语音 Agent

export interface VoiceBudget {  firstByteMs: number;  totalTurnMs: number;  bargeInMs: number;}export interface VoiceSessionState {  turnId: string;  phase: "idle" | "listening" | "thinking" | "speaking" | "interrupted";  partialTranscript: string | undefined;  budget: VoiceBudget;}export interface VoiceSession {  sessionId: string;  state: VoiceSessionState;  speech: SpeechOutcome | undefined;}export type SpeechOutcome =  | { kind: "spoken"; turnId: string; text: string }  | { kind: "interrupted"; turnId: string; spokenPrefix: string; reason: string };export type VoiceEvent =  | { type: "partial_transcript"; turnId: string; text: string }  | { type: "final_transcript"; turnId: string; text: string }  | { type: "agent_speech_started"; turnId: string }  | { type: "agent_speech_chunk"; turnId: string; text: string }  | { type: "barge_in"; turnId: string; reason: string }  | { type: "session_closed"; sessionId: string; reason: string };// 事件 type 继续使用主线状态章那种稳定字符串注册表，而不是另造一套联合类型；// appendEvent、nextEventSequence 和 recordedAt 都原样复用。export type VoiceEventType = VoiceEvent["type"];export function appendVoiceEvent(  session: VoiceSession,  event: VoiceEvent,): VoiceSession {  return {    ...session,    state: {      ...session.state,      partialTranscript:        event.type === "partial_transcript" ? event.text : session.state.partialTranscript,    },  };}

export interface AudioChunk {  sequence: number;  pcm: Uint8Array;  startedAt: number;  durationMs: number;}export interface SpeechRecognizer {  accept(chunk: AudioChunk): VoiceEvent[];  flush(): VoiceEvent[];}

function reduceTranscript(  state: AgentState,  session: VoiceSession,  event: VoiceEvent,  now: Date,): { session: VoiceSession; state: AgentState } {  if (event.type === "partial_transcript") {    return {      session: appendVoiceEvent(session, event),      state: appendEvent(state, "voice_partial_transcript", event.text, now),    };  }  if (event.type !== "final_transcript" || event.turnId === state.lastCommittedTurnId) {    return { session, state };  }  return {    session: { ...session, state: { ...session.state, partialTranscript: undefined } },    state: appendEvent(      { ...state, lastCommittedTurnId: event.turnId },      "voice_final_transcript",      event.text,      now,    ),  };}

export function handleBargeIn(  session: VoiceSession,  state: AgentState,  turnController: AbortController,  now = new Date(),): { session: VoiceSession; state: AgentState } {  turnController.abort(new Error("barge_in"));  const spokenPrefix =    state.messages.at(-1)?.role === "assistant" ? state.messages.at(-1)!.content : "";  return {    // 插话只结束当前回合：运行状态仍是 running，预算、计划与已完成的步骤都不动。    state: appendEvent(state, "barge_in", session.state.turnId, now),    session: {      ...session,      state: { ...session.state, phase: "interrupted", partialTranscript: undefined },      speech: { kind: "interrupted", turnId: session.state.turnId, spokenPrefix, reason: "interrupted" },    },  };}// “被打断”不是“失败”：interrupted 表示用户主动结束了当前回合，// 而不是模型错误或工具异常。// 但要注意区分两件事：回合被打断**不结束 run**（不上报运行级停止原因），// 只有用户随后明确要求终止整个运行时，才写运行级 interrupted。export interface StopReasonMap {  interrupted: true;}

export interface VoiceUsage {  firstByteMs: number;  totalTurnMs: number;  bargeInMs: number;  speechChunks: number;  degraded: boolean;}export interface RunUsage {  voice?: VoiceUsage | undefined;}

firstByteMs   从 final transcript 到第一个 agent_speech_chunk  目标 ≤ 700ms   硬上限 1200mstotalTurnMs   从 final transcript 到最后一个 agent_speech_chunk  目标 ≤ 2500ms  硬上限 4000msbargeInMs     从检测到插话到停止播放并取消模型请求              目标 ≤ 200ms   硬上限 500ms

export type DegradeAction =  | { kind: "acknowledge_first"; text: string }  | { kind: "shorten_answer"; maxSentences: number }  | { kind: "defer_deep_work"; note: string };export function planVoiceDegradation(  budget: VoiceBudget,  elapsed: { firstByteMs: number; totalTurnMs: number },): DegradeAction | undefined {  if (elapsed.firstByteMs > budget.firstByteMs) {    return { kind: "acknowledge_first", text: "我先确认一下，再继续处理。" };  }  if (elapsed.totalTurnMs > budget.totalTurnMs) {    return { kind: "shorten_answer", maxSentences: 2 };  }  return undefined;}

1 打断后仍收到旧回答：插话后旧的 agent_speech_chunk 不得继续进入输出；  停止原因必须是 interrupted，工具被取消，不得记入 failedAttempts，会话仍能开始下一个 turn。2 partial 覆盖 final：final 之后的 partial 不得修改已提交的用户消息，  也不得让 messages 出现第二条用户消息。3 超预算未降级：人造超过 firstByteMs 与 totalTurnMs 的延迟后，  必须触发 DegradeAction 且 degraded 为 true；没有降级即为失败。