[LLM/TTS 与对话闭环]：完成语音对话主链路，包含上下文管理、NewAPI 兼容 LLM、分句 TTS、状态机和错误恢复测试

openspec/changes/add-voice-pet-pipeline/tasks.md

@@ -26,12 +26,12 @@
 
 ## 4. LLM/TTS 与对话闭环
 
-- [ ] 4.1 实现 ConversationContext；前置条件：Message 模型已定义；验收标准：追加 user/assistant、构造 LLM messages、按预算截断；测试要点：上下文保留 system prompt 和最近轮次；优先级：P0；预计：45 分钟。
-- [ ] 4.2 实现 OpenAI/NewAPI 兼容 LLM Provider；前置条件：配置加载已完成；验收标准：base URL、API key、model 均配置化，支持流式或非流式解析，不记录密钥；测试要点：mock HTTP 和可选真实 smoke 测试通过；优先级：P0；预计：60 分钟。
-- [ ] 4.3 实现分句缓冲；前置条件：LLM delta 模型已定义；验收标准：中文标点、换行和长度阈值触发 TTS chunk；测试要点：空分句、超长句、连续 delta 测试通过；优先级：P0；预计：45 分钟。
-- [ ] 4.4 实现本地 TTS Provider；前置条件：AudioSegment 模型和 Transport 可用；验收标准：提供 macOS `say` 可选 Provider 和 deterministic 测试 Provider；测试要点：非空文本生成非空音频，空文本报错；优先级：P0；预计：60 分钟。
-- [ ] 4.5 实现 Pipeline 状态机和自抑制；前置条件：Wake/VAD/STT、Context、LLM、TTS、Transport 均可测试；验收标准：正常链路、空转写、LLM 失败、TTS 失败、播放期自抑制均可验证；测试要点：状态序列和错误恢复测试通过；优先级：P0；预计：60 分钟。
-- [ ] 4.6 完成“LLM/TTS 与对话闭环”模块提交；前置条件：4.1 至 4.5 已完成；验收标准：先通过相关测试、compileall、真实或 mock LLM smoke 和 OpenSpec strict 校验，再立即执行 Git commit；测试要点：提交信息使用“`[LLM/TTS 与对话闭环]：完成[具体功能描述]，包含[关键变更]`”格式；优先级：P0；预计：20 分钟。
+- [x] 4.1 实现 ConversationContext；前置条件：Message 模型已定义；验收标准：追加 user/assistant、构造 LLM messages、按预算截断；测试要点：上下文保留 system prompt 和最近轮次；优先级：P0；预计：45 分钟。
+- [x] 4.2 实现 OpenAI/NewAPI 兼容 LLM Provider；前置条件：配置加载已完成；验收标准：base URL、API key、model 均配置化，支持流式或非流式解析，不记录密钥；测试要点：mock HTTP 和可选真实 smoke 测试通过；优先级：P0；预计：60 分钟。
+- [x] 4.3 实现分句缓冲；前置条件：LLM delta 模型已定义；验收标准：中文标点、换行和长度阈值触发 TTS chunk；测试要点：空分句、超长句、连续 delta 测试通过；优先级：P0；预计：45 分钟。
+- [x] 4.4 实现本地 TTS Provider；前置条件：AudioSegment 模型和 Transport 可用；验收标准：提供 macOS `say` 可选 Provider 和 deterministic 测试 Provider；测试要点：非空文本生成非空音频，空文本报错；优先级：P0；预计：60 分钟。
+- [x] 4.5 实现 Pipeline 状态机和自抑制；前置条件：Wake/VAD/STT、Context、LLM、TTS、Transport 均可测试；验收标准：正常链路、空转写、LLM 失败、TTS 失败、播放期自抑制均可验证；测试要点：状态序列和错误恢复测试通过；优先级：P0；预计：60 分钟。
+- [x] 4.6 完成“LLM/TTS 与对话闭环”模块提交；前置条件：4.1 至 4.5 已完成；验收标准：先通过相关测试、compileall、真实或 mock LLM smoke 和 OpenSpec strict 校验，再立即执行 Git commit；测试要点：提交信息使用“`[LLM/TTS 与对话闭环]：完成[具体功能描述]，包含[关键变更]`”格式；优先级：P0；预计：20 分钟。
 
 ## 5. 桌宠 UI 与生图资产
 

src/owner_voice_pet/__init__.py

@@ -19,6 +19,10 @@ from .transport import AudioRingBuffer, FileReplayTransport, MemoryAudioTranspor
 from .wakeword import KeywordWakeWordProvider
 from .vad import EnergyVadProvider, VadRecorder
 from .stt import MetadataSttProvider, is_valid_transcript_text
+from .conversation import ConversationContext
+from .llm import MockLlmProvider, OpenAICompatibleLlmProvider
+from .pipeline import PipelineResult, VoicePipeline
+from .tts import MacSayTtsProvider, SentenceBuffer, SineTtsProvider
 
 __all__ = [
     "AppConfig",
@@ -33,6 +37,14 @@ __all__ = [
     "VadRecorder",
     "MetadataSttProvider",
     "is_valid_transcript_text",
+    "ConversationContext",
+    "MockLlmProvider",
+    "OpenAICompatibleLlmProvider",
+    "PipelineResult",
+    "VoicePipeline",
+    "MacSayTtsProvider",
+    "SentenceBuffer",
+    "SineTtsProvider",
     "Message",
     "PipelineState",
     "PlaybackResult",

src/owner_voice_pet/conversation.py

@@ -0,0 +1,48 @@
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass, field
+from typing import Iterable
+
+from .models import Message
+
+
+@dataclass(slots=True)
+class ConversationContext:
+    system_prompt: str = "你是一个中文桌宠助手，回答要简洁、自然、适合语音播报。"
+    max_messages: int = 12
+    max_chars: int = 12000
+    _messages: list[Message] = field(default_factory=list)
+
+    def append_user(self, text: str) -> None:
+        self._append("user", text)
+
+    def append_assistant(self, text: str) -> None:
+        self._append("assistant", text)
+
+    def build_llm_messages(self) -> list[Message]:
+        system = Message("system", self.system_prompt, 0.0)
+        return [system, *self._messages]
+
+    def reset(self) -> None:
+        self._messages.clear()
+
+    def messages(self) -> tuple[Message, ...]:
+        return tuple(self._messages)
+
+    def _append(self, role: str, text: str) -> None:
+        clean = text.strip()
+        if not clean:
+            return
+        self._messages.append(Message(role, clean, time.time()))
+        self.truncate()
+
+    def truncate(self) -> None:
+        while len(self._messages) > self.max_messages:
+            self._messages.pop(0)
+        while self._total_chars(self._messages) > self.max_chars and self._messages:
+            self._messages.pop(0)
+
+    @staticmethod
+    def _total_chars(messages: Iterable[Message]) -> int:
+        return sum(len(message.content) for message in messages)

src/owner_voice_pet/llm.py

@@ -0,0 +1,149 @@
+from __future__ import annotations
+
+import json
+import socket
+import urllib.error
+import urllib.request
+from collections.abc import Callable, Iterable, Sequence
+from typing import Any
+
+from .config import AppConfig
+from .models import ErrorCode, Message, ProviderError, ReplyDelta
+
+
+class MockLlmProvider:
+    def __init__(self, chunks: list[str] | None = None, fail: ProviderError | None = None) -> None:
+        self.chunks = chunks or ["好的，我听到了。"]
+        self.fail = fail
+        self.calls: list[list[Message]] = []
+
+    def stream_reply(self, messages: Sequence[Message]) -> Iterable[ReplyDelta]:
+        self.calls.append(list(messages))
+        if self.fail:
+            raise self.fail
+        for chunk in self.chunks:
+            yield ReplyDelta(chunk, is_sentence_boundary=_ends_sentence(chunk))
+        yield ReplyDelta("", finish_reason="stop")
+
+
+class OpenAICompatibleLlmProvider:
+    def __init__(
+        self,
+        config: AppConfig,
+        timeout_s: float = 20.0,
+        urlopen: Callable[..., Any] | None = None,
+    ) -> None:
+        self.config = config
+        self.timeout_s = timeout_s
+        self.urlopen = urlopen or urllib.request.urlopen
+
+    def stream_reply(self, messages: Sequence[Message]) -> Iterable[ReplyDelta]:
+        self.config.require_llm_credentials()
+        if self.config.llm_api_style == "responses":
+            yield from self._responses(messages)
+        else:
+            yield from self._chat_completions(messages)
+
+    def _chat_completions(self, messages: Sequence[Message]) -> Iterable[ReplyDelta]:
+        payload = {
+            "model": self.config.llm_model,
+            "messages": [{"role": m.role, "content": m.content} for m in messages],
+            "stream": True,
+        }
+        yield from self._post_stream(
+            f"{self.config.llm_base_url}/v1/chat/completions",
+            payload,
+            parser=_parse_chat_completion_sse,
+        )
+
+    def _responses(self, messages: Sequence[Message]) -> Iterable[ReplyDelta]:
+        payload = {
+            "model": self.config.llm_model,
+            "input": [{"role": m.role, "content": m.content} for m in messages],
+            "stream": True,
+        }
+        yield from self._post_stream(
+            f"{self.config.llm_base_url}/v1/responses",
+            payload,
+            parser=_parse_responses_sse,
+        )
+
+    def _post_stream(
+        self,
+        url: str,
+        payload: dict[str, Any],
+        parser: Callable[[dict[str, Any]], ReplyDelta | None],
+    ) -> Iterable[ReplyDelta]:
+        body = json.dumps(payload).encode("utf-8")
+        request = urllib.request.Request(
+            url,
+            data=body,
+            headers={
+                "Authorization": f"Bearer {self.config.llm_api_key}",
+                "Content-Type": "application/json",
+            },
+            method="POST",
+        )
+        try:
+            with self.urlopen(request, timeout=self.timeout_s) as response:
+                emitted = False
+                for raw_line in response:
+                    line = raw_line.decode("utf-8", errors="replace").strip()
+                    if not line or not line.startswith("data:"):
+                        continue
+                    data = line.removeprefix("data:").strip()
+                    if data == "[DONE]":
+                        yield ReplyDelta("", finish_reason="stop")
+                        return
+                    event = json.loads(data)
+                    delta = parser(event)
+                    if delta is not None:
+                        emitted = True
+                        yield delta
+                if not emitted:
+                    raise ProviderError(
+                        ErrorCode.LLM_EMPTY_REPLY,
+                        "LLM stream ended without text",
+                        True,
+                        "openai-compatible",
+                        "llm",
+                    )
+        except ProviderError:
+            raise
+        except urllib.error.HTTPError as exc:
+            code = ErrorCode.LLM_RATE_LIMITED if exc.code == 429 else ErrorCode.LLM_NETWORK_ERROR
+            raise ProviderError(code, f"LLM HTTP error {exc.code}", exc.code >= 500, "openai-compatible", "llm") from exc
+        except (urllib.error.URLError, TimeoutError, socket.timeout) as exc:
+            raise ProviderError(
+                ErrorCode.LLM_REQUEST_TIMEOUT,
+                f"LLM request failed or timed out: {exc}",
+                True,
+                "openai-compatible",
+                "llm",
+            ) from exc
+
+
+def _parse_chat_completion_sse(event: dict[str, Any]) -> ReplyDelta | None:
+    choices = event.get("choices") or []
+    if not choices:
+        return None
+    choice = choices[0]
+    text = str((choice.get("delta") or {}).get("content") or "")
+    finish = choice.get("finish_reason")
+    if not text and not finish:
+        return None
+    return ReplyDelta(text, is_sentence_boundary=_ends_sentence(text), finish_reason=finish)
+
+
+def _parse_responses_sse(event: dict[str, Any]) -> ReplyDelta | None:
+    event_type = event.get("type")
+    if event_type == "response.output_text.delta":
+        text = str(event.get("delta") or "")
+        return ReplyDelta(text, is_sentence_boundary=_ends_sentence(text))
+    if event_type in {"response.completed", "response.output_text.done"}:
+        return ReplyDelta("", finish_reason="stop")
+    return None
+
+
+def _ends_sentence(text: str) -> bool:
+    return text.endswith(("。", "！", "？", ".", "!", "?", "\n"))

src/owner_voice_pet/pipeline.py

@@ -0,0 +1,147 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from .conversation import ConversationContext
+from .models import AudioSegment, ErrorCode, PipelineState, ProviderError
+from .protocols import AudioTransport, LlmProvider, SttProvider, TtsProvider, WakeWordProvider
+from .tts import SentenceBuffer
+from .vad import VadRecorder
+
+
+@dataclass(slots=True)
+class PipelineResult:
+    success: bool
+    transcript: str = ""
+    assistant_text: str = ""
+    states: list[PipelineState] = field(default_factory=list)
+    error: ProviderError | None = None
+    played_segments: int = 0
+
+
+class VoicePipeline:
+    def __init__(
+        self,
+        transport: AudioTransport,
+        wakeword: WakeWordProvider,
+        vad_recorder: VadRecorder,
+        stt: SttProvider,
+        context: ConversationContext,
+        llm: LlmProvider,
+        tts: TtsProvider,
+        sentence_buffer: SentenceBuffer | None = None,
+    ) -> None:
+        self.transport = transport
+        self.wakeword = wakeword
+        self.vad_recorder = vad_recorder
+        self.stt = stt
+        self.context = context
+        self.llm = llm
+        self.tts = tts
+        self.sentence_buffer = sentence_buffer or SentenceBuffer()
+        self.states: list[PipelineState] = []
+        self.suppress_input = False
+
+    def load(self) -> None:
+        self.wakeword.load()
+        self.vad_recorder.provider.load()
+        self.stt.load()
+        self.tts.load()
+
+    def run_once(self, max_frames: int = 1000) -> PipelineResult:
+        self.states = []
+        try:
+            self._state(PipelineState.WAKE_LISTENING)
+            self.transport.start_input()
+            wake_found = False
+            frames_read = 0
+            while frames_read < max_frames:
+                frames = self.transport.read_frames(timeout_ms=20)
+                if not frames:
+                    break
+                for frame in frames:
+                    frames_read += 1
+                    if self.suppress_input:
+                        continue
+                    if not wake_found:
+                        if self.wakeword.detect(frame):
+                            wake_found = True
+                            self._state(PipelineState.SPEECH_DETECTING)
+                        continue
+                    result = self.vad_recorder.feed(frame)
+                    if isinstance(result, ProviderError):
+                        return self._recover(result)
+                    if isinstance(result, AudioSegment):
+                        return self._handle_segment(result)
+            return self._recover(
+                ProviderError(
+                    ErrorCode.VAD_TIMEOUT_NO_SPEECH,
+                    "no complete utterance was captured",
+                    True,
+                    "voice-pipeline",
+                    "pipeline",
+                )
+            )
+        finally:
+            self.transport.stop()
+
+    def _handle_segment(self, segment: AudioSegment) -> PipelineResult:
+        self._state(PipelineState.TRANSCRIBING)
+        try:
+            transcript = self.stt.transcribe(segment)
+        except ProviderError as exc:
+            if exc.code == ErrorCode.STT_EMPTY_TRANSCRIPT:
+                self._state(PipelineState.WAKE_LISTENING)
+                return PipelineResult(False, states=list(self.states), error=exc)
+            return self._recover(exc)
+        text = transcript.normalized_text
+        self.context.append_user(text)
+        self._state(PipelineState.THINKING)
+        assistant_text = ""
+        played = 0
+        try:
+            for delta in self.llm.stream_reply(self.context.build_llm_messages()):
+                assistant_text += delta.text_delta
+                for sentence in self.sentence_buffer.feed(delta.text_delta, bool(delta.finish_reason)):
+                    self._state(PipelineState.SPEAKING)
+                    self.suppress_input = True
+                    audio = self.tts.synthesize(sentence)
+                    playback = self.transport.play_pcm(audio)
+                    self.suppress_input = False
+                    if playback.error:
+                        return self._recover(playback.error)
+                    played += 1
+            for sentence in self.sentence_buffer.flush():
+                self._state(PipelineState.SPEAKING)
+                self.suppress_input = True
+                audio = self.tts.synthesize(sentence)
+                playback = self.transport.play_pcm(audio)
+                self.suppress_input = False
+                if playback.error:
+                    return self._recover(playback.error)
+                played += 1
+        except ProviderError as exc:
+            self.suppress_input = False
+            return self._recover(exc)
+        if not assistant_text.strip():
+            return self._recover(
+                ProviderError(
+                    ErrorCode.LLM_EMPTY_REPLY,
+                    "LLM returned no assistant text",
+                    True,
+                    "voice-pipeline",
+                    "llm",
+                )
+            )
+        self.context.append_assistant(assistant_text)
+        self._state(PipelineState.WAKE_LISTENING)
+        return PipelineResult(True, text, assistant_text, list(self.states), played_segments=played)
+
+    def _recover(self, error: ProviderError) -> PipelineResult:
+        self._state(PipelineState.ERROR_RECOVERING)
+        self.suppress_input = False
+        self._state(PipelineState.WAKE_LISTENING)
+        return PipelineResult(False, states=list(self.states), error=error)
+
+    def _state(self, state: PipelineState) -> None:
+        self.states.append(state)

src/owner_voice_pet/tts.py

@@ -0,0 +1,123 @@
+from __future__ import annotations
+
+import math
+import struct
+import subprocess
+import tempfile
+from pathlib import Path
+
+from .models import AudioSegment, ErrorCode, ProviderError
+
+
+class SentenceBuffer:
+    def __init__(self, max_chars: int = 80) -> None:
+        self.max_chars = max_chars
+        self._buffer = ""
+
+    def feed(self, text: str, force: bool = False) -> list[str]:
+        if text:
+            self._buffer += text
+        chunks: list[str] = []
+        while self._should_emit(force):
+            chunks.append(self._pop_chunk(force))
+            force = False
+        return [chunk for chunk in chunks if chunk.strip()]
+
+    def flush(self) -> list[str]:
+        return self.feed("", force=True)
+
+    def _should_emit(self, force: bool) -> bool:
+        stripped = self._buffer.strip()
+        if not stripped:
+            return False
+        return force or stripped.endswith(("。", "！", "？", ".", "!", "?", "\n")) or len(stripped) >= self.max_chars
+
+    def _pop_chunk(self, force: bool) -> str:
+        stripped = self._buffer.strip()
+        if force or len(stripped) <= self.max_chars:
+            self._buffer = ""
+            return stripped
+        chunk = stripped[: self.max_chars]
+        self._buffer = stripped[self.max_chars :]
+        return chunk
+
+
+class SineTtsProvider:
+    def __init__(self, sample_rate: int = 16000) -> None:
+        self.sample_rate = sample_rate
+        self.loaded = False
+
+    def load(self) -> None:
+        self.loaded = True
+
+    def synthesize(self, text: str) -> AudioSegment:
+        if not self.loaded:
+            raise ProviderError(
+                ErrorCode.TTS_SYNTHESIS_FAILED,
+                "TTS provider is not loaded",
+                False,
+                "sine-tts",
+                "tts",
+            )
+        clean = text.strip()
+        if not clean:
+            raise ProviderError(
+                ErrorCode.TTS_EMPTY_AUDIO,
+                "cannot synthesize empty text",
+                True,
+                "sine-tts",
+                "tts",
+            )
+        duration_ms = max(120, min(1200, len(clean) * 45))
+        samples = int(self.sample_rate * duration_ms / 1000)
+        pcm = bytearray()
+        for idx in range(samples):
+            sample = int(math.sin(2 * math.pi * 440 * idx / self.sample_rate) * 12000)
+            pcm.extend(struct.pack("<h", sample))
+        return AudioSegment(bytes(pcm), self.sample_rate, 1, 0, duration_ms, {"text": clean})
+
+
+class MacSayTtsProvider:
+    def __init__(self, voice: str | None = None) -> None:
+        self.voice = voice
+        self.loaded = False
+
+    def load(self) -> None:
+        self.loaded = True
+
+    def synthesize(self, text: str) -> AudioSegment:
+        clean = text.strip()
+        if not clean:
+            raise ProviderError(
+                ErrorCode.TTS_EMPTY_AUDIO,
+                "cannot synthesize empty text",
+                True,
+                "macos-say",
+                "tts",
+            )
+        with tempfile.TemporaryDirectory() as tmp:
+            output = Path(tmp) / "speech.aiff"
+            command = ["say", "-o", str(output)]
+            if self.voice:
+                command.extend(["-v", self.voice])
+            command.append(clean)
+            try:
+                subprocess.run(command, check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+            except (FileNotFoundError, subprocess.CalledProcessError) as exc:
+                raise ProviderError(
+                    ErrorCode.TTS_SYNTHESIS_FAILED,
+                    f"macOS say failed: {exc}",
+                    False,
+                    "macos-say",
+                    "tts",
+                ) from exc
+            data = output.read_bytes()
+        if not data:
+            raise ProviderError(
+                ErrorCode.TTS_EMPTY_AUDIO,
+                "macOS say produced empty audio",
+                True,
+                "macos-say",
+                "tts",
+            )
+        return AudioSegment(data, 16000, 1, 0, max(120, len(clean) * 45), {"text": clean, "format": "aiff"})

tests/test_pipeline_llm_tts.py

@@ -0,0 +1,151 @@
+from __future__ import annotations
+
+import io
+import json
+import unittest
+
+from owner_voice_pet.config import AppConfig
+from owner_voice_pet.conversation import ConversationContext
+from owner_voice_pet.llm import MockLlmProvider, OpenAICompatibleLlmProvider
+from owner_voice_pet.models import AudioFrame, ErrorCode, Message, PipelineState, ProviderError
+from owner_voice_pet.pipeline import VoicePipeline
+from owner_voice_pet.stt import MetadataSttProvider
+from owner_voice_pet.transport import MemoryAudioTransport
+from owner_voice_pet.tts import SentenceBuffer, SineTtsProvider
+from owner_voice_pet.vad import EnergyVadProvider, VadRecorder
+from owner_voice_pet.wakeword import KeywordWakeWordProvider
+
+
+def frame(idx: int, timestamp_ms: int, metadata: dict[str, object]) -> AudioFrame:
+    speech = bool(metadata.get("speech"))
+    return AudioFrame(
+        b"\xff\xff" if speech else b"\x80\x80",
+        16000,
+        1,
+        timestamp_ms,
+        idx,
+        {"duration_ms": 20, **metadata},
+    )
+
+
+def make_pipeline(frames: list[AudioFrame], llm: MockLlmProvider | None = None) -> tuple[VoicePipeline, MemoryAudioTransport]:
+    transport = MemoryAudioTransport(frames)
+    pipeline = VoicePipeline(
+        transport=transport,
+        wakeword=KeywordWakeWordProvider(),
+        vad_recorder=VadRecorder(EnergyVadProvider(), min_duration_ms=40, end_silence_ms=40),
+        stt=MetadataSttProvider(),
+        context=ConversationContext(max_messages=4, max_chars=200),
+        llm=llm or MockLlmProvider(["你好，我在。"]),
+        tts=SineTtsProvider(),
+    )
+    pipeline.load()
+    return pipeline, transport
+
+
+class PipelineLlmTtsTests(unittest.TestCase):
+    def test_context_truncates_old_messages(self) -> None:
+        context = ConversationContext(max_messages=2, max_chars=100)
+        context.append_user("一")
+        context.append_assistant("二")
+        context.append_user("三")
+        self.assertEqual([m.content for m in context.messages()], ["二", "三"])
+        self.assertEqual(context.build_llm_messages()[0].role, "system")
+
+    def test_sentence_buffer_chunks_on_chinese_punctuation(self) -> None:
+        buffer = SentenceBuffer(max_chars=20)
+        self.assertEqual(buffer.feed("你好"), [])
+        self.assertEqual(buffer.feed("。"), ["你好。"])
+        self.assertEqual(buffer.feed("剩余").copy(), [])
+        self.assertEqual(buffer.flush(), ["剩余"])
+
+    def test_sine_tts_generates_non_empty_audio(self) -> None:
+        provider = SineTtsProvider()
+        provider.load()
+        segment = provider.synthesize("你好")
+        self.assertGreater(len(segment.pcm), 0)
+        self.assertGreater(segment.duration_ms, 0)
+
+    def test_pipeline_runs_from_wake_to_playback(self) -> None:
+        frames = [
+            frame(0, 0, {"wake_word": "小杰小杰", "wake_confidence": 0.95}),
+            frame(1, 20, {"speech": True, "transcript": "你好"}),
+            frame(2, 40, {"speech": True}),
+            frame(3, 60, {"speech": False}),
+            frame(4, 80, {"speech": False}),
+        ]
+        pipeline, transport = make_pipeline(frames)
+        result = pipeline.run_once()
+        self.assertTrue(result.success)
+        self.assertEqual(result.transcript, "你好")
+        self.assertIn(PipelineState.SPEAKING, result.states)
+        self.assertEqual(len(transport.played_segments), 1)
+        self.assertEqual(pipeline.context.messages()[-1].role, "assistant")
+
+    def test_pipeline_skips_llm_on_empty_transcript(self) -> None:
+        llm = MockLlmProvider(["不应调用"])
+        frames = [
+            frame(0, 0, {"wake": True}),
+            frame(1, 20, {"speech": True, "transcript": "？！"}),
+            frame(2, 40, {"speech": True}),
+            frame(3, 60, {"speech": False}),
+            frame(4, 80, {"speech": False}),
+        ]
+        pipeline, _ = make_pipeline(frames, llm)
+        result = pipeline.run_once()
+        self.assertFalse(result.success)
+        self.assertEqual(result.error.code, ErrorCode.STT_EMPTY_TRANSCRIPT)
+        self.assertEqual(llm.calls, [])
+
+    def test_pipeline_recovers_from_llm_failure(self) -> None:
+        error = ProviderError(ErrorCode.LLM_NETWORK_ERROR, "boom", True, "mock", "llm")
+        frames = [
+            frame(0, 0, {"wake": True}),
+            frame(1, 20, {"speech": True, "transcript": "你好"}),
+            frame(2, 40, {"speech": True}),
+            frame(3, 60, {"speech": False}),
+            frame(4, 80, {"speech": False}),
+        ]
+        pipeline, _ = make_pipeline(frames, MockLlmProvider(fail=error))
+        result = pipeline.run_once()
+        self.assertFalse(result.success)
+        self.assertIn(PipelineState.ERROR_RECOVERING, result.states)
+        self.assertEqual(result.error.code, ErrorCode.LLM_NETWORK_ERROR)
+
+    def test_openai_chat_completion_sse_parser(self) -> None:
+        class FakeResponse:
+            def __enter__(self) -> "FakeResponse":
+                return self
+
+            def __exit__(self, *args: object) -> None:
+                return None
+
+            def __iter__(self):
+                event = {"choices": [{"delta": {"content": "你好。"}, "finish_reason": None}]}
+                done = {"choices": [{"delta": {}, "finish_reason": "stop"}]}
+                yield f"data: {json.dumps(event, ensure_ascii=False)}\n".encode()
+                yield f"data: {json.dumps(done, ensure_ascii=False)}\n".encode()
+                yield b"data: [DONE]\n"
+
+        requests = []
+
+        def fake_urlopen(request, timeout):
+            requests.append(request)
+            return FakeResponse()
+
+        config = AppConfig(
+            llm_base_url="https://newapi.mkbk.shop",
+            llm_api_key="secret",
+            llm_model="test-model",
+            llm_api_style="chat_completions",
+        )
+        provider = OpenAICompatibleLlmProvider(config, urlopen=fake_urlopen)
+        deltas = list(provider.stream_reply([Message("user", "你好", 1.0)]))
+        self.assertEqual(deltas[0].text_delta, "你好。")
+        body = json.loads(requests[0].data.decode())
+        self.assertTrue(body["stream"])
+        self.assertEqual(body["model"], "test-model")
+
+
+if __name__ == "__main__":
+    unittest.main()