feat: integrate OCI agent streaming and web voice client

.gitignore

@@ -21,6 +21,7 @@ venv/
 .ruff_cache/
 .coverage
 htmlcov/
+config/oci-e2e.yaml
 
 # IDE and operating system
 .vscode/

README.md

@@ -18,6 +18,7 @@ A modular voice pipeline for streaming audio input, speech transcription, agent
 - [Testing](#testing)
 - [Project structure](#project-structure)
 - [Known limitations](#known-limitations)
+- [Architecture Backbone & Abstraction Boundaries](#architecture-backbone--abstraction-boundaries)
 
 ## Architecture
 
@@ -38,6 +39,212 @@ flowchart TB
 
 The **Audio Gateway** handles transport and protocol messages. **VoicePipeline** coordinates processing and task ownership. **SessionState** tracks session and audio sequence; **TurnManager** enforces turn transitions. A new STT adapter is created for each turn, while Agent and TTS adapters are created per session.
 
+## Architecture Backbone & Abstraction Boundaries
+
+AgentVoice follows a **provider-independent architecture** built around explicit contracts, lifecycle ownership, and asynchronous streaming.
+
+The architecture separates five concerns:
+
+- **Transport:** WebSocket communication and protocol serialization.
+- **Orchestration:** Coordination of STT, Agent, and TTS execution.
+- **Lifecycle:** Session state, turn transitions, interruption, and task ownership.
+- **Provider contracts:** Abstract interfaces defining component behavior.
+- **Provider implementations:** Concrete integrations with NVIDIA, OCI Generative AI, and VoxCPM.
+
+### Architectural Backbone
+
+```mermaid
+flowchart TB
+    Client["Web Client / Audio Client"]
+    Gateway["Audio Gateway<br/>FastAPI · WebSocket"]
+    Pipeline["VoicePipeline<br/>Orchestration & Streaming"]
+    Session["SessionState<br/>Session Lifecycle"]
+    Turns["TurnManager<br/>Semantic Turn Lifecycle"]
+
+    subgraph Contracts["Provider Abstraction Layer"]
+        direction LR
+        STTA["STTAdapter<br/>Abstract Contract"]
+        AgentA["AgentAdapter<br/>Abstract Contract"]
+        TTSA["TTSAdapter<br/>Abstract Contract"]
+    end
+
+    subgraph Implementations["Provider Implementation Layer"]
+        direction LR
+        Nvidia["NvidiaSTTAdapter"]
+        OCI["OCIAgentAdapter"]
+        Mock["MockAgentAdapter"]
+        Vox["VoxCPMTTSAdapter"]
+    end
+
+    subgraph External["External AI Services"]
+        ASR["NVIDIA ASR"]
+        LLM["OCI Generative AI"]
+        TTS["VoxCPM"]
+    end
+
+    Client <-->|"JSON Events + Binary PCM"| Gateway
+    Gateway --> Pipeline
+
+    Pipeline --- Session
+    Pipeline --- Turns
+
+    Pipeline --> STTA
+    Pipeline --> AgentA
+    Pipeline --> TTSA
+
+    STTA -.->|implements| Nvidia
+    AgentA -.->|implements| OCI
+    AgentA -.->|implements| Mock
+    TTSA -.->|implements| Vox
+
+    Nvidia --> ASR
+    OCI --> LLM
+    Vox --> TTS
+```
+
+The `VoicePipeline` depends on provider contracts rather than concrete implementations. Provider selection is performed through configuration and dynamic factories.
+
+This allows implementations to evolve without requiring the Audio Gateway to understand provider-specific APIs.
+
+### Provider Contracts
+
+| Abstraction | Contract | Implementations | Responsibility |
+|---|---|---|---|
+| Speech-to-text | `STTAdapter` | `NvidiaSTTAdapter`, `MockSTTAdapter` | Audio ingestion and transcript events |
+| Agent | `AgentAdapter` | `OCIAgentAdapter`, `MockAgentAdapter` | Text processing and conversational responses |
+| Text-to-speech | `TTSAdapter` | `VoxCPMTTSAdapter`, `MockTTSAdapter` | Speech synthesis and audio streaming |
+
+The `AgentAdapter` defines the common text-processing interface. OCI-specific incremental generation is currently exposed through an additional `stream()` method, with streaming capability detected by the orchestration layer.
+
+### Runtime Composition
+
+Provider implementations are selected through environment configuration and resolved by the provider factory.
+
+```mermaid
+flowchart LR
+    Config["Environment YAML"]
+    Catalog["Provider Catalog"]
+    Factory["Provider Factories"]
+    Contracts["Adapter Contracts"]
+    Runtime["VoicePipeline"]
+
+    Config --> Factory
+    Catalog --> Factory
+    Factory -->|Validate implementation| Contracts
+    Factory -->|Instantiate providers| Runtime
+```
+
+The primary configuration components are:
+
+| Component | Location |
+|---|---|
+| Environment selection | `config/loader.py` |
+| Typed configuration | `config/models.py` |
+| Provider catalog | `config/providers.yaml` |
+| Dynamic provider loading | `providers.py` |
+| Session lifecycle | `audio_gateway/app/session.py` |
+| Semantic turn lifecycle | `turn_manager.py` |
+| Execution orchestration | `voice_pipeline.py` |
+
+### Streaming Execution Backbone
+
+The streaming pipeline allows OCI Generative AI to continue producing text while VoxCPM synthesizes previously completed text segments.
+
+```mermaid
+flowchart TB
+    Input["Microphone / PCM Input"]
+    Gateway["Audio Gateway"]
+    STT["NVIDIA STT"]
+    Transcript["Final Transcript"]
+
+    subgraph Pipeline["VoicePipeline · Concurrent Execution"]
+        direction TB
+        Agent["OCIAgentAdapter.stream()"]
+        Bridge["OCI SSE Async Bridge"]
+        Chunker["TextChunker"]
+        Queue["Bounded Segment Queue"]
+        Synth["VoxCPM TTS Streaming"]
+
+        Agent --> Bridge
+        Bridge --> Chunker
+        Chunker --> Queue
+        Queue --> Synth
+    end
+
+    OCI["OCI Generative AI"]
+    Vox["VoxCPM Service"]
+    Output["WebSocket PCM Output"]
+
+    Input --> Gateway
+    Gateway --> STT
+    STT --> Transcript
+    Transcript --> Agent
+
+    Agent <--> OCI
+    Synth <--> Vox
+    Synth --> Gateway
+    Gateway --> Output
+```
+
+The pipeline uses bounded queues and asynchronous task coordination to support incremental generation, ordered synthesis, and cooperative cancellation.
+
+### Lifecycle Ownership
+
+AgentVoice separates semantic state from asynchronous execution.
+
+```mermaid
+flowchart LR
+    Listening["LISTENING"]
+    Processing["PROCESSING"]
+    Speaking["SPEAKING"]
+    Completed["COMPLETED"]
+    Interrupted["INTERRUPTED"]
+
+    Listening -->|"STT final"| Processing
+    Processing -->|"First TTS segment"| Speaking
+    Speaking -->|"Synthesis complete"| Completed
+
+    Listening --> Interrupted
+    Processing --> Interrupted
+    Speaking --> Interrupted
+```
+
+**Ownership responsibilities:**
+
+| Component | Owns |
+|---|---|
+| Audio Gateway | WebSocket connection, protocol messages, provider cleanup coordination |
+| SessionState | Session lifecycle and audio sequence validation |
+| TurnManager | Semantic turn state transitions |
+| VoicePipeline | Turn execution task and Agent/TTS orchestration |
+| OCI streaming bridge | SSE reader, event queue, and stream closure |
+| Provider adapters | Provider-specific requests, connections, and data formats |
+
+Cancellation is cooperative. The system requests cancellation and waits for task termination, but termination of blocking external operations depends on the underlying provider and transport behavior.
+
+### Architectural Extension Points
+
+The abstraction layer allows additional implementations to be introduced through the provider catalog.
+
+For example, a future LiteLLM integration could implement the Agent contract without requiring a redesign of the Gateway or the semantic turn lifecycle.
+
+```mermaid
+flowchart TB
+    Pipeline["VoicePipeline"]
+    Contract["AgentAdapter"]
+
+    OCI["OCIAgentAdapter<br/>Current"]
+    Mock["MockAgentAdapter<br/>Testing"]
+    Lite["LiteLLMAgentAdapter<br/>Future"]
+
+    Pipeline --> Contract
+    Contract -.-> OCI
+    Contract -.-> Mock
+    Contract -.-> Lite
+```
+
+**Design principle:** transport, orchestration, lifecycle, and provider-specific integrations remain separate concerns. Provider independence is established through contracts and factories, while streaming and cancellation behavior are validated at their respective integration boundaries.
+
 ## Quickstart
 
 ### Requirements
@@ -197,3 +404,5 @@ tools/                          Integration and smoke scripts
 ---
 
 **Status:** V1 development baseline · Python 3.12+ · WebSocket protocol 1.0
+
+

config/providers.yaml

@@ -27,3 +27,9 @@ providers:
         realtime: false
         streaming: false
 
+    oci:
+      name: "OCI Generative AI"
+      implementation: "voice_assistant.agent.oci:OCIAgentAdapter"
+      capabilities:
+        realtime: true
+        streaming: true

src/voice_assistant/agent/oci.py

@@ -10,7 +10,6 @@ from __future__ import annotations
 import json
 from typing import Any
 
-import asyncio
 import logging
 from collections.abc import AsyncIterator
 from contextlib import aclosing
@@ -19,6 +18,7 @@ import oci
 
 from voice_assistant.agent.adapter import AgentAdapter
 from voice_assistant.agent.oci_stream import stream_sse_events
+from voice_assistant.agent.oci_connection import open_oci_stream
 
 
 def extract_text_delta(raw_data: str) -> str | None:
@@ -205,9 +205,8 @@ class OCIAgentAdapter(AgentAdapter):
             self.model_id,
         )
 
-        response = await asyncio.to_thread(
+        response = await open_oci_stream(
-            self.client.chat,
+            lambda: self.client.chat(details)
-            details,
         )
 
         if response.status != 200:
@@ -256,4 +255,3 @@ class OCIAgentAdapter(AgentAdapter):
             )
 
         return response_text
-

src/voice_assistant/agent/oci_connection.py

@@ -0,0 +1,107 @@
+"""
+Lifecycle management for blocking OCI HTTP connection setup.
+
+Ensures that responses arriving after coroutine cancellation
+are closed instead of being abandoned.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import threading
+from typing import Any, Callable
+
+
+logger = logging.getLogger(__name__)
+
+
+def _close_response(response: Any) -> None:
+    """Close the SSE response returned by the OCI SDK."""
+    stream = getattr(response, "data", None)
+
+    if stream is not None:
+        close = getattr(stream, "close", None)
+
+        if callable(close):
+            close()
+
+
+async def open_oci_stream(
+    open_connection: Callable[[], Any],
+) -> Any:
+    """
+    Open a blocking OCI connection with explicit response ownership.
+
+    The worker retains ownership until the async caller claims
+    the response. Cancellation closes any unclaimed response.
+
+    A blocking SDK call may continue after cancellation.
+    """
+    ownership_lock = threading.Lock()
+    abandoned = False
+    claimed = False
+    response = None
+
+    def worker() -> Any:
+        nonlocal response
+
+        opened = open_connection()
+
+        with ownership_lock:
+            if abandoned:
+                should_close = True
+            else:
+                response = opened
+                should_close = False
+
+        if should_close:
+            try:
+                _close_response(opened)
+            except Exception:
+                logger.exception(
+                    "Failed to close abandoned OCI response."
+                )
+
+        return opened
+
+    task = asyncio.create_task(
+        asyncio.to_thread(worker)
+    )
+
+    try:
+        opened = await asyncio.shield(task)
+
+        with ownership_lock:
+            if abandoned:
+                raise RuntimeError(
+                    "OCI connection was abandoned."
+                )
+
+            claimed = True
+
+        return opened
+
+    except asyncio.CancelledError:
+        with ownership_lock:
+            abandoned = True
+
+            if not claimed:
+                unclaimed = response
+                response = None
+            else:
+                unclaimed = None
+
+        if unclaimed is not None:
+            try:
+                await asyncio.to_thread(
+                    _close_response,
+                    unclaimed,
+                )
+            except Exception:
+                logger.exception(
+                    "Failed to close unclaimed OCI response."
+                )
+
+        raise
+

src/voice_assistant/agent/text_chunker.py

@@ -0,0 +1,85 @@
+"""
+Incremental text segmentation for streaming voice synthesis.
+
+Converts LLM text deltas into speech-ready segments while
+preserving word boundaries and punctuation.
+"""
+
+from __future__ import annotations
+
+
+class TextChunker:
+    """
+    Accumulate text deltas and emit segments suitable for TTS.
+
+    Segments are emitted on sentence-ending punctuation or
+    at a configurable maximum length.
+    """
+
+    def __init__(
+        self,
+        min_chars: int = 20,
+        max_chars: int = 160,
+    ) -> None:
+        if min_chars < 1:
+            raise ValueError("min_chars must be positive.")
+
+        if max_chars < min_chars:
+            raise ValueError(
+                "max_chars must be greater than or equal to min_chars."
+            )
+
+        self.min_chars = min_chars
+        self.max_chars = max_chars
+        self._buffer = ""
+
+    def feed(self, delta: str) -> list[str]:
+        """Accept a text delta and return completed segments."""
+        self._buffer += delta
+        segments: list[str] = []
+
+        while len(self._buffer) >= self.min_chars:
+            boundary = self._find_boundary()
+
+            if boundary is None:
+                break
+
+            segment = self._buffer[:boundary].strip()
+            self._buffer = self._buffer[boundary:]
+
+            if segment:
+                segments.append(segment)
+
+        return segments
+
+    def _find_boundary(self) -> int | None:
+        """Find a sentence boundary or safe length boundary."""
+        for index, char in enumerate(self._buffer):
+            if char in ".!?":
+                boundary = index + 1
+
+                if boundary >= self.min_chars:
+                    return boundary
+
+        if len(self._buffer) < self.max_chars:
+            return None
+
+        # Prefer breaking at a word boundary.
+        boundary = self._buffer.rfind(
+            " ",
+            0,
+            self.max_chars + 1,
+        )
+
+        if boundary >= self.min_chars:
+            return boundary + 1
+
+        return self.max_chars
+
+    def flush(self) -> str | None:
+        """Return any remaining text after generation completes."""
+        remaining = self._buffer.strip()
+        self._buffer = ""
+
+        return remaining or None
+

src/voice_assistant/audio_gateway/app/main.py

@@ -34,8 +34,13 @@ Architecture:
 
 import json
 import logging
+from pathlib import Path
+
+from fastapi.responses import FileResponse
+from fastapi.staticfiles import StaticFiles
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect
 from pydantic import BaseModel, ValidationError
+from contextlib import aclosing
 
 import asyncio
 
@@ -83,8 +88,16 @@ from voice_assistant.audio_gateway.app.protocol import (
     STTPartial,
     TurnEnded,
     TurnStarted,
+    AgentDelta,
 )
 
+from voice_assistant.voice_pipeline import (
+    PipelineTextDelta,
+    PipelineAgentResponse,
+    PipelineAudioChunk,
+)
+
+
 from voice_assistant.audio_gateway.app.session import SessionState
 
 # logger configuration
@@ -100,6 +113,18 @@ provider_catalog = load_provider_catalog()
 
 MAX_AUDIO_CHUNK_BYTES = 64 * 1024 # 64 KB
 
+WEB_DIR = Path(__file__).resolve().parents[1] / "web"
+
+app.mount(
+    "/web",
+    StaticFiles(directory=WEB_DIR),
+    name="web",
+)
+
+
+@app.get("/ui", include_in_schema=False)
+async def web_client() -> FileResponse:
+    return FileResponse(WEB_DIR / "index.html")
 
 
 def create_turn_manager() -> TurnManager:
@@ -176,6 +201,42 @@ async def forward_stt_events(
                 ),
             )
 
+            if callable(getattr(pipeline.agent, "stream", None)):
+                async with aclosing(
+                    pipeline.stream_turn_response(
+                        session_id=session_id,
+                        turn_id=turn_id,
+                        text=event.text,
+                    )
+                ) as response_events:
+                    async for response_event in response_events:
+                        if not pipeline.is_active_turn(turn_id):
+                            break
+
+                        if isinstance(response_event, PipelineTextDelta):
+                            await send_json(
+                                websocket,
+                                AgentDelta(
+                                    session_id=session_id,
+                                    turn_id=turn_id,
+                                    text=response_event.text,
+                                ),
+                            )
+
+                        elif isinstance(response_event, PipelineAgentResponse):
+                            await send_json(
+                                websocket,
+                                AgentResponse(
+                                    session_id=session_id,
+                                    turn_id=turn_id,
+                                    text=response_event.text,
+                                ),
+                            )
+
+                        elif isinstance(response_event, PipelineAudioChunk):
+                            await websocket.send_bytes(response_event.audio)
+
+            else:
                 result = await pipeline.process_transcript(
                     session_id=session_id,
                     turn_id=turn_id,

src/voice_assistant/audio_gateway/app/protocol.py

@@ -109,6 +109,20 @@ class STTFinal(BaseModel):
     text: str
 
 
+class AgentDelta(BaseModel):
+    """
+    Incremental text generated by the configured Agent.
+
+    Each event belongs to one session and turn.
+    Text fragments must be concatenated in arrival order.
+    """
+
+    type: Literal["agent.delta"] = "agent.delta"
+    session_id: str
+    turn_id: int
+    text: str
+
+
 class AgentResponse(BaseModel):
     """Text response produced by the configured Agent."""
 

src/voice_assistant/audio_gateway/web/app.js

@@ -0,0 +1,387 @@
+
+const $ = (id) => document.getElementById(id);
+
+const status = $("status");
+const talkButton = $("talkButton");
+const talkLabel = $("talkLabel");
+const connectButton = $("connectButton");
+const interruptButton = $("interruptButton");
+const messages = $("messages");
+
+let ws = null;
+let sessionId = null;
+let turnId = null;
+let sequence = 0;
+let phase = "disconnected";
+
+let microphone = null;
+let inputContext = null;
+let inputSource = null;
+let inputNode = null;
+let silentGain = null;
+
+let playbackContext = null;
+let playbackTime = 0;
+const scheduledAudio = new Set();
+
+let userMessage = null;
+let agentMessage = null;
+let pendingFinal = null;
+
+function setPhase(next) {
+  phase = next;
+
+  status.textContent = {
+    disconnected: "Desconectado",
+    connecting: "Conectando",
+    ready: "Conectado",
+    starting: "Iniciando turno",
+    recording: "Gravando",
+    processing: "Processando",
+    speaking: "Respondendo"
+  }[next] || next;
+
+  status.classList.toggle(
+    "connected",
+    next !== "disconnected" && next !== "connecting"
+  );
+
+  connectButton.disabled = next !== "disconnected";
+  talkButton.disabled = next !== "ready" && next !== "recording";
+  interruptButton.disabled = turnId === null;
+
+  talkButton.classList.toggle("recording", next === "recording");
+  talkLabel.textContent = next === "recording"
+    ? "Clique para terminar de falar"
+    : next === "ready"
+      ? "Clique para falar"
+      : status.textContent;
+}
+
+function addMessage(role, text = "") {
+  $("emptyState")?.remove();
+
+  const element = document.createElement("div");
+  element.className = `message ${role}`;
+
+  const label = document.createElement("small");
+  label.textContent = role === "user" ? "Você" : "AgentVoice";
+
+  const content = document.createElement("span");
+  content.textContent = text;
+
+  element.append(label, content);
+  messages.appendChild(element);
+  messages.scrollTop = messages.scrollHeight;
+
+  return content;
+}
+
+function sendJson(payload) {
+  if (ws?.readyState !== WebSocket.OPEN) {
+    throw new Error("WebSocket não está conectado.");
+  }
+  ws.send(JSON.stringify(payload));
+}
+
+async function ensurePlayback() {
+  if (!playbackContext) {
+    playbackContext = new AudioContext();
+  }
+  await playbackContext.resume();
+}
+
+async function playPCM(bytes) {
+  await ensurePlayback();
+
+  const rate = Number($("outputRate").value);
+  const sampleCount = Math.floor(bytes.byteLength / 2);
+  if (!sampleCount) return;
+
+  const view = new DataView(bytes);
+  const samples = new Float32Array(sampleCount);
+
+  for (let i = 0; i < sampleCount; i++) {
+    samples[i] = view.getInt16(i * 2, true) / 32768;
+  }
+
+  const audio = playbackContext.createBuffer(
+    1,
+    sampleCount,
+    rate
+  );
+  audio.copyToChannel(samples, 0);
+
+  const source = playbackContext.createBufferSource();
+  source.buffer = audio;
+  source.connect(playbackContext.destination);
+
+  const startAt = Math.max(
+    playbackContext.currentTime + 0.025,
+    playbackTime
+  );
+
+  source.start(startAt);
+  playbackTime = startAt + audio.duration;
+
+  scheduledAudio.add(source);
+  source.onended = () => scheduledAudio.delete(source);
+}
+
+function stopPlayback() {
+  for (const source of scheduledAudio) {
+    try { source.stop(); } catch {}
+  }
+  scheduledAudio.clear();
+  playbackTime = 0;
+}
+
+async function stopMicrophone() {
+  if (inputNode) {
+    inputNode.port.onmessage = null;
+    inputNode.disconnect();
+    inputNode = null;
+  }
+
+  inputSource?.disconnect();
+  silentGain?.disconnect();
+
+  inputSource = null;
+  silentGain = null;
+
+  microphone?.getTracks().forEach((track) => track.stop());
+  microphone = null;
+
+  if (inputContext) {
+    const context = inputContext;
+    inputContext = null;
+    await context.close();
+  }
+}
+
+async function startMicrophone() {
+  microphone = await navigator.mediaDevices.getUserMedia({
+    audio: {
+      channelCount: 1,
+      echoCancellation: true,
+      noiseSuppression: true,
+      autoGainControl: true
+    }
+  });
+
+  inputContext = new AudioContext();
+  await inputContext.audioWorklet.addModule("/web/pcm-worklet.js");
+
+  inputSource = inputContext.createMediaStreamSource(microphone);
+  inputNode = new AudioWorkletNode(inputContext, "pcm-input");
+
+  inputNode.port.onmessage = ({ data }) => {
+    if (
+      data.type !== "pcm" ||
+      phase !== "recording" ||
+      turnId === null
+    ) return;
+
+    if (ws?.readyState !== WebSocket.OPEN) return;
+
+    sendJson({
+      type: "audio.chunk",
+      turn_id: turnId,
+      sequence_number: sequence++
+    });
+
+    ws.send(data.buffer);
+  };
+
+  silentGain = inputContext.createGain();
+  silentGain.gain.value = 0;
+
+  inputSource.connect(inputNode);
+  inputNode.connect(silentGain);
+  silentGain.connect(inputContext.destination);
+
+  await inputContext.resume();
+}
+
+async function beginTurn() {
+  if (phase !== "ready") return;
+
+  try {
+    await ensurePlayback();
+    stopPlayback();
+
+    setPhase("starting");
+    sendJson({ type: "audio.start" });
+  } catch (error) {
+    console.error(error);
+    setPhase("ready");
+  }
+}
+
+async function endRecording() {
+  if (phase !== "recording") return;
+
+  const activeTurn = turnId;
+  setPhase("processing");
+
+  await stopMicrophone();
+
+  if (activeTurn !== null && ws?.readyState === WebSocket.OPEN) {
+    sendJson({ type: "audio.stop", turn_id: activeTurn });
+  }
+}
+
+async function interruptTurn() {
+  if (turnId === null) return;
+
+  const activeTurn = turnId;
+
+  setPhase("processing");
+  await stopMicrophone();
+  stopPlayback();
+
+  if (ws?.readyState === WebSocket.OPEN) {
+    sendJson({ type: "interrupt", turn_id: activeTurn });
+  }
+}
+
+async function handleEvent(event) {
+  switch (event.type) {
+    case "session.ready":
+      sessionId = event.session_id;
+      setPhase("ready");
+      break;
+
+    case "turn.started":
+      turnId = event.turn_id;
+      sequence = 0;
+      userMessage = addMessage("user");
+      agentMessage = addMessage("agent");
+      pendingFinal = null;
+
+      try {
+        await startMicrophone();
+        setPhase("recording");
+      } catch (error) {
+        console.error(error);
+        talkLabel.textContent = "Não foi possível acessar o microfone.";
+        await interruptTurn();
+      }
+      break;
+
+    case "stt.partial":
+      if (event.turn_id === turnId && userMessage) {
+        userMessage.textContent = event.text;
+      }
+      break;
+
+    case "stt.final":
+      if (event.turn_id === turnId && userMessage) {
+        userMessage.textContent = event.text;
+      }
+      break;
+
+    case "agent.delta":
+      if (event.turn_id === turnId && agentMessage) {
+        agentMessage.textContent += event.text;
+      }
+      break;
+
+    case "agent.response":
+      if (event.turn_id === turnId && agentMessage) {
+        agentMessage.textContent = event.text;
+      }
+      break;
+
+    case "turn.ended":
+      if (event.turn_id === turnId) {
+        turnId = null;
+        await stopMicrophone();
+        setPhase("ready");
+      }
+      break;
+
+    case "interrupted":
+      if (event.turn_id === turnId) {
+        turnId = null;
+        await stopMicrophone();
+        stopPlayback();
+        setPhase("ready");
+      }
+      break;
+
+    case "error":
+      console.error("Gateway error:", event);
+      talkLabel.textContent =
+        `Erro: ${event.code} — ${event.message}`;
+
+      if (phase === "starting") {
+        setPhase("ready");
+      } else if (turnId !== null && phase !== "disconnected") {
+        await interruptTurn();
+      }
+      break;
+
+    case "session.ended":
+      ws?.close();
+      break;
+  }
+}
+
+async function connect() {
+  if (ws && ws.readyState !== WebSocket.CLOSED) return;
+
+  setPhase("connecting");
+
+  const scheme = location.protocol === "https:" ? "wss" : "ws";
+  ws = new WebSocket(`${scheme}://${location.host}/ws/audio`);
+  ws.binaryType = "arraybuffer";
+
+  // Serialize event handling to preserve ordering.
+  let incoming = Promise.resolve();
+
+  ws.onopen = () => {
+    sendJson({
+      type: "session.start",
+      audio: {
+        encoding: "pcm_s16le",
+        sample_rate: 16000,
+        channels: 1
+      }
+    });
+  };
+
+  ws.onmessage = ({ data }) => {
+    incoming = incoming.then(async () => {
+      if (data instanceof ArrayBuffer) {
+        if (turnId !== null) {
+          await playPCM(data);
+        }
+      } else {
+        await handleEvent(JSON.parse(data));
+      }
+    }).catch(console.error);
+  };
+
+  ws.onerror = (error) => console.error("WebSocket error", error);
+
+  ws.onclose = async () => {
+    await stopMicrophone();
+    stopPlayback();
+    ws = null;
+    sessionId = null;
+    turnId = null;
+    setPhase("disconnected");
+  };
+}
+
+connectButton.addEventListener("click", connect);
+
+talkButton.addEventListener("click", async () => {
+  if (phase === "ready") await beginTurn();
+  else if (phase === "recording") await endRecording();
+});
+
+interruptButton.addEventListener("click", interruptTurn);
+
+setPhase("disconnected");

src/voice_assistant/audio_gateway/web/index.html

@@ -0,0 +1,83 @@
+
+<!doctype html>
+<html lang="pt-BR">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport"
+        content="width=device-width, initial-scale=1">
+  <title>AgentVoice · Voice Console</title>
+  <link rel="stylesheet" href="/web/styles.css">
+</head>
+<body>
+  <main class="console">
+    <header class="header">
+      <div class="brand">
+        <div class="brand-icon">◉</div>
+        <div>
+          <h1>AgentVoice</h1>
+          <p>Realtime Voice Console</p>
+        </div>
+      </div>
+      <span id="status" class="status">Desconectado</span>
+    </header>
+
+    <section class="hero">
+      <span class="eyebrow">VOICE ASSISTANT · V1</span>
+      <h2>Uma conversa, naturalmente.</h2>
+      <p>Fale pelo microfone e receba respostas
+         geradas por inteligência artificial.</p>
+    </section>
+
+    <section id="messages"
+             class="messages"
+             aria-live="polite">
+      <div class="empty-state" id="emptyState">
+        <div class="empty-icon">◎</div>
+        <p>Sua conversa aparecerá aqui.</p>
+      </div>
+    </section>
+
+    <section class="controls">
+      <div class="mic-container">
+        <button id="talkButton"
+                class="talk-button"
+                disabled
+                aria-label="Iniciar gravação">
+          <span id="micIcon">🎙</span>
+        </button>
+        <p id="talkLabel">Conecte para começar</p>
+      </div>
+
+      <div class="actions">
+        <button id="connectButton" class="primary">
+          Conectar
+        </button>
+        <button id="interruptButton"
+                class="secondary" disabled>
+          Interromper
+        </button>
+      </div>
+
+      <div class="settings">
+        <label for="outputRate">
+          Frequência de reprodução PCM
+        </label>
+        <select id="outputRate">
+          <option value="24000">24 kHz</option>
+          <option value="16000">16 kHz</option>
+          <option value="32000">32 kHz</option>
+          <option value="48000">48 kHz</option>
+        </select>
+      </div>
+    </section>
+
+    <footer>
+      <span>NVIDIA STT</span>
+      <span>OCI Generative AI</span>
+      <span>VoxCPM TTS</span>
+    </footer>
+  </main>
+
+  <script type="module" src="/web/app.js"></script>
+</body>
+</html>

src/voice_assistant/audio_gateway/web/pcm-worklet.js

@@ -0,0 +1,74 @@
+
+class PCMInputProcessor extends AudioWorkletProcessor {
+  constructor() {
+    super();
+    this.targetRate = 16000;
+    this.step = sampleRate / this.targetRate;
+    this.buffer = [];
+    this.position = 0;
+    this.sequence = 0;
+    this.chunkSamples = 1600;
+    this.output = [];
+  }
+
+  process(inputs) {
+    const channels = inputs[0];
+
+    if (!channels || channels.length === 0) {
+      return true;
+    }
+
+    const frameLength = channels[0].length;
+
+    for (let i = 0; i < frameLength; i++) {
+      let sample = 0;
+
+      for (const channel of channels) {
+        sample += channel[i];
+      }
+
+      this.buffer.push(sample / channels.length);
+    }
+
+    while (this.position + 1 < this.buffer.length) {
+      const index = Math.floor(this.position);
+      const fraction = this.position - index;
+
+      const sample =
+        this.buffer[index] * (1 - fraction) +
+        this.buffer[index + 1] * fraction;
+
+      const clipped = Math.max(-1, Math.min(1, sample));
+
+      this.output.push(
+        clipped < 0
+          ? Math.round(clipped * 32768)
+          : Math.round(clipped * 32767)
+      );
+
+      this.position += this.step;
+
+      if (this.output.length >= this.chunkSamples) {
+        const pcm = new Int16Array(this.output);
+        this.output = [];
+
+        this.port.postMessage({
+          type: "pcm",
+          sequence: this.sequence++,
+          buffer: pcm.buffer
+        }, [pcm.buffer]);
+      }
+    }
+
+    const consumed = Math.floor(this.position);
+
+    if (consumed > 0) {
+      this.buffer.splice(0, consumed);
+      this.position -= consumed;
+    }
+
+    return true;
+  }
+}
+
+registerProcessor("pcm-input", PCMInputProcessor);

src/voice_assistant/audio_gateway/web/styles.css

@@ -0,0 +1,164 @@
+
+:root {
+  font-family: Inter, system-ui, -apple-system, sans-serif;
+  color: #e8edf6;
+  background: #090e18;
+  font-synthesis: none;
+}
+
+* { box-sizing: border-box; }
+
+body {
+  margin: 0;
+  min-height: 100vh;
+  display: grid;
+  place-items: center;
+  padding: 24px;
+  background:
+    radial-gradient(ellipse at 50% 0%, #17294b, transparent 65%),
+    #090e18;
+}
+
+.console {
+  width: min(100%, 820px);
+  background: #111a2a;
+  border: 1px solid #26334a;
+  border-radius: 22px;
+  overflow: hidden;
+  box-shadow: 0 30px 90px #0006;
+}
+
+.header, .brand, .actions, footer {
+  display: flex;
+  align-items: center;
+}
+
+.header {
+  justify-content: space-between;
+  padding: 22px 28px;
+  border-bottom: 1px solid #26334a;
+}
+
+.brand { gap: 12px; }
+.brand-icon { color: #67e8f9; font-size: 34px; }
+h1 { font-size: 19px; margin: 0; }
+.brand p, .hero p, .empty-state, footer {
+  color: #93a4bd;
+}
+.brand p { margin: 3px 0 0; font-size: 12px; }
+
+.status {
+  font-size: 12px;
+  border: 1px solid #42506a;
+  border-radius: 20px;
+  padding: 7px 12px;
+}
+
+.status.connected {
+  color: #86efac;
+  border-color: #166534;
+}
+
+.hero { text-align: center; padding: 36px 24px 24px; }
+.eyebrow {
+  color: #67e8f9;
+  font-size: 11px;
+  letter-spacing: 2px;
+}
+h2 { font-size: clamp(23px, 4vw, 32px); margin: 14px 0; }
+.hero p { font-size: 14px; line-height: 1.6; }
+
+.messages {
+  height: 310px;
+  overflow-y: auto;
+  padding: 24px;
+  background: #0c1422;
+  border-block: 1px solid #26334a;
+}
+
+.empty-state { text-align: center; padding: 70px 0; }
+.empty-icon { font-size: 40px; color: #526782; }
+
+.message {
+  margin: 12px 0;
+  padding: 14px 18px;
+  border-radius: 14px;
+  max-width: 88%;
+  line-height: 1.55;
+  font-size: 14px;
+  white-space: pre-wrap;
+  overflow-wrap: anywhere;
+}
+
+.message.user { background: #20385b; margin-left: auto; }
+.message.agent { background: #1b2b36; margin-right: auto; }
+.message small {
+  display: block;
+  color: #91a8c5;
+  font-size: 11px;
+  margin-bottom: 5px;
+}
+
+.controls { padding: 24px; text-align: center; }
+.mic-container { padding: 8px 0 20px; }
+
+.talk-button {
+  width: 88px;
+  height: 88px;
+  border-radius: 50%;
+  border: 1px solid #3a6a9b;
+  background: linear-gradient(145deg, #23598c, #183755);
+  color: white;
+  font-size: 30px;
+  cursor: pointer;
+  transition: transform .15s, box-shadow .15s;
+}
+
+.talk-button:hover:not(:disabled) { transform: scale(1.04); }
+.talk-button.recording {
+  background: #9f2442;
+  border-color: #fb7185;
+  box-shadow: 0 0 0 12px #fb71851a;
+}
+
+#talkLabel { color: #a4b4ca; font-size: 13px; }
+.actions { justify-content: center; gap: 12px; }
+
+.actions button, select {
+  padding: 11px 20px;
+  border-radius: 9px;
+  border: 1px solid #344862;
+  color: #e8edf6;
+  background: #1b2b42;
+  cursor: pointer;
+}
+
+.actions .primary { background: #2563eb; border-color: #2563eb; }
+button:disabled { opacity: .45; cursor: not-allowed; }
+
+.settings {
+  display: flex;
+  justify-content: center;
+  align-items: center;
+  flex-wrap: wrap;
+  gap: 12px;
+  margin-top: 22px;
+  font-size: 12px;
+  color: #93a4bd;
+}
+
+footer {
+  justify-content: center;
+  flex-wrap: wrap;
+  gap: 24px;
+  border-top: 1px solid #26334a;
+  padding: 16px;
+  font-size: 11px;
+}
+
+@media (max-width: 520px) {
+  body { padding: 0; }
+  .console { min-height: 100vh; border-radius: 0; }
+  .header { padding: 18px; }
+  .messages { height: 36vh; }
+}

src/voice_assistant/voice_pipeline.py

@@ -41,7 +41,10 @@ import asyncio
 
 from collections.abc import AsyncIterator
 from dataclasses import dataclass
+from contextlib import aclosing
 
+from voice_assistant.agent.text_chunker import TextChunker
+from voice_assistant.turn_manager import TurnState
 from voice_assistant.agent.adapter import AgentAdapter
 from voice_assistant.audio_gateway.app.session import SessionState
 from voice_assistant.turn_manager import Turn, TurnManager
@@ -64,6 +67,21 @@ class PipelineResult:
     response_text: str
 
 
+@dataclass(slots=True)
+class PipelineTextDelta:
+    text: str
+
+
+@dataclass(slots=True)
+class PipelineAgentResponse:
+    text: str
+
+
+@dataclass(slots=True)
+class PipelineAudioChunk:
+    audio: bytes
+
+
 class VoicePipeline:
     """
     Coordinate session, turn, Agent, and TTS processing.
@@ -168,6 +186,184 @@ class VoicePipeline:
         ):
             yield audio_chunk
 
+
+    async def stream_turn_response(
+        self,
+        session_id: str,
+        turn_id: int,
+        text: str,
+    ) -> AsyncIterator[
+        PipelineTextDelta | PipelineAgentResponse | PipelineAudioChunk
+    ]:
+        """
+        Stream Agent text and TTS audio concurrently.
+
+        The producer generates text segments while the consumer
+        synthesizes completed segments in their original order.
+        """
+        turn = self.turn_manager.active_turn
+
+        if turn is None or turn.turn_id != turn_id:
+            raise ValueError(f"Unexpected turn_id={turn_id}.")
+
+        if turn.state != TurnState.PROCESSING:
+            raise RuntimeError(
+                f"Cannot stream response while turn is {turn.state.value}."
+            )
+
+        stream_method = getattr(self.agent, "stream", None)
+
+        if not callable(stream_method):
+            raise RuntimeError(
+                "Configured Agent does not support streaming."
+            )
+
+        events: asyncio.Queue[
+            PipelineTextDelta | PipelineAgentResponse |
+            PipelineAudioChunk | None
+        ] = asyncio.Queue(maxsize=16)
+
+        segments: asyncio.Queue[str | None] = asyncio.Queue(maxsize=4)
+
+        async def produce() -> None:
+            chunker = TextChunker()
+            parts: list[str] = []
+
+            async with aclosing(
+                stream_method(
+                    session_id=session_id,
+                    turn_id=turn_id,
+                    text=text,
+                )
+            ) as deltas:
+                async for delta in deltas:
+                    parts.append(delta)
+
+                    await events.put(
+                        PipelineTextDelta(text=delta)
+                    )
+
+                    for segment in chunker.feed(delta):
+                        await segments.put(segment)
+
+            remaining = chunker.flush()
+
+            if remaining:
+                await segments.put(remaining)
+
+            response_text = "".join(parts).strip()
+
+            if not response_text:
+                raise RuntimeError(
+                    "Agent returned an empty streaming response."
+                )
+
+            await events.put(
+                PipelineAgentResponse(text=response_text)
+            )
+
+            await segments.put(None)
+
+        async def consume() -> None:
+            first_segment = True
+
+            while True:
+                segment = await segments.get()
+
+                if segment is None:
+                    break
+
+                if first_segment:
+                    self.turn_manager.on_agent_response(turn_id)
+                    first_segment = False
+
+                async with aclosing(
+                    self.tts.synthesize(segment)
+                ) as audio_stream:
+                    async for audio_chunk in audio_stream:
+                        await events.put(
+                            PipelineAudioChunk(audio=audio_chunk)
+                        )
+
+            if first_segment:
+                raise RuntimeError(
+                    "No text segments were available for TTS."
+                )
+
+            self.turn_manager.on_tts_complete(turn_id)
+
+        async def run_pipeline() -> None:
+            async with asyncio.TaskGroup() as group:
+                group.create_task(produce())
+                group.create_task(consume())
+
+        execution = asyncio.create_task(run_pipeline())
+
+        pending_get: asyncio.Task | None = None
+
+        try:
+            while True:
+                if execution.done() and events.empty():
+                    await execution
+                    break
+
+                pending_get = asyncio.create_task(events.get())
+
+                done, _ = await asyncio.wait(
+                    {pending_get, execution},
+                    return_when=asyncio.FIRST_COMPLETED,
+                )
+
+                if pending_get in done:
+                    event = pending_get.result()
+                    pending_get = None
+
+                    if event is not None:
+                        yield event
+
+                    continue
+
+                # Supervisor completed before another event arrived.
+                # Cancel the pending queue read before inspecting errors.
+                pending_get.cancel()
+
+                await asyncio.gather(
+                    pending_get,
+                    return_exceptions=True,
+                )
+
+                pending_get = None
+
+                # Propagate failures, including ExceptionGroup.
+                await execution
+
+                # Normal completion: drain remaining events.
+                while not events.empty():
+                    event = events.get_nowait()
+
+                    if event is not None:
+                        yield event
+
+                break
+
+        finally:
+            if pending_get is not None:
+                pending_get.cancel()
+
+                await asyncio.gather(
+                    pending_get,
+                    return_exceptions=True,
+                )
+
+            if not execution.done():
+                execution.cancel()
+
+            await asyncio.gather(
+                execution,
+                return_exceptions=True,
+            )
+
+
     def complete_turn(
         self,
         turn_id: int,

tests/test_gateway.py

@@ -1698,3 +1698,105 @@ def test_websocket_registration_failure_awaits_task_before_abort():
     assert len(tasks) == 1
     assert tasks[0].done()
     assert abort_checked
+
+
+def test_websocket_streaming_agent_and_tts():
+    """
+    Verify streaming Agent text and TTS audio through WebSocket.
+
+    The first audio segment must be delivered before the
+    Agent's complete response.
+    """
+    created_stt.clear()
+
+    class StreamingTestAgent:
+        async def stream(self, session_id, turn_id, text):
+            yield "Olá, como posso ajudar?"
+            yield " Estou à disposição."
+
+    class StreamingTestTTS:
+        async def synthesize(self, text):
+            yield text.encode("utf-8")
+
+    with patch.object(
+        main,
+        "create_agent_adapter",
+        return_value=StreamingTestAgent(),
+    ):
+        with patch.object(
+            main,
+            "create_tts_adapter",
+            return_value=StreamingTestTTS(),
+        ):
+            with client.websocket_connect("/ws/audio") as websocket:
+                websocket.send_json(
+                    {
+                        "type": "session.start",
+                        "audio": {
+                            "encoding": AUDIO_ENCODING,
+                            "sample_rate": AUDIO_SAMPLE_RATE,
+                            "channels": AUDIO_CHANNELS,
+                        },
+                    }
+                )
+
+                session_ready = websocket.receive_json()
+                assert session_ready["type"] == "session.ready"
+
+                websocket.send_json({"type": "audio.start"})
+
+                turn_started = websocket.receive_json()
+                assert turn_started["type"] == "turn.started"
+
+                turn_id = turn_started["turn_id"]
+                stt = created_stt[0]
+
+                stt.emit_final_now("Olá")
+
+                received_text = []
+                received_audio = []
+                received_complete_response = False
+
+                # The streaming path emits two deltas and
+                # two TTS audio segments.
+                for _ in range(20):
+                    message = websocket.receive()
+
+                    if message.get("bytes") is not None:
+                        received_audio.append(message["bytes"])
+                    else:
+                        import json
+
+                        event = json.loads(message["text"])
+
+                        if event["type"] == "agent.delta":
+                            received_text.append(event["text"])
+
+                        elif event["type"] == "agent.response":
+                            received_complete_response = True
+
+                        elif event["type"] == "error":
+                            raise AssertionError(
+                                f"Gateway returned error: {event}"
+                            )
+
+                    if received_complete_response and len(received_audio) >= 2:
+                        break
+
+                assert "".join(received_text) == (
+                    "Olá, como posso ajudar?"
+                    " Estou à disposição."
+                )
+
+                assert received_audio
+                assert received_complete_response
+
+                websocket.send_json(
+                    {
+                        "type": "audio.stop",
+                        "turn_id": turn_id,
+                    }
+                )
+
+                turn_ended = websocket.receive_json()
+                assert turn_ended["type"] == "turn.ended"

tests/test_oci_connection.py

@@ -0,0 +1,132 @@
+import asyncio
+import threading
+from types import SimpleNamespace
+
+import pytest
+
+from voice_assistant.agent.oci_connection import open_oci_stream
+
+
+class FakeStream:
+    def __init__(self):
+        self.closed = False
+
+    def close(self):
+        self.closed = True
+
+
+@pytest.mark.anyio
+async def test_open_oci_stream_success():
+    stream = FakeStream()
+
+    response = SimpleNamespace(data=stream)
+
+    result = await open_oci_stream(lambda: response)
+
+    assert result is response
+    assert stream.closed is False
+
+    stream.close()
+
+
+@pytest.mark.anyio
+async def test_open_oci_stream_closes_late_response():
+    started = threading.Event()
+    release = threading.Event()
+    finished = threading.Event()
+
+    stream = FakeStream()
+    response = SimpleNamespace(data=stream)
+
+    def open_connection():
+        started.set()
+        release.wait(timeout=5.0)
+
+        try:
+            return response
+        finally:
+            finished.set()
+
+    task = asyncio.create_task(
+        open_oci_stream(open_connection)
+    )
+
+    try:
+        assert await asyncio.to_thread(
+            started.wait,
+            1.0,
+        )
+
+        task.cancel()
+
+        with pytest.raises(asyncio.CancelledError):
+            await task
+
+    finally:
+        release.set()
+
+    assert await asyncio.to_thread(
+        finished.wait,
+        1.0,
+    )
+
+    # The worker may still be completing its cleanup.
+    for _ in range(100):
+        if stream.closed:
+            break
+        await asyncio.sleep(0.001)
+
+    assert stream.closed is True
+
+
+@pytest.mark.anyio
+async def test_open_oci_stream_cancellation_during_handoff():
+    """
+    Verify that a response is not abandoned when cancellation
+    happens around completion of the opening operation.
+    """
+    stream = FakeStream()
+    response = SimpleNamespace(data=stream)
+
+    opened = threading.Event()
+    release = threading.Event()
+
+    def open_connection():
+        opened.set()
+        release.wait(timeout=5.0)
+        return response
+
+    task = asyncio.create_task(
+        open_oci_stream(open_connection)
+    )
+
+    try:
+        assert await asyncio.to_thread(
+            opened.wait,
+            1.0,
+        )
+
+        release.set()
+        task.cancel()
+
+        with pytest.raises(asyncio.CancelledError):
+            await task
+
+        for _ in range(100):
+            if stream.closed:
+                break
+            await asyncio.sleep(0.001)
+
+        assert stream.closed is True
+
+    finally:
+        release.set()
+
+        if not task.done():
+            task.cancel()
+
+        await asyncio.gather(
+            task,
+            return_exceptions=True,
+        )
+

tests/test_protocol.py

@@ -20,6 +20,7 @@ import pytest
 from pydantic import ValidationError
 from voice_assistant.audio.models import CANONICAL_AUDIO_FORMAT
 from voice_assistant.audio_gateway.app.protocol import (
+    AgentDelta,
     AgentResponse,
     AudioChunk,
     AudioStart,
@@ -38,6 +39,7 @@ from voice_assistant.audio_gateway.app.protocol import (
 )
 
 
+
 @pytest.mark.parametrize(
     ("model", "expected_type"),
     [
@@ -101,6 +103,7 @@ def test_session_ready_output():
         (TurnStarted, "turn.started"),
         (STTPartial, "stt.partial"),
         (STTFinal, "stt.final"),
+        (AgentDelta, "agent.delta"),
         (AgentResponse, "agent.response"),
         (TurnEnded, "turn.ended"),
         (Interrupted, "interrupted"),
@@ -115,7 +118,7 @@ def test_turn_output_event_types(
         "turn_id": 1,
     }
 
-    if model in (STTPartial, STTFinal, AgentResponse):
+    if model in (STTPartial, STTFinal, AgentDelta, AgentResponse):
         data["text"] = "test"
 
     event = model(**data)
@@ -168,3 +171,30 @@ def test_error_event_defaults():
     assert event.retryable is False
     assert event.details == {}
 
+
+def test_agent_delta_serialization():
+    event = AgentDelta(
+        session_id="session-1",
+        turn_id=7,
+        text=" mundo",
+    )
+
+    assert event.model_dump() == {
+        "type": "agent.delta",
+        "session_id": "session-1",
+        "turn_id": 7,
+        "text": " mundo",
+    }
+
+
+def test_agent_delta_rejects_wrong_type():
+    with pytest.raises(ValidationError):
+        AgentDelta.model_validate(
+            {
+                "type": "agent.response",
+                "session_id": "session-1",
+                "turn_id": 7,
+                "text": "Olá",
+            }
+        )
+

tests/test_text_chunker.py

@@ -0,0 +1,58 @@
+from voice_assistant.agent.text_chunker import TextChunker
+
+
+def test_chunker_emits_sentence():
+    chunker = TextChunker(min_chars=5)
+
+    assert chunker.feed("Olá, tudo bem?") == [
+        "Olá, tudo bem?"
+    ]
+
+
+def test_chunker_handles_incremental_deltas():
+    chunker = TextChunker(min_chars=5)
+
+    assert chunker.feed("Olá,") == []
+    assert chunker.feed(" tudo") == []
+    assert chunker.feed(" bem?") == [
+        "Olá, tudo bem?"
+    ]
+
+
+def test_chunker_preserves_remaining_text():
+    chunker = TextChunker(min_chars=5)
+
+    assert chunker.feed("Olá, tudo bem? Posso ajudar") == [
+        "Olá, tudo bem?"
+    ]
+
+    assert chunker.flush() == "Posso ajudar"
+
+
+def test_chunker_breaks_long_text():
+    chunker = TextChunker(
+        min_chars=5,
+        max_chars=20,
+    )
+
+    segments = chunker.feed(
+        "Esta é uma frase bastante longa sem pontuação"
+    )
+
+    assert segments
+    assert all(
+        len(segment) <= 20
+        for segment in segments
+    )
+
+
+def test_chunker_rejects_invalid_limits():
+    try:
+        TextChunker(min_chars=20, max_chars=10)
+    except ValueError:
+        pass
+    else:
+        raise AssertionError(
+            "Expected ValueError for invalid limits."
+        )
+

tests/test_voice_pipeline.py

@@ -18,10 +18,17 @@ Non-responsibilities:
 
 import pytest
 import asyncio
+from contextlib import aclosing
+
+from voice_assistant.voice_pipeline import (
+    PipelineAudioChunk,
+    PipelineAgentResponse,
+    PipelineTextDelta,
+    VoicePipeline,
+)
 
 from voice_assistant.audio_gateway.app.session import SessionState
 from voice_assistant.turn_manager import TurnManager, TurnState
-from voice_assistant.voice_pipeline import VoicePipeline
 from voice_assistant.agent.mock import MockAgentAdapter
 from voice_assistant.tts.mock import MockTTSAdapter
 
@@ -818,3 +825,268 @@ async def test_pipeline_rejects_new_turn_while_previous_task_runs():
     assert second_turn.turn_id == 2
     assert session.active_turn_id == 2
 
+@pytest.mark.anyio
+async def test_stream_turn_response_produces_audio_before_agent_finishes():
+    """
+    Verify that TTS audio can be delivered while the Agent
+    is still generating its response.
+    """
+
+    class StreamingAgent:
+        def __init__(self):
+            self.first_delta_sent = asyncio.Event()
+            self.release_generation = asyncio.Event()
+            self.finished = False
+
+        async def stream(self, session_id, turn_id, text):
+            yield "Olá, como posso ajudar?"
+
+            self.first_delta_sent.set()
+
+            await self.release_generation.wait()
+
+            yield " Estou à disposição."
+            self.finished = True
+
+    class StreamingTTS:
+        async def synthesize(self, text):
+            yield text.encode("utf-8")
+
+    agent = StreamingAgent()
+
+    pipeline = VoicePipeline(
+        session=SessionState(session_id="stream-session"),
+        turn_manager=TurnManager(),
+        agent=agent,
+        tts=StreamingTTS(),
+    )
+
+    turn = pipeline.start_turn()
+    pipeline.on_final(turn.turn_id, "Olá")
+
+    first_audio_received = asyncio.Event()
+    received_events = []
+
+    async def consume():
+        async with aclosing(
+            pipeline.stream_turn_response(
+                session_id="stream-session",
+                turn_id=turn.turn_id,
+                text="Olá",
+            )
+        ) as stream:
+            async for event in stream:
+                received_events.append(event)
+
+                if isinstance(event, PipelineAudioChunk):
+                    first_audio_received.set()
+
+    task = asyncio.create_task(consume())
+
+    try:
+        await asyncio.wait_for(
+            first_audio_received.wait(),
+            timeout=2.0,
+        )
+
+        assert agent.finished is False
+
+        assert any(
+            isinstance(event, PipelineTextDelta)
+            for event in received_events
+        )
+
+        assert any(
+            isinstance(event, PipelineAudioChunk)
+            for event in received_events
+        )
+
+        agent.release_generation.set()
+
+        await asyncio.wait_for(task, timeout=2.0)
+
+        assert agent.finished is True
+
+        assert any(
+            isinstance(event, PipelineAgentResponse)
+            for event in received_events
+        )
+
+        assert (
+            pipeline.turn_manager.active_turn.state
+            == TurnState.COMPLETED
+        )
+
+    finally:
+        agent.release_generation.set()
+
+        if not task.done():
+            task.cancel()
+
+        await asyncio.gather(
+            task,
+            return_exceptions=True,
+        )
+
+
+@pytest.mark.anyio
+async def test_stream_turn_response_propagates_agent_failure():
+    class FailingStreamingAgent:
+        async def stream(self, session_id, turn_id, text):
+            yield "Olá, como posso ajudar?"
+            raise RuntimeError("Simulated OCI streaming failure.")
+
+    class StreamingTTS:
+        async def synthesize(self, text):
+            yield text.encode("utf-8")
+
+    pipeline = VoicePipeline(
+        session=SessionState(session_id="agent-failure-session"),
+        turn_manager=TurnManager(),
+        agent=FailingStreamingAgent(),
+        tts=StreamingTTS(),
+    )
+
+    turn = pipeline.start_turn()
+    pipeline.on_final(turn.turn_id, "Olá")
+
+    async def consume():
+        async with aclosing(
+            pipeline.stream_turn_response(
+                session_id="agent-failure-session",
+                turn_id=turn.turn_id,
+                text="Olá",
+            )
+        ) as stream:
+            return [event async for event in stream]
+
+    with pytest.raises(ExceptionGroup) as exc_info:
+        await asyncio.wait_for(consume(), timeout=2.0)
+
+    assert any(
+        isinstance(error, RuntimeError)
+        and "Simulated OCI streaming failure" in str(error)
+        for error in exc_info.value.exceptions
+    )
+
+@pytest.mark.anyio
+async def test_stream_turn_response_cancels_agent_on_tts_failure():
+    class BlockingStreamingAgent:
+        def __init__(self):
+            self.cancelled = asyncio.Event()
+
+        async def stream(self, session_id, turn_id, text):
+            try:
+                yield "Olá, como posso ajudar?"
+                await asyncio.Event().wait()
+            finally:
+                self.cancelled.set()
+
+    class FailingTTS:
+        async def synthesize(self, text):
+            raise RuntimeError("Simulated TTS failure.")
+            yield b""  # Keep this method an async generator.
+
+    agent = BlockingStreamingAgent()
+
+    pipeline = VoicePipeline(
+        session=SessionState(session_id="tts-failure-session"),
+        turn_manager=TurnManager(),
+        agent=agent,
+        tts=FailingTTS(),
+    )
+
+    turn = pipeline.start_turn()
+    pipeline.on_final(turn.turn_id, "Olá")
+
+    async def consume():
+        async with aclosing(
+            pipeline.stream_turn_response(
+                session_id="tts-failure-session",
+                turn_id=turn.turn_id,
+                text="Olá",
+            )
+        ) as stream:
+            async for _ in stream:
+                pass
+
+    with pytest.raises(ExceptionGroup) as exc_info:
+        await asyncio.wait_for(consume(), timeout=2.0)
+
+    assert any(
+        isinstance(error, RuntimeError)
+        and "Simulated TTS failure" in str(error)
+        for error in exc_info.value.exceptions
+    )
+
+    assert agent.cancelled.is_set()
+
+@pytest.mark.anyio
+async def test_stream_turn_response_external_cancellation():
+    class BlockingAgent:
+        def __init__(self):
+            self.closed = asyncio.Event()
+
+        async def stream(self, session_id, turn_id, text):
+            try:
+                yield "Olá, como posso ajudar?"
+                await asyncio.Event().wait()
+            finally:
+                self.closed.set()
+
+    class BlockingTTS:
+        def __init__(self):
+            self.started = asyncio.Event()
+            self.closed = asyncio.Event()
+
+        async def synthesize(self, text):
+            try:
+                self.started.set()
+                await asyncio.Event().wait()
+                yield b"audio"
+            finally:
+                self.closed.set()
+
+    agent = BlockingAgent()
+    tts = BlockingTTS()
+
+    pipeline = VoicePipeline(
+        session=SessionState(session_id="cancel-session"),
+        turn_manager=TurnManager(),
+        agent=agent,
+        tts=tts,
+    )
+
+    turn = pipeline.start_turn()
+    pipeline.on_final(turn.turn_id, "Olá")
+
+    async def consume():
+        async with aclosing(
+            pipeline.stream_turn_response(
+                session_id="cancel-session",
+                turn_id=turn.turn_id,
+                text="Olá",
+            )
+        ) as stream:
+            async for _ in stream:
+                pass
+
+    task = asyncio.create_task(consume())
+
+    try:
+        await asyncio.wait_for(tts.started.wait(), timeout=2.0)
+
+        task.cancel()
+
+        with pytest.raises(asyncio.CancelledError):
+            await asyncio.wait_for(task, timeout=2.0)
+
+        assert agent.closed.is_set()
+        assert tts.closed.is_set()
+
+    finally:
+        if not task.done():
+            task.cancel()
+
+        await asyncio.gather(task, return_exceptions=True)
+

tests/test_web_client.py

@@ -0,0 +1,32 @@
+
+from fastapi.testclient import TestClient
+
+from voice_assistant.audio_gateway.app.main import app
+
+
+client = TestClient(app)
+
+
+def test_web_client_page():
+    response = client.get("/ui")
+
+    assert response.status_code == 200
+    assert "AgentVoice" in response.text
+    assert "/web/app.js" in response.text
+
+
+def test_web_client_javascript():
+    response = client.get("/web/app.js")
+
+    assert response.status_code == 200
+    assert "audio.start" in response.text
+    assert "audio.stop" in response.text
+    assert "agent.delta" in response.text
+
+
+def test_web_client_audio_worklet():
+    response = client.get("/web/pcm-worklet.js")
+
+    assert response.status_code == 200
+    assert "registerProcessor" in response.text
+    assert "pcm-input" in response.text