await ai.models.generateContent({
+    const contents = createUserContent([buildPrompt(tracks.map((t) => t.name), startedAt), ...parts]);
-      model: MODEL,
+    const res = await generateWithRetry(contents);
-      contents: createUserContent([buildPrompt(tracks.map((t) => t.name), startedAt), ...parts]),
-    });
     const text = res.text?.trim();
     if (!text) throw new Error('Gemini devolvió una respuesta vacía');
     return text.replace(/^```(?:markdown|md)?\n([\s\S]*)\n```$/, '$1');

src/index.ts

@@ -9,7 +9,7 @@ import {
 } from 'discord.js';
 import { VoiceConnectionStatus, entersState } from '@discordjs/voice';
 import { Recording } from './recorder.js';
-import { MODEL, generateNotes } from './gemini.js';
+import { FALLBACK_MODEL, MODEL, generateNotes } from './gemini.js';
 
 for (const key of ['DISCORD_TOKEN', 'GEMINI_API_KEY', 'GEMINI_MODEL']) {
   if (!process.env[key]) throw new Error(`Falta la variable de entorno ${key}`);
@@ -62,7 +62,7 @@ async function autoFinish(guildId: string, reason: string) {
 
 client.once(Events.ClientReady, async (c) => {
   await c.application.commands.set(commands.map((cmd) => cmd.toJSON()));
-  console.log(`ready as ${c.user.tag} | modelo: ${MODEL}`);
+  console.log(`ready as ${c.user.tag} | modelo: ${MODEL} | respaldo: ${FALLBACK_MODEL ?? '(ninguno)'}`);
 });
 
 client.on(Events.InteractionCreate, async (interaction) => {