@alexkroman1/aai 1.4.0 → 1.4.2

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @alexkroman1/aai@1.4.0 build /home/runner/work/agent/agent/packages/aai
+> @alexkroman1/aai@1.4.2 build /home/runner/work/agent/agent/packages/aai
 > tsdown && tsc -p tsconfig.build.json
 
 ℹ tsdown v0.21.7 powered by rolldown v1.0.0-rc.12
@@ -8,7 +8,7 @@
 ℹ target: node22
 ℹ tsconfig: tsconfig.json
 ℹ Build start
-ℹ dist/host/runtime-barrel.js       74.18 kB │ gzip: 22.09 kB
+ℹ dist/host/runtime-barrel.js       75.92 kB │ gzip: 22.48 kB
 ℹ dist/sdk/protocol.js               4.75 kB │ gzip:  1.76 kB
 ℹ dist/index.js                      2.88 kB │ gzip:  1.24 kB
 ℹ dist/sdk/manifest-barrel.js        0.36 kB │ gzip:  0.20 kB
@@ -22,5 +22,5 @@
 ℹ dist/assemblyai-Cxg9eobY.js        0.53 kB │ gzip:  0.35 kB
 ℹ dist/anthropic-BrUCPKUc.js         0.23 kB │ gzip:  0.18 kB
 ℹ dist/cartesia-DwDk2tEu.js          0.22 kB │ gzip:  0.17 kB
-ℹ 14 files, total: 99.00 kB
-✔ Build complete in 39ms
+ℹ 14 files, total: 100.74 kB
+✔ Build complete in 48ms

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # @alexkroman1/aai
 
+## 1.4.2
+
+### Patch Changes
+
+- f877a6f: Fix pipeline mode: play greeting, emit a single agent_transcript per turn, open TTS at the client's playback sample rate, and stop the Cartesia adapter from eagerly rotating its context (which was silently dropping in-flight audio chunks).
+
+## 1.4.1
+
+### Patch Changes
+
+- 63de397: Pass explicit baseURL to createAnthropic so the SDK's loadOptionalSetting returns before reading process.env['ANTHROPIC_BASE_URL']. The Deno platform server runs without --allow-env, and the missing baseURL caused pipeline-mode sessions to crash on first use.
+
 ## 1.4.0
 
 ## 1.3.2

package/dist/host/pipeline-session.d.ts

@@ -37,8 +37,12 @@ export interface PipelineSessionOptions {
     sttApiKey: string;
     /** TTS API key. */
     ttsApiKey: string;
-    /** Audio sample rate (PCM16, Hz). Defaults to {@link DEFAULT_STT_SAMPLE_RATE}. */
-    sampleRate?: number | undefined;
+    /** STT audio sample rate (PCM16, Hz). Defaults to {@link DEFAULT_STT_SAMPLE_RATE}. */
+    sttSampleRate?: number | undefined;
+    /** TTS audio sample rate (PCM16, Hz). Must match the client's playback AudioContext rate. Defaults to {@link DEFAULT_TTS_SAMPLE_RATE}. */
+    ttsSampleRate?: number | undefined;
+    /** Skip the initial greeting audio on connect (used for session resume). */
+    skipGreeting?: boolean | undefined;
     /** Logger. Defaults to the console logger. */
     logger?: Logger | undefined;
     /** Sliding-window conversation history size. */

package/dist/host/runtime-barrel.js

@@ -560,10 +560,6 @@ function handleStreamPart(part, deps) {
 			if (delta.length === 0) return;
 			deps.onTextDelta(delta);
 			deps.tts?.sendText(delta);
-			deps.client.event({
-				type: "agent_transcript",
-				text: delta
-			});
 			return;
 		}
 		case "tool-call": {
@@ -601,7 +597,8 @@ function handleStreamPart(part, deps) {
 /** Create a pluggable-provider voice session. */
 function createPipelineSession(opts) {
 	const log = opts.logger ?? consoleLogger;
-	const sampleRate = opts.sampleRate ?? 16e3;
+	const sttSampleRate = opts.sttSampleRate ?? 16e3;
+	const ttsSampleRate = opts.ttsSampleRate ?? 24e3;
 	const { client, agentConfig, toolSchemas, executeTool } = opts;
 	const systemPrompt = buildSystemPrompt(agentConfig, {
 		hasTools: toolSchemas.length > 0 || (agentConfig.builtinTools?.length ?? 0) > 0,
@@ -797,15 +794,44 @@ function createPipelineSession(opts) {
 			if (turnController === ctl) turnController = null;
 			return;
 		}
+		if (accumulated.length > 0) {
+			client.event({
+				type: "agent_transcript",
+				text: accumulated
+			});
+			ctx.pushMessages({
+				role: "assistant",
+				content: accumulated
+			});
+		}
 		await flushTtsAndWait(ctl.signal);
 		if (ctl.signal.aborted) {
 			if (turnController === ctl) turnController = null;
 			return;
 		}
-		if (accumulated.length > 0) ctx.pushMessages({
+		client.playAudioDone();
+		client.event({ type: "reply_done" });
+		if (turnController === ctl) turnController = null;
+	}
+	async function runGreeting(text) {
+		const replyId = `pipeline-greeting-${++nextReplyId}`;
+		ctx.beginReply(replyId);
+		const ctl = new AbortController();
+		turnController = ctl;
+		client.event({
+			type: "agent_transcript",
+			text
+		});
+		ctx.pushMessages({
 			role: "assistant",
-			content: accumulated
+			content: text
 		});
+		ctx.tts?.sendText(text);
+		await flushTtsAndWait(ctl.signal);
+		if (ctl.signal.aborted) {
+			if (turnController === ctl) turnController = null;
+			return;
+		}
 		client.playAudioDone();
 		client.event({ type: "reply_done" });
 		if (turnController === ctl) turnController = null;
@@ -841,12 +867,12 @@ function createPipelineSession(opts) {
 	}
 	async function openProviders() {
 		const [sttResult, ttsResult] = await Promise.allSettled([opts.stt.open({
-			sampleRate,
+			sampleRate: sttSampleRate,
 			apiKey: opts.sttApiKey,
 			sttPrompt: agentConfig.sttPrompt,
 			signal: sessionAbort.signal
 		}), opts.tts.open({
-			sampleRate,
+			sampleRate: ttsSampleRate,
 			apiKey: opts.ttsApiKey,
 			signal: sessionAbort.signal
 		})]);
@@ -890,7 +916,18 @@ function createPipelineSession(opts) {
 			ctx.stt?.sendAudio(pcm);
 		},
 		onAudioReady() {
+			if (audioReady || terminated) return;
 			audioReady = true;
+			if (opts.skipGreeting) return;
+			const greeting = agentConfig.greeting;
+			if (!greeting) return;
+			const turn = runGreeting(greeting).catch((err) => {
+				log.error("Pipeline greeting failed", {
+					error: errorMessage(err),
+					sessionId: opts.id
+				});
+			});
+			ctx.chainTurn(turn);
 		},
 		onCancel() {
 			if (terminated) return;
@@ -1076,9 +1113,21 @@ function openCartesia(opts) {
 			* Reset whenever a fresh context is minted (i.e. at turn boundaries).
 			*/
 			let doneEmitted = false;
+			/**
+			* After `flush()` or `cancel()`, the current context is done accepting
+			* input. We defer minting a fresh one until the next `sendText()` so
+			* that late audio chunks + Cartesia's real `done` event (both tagged
+			* with the flushed context's id) still pass the filter below. Rotating
+			* eagerly would silently drop all audio still in flight.
+			*/
+			let rotatePending = false;
 			const rotateContext = () => {
 				context = mintContext();
 				doneEmitted = false;
+				rotatePending = false;
+			};
+			const rotateIfPending = () => {
+				if (rotatePending) rotateContext();
 			};
 			const emitDoneOnce = () => {
 				if (doneEmitted || closed) return;
@@ -1133,6 +1182,7 @@ function openCartesia(opts) {
 			return {
 				sendText(text) {
 					if (closed || text.length === 0) return;
+					rotateIfPending();
 					context.send({
 						...baseRequest,
 						transcript: text,
@@ -1140,20 +1190,19 @@ function openCartesia(opts) {
 					}).catch(ignoreRejection);
 				},
 				flush() {
-					if (closed) return;
+					if (closed || rotatePending) return;
 					context.send({
 						...baseRequest,
 						transcript: "",
 						continue: false
 					}).catch(ignoreRejection);
-					queueMicrotask(emitDoneOnce);
-					rotateContext();
+					rotatePending = true;
 				},
 				cancel() {
 					if (closed) return;
 					context.cancel().catch(ignoreRejection);
 					emitDoneOnce();
-					rotateContext();
+					rotatePending = true;
 				},
 				on(event, fn) {
 					return emitter.on(event, fn);
@@ -1216,7 +1265,10 @@ function resolveLlm(descriptor, env) {
 			const options = descriptor.options;
 			const apiKey = resolveApiKey("ANTHROPIC_API_KEY", env);
 			if (!apiKey) throw new Error("Anthropic LLM: missing API key. Set ANTHROPIC_API_KEY in the agent env.");
-			return createAnthropic({ apiKey })(options.model);
+			return createAnthropic({
+				apiKey,
+				baseURL: "https://api.anthropic.com/v1"
+			})(options.model);
 		}
 		default: throw new Error(`Unknown LLM provider kind: "${descriptor.kind}". Supported: ${ANTHROPIC_KIND}.`);
 	}
@@ -2150,6 +2202,9 @@ function createRuntime(opts) {
 			tts: pipelineProviders.tts,
 			sttApiKey: resolveApiKey("ASSEMBLYAI_API_KEY", env),
 			ttsApiKey: resolveApiKey("CARTESIA_API_KEY", env),
+			sttSampleRate: s2sConfig.inputSampleRate,
+			ttsSampleRate: s2sConfig.outputSampleRate,
+			skipGreeting: sessionOpts.skipGreeting ?? false,
 			logger
 		});
 		const apiKey = env.ASSEMBLYAI_API_KEY ?? "";

package/host/pipeline-session.test.ts

@@ -43,7 +43,8 @@ function makeOpts(overrides: Partial<PipelineSessionOptions> = {}): {
     tts,
     sttApiKey: "stt-key",
     ttsApiKey: "tts-key",
-    sampleRate: 16_000,
+    sttSampleRate: 16_000,
+    ttsSampleRate: 24_000,
     logger: silentLogger,
     ...overrides,
   };
@@ -81,20 +82,112 @@ describe("createPipelineSession — happy path", () => {
     expect(ttsSession.textChunks).toEqual(["Hello", " there"]);
     expect(ttsSession.flush).toHaveBeenCalledTimes(1);
 
-    // Verify wire events in order
+    // Verify wire events in order — the pipeline emits a single
+    // `agent_transcript` with the full accumulated reply (not one per
+    // delta) so the UI renders one assistant message per turn.
     const types = eventTypes(client.events);
-    expect(types).toEqual([
-      "user_transcript",
-      "agent_transcript", // "Hello"
-      "agent_transcript", // " there"
-      "reply_done",
-    ]);
+    expect(types).toEqual(["user_transcript", "agent_transcript", "reply_done"]);
 
     // user_transcript text matches
     expect(client.events[0]).toMatchObject({
       type: "user_transcript",
       text: "Hello there, how are you?",
     });
+    expect(client.events[1]).toMatchObject({
+      type: "agent_transcript",
+      text: "Hello there",
+    });
+
+    await session.stop();
+  });
+});
+
+describe("createPipelineSession — greeting", () => {
+  test("onAudioReady sends greeting to TTS and emits agent_transcript + reply_done", async () => {
+    const { opts, tts, client } = makeOpts({
+      agentConfig: {
+        name: "pipeline-agent",
+        systemPrompt: DEFAULT_SYSTEM_PROMPT,
+        greeting: "Hi! I'm pipeline mode.",
+      },
+    });
+
+    const session = createPipelineSession(opts);
+    await session.start();
+
+    const ttsSession = tts.last();
+    if (!ttsSession) throw new Error("TTS didn't open");
+
+    session.onAudioReady();
+    await session.waitForTurn();
+
+    expect(ttsSession.textChunks).toEqual(["Hi! I'm pipeline mode."]);
+    expect(ttsSession.flush).toHaveBeenCalledTimes(1);
+
+    const types = eventTypes(client.events);
+    expect(types).toEqual(["agent_transcript", "reply_done"]);
+    expect(client.events[0]).toMatchObject({
+      type: "agent_transcript",
+      text: "Hi! I'm pipeline mode.",
+    });
+
+    await session.stop();
+  });
+
+  test("skipGreeting=true suppresses the greeting turn", async () => {
+    const { opts, tts, client } = makeOpts({
+      agentConfig: {
+        name: "pipeline-agent",
+        systemPrompt: DEFAULT_SYSTEM_PROMPT,
+        greeting: "Hello there.",
+      },
+      skipGreeting: true,
+    });
+
+    const session = createPipelineSession(opts);
+    await session.start();
+
+    const ttsSession = tts.last();
+    if (!ttsSession) throw new Error("TTS didn't open");
+
+    session.onAudioReady();
+    await session.waitForTurn();
+
+    expect(ttsSession.sendText).not.toHaveBeenCalled();
+    expect(ttsSession.flush).not.toHaveBeenCalled();
+    expect(client.events).toEqual([]);
+
+    await session.stop();
+  });
+
+  test("empty greeting is a no-op", async () => {
+    const { opts, tts, client } = makeOpts();
+    // CONFIG already has greeting: ""
+    const session = createPipelineSession(opts);
+    await session.start();
+
+    const ttsSession = tts.last();
+    if (!ttsSession) throw new Error("TTS didn't open");
+
+    session.onAudioReady();
+    await session.waitForTurn();
+
+    expect(ttsSession.sendText).not.toHaveBeenCalled();
+    expect(client.events).toEqual([]);
+
+    await session.stop();
+  });
+
+  test("passes sttSampleRate / ttsSampleRate through to providers", async () => {
+    const { opts, stt, tts } = makeOpts({
+      sttSampleRate: 16_000,
+      ttsSampleRate: 24_000,
+    });
+    const session = createPipelineSession(opts);
+    await session.start();
+
+    expect(stt.last()?.opts.sampleRate).toBe(16_000);
+    expect(tts.last()?.opts.sampleRate).toBe(24_000);
 
     await session.stop();
   });
@@ -163,8 +256,10 @@ describe("createPipelineSession — barge-in", () => {
 
     // TTS.cancel must have been called exactly once.
     expect(ttsSession.cancel).toHaveBeenCalledTimes(1);
-    // Wire events: user_transcript, some agent_transcript(s), then cancelled.
-    // No reply_done — barge-in short-circuits the drain.
+    // Wire events: user_transcript then cancelled. No agent_transcript
+    // (the pipeline only emits it after the LLM stream finishes cleanly)
+    // and no reply_done — barge-in short-circuits both the stream and
+    // the drain.
     const types = eventTypes(client.events);
     expect(types).toContain("user_transcript");
     expect(types).toContain("cancelled");
@@ -212,12 +307,17 @@ describe("createPipelineSession — tool calls", () => {
     const types = eventTypes(client.events);
     expect(types).toEqual([
       "user_transcript",
-      "agent_transcript", // "Let me check"
       "tool_call",
       "tool_call_done",
-      "agent_transcript", // " — it's sunny."
+      "agent_transcript", // combined: "Let me check — it's sunny."
       "reply_done",
     ]);
+    expect(client.events.find((e) => (e as ClientEvent).type === "agent_transcript")).toMatchObject(
+      {
+        type: "agent_transcript",
+        text: "Let me check — it's sunny.",
+      },
+    );
 
     const toolCall = client.events.find((e) => (e as ClientEvent).type === "tool_call");
     expect(toolCall).toMatchObject({

package/host/pipeline-session.ts

@@ -10,7 +10,11 @@
 import type { LanguageModel, ModelMessage } from "ai";
 import { stepCountIs, streamText } from "ai";
 import type { AgentConfig, ExecuteTool, ToolSchema } from "../sdk/_internal-types.ts";
-import { DEFAULT_STT_SAMPLE_RATE, PIPELINE_FLUSH_TIMEOUT_MS } from "../sdk/constants.ts";
+import {
+  DEFAULT_STT_SAMPLE_RATE,
+  DEFAULT_TTS_SAMPLE_RATE,
+  PIPELINE_FLUSH_TIMEOUT_MS,
+} from "../sdk/constants.ts";
 import type { ClientSink, SessionErrorCode } from "../sdk/protocol.ts";
 import type {
   SttError,
@@ -55,8 +59,12 @@ export interface PipelineSessionOptions {
   sttApiKey: string;
   /** TTS API key. */
   ttsApiKey: string;
-  /** Audio sample rate (PCM16, Hz). Defaults to {@link DEFAULT_STT_SAMPLE_RATE}. */
-  sampleRate?: number | undefined;
+  /** STT audio sample rate (PCM16, Hz). Defaults to {@link DEFAULT_STT_SAMPLE_RATE}. */
+  sttSampleRate?: number | undefined;
+  /** TTS audio sample rate (PCM16, Hz). Must match the client's playback AudioContext rate. Defaults to {@link DEFAULT_TTS_SAMPLE_RATE}. */
+  ttsSampleRate?: number | undefined;
+  /** Skip the initial greeting audio on connect (used for session resume). */
+  skipGreeting?: boolean | undefined;
   /** Logger. Defaults to the console logger. */
   logger?: Logger | undefined;
   /** Sliding-window conversation history size. */
@@ -99,7 +107,6 @@ function handleStreamPart(
       if (delta.length === 0) return;
       deps.onTextDelta(delta);
       deps.tts?.sendText(delta);
-      deps.client.event({ type: "agent_transcript", text: delta });
       return;
     }
     case "tool-call": {
@@ -136,7 +143,8 @@ function handleStreamPart(
 /** Create a pluggable-provider voice session. */
 export function createPipelineSession(opts: PipelineSessionOptions): Session {
   const log = opts.logger ?? consoleLogger;
-  const sampleRate = opts.sampleRate ?? DEFAULT_STT_SAMPLE_RATE;
+  const sttSampleRate = opts.sttSampleRate ?? DEFAULT_STT_SAMPLE_RATE;
+  const ttsSampleRate = opts.ttsSampleRate ?? DEFAULT_TTS_SAMPLE_RATE;
   const { client, agentConfig, toolSchemas, executeTool } = opts;
 
   const hasTools = toolSchemas.length > 0 || (agentConfig.builtinTools?.length ?? 0) > 0;
@@ -342,6 +350,14 @@ export function createPipelineSession(opts: PipelineSessionOptions): Session {
       return;
     }
 
+    // Emit the complete transcript once the LLM finishes streaming, so the
+    // UI renders a single assistant message (vs. one per delta) and the user
+    // sees the text while TTS drains the synthesized audio.
+    if (accumulated.length > 0) {
+      client.event({ type: "agent_transcript", text: accumulated });
+      ctx.pushMessages({ role: "assistant", content: accumulated });
+    }
+
     await flushTtsAndWait(ctl.signal);
 
     if (ctl.signal.aborted) {
@@ -349,9 +365,29 @@ export function createPipelineSession(opts: PipelineSessionOptions): Session {
       return;
     }
 
-    if (accumulated.length > 0) {
-      ctx.pushMessages({ role: "assistant", content: accumulated });
+    client.playAudioDone();
+    client.event({ type: "reply_done" });
+    if (turnController === ctl) turnController = null;
+  }
+
+  async function runGreeting(text: string): Promise<void> {
+    const replyId = `pipeline-greeting-${++nextReplyId}`;
+    ctx.beginReply(replyId);
+
+    const ctl = new AbortController();
+    turnController = ctl;
+
+    client.event({ type: "agent_transcript", text });
+    ctx.pushMessages({ role: "assistant", content: text });
+    ctx.tts?.sendText(text);
+
+    await flushTtsAndWait(ctl.signal);
+
+    if (ctl.signal.aborted) {
+      if (turnController === ctl) turnController = null;
+      return;
     }
+
     client.playAudioDone();
     client.event({ type: "reply_done" });
     if (turnController === ctl) turnController = null;
@@ -394,13 +430,13 @@ export function createPipelineSession(opts: PipelineSessionOptions): Session {
   async function openProviders(): Promise<void> {
     const [sttResult, ttsResult] = await Promise.allSettled([
       opts.stt.open({
-        sampleRate,
+        sampleRate: sttSampleRate,
         apiKey: opts.sttApiKey,
         sttPrompt: agentConfig.sttPrompt,
         signal: sessionAbort.signal,
       }),
       opts.tts.open({
-        sampleRate,
+        sampleRate: ttsSampleRate,
         apiKey: opts.ttsApiKey,
         signal: sessionAbort.signal,
       }),
@@ -458,7 +494,15 @@ export function createPipelineSession(opts: PipelineSessionOptions): Session {
       ctx.stt?.sendAudio(pcm);
     },
     onAudioReady(): void {
+      if (audioReady || terminated) return;
       audioReady = true;
+      if (opts.skipGreeting) return;
+      const greeting = agentConfig.greeting;
+      if (!greeting) return;
+      const turn = runGreeting(greeting).catch((err: unknown) => {
+        log.error("Pipeline greeting failed", { error: errorMessage(err), sessionId: opts.id });
+      });
+      ctx.chainTurn(turn);
     },
     onCancel(): void {
       if (terminated) return;

package/host/providers/resolve.ts

@@ -77,7 +77,10 @@ export function resolveLlm(descriptor: LlmProvider, env: Record<string, string>)
       if (!apiKey) {
         throw new Error("Anthropic LLM: missing API key. Set ANTHROPIC_API_KEY in the agent env.");
       }
-      return createAnthropic({ apiKey })(options.model);
+      // Pass baseURL explicitly so the SDK's loadOptionalSetting returns
+      // before reading process.env["ANTHROPIC_BASE_URL"]. Without this,
+      // the Deno platform server needs --allow-env to start a session.
+      return createAnthropic({ apiKey, baseURL: "https://api.anthropic.com/v1" })(options.model);
     }
     default:
       throw new Error(

package/host/providers/tts/cartesia.test.ts

@@ -151,12 +151,15 @@ describe("cartesia TTS adapter", () => {
       },
     ]);
 
-    // After flush(), the adapter has rotated to a new context.
-    const turn2 = session._currentContextId();
-    expect(turn2).not.toBe(turn1);
+    // Rotation is deferred until the next sendText so Cartesia's late
+    // audio chunks + real `done` event (both tagged with turn1's id) still
+    // pass the context-id filter.
+    expect(session._currentContextId()).toBe(turn1);
 
-    // Subsequent sendText targets the new context.
+    // Subsequent sendText rotates to a fresh context.
     session.sendText("next");
+    const turn2 = session._currentContextId();
+    expect(turn2).not.toBe(turn1);
     await flush();
     expect(sends.filter((s) => s.contextId === turn2)).toEqual([
       {
@@ -201,8 +204,15 @@ describe("cartesia TTS adapter", () => {
       { kind: "cancel", contextId: turn1 },
     ]);
 
-    // Cancelling rotates the context so the next turn is unambiguous.
-    expect(session._currentContextId()).not.toBe(turn1);
+    // Rotation is deferred until the next sendText — cancel() halts the
+    // old context on Cartesia's side, so late events for turn1 can safely
+    // keep passing the filter until the next turn actually begins.
+    expect(session._currentContextId()).toBe(turn1);
+
+    // A subsequent sendText mints a fresh context for turn2.
+    session.sendText("again");
+    const turn2 = session._currentContextId();
+    expect(turn2).not.toBe(turn1);
 
     controller.abort();
     await session.close();

package/host/providers/tts/cartesia.ts

@@ -107,9 +107,21 @@ export function openCartesia(opts: CartesiaOptions): TtsOpener {
        * Reset whenever a fresh context is minted (i.e. at turn boundaries).
        */
       let doneEmitted = false;
+      /**
+       * After `flush()` or `cancel()`, the current context is done accepting
+       * input. We defer minting a fresh one until the next `sendText()` so
+       * that late audio chunks + Cartesia's real `done` event (both tagged
+       * with the flushed context's id) still pass the filter below. Rotating
+       * eagerly would silently drop all audio still in flight.
+       */
+      let rotatePending = false;
       const rotateContext = () => {
         context = mintContext();
         doneEmitted = false;
+        rotatePending = false;
+      };
+      const rotateIfPending = () => {
+        if (rotatePending) rotateContext();
       };
       const emitDoneOnce = () => {
         if (doneEmitted || closed) return;
@@ -179,33 +191,35 @@ export function openCartesia(opts: CartesiaOptions): TtsOpener {
       const session: CartesiaSession = {
         sendText(text: string) {
           if (closed || text.length === 0) return;
+          // First sendText after a flush/cancel starts a fresh context so
+          // we don't append to one that's already been finalized.
+          rotateIfPending();
           void context
             .send({ ...baseRequest, transcript: text, continue: true })
             .catch(ignoreRejection);
         },
         flush() {
-          if (closed) return;
+          if (closed || rotatePending) return;
           // Empty transcript with `continue: false` is the canonical
-          // end-of-turn signal. Cartesia replies with a `done` tagged
-          // by context_id, driving `emitDoneOnce`. The microtask
-          // fallback guards against a dropped server event wedging
-          // the orchestrator's state machine.
-          // TODO: drop the microtask fallback once we've verified
-          // Cartesia always emits `done` for cleanly-flushed contexts.
+          // end-of-turn signal. Cartesia finishes synthesizing whatever
+          // is queued and then emits a `done` tagged with the same
+          // context_id — at that point `emitDoneOnce` fires for real.
+          // Defer rotation so the filter below still accepts in-flight
+          // audio chunks and the real `done` event.
           void context
             .send({ ...baseRequest, transcript: "", continue: false })
             .catch(ignoreRejection);
-          queueMicrotask(emitDoneOnce);
-          rotateContext();
+          rotatePending = true;
         },
         cancel() {
           if (closed) return;
           void context.cancel().catch(ignoreRejection);
           // Emit synchronously: barge-in advances the orchestrator's
           // state machine on `done`, and delaying it would audibly
-          // stall subsequent turns.
+          // stall subsequent turns. Cartesia stops producing audio
+          // after cancel, so dropping any late chunks is fine.
           emitDoneOnce();
-          rotateContext();
+          rotatePending = true;
         },
         on(event, fn) {
           return emitter.on(event, fn);

package/host/runtime.ts

@@ -326,6 +326,9 @@ export function createRuntime(opts: RuntimeOptions): Runtime {
         tts: pipelineProviders.tts,
         sttApiKey: resolveApiKey("ASSEMBLYAI_API_KEY", env),
         ttsApiKey: resolveApiKey("CARTESIA_API_KEY", env),
+        sttSampleRate: s2sConfig.inputSampleRate,
+        ttsSampleRate: s2sConfig.outputSampleRate,
+        skipGreeting: sessionOpts.skipGreeting ?? false,
         logger,
       });
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@alexkroman1/aai",
-  "version": "1.4.0",
+  "version": "1.4.2",
   "type": "module",
   "exports": {
     ".": {