@vellumai/assistant 0.11.3 → 0.11.4-staging.2

package/src/live-voice/__tests__/live-voice-flux-turn-end.test.ts

@@ -0,0 +1,932 @@
+import { afterEach, beforeEach, describe, expect, mock, test } from "bun:test";
+
+import { setConfig } from "../../__tests__/helpers/set-config.js";
+import type {
+  VoiceTurnCallbacks,
+  VoiceTurnOptions,
+} from "../../calls/voice-session-bridge.js";
+import type { LiveVoiceFluxConfig } from "../../config/schemas/live-voice.js";
+import type {
+  StreamingTranscriber,
+  SttProviderId,
+  SttStreamServerEvent,
+} from "../../stt/types.js";
+import {
+  LiveVoiceSession,
+  type LiveVoiceTtsStreamer,
+} from "../live-voice-session.js";
+import type { LiveVoiceSessionFactoryContext } from "../live-voice-session-manager.js";
+import type {
+  LiveVoiceTtsOptions,
+  LiveVoiceTtsResult,
+} from "../live-voice-tts.js";
+import {
+  createLiveVoiceServerFrameSequencer,
+  type LiveVoiceClientStartFrame,
+  type LiveVoiceServerFrame,
+} from "../protocol.js";
+
+const SAMPLE_RATE = 24_000;
+
+const VAD_START_FRAME = {
+  type: "start",
+  conversationId: "conversation-123",
+  turnDetection: "server_vad",
+  audio: {
+    mimeType: "audio/pcm",
+    sampleRate: SAMPLE_RATE,
+    channels: 1,
+  },
+} as const satisfies LiveVoiceClientStartFrame;
+
+function pcm(amplitude: number, sampleCount = 240): Uint8Array {
+  const buffer = Buffer.alloc(sampleCount * 2);
+  for (let index = 0; index < sampleCount; index += 1) {
+    buffer.writeInt16LE(amplitude, index * 2);
+  }
+  return new Uint8Array(buffer);
+}
+
+// 10 ms of speech at 24 kHz.
+const LOUD_CHUNK = pcm(8_000);
+// 300 ms of speech at 24 kHz, comfortably past the default sustained-speech
+// barge-in guard (250 ms) in a single chunk.
+const SUSTAINED_LOUD_CHUNK = pcm(8_000, 7_200);
+
+/**
+ * Flux-shaped streaming transcriber: no `finalizeUtterance` (Flux owns turn
+ * boundaries, so the session runs it per cycle), and every transcript event is
+ * scripted by the test. `stop()` closes without inventing a trailing final,
+ * exactly as Flux does: its transcript is committed by EndOfTurn.
+ */
+class MockFluxTranscriber implements StreamingTranscriber {
+  readonly boundaryId = "daemon-streaming" as const;
+  readonly received: Buffer[] = [];
+  stopped = false;
+  private onEvent: ((event: SttStreamServerEvent) => void) | null = null;
+
+  constructor(readonly providerId: SttProviderId) {}
+
+  async start(onEvent: (event: SttStreamServerEvent) => void): Promise<void> {
+    this.onEvent = onEvent;
+  }
+
+  sendAudio(chunk: Buffer): void {
+    this.received.push(Buffer.from(chunk));
+  }
+
+  stop(): void {
+    if (this.stopped) {
+      return;
+    }
+    this.stopped = true;
+    this.onEvent?.({ type: "closed" });
+  }
+
+  emit(event: SttStreamServerEvent): void {
+    this.onEvent?.(event);
+  }
+
+  // Flux's StartOfTurn frame, numbering the turn it opens.
+  startOfTurn(turnIndex: number): void {
+    this.emit({ type: "turn-start", turnIndex });
+  }
+
+  // Flux's EndOfTurn frame maps onto a `final` followed by a `turn-end`.
+  endOfTurn(text: string, turnIndex?: number): void {
+    this.emit({ type: "final", text });
+    this.emit({
+      type: "turn-end",
+      text,
+      confidence: 0.9,
+      ...(turnIndex !== undefined ? { turnIndex } : {}),
+    });
+  }
+}
+
+function createHarness(options: {
+  providerId?: SttProviderId;
+  fluxConfig?: Partial<LiveVoiceFluxConfig>;
+  silenceThresholdMs?: number;
+  startVoiceTurn?: (options: VoiceTurnOptions) => Promise<{
+    turnId: string;
+    abort: () => void;
+  }>;
+  streamTtsAudio?: LiveVoiceTtsStreamer | null;
+  emitMetrics?: boolean;
+  // Holds the STT dial open so a test can drive a whole utterance through the
+  // window between `ready` and the resolved provider.
+  resolveGate?: Promise<unknown>;
+}) {
+  const sequencer = createLiveVoiceServerFrameSequencer();
+  const frames: LiveVoiceServerFrame[] = [];
+  const context: LiveVoiceSessionFactoryContext = {
+    sessionId: "session-123",
+    startFrame: VAD_START_FRAME,
+    sendFrame: mock(async (payload) => {
+      const frame = sequencer.next(payload);
+      frames.push(frame);
+      return frame;
+    }),
+  };
+
+  const transcribers: MockFluxTranscriber[] = [];
+  const resolveTranscriber = mock(async () => {
+    if (options.resolveGate) {
+      await options.resolveGate;
+    }
+    const transcriber = new MockFluxTranscriber(
+      options.providerId ?? "deepgram-flux",
+    );
+    transcribers.push(transcriber);
+    return transcriber;
+  });
+
+  const turnCalls: VoiceTurnOptions[] = [];
+  const startVoiceTurn = mock(async (turnOptions: VoiceTurnOptions) => {
+    turnCalls.push(turnOptions);
+    if (options.startVoiceTurn) {
+      return await options.startVoiceTurn(turnOptions);
+    }
+    return { turnId: `bridge-turn-${turnCalls.length}`, abort: mock() };
+  });
+
+  const session = new LiveVoiceSession(context, {
+    resolveTranscriber,
+    startVoiceTurn,
+    streamTtsAudio: options.streamTtsAudio ?? null,
+    emitMetrics: options.emitMetrics ?? false,
+    spawnBackgroundContinuation: mock(async () => ""),
+    turnDetectorConfig: {
+      silenceThresholdMs: options.silenceThresholdMs ?? 40,
+    },
+    ...(options.fluxConfig ? { fluxConfig: options.fluxConfig } : {}),
+  });
+
+  return { frames, session, transcribers, turnCalls };
+}
+
+function makeTextDelta(
+  text: string,
+): Parameters<NonNullable<VoiceTurnCallbacks["assistant_text_delta"]>>[0] {
+  return {
+    type: "assistant_text_delta",
+    text,
+    conversationId: "conversation-123",
+  };
+}
+
+function makeMessageComplete(): Parameters<
+  NonNullable<VoiceTurnCallbacks["message_complete"]>
+>[0] {
+  return {
+    type: "message_complete",
+    conversationId: "conversation-123",
+    messageId: "assistant-message-123",
+  };
+}
+
+// Answers immediately: one text delta (which commits a speculative leg), then
+// message_complete.
+function autoCompletingTurn(
+  reply = "Okay.",
+): (
+  options: VoiceTurnOptions,
+) => Promise<{ turnId: string; abort: () => void }> {
+  let count = 0;
+  return async (options) => {
+    count += 1;
+    options.callbacks?.assistant_text_delta?.(makeTextDelta(reply));
+    options.callbacks?.message_complete?.(makeMessageComplete());
+    return { turnId: `bridge-turn-${count}`, abort: mock() };
+  };
+}
+
+async function waitFor(
+  predicate: () => boolean,
+  message = "Timed out waiting for a live voice Flux test condition",
+): Promise<void> {
+  const deadline = Date.now() + 4_000;
+  while (Date.now() < deadline) {
+    if (predicate()) {
+      return;
+    }
+    await new Promise((resolve) => setTimeout(resolve, 5));
+  }
+  throw new Error(message);
+}
+
+async function flushAsyncCallbacks(): Promise<void> {
+  await new Promise((resolve) => setTimeout(resolve, 10));
+}
+
+async function sleep(ms: number): Promise<void> {
+  await new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+// Holds the STT dial open until the test opens it, so the whole
+// ready-before-resolve window is under the test's control.
+function createDialGate(): { promise: Promise<void>; open: () => void } {
+  let open = (): void => {};
+  const promise = new Promise<void>((resolve) => {
+    open = () => resolve();
+  });
+  return { promise, open };
+}
+
+function countFrames(
+  frames: LiveVoiceServerFrame[],
+  type: LiveVoiceServerFrame["type"],
+): number {
+  return frames.filter((frame) => frame.type === type).length;
+}
+
+// Waits for the completed turn's metrics frame and hands it back narrowed, so
+// the assertions below read the endpoint fields without a type guard.
+async function waitForTurnMetrics(
+  frames: LiveVoiceServerFrame[],
+): Promise<Extract<LiveVoiceServerFrame, { type: "metrics" }>> {
+  const isTurnMetrics = (
+    frame: LiveVoiceServerFrame,
+  ): frame is Extract<LiveVoiceServerFrame, { type: "metrics" }> =>
+    frame.type === "metrics" && frame.event === "turn_completed";
+  await waitFor(
+    () => frames.some(isTurnMetrics),
+    "The completed turn never reported its metrics",
+  );
+  const metricsFrame = frames.find(isTurnMetrics);
+  if (!metricsFrame) {
+    throw new Error("The completed turn never reported its metrics");
+  }
+  return metricsFrame;
+}
+
+// Long enough for the 40 ms local silence timer to fire and hand the boundary
+// to Flux (which arms the fail-open deadline and commits nothing).
+const PAST_SILENCE_BOUNDARY_MS = 150;
+
+const FLUX_ON = {
+  turnEnd: { enabled: true },
+  eotTimeoutMs: 500,
+} as const satisfies Partial<LiveVoiceFluxConfig>;
+
+describe("LiveVoiceSession Flux end-of-turn", () => {
+  test("commits the turn on turn-end without an endpoint-decision leg", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      // Long enough that the local silence boundary cannot be what commits.
+      silenceThresholdMs: 10_000,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.endOfTurn("what is the weather");
+
+    await waitFor(() => turnCalls.length === 1);
+    expect(turnCalls[0]?.content).toBe("what is the weather");
+    // Non-speculative dispatch: the front-door leg's decision rule is built
+    // with includeHold false, so the hold token is never taught.
+    expect(turnCalls[0]?.unifiedVerdict).toBeUndefined();
+    expect(turnCalls[0]?.routingLeg).toBe("front-door");
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(true);
+
+    await session.close("client_end");
+  });
+
+  test("still escalates on the [1] verdict after a flux commit", async () => {
+    const { session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 10_000,
+      startVoiceTurn: async (turnOptions) => {
+        if (turnOptions.routingLeg === "front-door") {
+          turnOptions.callbacks?.assistant_text_delta?.(
+            makeTextDelta("[1] Let me check that for you."),
+          );
+        }
+        return { turnId: "bridge-turn", abort: mock() };
+      },
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.endOfTurn("what is on my calendar");
+
+    await waitFor(
+      () => turnCalls.some((call) => call.routingLeg === "escalated"),
+      "The escalate verdict never handed off to the escalated leg",
+    );
+
+    await session.close("client_end");
+  });
+
+  test("ignores turn-end when the flag is off and keeps the hold path", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.emit({ type: "partial", text: "hello there" });
+    transcribers[0]?.endOfTurn("hello there");
+    await flushAsyncCallbacks();
+
+    // The turn-end event committed nothing: the silence boundary still owns
+    // the commit, and it dispatches speculatively (the hold path).
+    await waitFor(() => turnCalls.length === 1);
+    expect(turnCalls[0]?.unifiedVerdict).toBe(true);
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(true);
+
+    await session.close("client_end");
+  });
+
+  test("ignores turn-end from a provider that is not deepgram-flux", async () => {
+    const { session, transcribers, turnCalls } = createHarness({
+      providerId: "deepgram",
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 10_000,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.endOfTurn("hello there");
+    await flushAsyncCallbacks();
+
+    expect(turnCalls).toHaveLength(0);
+
+    await session.close("client_end");
+  });
+
+  test("reports the endpoint decision as flux", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 10_000,
+      emitMetrics: true,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.endOfTurn("hello there");
+    await waitFor(() => turnCalls.length === 1);
+
+    const metricsFrame = await waitForTurnMetrics(frames);
+    expect(metricsFrame.endpointDecisionSource).toBe("provider");
+    expect(metricsFrame.endpointHoldCount).toBe(0);
+
+    await session.close("client_end");
+  });
+
+  test("anchors the flux commit latency at the local speech-stop mark", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: { turnEnd: { enabled: true }, eotTimeoutMs: 30_000 },
+      // Long enough that the local silence boundary cannot be what commits,
+      // and that the fail-open deadline stays far away.
+      silenceThresholdMs: 10_000,
+      emitMetrics: true,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    // Silence between the caller's last above-gate chunk and the commit. A
+    // dispatch-anchored number would not see it; the speech-stop anchor must.
+    await sleep(250);
+    transcribers[0]?.endOfTurn("hello there");
+    await waitFor(() => turnCalls.length === 1);
+
+    const metricsFrame = await waitForTurnMetrics(frames);
+    expect(metricsFrame.endpointCommitLatencyMs).toBeGreaterThanOrEqual(250);
+
+    await session.close("client_end");
+  });
+
+  test("records the commit latency on the front-door path too, with no hold", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      // No flux config: the latch is down and the local silence boundary owns
+      // the commit, which is arm A of the measurement run.
+      silenceThresholdMs: 40,
+      emitMetrics: true,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.emit({ type: "partial", text: "hello there" });
+    await waitFor(() => turnCalls.length === 1);
+
+    // The turn committed straight through, so it records no endpoint decision
+    // at all. The commit latency is recorded regardless, which is what makes
+    // the two arms one population rather than two.
+    const metricsFrame = await waitForTurnMetrics(frames);
+    expect(metricsFrame.endpointDecisionSource).toBeUndefined();
+    expect(metricsFrame.endpointCommitLatencyMs).toBeGreaterThanOrEqual(40);
+
+    await session.close("client_end");
+  });
+
+  test("records no commit latency for a turn that never committed", async () => {
+    const { frames, session, transcribers } = createHarness({
+      fluxConfig: { turnEnd: { enabled: true }, eotTimeoutMs: 30_000 },
+      silenceThresholdMs: 10_000,
+      emitMetrics: true,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.emit({ type: "partial", text: "hello the" });
+    // Nothing ends this turn: no flux turn-end, no silence boundary, and the
+    // fail-open deadline is 30 s away.
+    await flushAsyncCallbacks();
+    await session.close("client_end");
+
+    const metricsFrames = frames.filter((frame) => frame.type === "metrics");
+    expect(metricsFrames.length).toBeGreaterThan(0);
+    for (const frame of metricsFrames) {
+      expect(frame).not.toHaveProperty("endpointCommitLatencyMs");
+    }
+  });
+
+  test("falls back to the silence-boundary path when no turn-end arrives", async () => {
+    const { session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    // Flux is transcribing but its EndOfTurn never arrives.
+    transcribers[0]?.emit({ type: "partial", text: "are you still there" });
+
+    // The local silence boundary (40 ms) passes without committing anything:
+    // the fallback deadline (eotTimeoutMs plus the margin) owns this
+    // utterance.
+    await new Promise((resolve) => setTimeout(resolve, 300));
+    expect(turnCalls).toHaveLength(0);
+
+    await waitFor(
+      () => turnCalls.length === 1,
+      "Flux fallback never replayed the silence boundary",
+    );
+    // Back on the existing path: a speculative dispatch that knows the hold
+    // token, judging the partial the silence boundary saw.
+    expect(turnCalls[0]?.content).toBe("are you still there");
+    expect(turnCalls[0]?.unifiedVerdict).toBe(true);
+
+    await session.close("client_end");
+  }, 10_000);
+
+  test("commits a turn-end that arrives after the local silence boundary", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+
+    // The boundary passes and commits nothing: Flux owns it.
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+    expect(turnCalls).toHaveLength(0);
+
+    // No resumed speech, so this turn-end is the one the boundary was waiting
+    // for and it commits exactly as before.
+    transcribers[0]?.endOfTurn("what is the weather");
+    await waitFor(
+      () => turnCalls.length === 1,
+      "A non-stale turn-end after the silence boundary never committed",
+    );
+    expect(turnCalls[0]?.content).toBe("what is the weather");
+    expect(turnCalls[0]?.unifiedVerdict).toBeUndefined();
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(true);
+
+    await session.close("client_end");
+  });
+
+  test("drops a turn-end that arrives after the caller resumed speaking", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    const transcriber = transcribers[0];
+    transcriber?.emit({ type: "partial", text: "what is the" });
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+
+    // The caller draws breath and keeps going before the provider's turn-end
+    // for the speech that boundary closed reaches the session.
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(
+      () => countFrames(frames, "speech_started") === 2,
+      "The resumed speech never re-opened a detector turn",
+    );
+    const receivedBeforeResume = transcriber?.received.length ?? 0;
+
+    transcriber?.emit({
+      type: "turn-end",
+      text: "what is the",
+      confidence: 0.9,
+    });
+    await flushAsyncCallbacks();
+
+    // Dropped: no boundary frame, the transcriber keeps running, no turn.
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(false);
+    expect(transcriber?.stopped).toBe(false);
+    expect(turnCalls).toHaveLength(0);
+
+    // ...and the resumed speech still routes into the same open utterance.
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(
+      () => (transcriber?.received.length ?? 0) > receivedBeforeResume,
+      "The resumed speech stopped reaching the transcriber",
+    );
+
+    await session.close("client_end");
+  });
+
+  test("commits a fast turn-end for the still-open turn after a mid-thought pause", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      emitMetrics: true,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    const transcriber = transcribers[0];
+    transcriber?.startOfTurn(0);
+    transcriber?.emit({ type: "partial", text: "what is the" });
+
+    // A mid-thought pause, long enough for the local silence boundary to hand
+    // the cycle to Flux, and then the caller keeps going.
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(
+      () => countFrames(frames, "speech_started") === 2,
+      "The resumed speech never re-opened a detector turn",
+    );
+
+    // Flux ends the turn it kept open across the pause, beating the next local
+    // silence boundary: the fast commit this feature exists for. Nothing here
+    // sleeps past that boundary, on purpose.
+    const emittedAtMs = Date.now();
+    transcriber?.endOfTurn("what is the weather today", 0);
+    await waitFor(
+      () => turnCalls.length === 1,
+      "The fast turn-end after a mid-thought pause never committed",
+    );
+    const commitLatencyMs = Date.now() - emittedAtMs;
+
+    // The fail-open budget is eotTimeoutMs (500 ms) plus the margin (1000 ms).
+    // A commit anywhere near it means the turn-end was dropped and the
+    // deadline committed in its place.
+    expect(commitLatencyMs).toBeLessThan(300);
+    expect(turnCalls[0]?.content).toBe("what is the weather today");
+    // The Flux path, not the speculative hold path the deadline falls back to.
+    expect(turnCalls[0]?.unifiedVerdict).toBeUndefined();
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(true);
+
+    const metricsFrame = await waitForTurnMetrics(frames);
+    expect(metricsFrame.endpointDecisionSource).toBe("provider");
+
+    await session.close("client_end");
+  });
+
+  test("drops a turn-end for a turn flux has already superseded", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    const transcriber = transcribers[0];
+    transcriber?.startOfTurn(0);
+    transcriber?.emit({ type: "partial", text: "what is the" });
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(
+      () => countFrames(frames, "speech_started") === 2,
+      "The resumed speech never re-opened a detector turn",
+    );
+    const receivedBeforeResume = transcriber?.received.length ?? 0;
+
+    // Flux closed turn 0 and opened turn 1 for the resumed speech, so the
+    // end-of-turn for turn 0 trailing behind describes speech the caller has
+    // moved past.
+    transcriber?.startOfTurn(1);
+    transcriber?.emit({
+      type: "turn-end",
+      text: "what is the",
+      confidence: 0.9,
+      turnIndex: 0,
+    });
+    await flushAsyncCallbacks();
+
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(false);
+    expect(transcriber?.stopped).toBe(false);
+    expect(turnCalls).toHaveLength(0);
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(
+      () => (transcriber?.received.length ?? 0) > receivedBeforeResume,
+      "The resumed speech stopped reaching the transcriber",
+    );
+
+    // Turn 1's own end-of-turn is the one that commits.
+    transcriber?.endOfTurn("what is the weather today", 1);
+    await waitFor(
+      () => turnCalls.length === 1,
+      "The superseding turn never committed",
+    );
+    expect(turnCalls[0]?.content).toBe("what is the weather today");
+    expect(turnCalls[0]?.unifiedVerdict).toBeUndefined();
+
+    await session.close("client_end");
+  });
+
+  test("commits the resumed turn on the next turn-end after dropping a stale one", async () => {
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => countFrames(frames, "speech_started") === 2);
+    transcribers[0]?.emit({
+      type: "turn-end",
+      text: "what is the",
+      confidence: 0.9,
+    });
+    await flushAsyncCallbacks();
+    expect(turnCalls).toHaveLength(0);
+
+    // The caller finishes for real: the next silence boundary re-stamps the
+    // cycle, so Flux's turn-end for the resumed speech commits it.
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+    transcribers[0]?.endOfTurn("what is the weather today");
+
+    await waitFor(
+      () => turnCalls.length === 1,
+      "The resumed turn never committed after the stale turn-end was dropped",
+    );
+    expect(turnCalls[0]?.content).toBe("what is the weather today");
+    expect(turnCalls[0]?.unifiedVerdict).toBeUndefined();
+    expect(frames.some((frame) => frame.type === "utterance_end")).toBe(true);
+
+    await session.close("client_end");
+  });
+
+  test("falls back to the deadline when nothing follows the dropped turn-end", async () => {
+    const { session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.emit({ type: "partial", text: "are you still there" });
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    transcribers[0]?.emit({
+      type: "turn-end",
+      text: "are you still there",
+      confidence: 0.9,
+    });
+    await flushAsyncCallbacks();
+    expect(turnCalls).toHaveLength(0);
+
+    // Flux says nothing more. The next silence boundary re-arms the fail-open
+    // deadline, so the utterance still commits on the silence path instead of
+    // hanging open forever.
+    await waitFor(
+      () => turnCalls.length === 1,
+      "The utterance hung after the stale turn-end was dropped",
+    );
+    expect(turnCalls[0]?.content).toBe("are you still there");
+    expect(turnCalls[0]?.unifiedVerdict).toBe(true);
+
+    await session.close("client_end");
+  }, 10_000);
+
+  for (const flagOn of [false, true]) {
+    test(`barge-in during playback fires from local VAD (flux turn end ${
+      flagOn ? "on" : "off"
+    })`, async () => {
+      const streamTtsAudio = mock(async (ttsOptions: LiveVoiceTtsOptions) => {
+        ttsOptions.onAudioChunk({
+          type: "tts_audio",
+          contentType: "audio/pcm",
+          sampleRate: SAMPLE_RATE,
+          dataBase64: Buffer.from("assistant audio").toString("base64"),
+        });
+        return {
+          provider: "fish-audio",
+          contentType: "audio/pcm",
+          sampleRate: SAMPLE_RATE,
+          chunks: 1,
+          bytes: 15,
+        } satisfies LiveVoiceTtsResult;
+      });
+      const { frames, session, transcribers } = createHarness({
+        ...(flagOn ? { fluxConfig: FLUX_ON } : {}),
+        silenceThresholdMs: 10_000,
+        streamTtsAudio,
+        // The leg never completes, so the turn is still in flight when the
+        // caller speaks over it.
+        startVoiceTurn: async (turnOptions) => {
+          turnOptions.callbacks?.assistant_text_delta?.(
+            makeTextDelta("Let me tell you a long story."),
+          );
+          return { turnId: "bridge-turn-1", abort: mock() };
+        },
+      });
+
+      await session.start();
+      await session.handleBinaryAudio(LOUD_CHUNK);
+      await waitFor(() => transcribers.length > 0);
+      if (flagOn) {
+        transcribers[0]?.endOfTurn("tell me a story");
+      } else {
+        transcribers[0]?.emit({ type: "final", text: "tell me a story" });
+        await session.handleClientFrame({ type: "ptt_release" });
+      }
+      await waitFor(() => frames.some((frame) => frame.type === "thinking"));
+
+      // Local energy detection, not the provider's turn model, is what
+      // interrupts the assistant.
+      await session.handleBinaryAudio(SUSTAINED_LOUD_CHUNK);
+      await waitFor(
+        () => frames.some((frame) => frame.type === "speech_started"),
+        "Local VAD barge-in never fired",
+      );
+      await waitFor(() =>
+        frames.some((frame) => frame.type === "turn_cancelled"),
+      );
+
+      await session.close("client_end");
+    });
+  }
+});
+
+/**
+ * The window between `ready` and the resolved transcriber. `start()` arms the
+ * utterance in the background so the client's mic acquisition overlaps the STT
+ * handshake, which means a fast caller can speak AND fall silent before the
+ * dial answers. These tests seed `services.stt.provider` for real, because the
+ * latch is armed from the configured provider before the dial and reconciled
+ * with the resolved one after it.
+ */
+describe("LiveVoiceSession Flux end-of-turn during the STT dial", () => {
+  beforeEach(() => {
+    setConfig("services", {
+      stt: { provider: "deepgram-flux", providers: {} },
+    });
+  });
+
+  afterEach(() => {
+    setConfig("services", { stt: { provider: "deepgram", providers: {} } });
+  });
+
+  test("waits for turn-end when the boundary lands before the dial resolves", async () => {
+    const gate = createDialGate();
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      fluxConfig: FLUX_ON,
+      silenceThresholdMs: 40,
+      resolveGate: gate.promise,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    expect(transcribers).toHaveLength(0);
+
+    // The caller speaks and falls silent entirely inside the handshake
+    // window, so the whole first turn is decided before any provider answers.
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+
+    // The silence boundary passed and committed nothing: the latch was seeded
+    // from the configured provider, so Flux owns this boundary like any other.
+    expect(transcribers).toHaveLength(0);
+    expect(countFrames(frames, "utterance_end")).toBe(0);
+    expect(turnCalls).toHaveLength(0);
+
+    gate.open();
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.endOfTurn("what is the weather");
+
+    await waitFor(
+      () => turnCalls.length === 1,
+      "The first utterance never committed on the Flux end-of-turn",
+    );
+    // Committed by Flux, not by the silence path: non-speculative dispatch.
+    expect(turnCalls[0]?.content).toBe("what is the weather");
+    expect(turnCalls[0]?.unifiedVerdict).toBeUndefined();
+    expect(countFrames(frames, "utterance_end")).toBe(1);
+
+    await session.close("client_end");
+  });
+
+  test("unwinds the seeded latch when the dial resolves to another provider", async () => {
+    const gate = createDialGate();
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      // Config names Flux, so the latch is seeded, but the dial answers with
+      // another provider: a fallback, or a resolver reading a config the
+      // session no longer matches. That provider never sends an end-of-turn.
+      providerId: "deepgram",
+      // Far enough out that the fail-open deadline alone cannot be what
+      // releases this utterance inside the test's budget.
+      fluxConfig: { turnEnd: { enabled: true }, eotTimeoutMs: 30_000 },
+      silenceThresholdMs: 40,
+      resolveGate: gate.promise,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+    await sleep(PAST_SILENCE_BOUNDARY_MS);
+    // Deferred under the seeded latch, exactly as a real Flux session would.
+    expect(countFrames(frames, "utterance_end")).toBe(0);
+
+    gate.open();
+    await waitFor(() => transcribers.length > 0);
+
+    // The deferred boundary is replayed as soon as the dial disproves the
+    // seed, instead of being stranded on a deadline 30 s away.
+    await waitFor(
+      () => countFrames(frames, "utterance_end") === 1,
+      "The utterance deferred under the seeded latch never released",
+    );
+    await waitFor(
+      () => transcribers[0]?.stopped === true,
+      "The utterance never finished releasing",
+    );
+    // ...and nothing routed through the Flux commit path.
+    expect(turnCalls.every((call) => call.unifiedVerdict !== undefined)).toBe(
+      true,
+    );
+
+    await session.close("client_end");
+  });
+
+  test("never seeds the latch when the flag is off", async () => {
+    const gate = createDialGate();
+    const { frames, session, transcribers, turnCalls } = createHarness({
+      // No fluxConfig: `turnEnd.enabled` keeps its schema default of false
+      // while config still names deepgram-flux as the provider.
+      silenceThresholdMs: 40,
+      resolveGate: gate.promise,
+      startVoiceTurn: autoCompletingTurn(),
+    });
+
+    await session.start();
+    await session.handleBinaryAudio(LOUD_CHUNK);
+
+    // Unchanged from today: the silence boundary releases during the dial.
+    await waitFor(
+      () => countFrames(frames, "utterance_end") === 1,
+      "The flag-off silence boundary stopped releasing during the dial",
+    );
+    expect(turnCalls).toHaveLength(0);
+
+    gate.open();
+    await waitFor(() => transcribers.length > 0);
+    transcribers[0]?.endOfTurn("hello there");
+    await flushAsyncCallbacks();
+
+    // The turn-end commits nothing on top of the boundary that already ran.
+    expect(countFrames(frames, "utterance_end")).toBe(1);
+
+    await session.close("client_end");
+  });
+});