diff --git a/agents/src/inference/tts_pool_drain_reuse.test.ts b/agents/src/inference/tts_pool_drain_reuse.test.ts
new file mode 100644
index 000000000..a356c4532
--- /dev/null
+++ b/agents/src/inference/tts_pool_drain_reuse.test.ts
@@ -0,0 +1,199 @@
+// SPDX-FileCopyrightText: 2026 LiveKit, Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+import type { AddressInfo } from 'node:net';
+import { afterEach, describe, expect, it } from 'vitest';
+import { WebSocketServer } from 'ws';
+import type { WebSocket as WsSocket } from 'ws';
+import { initializeLogger } from '../log.js';
+import { TTS } from './tts.js';
+
+/**
+ * Treating a `done` as a generation boundary means a run now ends on a quiet timeout rather
+ * than on the `done` itself, and returning the socket to the pool is gated on having reached
+ * that end. Those two rules meet in one place, so the pooled socket is either kept on every
+ * reply or on none, and nothing else pins which.
+ *
+ * These tests pin both halves as a count of gateway handshakes: a conversation of clean
+ * multi-generation replies must dial once, and a session dropped mid-synthesis must dial
+ * again rather than hand its socket to the next reply.
+ */
+
+initializeLogger({ pretty: false });
+
+const SAMPLE_RATE = 16000;
+/** `AudioByteStream` frames at 100ms; keep the gateway's chunks frame-aligned. */
+const CHUNK_MS = 100;
+const SAMPLES_PER_CHUNK = (SAMPLE_RATE * CHUNK_MS) / 1000;
+
+const CHUNKS_PER_GENERATION = 5;
+const GENERATION_AUDIO_MS = CHUNKS_PER_GENERATION * CHUNK_MS;
+/** Generations per flush, matching the several `done` events the gateway answers one with. */
+const GENERATIONS_PER_REPLY = 3;
+const GENERATION_GAP_MS = 150;
+
+const REPLIES = 5;
+
+function audioEvent(sessionId: string, sample: number): string {
+  const pcm = Buffer.alloc(SAMPLES_PER_CHUNK * 2);
+  for (let i = 0; i < SAMPLES_PER_CHUNK; i++) {
+    pcm.writeInt16LE(sample, i * 2);
+  }
+  return JSON.stringify({
+    type: 'output_audio',
+    session_id: sessionId,
+    audio: pcm.toString('base64'),
+  });
+}
+
+const sleep = (ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms));
+
+/**
+ * Gateway stand-in. `dropMidSynthesis` reproduces a session torn down while synthesis is
+ * still outstanding: a prefix, then `session.closed` with no `done` to close the flush.
+ * Otherwise every flush is served in full and the session then goes quiet, which is what a
+ * probe of `inworld/inworld-tts-2` showed — several `done` events, the last one final.
+ */
+async function startFakeGateway(options: { dropMidSynthesis?: boolean } = {}) {
+  const wss = new WebSocketServer({ port: 0, host: '127.0.0.1' });
+  await new Promise<void>((resolve) => wss.once('listening', () => resolve()));
+
+  const sockets: WsSocket[] = [];
+  let connections = 0;
+  let flushes = 0;
+
+  wss.on('connection', (ws: WsSocket) => {
+    sockets.push(ws);
+    const sessionId = `session-${++connections}`;
+
+    const send = (payload: string) => {
+      if (ws.readyState === ws.OPEN) ws.send(payload);
+    };
+
+    ws.on('message', (raw: Buffer) => {
+      const event = JSON.parse(raw.toString()) as { type: string };
+      if (event.type === 'session.create') {
+        send(JSON.stringify({ type: 'session.created', session_id: sessionId }));
+        return;
+      }
+      if (event.type !== 'session.flush') return;
+
+      // Each reply carries a distinct constant sample so a frame can be attributed to it.
+      const sample = 1000 + flushes++;
+
+      void (async () => {
+        if (options.dropMidSynthesis) {
+          for (let i = 0; i < CHUNKS_PER_GENERATION; i++) send(audioEvent(sessionId, sample));
+          send(JSON.stringify({ type: 'session.closed', session_id: sessionId }));
+          return;
+        }
+
+        for (let generation = 0; generation < GENERATIONS_PER_REPLY; generation++) {
+          if (generation > 0) await sleep(GENERATION_GAP_MS);
+          for (let i = 0; i < CHUNKS_PER_GENERATION; i++) send(audioEvent(sessionId, sample));
+          send(JSON.stringify({ type: 'done', session_id: sessionId }));
+        }
+      })();
+    });
+  });
+
+  const { port } = wss.address() as AddressInfo;
+  return {
+    baseURL: `http://127.0.0.1:${port}/v1`,
+    close: () => {
+      for (const socket of sockets) socket.terminate();
+      return new Promise<void>((resolve) => wss.close(() => resolve()));
+    },
+  };
+}
+
+/** A TTS whose gateway handshakes are counted, so pool misses are observable. */
+function createCountedTTS(baseURL: string, connOptions?: { maxRetry: number }) {
+  const tts = new TTS({
+    model: 'inworld/inworld-tts-2',
+    voice: 'Sarah',
+    sampleRate: SAMPLE_RATE,
+    baseURL,
+    apiKey: 'devkey',
+    apiSecret: 'secret'.padEnd(32, 'x'),
+    connOptions: connOptions
+      ? { maxRetry: connOptions.maxRetry, retryIntervalMs: 0, timeoutMs: 5_000 }
+      : undefined,
+  });
+
+  const counter = { dials: 0 };
+  const connect = tts.connectWs.bind(tts);
+  tts.connectWs = async (timeout: number) => {
+    counter.dials++;
+    return connect(timeout);
+  };
+  return { tts, counter };
+}
+
+/**
+ * Consumes the stream the way `Agent.default.ttsNode` does — break at END_OF_STREAM and
+ * close right away — because that is the shape reuse has to survive.
+ */
+async function synthesize(tts: TTS<string>, text: string) {
+  const stream = tts.stream();
+  stream.pushText(text);
+  stream.endInput();
+
+  const samples = new Set<number>();
+  let audioMs = 0;
+  let error: Error | undefined;
+  try {
+    for await (const event of stream) {
+      if (typeof event === 'symbol') break;
+      if (event.frame.samplesPerChannel === 0) continue;
+      samples.add(event.frame.data[0]!);
+      audioMs += (event.frame.samplesPerChannel / event.frame.sampleRate) * 1000;
+    }
+  } catch (e) {
+    error = e as Error;
+  }
+  stream.close();
+  return { samples, audioMs, error };
+}
+
+describe('inference TTS pooled socket reuse across a drained session', () => {
+  let gateway: Awaited<ReturnType<typeof startFakeGateway>>;
+
+  afterEach(async () => {
+    await gateway.close();
+  });
+
+  it('dials once for a conversation of clean multi-generation replies', async () => {
+    gateway = await startFakeGateway();
+    const { tts, counter } = createCountedTTS(gateway.baseURL);
+
+    for (let reply = 0; reply < REPLIES; reply++) {
+      const spoken = await synthesize(tts, `Tell me story number ${reply}.`);
+
+      // Guards the count below against passing on replies that never got their audio.
+      expect(spoken.error).toBeUndefined();
+      expect(spoken.samples).toEqual(new Set([1000 + reply]));
+      expect(spoken.audioMs).toBe(GENERATIONS_PER_REPLY * GENERATION_AUDIO_MS);
+    }
+
+    // One handshake for the whole conversation: every reply after the first reused the
+    // pooled socket. Waiting for the session to go quiet is what makes that safe, so
+    // recycling on the quiet timeout has to keep the socket rather than evict it.
+    expect(counter.dials).toBe(1);
+
+    await tts.close();
+  }, 60_000);
+
+  it('dials again when the gateway drops the session mid-synthesis', async () => {
+    gateway = await startFakeGateway({ dropMidSynthesis: true });
+    const { tts, counter } = createCountedTTS(gateway.baseURL, { maxRetry: 1 });
+
+    await synthesize(tts, 'Tell me a long story about the lighthouse.');
+
+    // A session that never closed its flush still owes audio, so its socket must not go
+    // back to the pool no matter how the reuse above is arranged: the retry has to dial.
+    expect(counter.dials).toBe(2);
+
+    await tts.close();
+  }, 60_000);
+});