cotutor 0.1.0 → 0.3.0

package/dist/runtime/live/call.js

@@ -0,0 +1,1211 @@
+/**
+ * 实时老师的一次通话:孩子端的 WebSocket ↔ 嘴(百炼 realtime,手动轮次)+ 脑(老师 CLI,Runner)。
+ * 一轮:孩子按住说 → 松手(嘴不接话,拍照才先说一句「我看一看图片」,src/realtime/cues.ts)→ 孩子那句的转写交给脑(实时轮:讲稿不配音)→ 脑的讲稿定稿一句交一句,
+ * 嘴一空就把排着的几句一起交给它照念(src/runtime/delivery/queue.ts,流水线:嘴出完上一段就交下一段,声音在孩子端接着排)→ 讲完就停。
+ * 按句不按拍(2026-10-19):卡宁缺毋滥以后一节常常只有一拍、甚至没卡,按拍要等脑整段写完才念;卡只管自己闭合了就铺,和念哪句不挂钩。
+ * 转写是空的(按了没说清):嘴在说就掐掉,改说「没听清」,不交给脑;按着时声音不到 100 毫秒(不 commit)、commit 没收下也一样。孩子在嘴的声音还在放时按下 = 打断:掐掉、孩子端清播放队列,
+ * 没念的拍丢掉,孩子听到第几句(src/runtime/delivery/playback.ts 按声音排到哪算)进下一轮的上下文包 live:。嘴只接住、只照念,不讲新内容(2026-10-19 P1);画板开着时例外(边画边聊,见下)。
+ *
+ * 孩子端 → 这里:二进制帧 = PCM16 16k 单声道一块(按着时才推);文本帧 = JSON
+ *   {t:'hold'} 按下 · {t:'release', body} 松手(body 同 POST /api/kid/conversations/<老师>/messages 少 text:thread / newThread / device…)
+ *   {t:'hush'} 停下不开新一轮(上滑取消):嘴在出就掐、孩子端清了、没念的丢掉
+ *   {t:'send', body} 不是说的(打字、拍照、交卡、首页按钮;舞台开着带 focus.card):body 同那条 POST · {t:'view', w, h} 屏幕多大 · {t:'bye'}
+ *   {t:'staged'} 讲稿 [[play]] 的小课堂那一段放完了(或放不了、被关):后面的讲稿接着交给嘴
+ *   边画边聊:画板开着时按头像说的,松手带 chat {canvas, image, ask}——只给嘴(图跟着声音进对话、嘴接住),不起脑;
+ *   {t:'chat', canvas, image, ask} 孩子端觉得该说一句了(换了颜色、停笔很久,src/board/canvas-chat.ts):安静时 0.5 秒静音 + 图交给嘴、嘴主动说一句,回 {t:'chat', canvas, said};不安静回 {t:'chat', canvas, skipped}。
+ *   这张画上孩子说的、嘴说的记下来,交画那一轮(交画板卡 / 纸上的 drawing)带给脑(live.draw)
+ * 这里 → 孩子端:二进制帧 = PCM16 24k 嘴的声音;文本帧 = JSON
+ *   {t:'ready'} 嘴连上了 · {t:'down', why} 连不上(孩子端回到「正在叫老师」,过一会儿重连)· {t:'kid', text} 孩子那句的转写 · {t:'tutor', of, delta} / {t:'tutor', of, done} 嘴的字(of:catch 接住 / read 照念)
+ *   {t:'turn', job, thread} 脑开始写这一轮 · {t:'read', job, line, text, play?} 讲稿这一句开始放(play = 念完放小课堂那一段;一句一个,按孩子端放到的时刻发:一段里第一句在段首,
+ *   后面几句等这段的声音出齐、按字数比例摊)· {t:'flush'} 清播放队列
+ *   {t:'state', speaking, holding} · {t:'err', status, error} 发不出去(400 / 409 …)· {t:'end', why}
+ * 嘴的会话断了(百炼 300 秒没输入就关,2026-10-20 照 Qwen-Live-Harness 的换连接):电话不挂、孩子端不知道;下次按下或发来时现连,
+ *   连上之前来的照序攒着补(录像里记一条 {t:'lost'},只记不发)。平时孩子没按着就 30 秒一次心跳,最后一次动静 15 分钟后停跳。
+ * 换连接:嘴的历史越长每次回答越贵,输入过了 SWAP_TOKENS 就在安静时换一条新的(录像里记 {t:'swap'},只记不发)。
+ * 一直听(孩子端只在口语课、或 live.listen 是 open 时连接带 ?listen=open,中途发 {t:'listen', open} 也能换):孩子端一直推声音(卡的声音、录音、小课堂、关麦时不推,{t:'gate', on} 告诉这里),
+ *   嘴的会话开 VAD(live.vad 缺省 server_vad、live.silenceMs),create_response / interrupt_response 关着:开口(speech_started)先暂停照念、嘴在放就让孩子端压低({t:'duck'}),
+ *   转写回来 src/realtime/listen.ts 判收下 / 放过——收下才算按下 + 松手(打断、接住、交给脑;交给脑之前孩子停一下又开口的并成一句),放过的恢复音量接着念({t:'skip', why} / {t:'unduck'})。
+ *   孩子那一句的上下文(话题、设备、舞台上的卡)孩子端开口时报一次 {t:'ctx', body}。换模式 = 安静时换一条连接(silence 只有新连接能给)。
+ *   commit 了不回答 300 秒百炼就关(response_idle_timeout,探针 probe-live-vad):放过的多了,离上次回答 IDLE_SWAP_MS 就在安静时换连接。
+ * 口语课:先热身、再打电话。热身是脑一轮写完的跟读,讲稿的 [[echo 词]] / [[say 词]] 那句念完关闸(src/runtime/delivery/queue.ts),
+ *   孩子说完(像跟读,src/scenes/talk/warmup.ts)开闸下一拍,不接住、不交脑;对老师说的照收下办(打断、接住、交脑);ECHO_WAIT_MS 没开口接着走,连着两拍没开口停下交脑(app: quiet)。
+ *   每拍记进摸底(录像 {t:'echo'},只记不发),交脑时带上(live.warmup)。[[call]] 那句念完孩子端响铃,这里马上交脑一轮(app: call,不接住),
+ *   之后的接住换英文(cue.ts 的 CATCH_EN),交脑带打电话几分钟(live.phone)。孩子端连上发来的第一条带 talk(口语单)时嘴先说「好,我们先热热身!」。
+ * 录像:进出的消息、两边的声音都记(CallRecorder),挂了落在 conversations/<老师>/<日期>.live/<时分秒>.reel.json / .wav。
+ * 字幕重放:嘴照念的声音按轮另留一份,出了的都留(孩子按下时清掉、没放出来的也在),
+ *   一段出完(或被掐、断了)就落 <日期>.<job>.mouth.wav 与每句的起点 .mouth.json(src/runtime/journal/mouth-index.ts);接住那句不在里面。孩子端从 /api/kid/conversations/<老师>/mouth/<日期>/<job> 取。
+ * 事件:嘴这一道(mouth:按下、松手 / 发了、接住开始放、第几句开始放、打断)补进那一轮的 events.jsonl(Runner.noteEvent),工作台的时间线上看;
+ * 按下、松手、接住在脑起来之前,先攒着,等那一轮的老师进程起来(main:start)再按它的起点换成相对毫秒(负的)。
+ */
+import { mkdir, writeFile } from 'node:fs/promises';
+import { join } from 'node:path';
+import { localDate } from "../../shared/time.js";
+import { catchCue, chatCue, readCue } from "../../realtime/cues.js";
+import { judge } from "../../realtime/listen.js";
+import { echoKind, warmLines } from "../../scenes/talk/warmup.js";
+import { coverage } from "../../scenes/talk/en-words.js";
+import { Delivery } from "../delivery/delivery.js";
+import { MIX_RATE } from "../journal/mix.js";
+import { wavOfPcm } from "../../shared/wav.js";
+import { mouthInstructions } from "../../realtime/prompt.js";
+import { mouthVoice } from "../../schema/index.js";
+import { INPUT_RATE, MAX_IMAGE_B64, OUTPUT_RATE, RealtimeSession } from "../../realtime/bailian.js";
+import { CallRecorder } from "../journal/recorder.js";
+import { MouthEvents, MouthTakes } from "../journal/mouth.js";
+import { Turns } from "./turns.js";
+/** 嘴接住的那句最多等多久再把孩子的话交给脑(接住的字一般 1 秒内齐;脑知道嘴说了什么,第一句就不再重复) */
+const CATCH_WAIT_MS = 1500;
+const BYTES_PER_MS = (OUTPUT_RATE * 2) / 1000;
+/** 按着时到的声音不到 100 毫秒就不 commit:百炼回「buffer too small, or have no audio」,不会有转写(2026-10-19 iPad 麦克风没开,嘴照样接住说了「你想问什么」) */
+const MIN_COMMIT_BYTES = (INPUT_RATE * 2 * 100) / 1000;
+/** 心跳间隔(Qwen-Live-Harness 也是 30 秒);孩子最后一次动静之后跳多久(过了就让百炼 300 秒后关,回来按下再连,连一次约半秒) */
+const HEARTBEAT_MS = 30_000;
+const KEEP_ALIVE_MS = 15 * 60_000;
+/** 嘴最近一次回答的输入过了这么多 token,安静时换一条新连接(约 9 轮一换) */
+const SWAP_TOKENS = 3000;
+/** 讲稿的 [[play]]:孩子端放小课堂那一段,最多等它这么久就接着念(孩子端没回 staged 时不卡死) */
+const GATE_MS = 180_000;
+/** 现连时攒着的声音最多一分钟(同 Qwen-Live-Harness 的 MAX_RECOVERY_AUDIO_BYTES);一直听时只留最近 10 秒(更早的孩子早不等了) */
+const MAX_BACKLOG_BYTES = INPUT_RATE * 2 * 60;
+const MAX_OPEN_BACKLOG_BYTES = INPUT_RATE * 2 * 10;
+/** 一直听:收下一句以后,孩子停一下又开口(或刚说完在等转写),交给脑最多再等这么久,收下的并进来 */
+const MERGE_WAIT_MS = 8000;
+/** 一直听:一句收下以后先等这么久,看孩子是不是又开口(没有接住那句垫着了) */
+const OPEN_PEEK_MS = 800;
+/** 一直听:commit 了没回答,离上次回答这么久就在安静时换连接(百炼 300 秒报 response_idle_timeout) */
+const IDLE_SWAP_MS = 240_000;
+/** 热身:[[echo]] / [[say]] 那句在孩子端放完,这么久没开口就接着下一拍(不催);连着这么多拍没开口就停下交脑 */
+const ECHO_WAIT_MS = 8000;
+const QUIET_BEATS = 2;
+/** 打电话:[[call]] 那句放完先响这么久的铃,脑的英文开场来得再早也等它(孩子听得到在打电话) */
+const RING_MS = 3000;
+/**
+ * 要嘴说了,这么久一块声音、一个字都没出就算卡住(真百炼首包约 0.5 秒;2026-10-31 照念第二段回了 created 就再没动静,孩子只听到第一句):
+ * 当这条连接坏了,照念的那段放回排头、换一条接着念(一个字没出,孩子没听过,再念不重复)
+ */
+const STALL_MS = 10_000;
+/** 字幕重放的声音最多留几轮在内存里(一轮约 1–2 MB;落了盘就只为接着往后接) */
+const MOUTH_KEEP = 3;
+/** 嘴主动说一句前垫的静音(百炼要先有声音再有图,commit 时一起进对话):0.5 秒 */
+const CHAT_SILENCE = Buffer.alloc((INPUT_RATE * 2) / 2);
+/** 静音那段的转写(空的)多久之内来算它的;过了就不吞了,免得吞掉孩子后来说的 */
+const CHAT_QUIET_MS = 5000;
+/** 一张画上边画边说最多记几句给脑 */
+const DRAW_TALK_MAX = 24;
+export class LiveCall {
+    session = null;
+    rec;
+    /** 交付:讲稿排队、嘴的声音在孩子端排到哪、第几句开始放、打断时听到第几句(./delivery.ts) */
+    d;
+    holding = false;
+    /** 这一次按着推进去多少字节的声音 */
+    heldBytes = 0;
+    /** 要了还没出完的那次回答;afterDone = 它出完(或被掐)之后再做的 */
+    said = null;
+    afterDone = null;
+    cancelOnCreate = false;
+    /** 这次回答出过声音或字没有(看门狗 STALL_MS 看它) */
+    saidOut = false;
+    /** 这一轮:嘴接住时说的、接住出完没有 */
+    caught = '';
+    caughtDone = false;
+    /** 嘴这一道的事件(补进那一轮的 events.jsonl)与字幕重放的声音(./journal.ts) */
+    events;
+    takes;
+    /** 交给脑:验与拼、runner.send、脑忙时排着(./turns.ts) */
+    turns;
+    turnSeq = 0;
+    /** 按住说的这一轮还在等转写 */
+    pendingVoice = null;
+    timers = [];
+    /** 孩子最近一次按下、松手、发来,或嘴说完一次的墙钟(心跳看它) */
+    lastActive = Date.now();
+    /** 嘴最近一次回答的输入 token(连接的历史越长越多);过了 SWAP_TOKENS 安静时换连接 */
+    lastInput = 0;
+    swapping = false;
+    /** 第一次连上了:之后嘴没了是断了要现连,不是连不上 */
+    opened = false;
+    /** 正在现连;连上之前孩子端来的按下、松手、发来、按着的声音照到达的顺序攒着 */
+    reviving = null;
+    backlog = [];
+    backlogBytes = 0;
+    backlogHolding = false;
+    /** 一直听:孩子端要的(open)与嘴这条会话开着 VAD 没有(mouthOpen);不一样就在安静时换连接 */
+    open;
+    mouthOpen = false;
+    /** 一直听:VAD 判着孩子在说、开口的时刻(通话毫秒);说完了在等转写的(说完的时刻) */
+    vadOn = false;
+    vadFrom = 0;
+    /** 孩子开口那一刻嘴的声音还在放(或刚放完 300 毫秒内):只有这时才可能是回声 */
+    vadOverMouth = false;
+    vadAwait = null;
+    /** 一直听:嘴的声音让孩子端压低着 */
+    ducked = false;
+    /** 一直听:孩子端开口时报的这一轮上下文(话题、设备、via、舞台上的卡),收下时当松手的 body */
+    ctxBody = {};
+    /** 嘴最近说过的话(比回声),at = 那段在孩子端放完的墙钟 */
+    saidLog = [];
+    /** 一直听:收下了还没交给脑的几句(孩子停一下接着说的并进来) */
+    draft = null;
+    /** 一直听:孩子的声音最后到的墙钟、上次回答以后有没有 commit 过、上次回答(或连上)的墙钟、孩子端关着麦 */
+    lastAudioAt = 0;
+    committedSince = false;
+    lastResponseAt = Date.now();
+    gated = false;
+    /** 热身:交给嘴的那段最后一句要孩子跟。endAt = 那句在孩子端放完的通话毫秒(嘴出完才知道);heardAt = 孩子开口的通话毫秒;overdue = 等够了时孩子正在说 */
+    echoing = null;
+    /** 热身:按住说时松手了、在等这一下的转写(跟读还是对老师说的);那一下松手的 body */
+    echoHeld = null;
+    /** 热身每拍记的(摸底)、连着没开口的拍数 */
+    warm = [];
+    quietBeats = 0;
+    /** 打电话从哪一刻起(墙钟,[[call]] 那句在孩子端放完、开始响铃);null = 还没打 */
+    phoneAt = null;
+    /** 边画边聊:每张画(画板卡 <job>/<n>,纸 paper)上孩子说的、嘴说的,一句一行;交画那轮给脑后清掉 */
+    drawTalk = new Map();
+    /** 边画边聊:按着头像说的那句在等转写(记到哪张画上) */
+    chatHeard = null;
+    chatHeardUntil = 0;
+    /** 嘴主动说那几次垫的静音:几段转写还没回来、等到什么时候(空的转写吞掉,不说没听清) */
+    chatQuiet = 0;
+    chatQuietUntil = 0;
+    offEvents = null;
+    ended = false;
+    sock;
+    deps;
+    constructor(sock, deps) {
+        this.sock = sock;
+        this.deps = deps;
+        this.rec = new CallRecorder(deps.ctx.now);
+        this.events = new MouthEvents((date, job, ev) => void deps.ctx.runner.noteEvent(deps.tutor, date, job, ev));
+        this.d = new Delivery({ now: () => this.ms(), send: (o) => this.send(o), event: (job, at, ev) => this.events.record(job, at, ev), onRead: (beat) => this.read(beat), ended: () => this.ended });
+        this.turns = new Turns({
+            runner: deps.ctx.runner,
+            now: deps.ctx.now,
+            tutor: deps.tutor,
+            deviceId: deps.deviceId,
+            live: (body) => this.liveOf(body),
+            send: (o) => this.send(o),
+            started: (st) => {
+                this.d.begin(st.job);
+                // 字幕重放:这一轮的声音从这里接
+                this.takes.open(st.job, st.date);
+                this.send({ t: 'turn', job: st.job, thread: st.thread });
+            },
+            lines: (job, lines) => { if (this.d.job === job)
+                this.d.lines(lines); },
+            settled: () => this.maybeSwap(),
+            timer: (t) => this.timers.push(t),
+            ended: () => this.ended,
+        });
+        this.takes = new MouthTakes({ conversations: () => deps.ctx.runner.ws.dirs.conversations, tutor: deps.tutor, keep: MOUTH_KEEP, bytesPerMs: BYTES_PER_MS, rate: OUTPUT_RATE });
+        this.open = deps.listen === 'open';
+    }
+    ms() { return this.rec.ms(); }
+    /** 通话里的毫秒 → 墙钟 */
+    wall(ms) { return this.rec.startedAt.getTime() + ms; }
+    /** 记嘴这一道的一条:那一轮起来了就补进它的事件,还不知道是哪一轮(job = null)就先攒着 */
+    mouth(job, at, ev) {
+        this.events.record(job, at, ev);
+    }
+    send(o) {
+        this.rec.out(o);
+        this.sock.sendText(JSON.stringify(o));
+    }
+    playing() { return Boolean(this.session?.responding) || this.ms() < this.d.end; }
+    state() { this.send({ t: 'state', speaking: this.playing(), holding: this.holding }); }
+    /** 排队说这一拍该交给嘴了:要嘴照念;热身那句念完要等孩子跟 */
+    read(beat) {
+        if (beat.echo)
+            this.echoing = { echo: beat.echo, endAt: null, heardAt: null, timer: null, overdue: false };
+        this.ask({ kind: 'read', beat }, readCue(beat.texts));
+        // 这段最后一句念完孩子端放小课堂那一段:排队关着闸,等 staged;孩子端一直没回也别卡死
+        if (beat.play)
+            this.timers.push(setTimeout(() => { if (!this.ended && this.d.gated)
+                this.d.step({ e: 'staged' }); }, GATE_MS));
+    }
+    /** 让嘴说一次;接住要先告诉排队,嘴出完之前别交讲稿 */
+    ask(said, instructions) {
+        this.said = said;
+        this.saidOut = false;
+        if (said.kind !== 'read')
+            this.d.note({ e: 'asked' });
+        this.session?.createResponse(instructions);
+        this.timers.push(setTimeout(() => this.stalled(said), this.deps.stallMs ?? STALL_MS));
+    }
+    /** 要嘴说的这次一个字、一块声音都没出(STALL_MS):当这条连接坏了,照念的那段放回排头、换一条接着念;接住、边画边聊那句不补 */
+    stalled(said) {
+        if (this.ended || this.said !== said || this.saidOut)
+            return;
+        const old = this.session;
+        this.rec.out({ t: 'stall', kind: said.kind });
+        this.lost('stall');
+        old?.close('stall');
+        // 放回排头以后排队不交(连接还没换好),现连好了 revive 里 done 再交
+        if (said.kind === 'read') {
+            this.d.retry(said.beat);
+            void this.revive();
+        }
+    }
+    /** 连嘴;连不上告诉孩子端(它灰着「正在叫老师」、过一会儿重连)、挂掉 */
+    async start() {
+        const why = await this.connectMouth();
+        if (this.ended)
+            return;
+        if (why !== null) {
+            this.send({ t: 'down', why });
+            await this.end(`connect:${why}`);
+            return;
+        }
+        this.opened = true;
+        const { ctx, tutor } = this.deps;
+        this.offEvents = ctx.runner.onEvent((e) => {
+            if (e.tutor !== tutor || e.job !== this.d.job)
+                return;
+            // 这一轮的老师进程起来了:攒着的按下、松手、接住按它的起点补进它的事件
+            if (e.event.lane === 'main' && e.event.kind === 'start')
+                this.events.started(e.job, e.date, Date.now() - e.event.t);
+            // 讲稿定稿一句(partial 解析器压着没换行的最后一句,发出来的都是定稿的)就交给排队
+            if (e.event.lane === 'main' && e.event.kind === 'line')
+                this.d.lines([{ i: e.event.line, text: e.event.text, play: e.event.play === true, echo: e.event.echo, call: e.event.call }]);
+        });
+        this.timers.push(setInterval(() => this.keepAlive(), this.deps.heartbeatMs ?? HEARTBEAT_MS));
+        this.send({ t: 'ready' });
+    }
+    /** 开一条嘴的会话(第一次、断了以后现连、换连接都走这里);不是 this.session 的会话,回调都不认 */
+    async openMouth() {
+        const { ctx, tutor } = this.deps;
+        const cfg = ctx.runner.ws.config;
+        const t = cfg.tutors[tutor];
+        let firstOfAnswer = true;
+        const open = this.open;
+        const s = new RealtimeSession({ instructions: mouthInstructions({ display: t?.display ?? tutor, kidName: cfg.kid.name ?? cfg.kid.slug }), voice: mouthVoice(cfg.live, tutor), model: cfg.live.model, turn: open ? cfg.live.vad : 'manual', silenceMs: cfg.live.silenceMs, vad: { createResponse: false, interruptResponse: false }, endpoint: this.deps.endpoint, env: this.deps.env }, {
+            onAudio: (pcm) => {
+                if (this.session !== s)
+                    return;
+                this.saidOut = true;
+                const said = this.said ?? { kind: 'catch' };
+                const start = this.d.add(said, pcm.length / BYTES_PER_MS);
+                // 照念的第一块:到它在孩子端开始放的那一刻,告诉孩子端这段的第一句;接住的第一块记一条(开始放的时刻)
+                if (firstOfAnswer && said.kind === 'read')
+                    this.d.readAt(said.beat, 0, start);
+                if (firstOfAnswer && said.kind === 'catch' && this.holding === false)
+                    this.mouth(null, this.wall(start), { lane: 'mouth', kind: 'catch', text: '' });
+                firstOfAnswer = false;
+                // 字幕重放:照念的声音另留一份(孩子端放没放都留)
+                const seg = this.mouthSeg(said);
+                if (seg) {
+                    seg.pcm.push(pcm);
+                    seg.bytes += pcm.length;
+                }
+                this.rec.tutor(pcm);
+                this.sock.sendBinary(pcm);
+            },
+            onTutorText: (delta, done) => {
+                if (this.session !== s)
+                    return;
+                this.saidOut = true;
+                const kind = this.said?.kind ?? 'catch';
+                // 垫的那句在孩子端同接住的字一样上字幕
+                const of = kind === 'read' ? 'read' : 'catch';
+                if (done !== null) {
+                    if (done.trim())
+                        this.saidLog = [...this.saidLog.slice(-5), { text: done.trim(), at: Date.now() }];
+                    if (this.said?.kind === 'chat' && done.trim())
+                        this.noteDraw(this.said.canvas, `你:${done.trim()}`);
+                    if (kind === 'catch') {
+                        this.caught = done.trim();
+                        this.caughtDone = true;
+                        this.events.caught(this.caught);
+                    }
+                    this.send({ t: 'tutor', of, done: done.trim() });
+                    return;
+                }
+                if (delta) {
+                    // 字幕重放:这段没出齐就停了的话,按出了几个字摊每句的起点
+                    const seg = this.said ? this.mouthSeg(this.said, false) : null;
+                    if (seg)
+                        seg.heard += delta.length;
+                    this.send({ t: 'tutor', of, delta });
+                }
+            },
+            onKidText: (text) => { if (this.session !== s)
+                return; if (this.mouthOpen)
+                this.verdict(text);
+            else
+                this.heard(text); },
+            onSpeech: (st) => { if (this.session === s && this.mouthOpen)
+                this.speech(st); },
+            onRejected: () => { if (this.session === s && this.mouthOpen)
+                this.verdict(null, 'rejected'); },
+            onKidFailed: () => { if (this.session === s)
+                this.kidFailed(); },
+            onResponse: (st, status, usage) => {
+                if (this.session !== s)
+                    return;
+                if (st === 'done' && usage)
+                    this.lastInput = usage.input;
+                if (st === 'created') {
+                    firstOfAnswer = true;
+                    if (this.cancelOnCreate) {
+                        this.cancelOnCreate = false;
+                        this.session?.cancel();
+                    }
+                    this.state();
+                    return;
+                }
+                if (this.said?.kind === 'read')
+                    this.mouthEnd(this.said, status !== 'cancelled');
+                if (status === 'cancelled')
+                    this.d.flush();
+                else if (this.said?.kind === 'read') {
+                    this.d.readRest(this.said);
+                    // 热身那句出齐了:知道它在孩子端什么时候放完,从那时起等孩子;[[call]] 那句出齐了:开始打电话
+                    if (this.said.beat.echo)
+                        this.armEcho(this.d.end);
+                    if (this.said.beat.call)
+                        this.startPhone();
+                }
+                // 比回声按这段在孩子端放完的时刻算(字早就出完了,声音要放十几秒)
+                const last = this.saidLog[this.saidLog.length - 1];
+                if (last && status !== 'cancelled')
+                    last.at = Math.max(last.at, this.wall(this.d.end));
+                this.lastResponseAt = Date.now();
+                this.committedSince = false;
+                if (this.said?.kind === 'catch')
+                    this.caughtDone = true;
+                this.said = null;
+                this.lastActive = Date.now();
+                this.d.idle();
+                const next = this.afterDone;
+                this.afterDone = null;
+                if (next)
+                    next();
+                else
+                    this.d.step({ e: 'done' });
+                this.state();
+                this.maybeSwap();
+            },
+            onError: (m) => {
+                if (this.session !== s)
+                    return;
+                this.send({ t: 'err', status: 502, error: m });
+                // commit 没收下(声音太少):这一句不会有转写,照转写是空的办
+                if ((this.pendingVoice || this.echoHeld) && /committing input audio buffer/i.test(m))
+                    this.heard('');
+            },
+            // 连上以后断了才收;还在连的时候断(配会话报错)由调用方处理
+            onClose: (r) => { if (!this.ended && this.session === s)
+                this.lost(r); },
+        });
+        const why = await s.connect();
+        if (this.ended) {
+            s.close();
+            return { why: 'ended', s, open };
+        }
+        return { why, s, open };
+    }
+    /** 开一条嘴的会话并换上(第一次、断了以后现连) */
+    async connectMouth() {
+        const { why, s, open } = await this.openMouth();
+        if (why === null)
+            this.mounted(s, open);
+        return why;
+    }
+    /** 换上一条嘴的会话:一直听的状态都从头来 */
+    mounted(s, open) {
+        this.session = s;
+        if (this.mouthOpen !== open)
+            this.send({ t: 'listen', open });
+        this.mouthOpen = open;
+        this.vadOn = false;
+        this.vadAwait = null;
+        this.committedSince = false;
+        this.lastResponseAt = Date.now();
+        if (this.ducked) {
+            this.ducked = false;
+            this.send({ t: 'unduck' });
+        }
+    }
+    /** 安静:孩子没按着、嘴没在说也没要说的、讲稿没排着、脑没在写、没在现连 / 换连接 */
+    quiet() {
+        return !this.holding && !this.vadOn && this.vadAwait === null && !this.said && !this.session?.responding && !this.d.pending && !this.d.gated && !this.turns.busy && !this.reviving && !this.swapping;
+    }
+    /**
+     * 换连接:嘴的输入每次回答都带着整条连接的历史,一轮约多 230 token(2026-10-20 真连 15 轮:843 涨到 5468,
+     * 首包不变,只是钱);嘴每次都给词、不靠前情,安静时开一条新的换上就行,不用摘要。开的时候孩子动了就作废这次。
+     */
+    maybeSwap() {
+        if (this.ended || !this.session)
+            return;
+        // 换的三种:历史长了;孩子端换了模式(按住说 ↔ 一直听);一直听放过的句子攒着没回答,快到百炼的 response_idle_timeout
+        const why = this.open !== this.mouthOpen ? 'listen' : this.lastInput >= (this.deps.swapTokens ?? SWAP_TOKENS) ? 'tokens' : this.mouthOpen && this.committedSince && Date.now() - this.lastResponseAt > (this.deps.idleSwapMs ?? IDLE_SWAP_MS) ? 'idle' : null;
+        if (!why || !this.quiet())
+            return;
+        this.swapping = true;
+        void this.openMouth().then(({ why: failed, s, open }) => {
+            this.swapping = false;
+            if (failed !== null)
+                return;
+            if (this.ended || !this.session || !this.quiet()) {
+                s.close('swap-abort');
+                return;
+            }
+            const old = this.session;
+            this.rec.out({ t: 'swap', input: this.lastInput, why });
+            this.mounted(s, open);
+            this.lastInput = 0;
+            old.close('swap');
+        });
+    }
+    /** 嘴的会话断了(百炼 300 秒没输入就关、网断了):电话不挂;讲稿还排着就马上连、接着念,不然等孩子下次按下或发来再连 */
+    lost(why) {
+        this.session = null;
+        // 只记进录像,孩子端不知道
+        this.rec.out({ t: 'lost', why });
+        // 照念到一半断了:出了的那截留着
+        if (this.said?.kind === 'read')
+            this.mouthEnd(this.said, false);
+        if (this.said?.kind === 'catch')
+            this.caughtDone = true;
+        if (this.said) {
+            this.said = null;
+            this.afterDone = null;
+            this.cancelOnCreate = false;
+            this.d.idle();
+        }
+        // 热身那句没出齐就断了:从现在起等孩子,别卡死
+        if (this.echoing && this.echoing.endAt === null)
+            this.armEcho(this.ms());
+        // 按着时断的:推过去的声音跟着那条会话没了,松手就说没听清
+        this.heldBytes = 0;
+        // 边画边聊在等的转写跟着那条会话没了
+        this.chatHeard = null;
+        this.chatQuiet = 0;
+        // 一直听时孩子说到一半断的:那一句跟着会话没了
+        this.vadOn = false;
+        this.vadAwait = null;
+        if (this.ducked) {
+            this.ducked = false;
+            this.send({ t: 'unduck' });
+        }
+        // 一直听、声音还在推:没有「下次按下」,马上连
+        if (this.d.pending || (this.open && Date.now() - this.lastAudioAt < 5000))
+            void this.revive();
+    }
+    /** 现连嘴(最多两次);连上了把攒着的照到达的顺序补上、排着的讲稿接着念;连不上告诉孩子端、挂掉 */
+    revive() {
+        this.reviving ??= (async () => {
+            let why = 'ended';
+            for (let i = 0; i < 2 && !this.ended && why !== null; i++)
+                why = await this.connectMouth();
+            this.reviving = null;
+            if (this.ended)
+                return false;
+            if (why !== null) {
+                this.send({ t: 'down', why });
+                await this.end(`reconnect:${why}`);
+                return false;
+            }
+            const items = this.backlog.splice(0);
+            this.backlogBytes = 0;
+            this.backlogHolding = false;
+            for (const x of items) {
+                if ('pcm' in x)
+                    this.pushAudio(x.pcm);
+                else
+                    this.handle(x.m, x.at);
+            }
+            if (!this.said)
+                this.d.step({ e: 'done' });
+            return true;
+        })();
+        return this.reviving;
+    }
+    /** 心跳(RealtimeSession.heartbeat):嘴连着、孩子没按着、嘴没在说、最近有动静;孩子走开久了就不跳,让百炼关,回来按下再连 */
+    keepAlive() {
+        const s = this.session;
+        this.maybeSwap();
+        if (!s?.isReady || this.holding || this.vadOn || s.responding || Date.now() - this.lastActive > (this.deps.keepAliveMs ?? KEEP_ALIVE_MS))
+            return;
+        // 一直听、声音一直在推:本来就有输入,不用跳
+        if (this.mouthOpen && Date.now() - this.lastAudioAt < 5000)
+            return;
+        s.heartbeat();
+    }
+    /** 字幕重放:这次照念在它那一轮里的那一段(第一块声音到时开;make = false 只找不开);接住、那一轮不在内存里的不留 */
+    mouthSeg(said, make = true) {
+        return said.kind === 'read' ? this.takes.seg(said, said.beat, make) : null;
+    }
+    /** 字幕重放:一段出完了(done)/ 被掐了、断了:这一轮整份重落一次 */
+    mouthEnd(said, done) {
+        if (said.kind === 'read')
+            this.takes.end(said, said.beat, done);
+    }
+    /** 孩子按下(或不是说的那一轮开始):嘴在出就掐、还在放就清孩子端;排着的丢掉;记下孩子听到第几句 */
+    interrupt() {
+        const now = this.ms();
+        const playing = this.playing();
+        // 排着的丢掉、记下孩子听到第几句、撤掉还没发的 read
+        this.d.interrupt(playing, this.said, now);
+        // 热身不等了(排着的那几拍跟着丢了;摸底留着)
+        if (this.echoing?.timer)
+            clearTimeout(this.echoing.timer);
+        this.echoing = null;
+        this.echoHeld = null;
+        this.afterDone = null;
+        if (this.session?.responding)
+            this.session.cancel();
+        else if (this.said)
+            this.cancelOnCreate = true;
+        if (playing) {
+            this.d.flush(now);
+            this.send({ t: 'flush' });
+        }
+    }
+    /** 孩子的一条文本消息 */
+    onText(raw) {
+        if (this.ended)
+            return;
+        let m;
+        try {
+            m = JSON.parse(raw);
+        }
+        catch {
+            return;
+        }
+        if (typeof m.t !== 'string')
+            return;
+        // 录像不存图(一张一两百 KB,画在纸上的那张交画时另存了)
+        this.rec.in(withoutImages(m));
+        // 嘴断了:按下、松手、发来的先攒着,现连,连上了照序补
+        if (!this.session && this.opened && INPUTS.has(m.t)) {
+            this.backlog.push({ m, at: Date.now() });
+            if (m.t === 'hold')
+                this.backlogHolding = true;
+            else if (m.t === 'release' || m.t === 'hush')
+                this.backlogHolding = false;
+            void this.revive();
+            return;
+        }
+        this.handle(m, Date.now());
+    }
+    /** 一条孩子端的消息;at = 到的墙钟(现连时补的那几条照原来到的时刻记事件) */
+    handle(m, at) {
+        if (typeof m.t === 'string' && INPUTS.has(m.t))
+            this.lastActive = Date.now();
+        switch (m.t) {
+            case 'hold': {
+                if (!this.session)
+                    return;
+                this.events.begin([{ at, ev: { lane: 'mouth', kind: 'hold' } }]);
+                this.holding = true;
+                this.heldBytes = 0;
+                this.session.clearAudio();
+                // 热身在等孩子跟:按下就是在跟,不打断(排着的几拍留着);那句的声音还没放完就清掉
+                if (this.echoing) {
+                    if (this.echoing.heardAt === null)
+                        this.echoing.heardAt = this.ms();
+                    if (this.playing()) {
+                        this.d.flush();
+                        this.send({ t: 'flush' });
+                    }
+                }
+                else
+                    this.interrupt();
+                this.state();
+                break;
+            }
+            case 'release': {
+                if (!this.session || !this.holding)
+                    return;
+                this.holding = false;
+                const body = isObj(m.body) ? m.body : {};
+                // 边画边聊:画板开着时按头像说的只给嘴;图跟着这段声音进对话,嘴接住,不起脑
+                const chat = isObj(m.chat) && typeof m.chat.canvas === 'string' && m.chat.canvas ? m.chat : null;
+                if (chat && !this.echoing && !this.mouthOpen) {
+                    this.chatRelease(chat);
+                    break;
+                }
+                this.events.record(null, at, { lane: 'mouth', kind: 'release', how: 'voice' });
+                // 热身在等孩子跟:等转写回来再判是跟读还是对老师说的(heard)
+                if (this.echoing) {
+                    if (this.heldBytes < MIN_COMMIT_BYTES) {
+                        this.session.clearAudio();
+                        this.echoHeard('');
+                        break;
+                    }
+                    this.session.commit();
+                    this.echoHeld = body;
+                    this.state();
+                    break;
+                }
+                if (this.heldBytes < MIN_COMMIT_BYTES) {
+                    this.session.clearAudio();
+                    // 按住「给老师看」没说出话:照交画(画板卡已经锁了、纸上的图传好了),嘴不先说
+                    if (handsIn(body)) {
+                        this.turn('card', body, '');
+                        break;
+                    }
+                    this.send({ t: 'kid', text: '' });
+                    this.turn('empty', body);
+                    break;
+                }
+                this.session.commit();
+                this.turn(this.d.asked && body.newThread !== true ? 'answer' : 'voice', body);
+                break;
+            }
+            case 'hush': {
+                if (!this.session)
+                    return;
+                if (this.holding) {
+                    this.holding = false;
+                    this.session.clearAudio();
+                }
+                this.interrupt();
+                this.d.note({ e: 'release' });
+                this.state();
+                break;
+            }
+            case 'send': {
+                if (!this.session || !isObj(m.body))
+                    return;
+                const body = m.body;
+                // 纸上交的画同交卡:嘴不先说
+                const kind = body.action === 'submit' || isObj(body.drawing) ? 'card' : Array.isArray(body.photos) && body.photos.length ? 'photo' : isObj(body.talk) ? 'warm' : isObj(body.via) ? 'start' : 'text';
+                this.interrupt();
+                this.events.begin([{ at, ev: { lane: 'mouth', kind: 'release', how: kind === 'card' || kind === 'photo' ? kind : 'text' } }]);
+                const text = typeof body.text === 'string' ? body.text : '';
+                this.turn(kind, body, text);
+                break;
+            }
+            // 孩子端放完了小课堂那一段(或放不了、被关):后面的讲稿接着交给嘴
+            case 'staged':
+                this.d.step({ e: 'staged' });
+                break;
+            // 边画边聊:孩子端觉得该说一句了
+            case 'chat':
+                this.chatLook(m);
+                break;
+            // 一直听:孩子端开口时报的上下文(收下时当松手的 body)
+            case 'ctx':
+                if (isObj(m.body))
+                    this.ctxBody = m.body;
+                break;
+            // 一直听:孩子端停推(关麦、卡的声音、录音、小课堂)/ 接着推;停的时候孩子正说到一半,那一句作废
+            case 'gate': {
+                this.gated = m.on === false;
+                if (!this.gated)
+                    break;
+                this.lastActive = Date.now();
+                if (this.vadOn || this.vadAwait !== null) {
+                    this.session?.clearAudio();
+                    this.vadOn = false;
+                    this.vadAwait = null;
+                    this.send({ t: 'vad', on: false });
+                    this.resume();
+                }
+                break;
+            }
+            // 从首页「一直听」进来:嘴先招呼一句(不进脑、不算一条);嘴在说、孩子在说、按住说时都不招呼
+            // 孩子端换了模式:安静时换一条连接(silence 只有新连接能给)
+            case 'listen': {
+                if (typeof m.open !== 'boolean')
+                    break;
+                this.open = m.open;
+                this.lastActive = Date.now();
+                if (!this.session && this.opened)
+                    void this.revive();
+                else
+                    this.maybeSwap();
+                break;
+            }
+            case 'view':
+                this.rec.view(m.w, m.h);
+                break;
+            case 'bye':
+                void this.end('bye');
+                break;
+            default: break;
+        }
+    }
+    /**
+     * 一轮开始:多半嘴不出声,等脑的第一句(孩子端左栏头像在想);拍照、热身这几种才先说一句(src/realtime/cues.ts;上一次回答被掐了还没回 done 就等它回了再说)。
+     * 说的等转写回来再交给脑,不是说的直接交
+     */
+    turn(kind, body, text) {
+        const seq = ++this.turnSeq;
+        this.caught = '';
+        this.d.note({ e: 'release' });
+        // 按住说的等转写回来(heard);一直听的转写已经有了(text)
+        this.pendingVoice = (kind === 'voice' || kind === 'answer') && text === undefined ? { seq, body } : null;
+        const cue = catchCue(kind, this.lang());
+        this.caughtDone = cue === null;
+        if (cue !== null) {
+            const go = () => this.ask({ kind: 'catch' }, cue);
+            if (this.said)
+                this.afterDone = go;
+            else
+                go();
+        }
+        if (text !== undefined)
+            void this.toBrain(seq, { ...body, text });
+    }
+    /** 孩子那句的转写回来了:空的改说没听清、不交给脑;不空交给脑 */
+    heard(text) {
+        if (this.echoHeld) {
+            this.echoHeard(text);
+            return;
+        }
+        // 嘴主动说那次垫的静音:那段的转写(真百炼回空的;转写按 commit 的先后到)不是孩子说的
+        if (this.chatQuiet > 0 && Date.now() < this.chatQuietUntil) {
+            this.chatQuiet--;
+            return;
+        }
+        if (this.chatHeard && Date.now() < this.chatHeardUntil) {
+            const canvas = this.chatHeard;
+            this.chatHeard = null;
+            this.send({ t: 'kid', text });
+            if (text.trim())
+                this.noteDraw(canvas, `孩子:${text.trim()}`);
+            return;
+        }
+        this.send({ t: 'kid', text });
+        const v = this.pendingVoice;
+        this.pendingVoice = null;
+        if (!v || v.seq !== this.turnSeq)
+            return;
+        if (text.trim()) {
+            void this.toBrain(v.seq, { ...v.body, text });
+            return;
+        }
+        // 按住「给老师看」说的转写是空的:画照交,不说没听清
+        if (handsIn(v.body)) {
+            void this.toBrain(v.seq, { ...v.body, text: '' });
+            return;
+        }
+        const sorry = () => this.ask({ kind: 'catch' }, catchCue('empty', this.lang()));
+        if (!this.said) {
+            sorry();
+            return;
+        }
+        // 接住还在要 / 在出:掐掉,出完(被掐)再说没听清
+        this.afterDone = sorry;
+        if (this.session?.responding)
+            this.session.cancel();
+        else
+            this.cancelOnCreate = true;
+        if (this.playing()) {
+            this.d.flush();
+            this.send({ t: 'flush' });
+        }
+    }
+    /** 边画边聊:这张画上记一句(孩子说的、嘴说的),交画那轮给脑 */
+    noteDraw(canvas, line) {
+        const lines = [...(this.drawTalk.get(canvas) ?? []), line];
+        this.drawTalk.set(canvas, lines.slice(-DRAW_TALK_MAX));
+    }
+    /**
+     * 边画边聊:画板开着时按着头像说完松手。声音太少照「没听清」;不然这一刻的画跟着这段声音 commit,嘴接住(看图、顺着他的话),不起脑。
+     * 按下时攒的那几条嘴的事件不归哪一轮(没有这一轮),丢掉
+     */
+    chatRelease(chat) {
+        const s = this.session;
+        const canvas = String(chat.canvas);
+        this.events.begin([]);
+        this.d.note({ e: 'release' });
+        if (this.heldBytes < MIN_COMMIT_BYTES) {
+            s.clearAudio();
+            this.send({ t: 'kid', text: '' });
+            const sorry = () => this.ask({ kind: 'catch' }, catchCue('empty', this.lang()));
+            if (this.said)
+                this.afterDone = sorry;
+            else
+                sorry();
+            this.state();
+            return;
+        }
+        const img = jpegOf(chat.image);
+        if (img)
+            s.appendImage(img);
+        s.commit();
+        this.chatHeard = canvas;
+        this.chatHeardUntil = Date.now() + 10_000;
+        const go = () => this.ask({ kind: 'chat', canvas }, chatCue('reply', askOf(chat.ask), this.drawTalk.get(canvas) ?? []));
+        if (this.said)
+            this.afterDone = go;
+        else
+            go();
+        this.state();
+    }
+    /**
+     * 边画边聊:孩子端觉得该说一句了(src/board/canvas-chat.ts 的规矩)。只在安静时说(孩子没按着、嘴没在说也没要说、脑没在写、讲稿没排着),
+     * 一直听、热身时不说;0.5 秒静音 + 这一刻的画 commit(百炼要先有声音再有图),嘴照应用给的题说一两句。说不了回 {t:'chat', skipped}
+     */
+    chatLook(m) {
+        const s = this.session;
+        const canvas = typeof m.canvas === 'string' ? m.canvas : '';
+        const img = jpegOf(m.image);
+        if (!s?.isReady || !canvas || !img || this.mouthOpen || this.echoing || !this.quiet()) {
+            this.send({ t: 'chat', canvas, skipped: true });
+            return;
+        }
+        s.clearAudio();
+        s.appendAudio(CHAT_SILENCE);
+        if (!s.appendImage(img)) {
+            s.clearAudio();
+            this.send({ t: 'chat', canvas, skipped: true });
+            return;
+        }
+        s.commit();
+        this.chatQuiet++;
+        this.chatQuietUntil = Date.now() + CHAT_QUIET_MS;
+        this.lastActive = Date.now();
+        this.ask({ kind: 'chat', canvas }, chatCue('look', askOf(m.ask), this.drawTalk.get(canvas) ?? []));
+        this.send({ t: 'chat', canvas, said: true });
+    }
+    /** 一直听:VAD 判出孩子开口 / 说完。开口先暂停照念(不丢)、嘴在放就让孩子端压低;说完了等转写 */
+    speech(st) {
+        this.lastActive = Date.now();
+        if (st === 'started') {
+            this.vadOn = true;
+            this.vadFrom = this.ms();
+            if (this.echoing && this.echoing.heardAt === null)
+                this.echoing.heardAt = this.vadFrom;
+            this.vadOverMouth = this.playing() || this.ms() < this.d.end + 300;
+            this.d.note({ e: 'pause' });
+            if (this.playing() && !this.ducked) {
+                this.ducked = true;
+                this.send({ t: 'duck' });
+            }
+            this.send({ t: 'vad', on: true });
+            return;
+        }
+        if (!this.vadOn)
+            return;
+        this.vadOn = false;
+        this.vadAwait = this.ms();
+        this.committedSince = true;
+        this.send({ t: 'vad', on: false });
+    }
+    /** 一直听:放过一句(或孩子端关麦作废了一句),照念接着走、压低的恢复 */
+    resume() {
+        if (this.ducked) {
+            this.ducked = false;
+            this.send({ t: 'unduck' });
+        }
+        if (!this.holding)
+            this.d.step({ e: 'release' });
+        this.maybeSwap();
+    }
+    /** 一直听:一句的转写回来了(null = VAD 不收、转写失败)。收下 = 按下 + 松手;放过的什么都不做,照念接着走 */
+    verdict(text, failed) {
+        const from = this.vadFrom;
+        const to = this.vadAwait ?? this.ms();
+        this.vadOn = false;
+        this.vadAwait = null;
+        // 热身在等孩子跟:「嗯」不算回答(那句是问句也一样);那句放完以后才开口的,字和嘴说的一样就是在跟,不比回声(放完前 300 毫秒的尾巴也不算嘴)
+        const after = this.echoing?.endAt != null && from >= this.echoing.endAt;
+        const v = text === null ? { take: false, why: failed ?? 'rejected' } : judge(text, { asked: this.d.asked && !this.echoing, said: this.saidLog, now: Date.now(), mouthOn: this.vadOverMouth && !after });
+        if (!v.take) {
+            this.send({ t: 'skip', why: v.why, text: text ?? '' });
+            this.resume();
+            if (this.echoing?.overdue)
+                this.echoDue();
+            return;
+        }
+        const said = text ?? '';
+        this.lastActive = Date.now();
+        // 热身:像跟读就开闸下一拍(不接住、不交脑);对老师说的记一笔,照收下往下走
+        if (this.echoing) {
+            const kind = echoKind(said, this.echoing.echo.word);
+            this.noteEcho(kind, said);
+            if (kind === 'said') {
+                this.rec.in({ t: 'hold', vad: true }, from);
+                this.rec.in({ t: 'release', vad: true, echo: true }, to);
+                this.echoed();
+                return;
+            }
+        }
+        this.send({ t: 'kid', text: said });
+        // 录像按孩子端的按下 / 松手画这一句:补记开口、说完两条(当时的时刻)
+        this.rec.in({ t: 'hold', vad: true }, from);
+        this.rec.in({ t: 'release', vad: true, body: this.ctxBody }, to);
+        // 停一下接着说的:上一句还没交给脑,并进去
+        if (this.draft && this.draft.seq === this.turnSeq) {
+            this.draft.texts.push(said);
+            if (this.ducked) {
+                this.ducked = false;
+                this.send({ t: 'unduck' });
+            }
+            if (!this.holding)
+                this.d.step({ e: 'release' });
+            return;
+        }
+        this.events.begin([{ at: this.wall(from), ev: { lane: 'mouth', kind: 'hold' } }, { at: this.wall(to), ev: { lane: 'mouth', kind: 'release', how: 'vad' } }]);
+        this.ducked = false;
+        this.interrupt();
+        const body = { ...this.ctxBody };
+        // draft 先占上(turn 里交给脑是同步起步的,停一下接着说的要并得进来)
+        this.draft = { seq: this.turnSeq + 1, texts: [said] };
+        this.turn(this.d.asked && body.newThread !== true ? 'answer' : 'voice', body, said);
+    }
+    /** 孩子那句转写失败:按住说的照转写是空的办;一直听时不到 1 秒的放过,长的嘴说没听清 */
+    kidFailed() {
+        if (!this.mouthOpen) {
+            this.heard('');
+            return;
+        }
+        const long = (this.vadAwait ?? this.ms()) - this.vadFrom >= 1000;
+        this.verdict(null, 'failed');
+        if (long && !this.said && this.session)
+            this.ask({ kind: 'catch' }, catchCue('empty', this.lang()));
+    }
+    /** 等嘴接住(最多 CATCH_WAIT_MS),交给脑;脑在忙就排着,它写完合并发 */
+    async toBrain(seq, body) {
+        const until = Date.now() + CATCH_WAIT_MS;
+        while (!this.caughtDone && Date.now() < until && !this.ended && seq === this.turnSeq)
+            await new Promise((r) => setTimeout(r, 20));
+        // 一直听:嘴不先说了,先留一小会儿看孩子是不是停一下又开口(原来接住那句正好留出这段)
+        const peek = Date.now() + OPEN_PEEK_MS;
+        while (this.mouthOpen && this.draft?.seq === seq && !this.vadOn && this.vadAwait === null && Date.now() < peek && !this.ended && seq === this.turnSeq)
+            await new Promise((r) => setTimeout(r, 20));
+        // 一直听:孩子停一下又开口了(或刚说完在等转写),等这一句判完,收下的并进来(probe-live-vad:句与句之间常常只隔 0–300 毫秒)
+        const more = Date.now() + MERGE_WAIT_MS;
+        while (this.draft?.seq === seq && (this.vadOn || this.vadAwait !== null) && Date.now() < more && !this.ended && seq === this.turnSeq)
+            await new Promise((r) => setTimeout(r, 20));
+        if (this.ended || seq !== this.turnSeq)
+            return;
+        if (this.draft?.seq === seq) {
+            body = { ...body, text: this.draft.texts.join(',') };
+            this.draft = null;
+        }
+        await this.turns.submit(body);
+    }
+    /**
+     * 这一轮实时的那些(收下了才算,交给脑时带):口语课热身的摸底一直带着,打了电话带打了几分钟,应用起的这一轮说明为什么;
+     * 交画那一轮(交画板卡、纸上的画):这张画上边画边说的带给脑,之后清掉;上一轮念到第几句被打断,带了就清
+     */
+    liveOf(body) {
+        const { app, ...rest } = body;
+        const drawKey = rest.action === 'submit' && isObj(rest.focus) && typeof rest.focus.card === 'string' ? rest.focus.card : isObj(rest.drawing) ? 'paper' : null;
+        const draw = drawKey ? this.drawTalk.get(drawKey) : undefined;
+        if (drawKey)
+            this.drawTalk.delete(drawKey);
+        const cut = this.d.takeCut();
+        return {
+            ...(draw?.length ? { draw } : {}),
+            ...(this.caught ? { caught: this.caught } : {}),
+            ...(cut !== undefined ? { cut } : {}),
+            ...(this.warm.length ? { warmup: warmLines(this.warm) } : {}),
+            ...(this.phoneAt !== null ? { phone: Math.max(0, Math.floor((Date.now() - this.phoneAt) / 60_000)) } : {}),
+            ...(app === 'call' || app === 'quiet' ? { app: app } : {}),
+        };
+    }
+    /** 接住说中文还是英文:口语课打了电话以后说英文 */
+    lang() {
+        return this.phoneAt !== null ? 'en' : 'zh';
+    }
+    /** 热身那句在孩子端放完的时刻(通话毫秒)知道了:从那时起等 ECHO_WAIT_MS */
+    armEcho(endAt) {
+        const e = this.echoing;
+        if (!e || e.endAt !== null)
+            return;
+        e.endAt = endAt;
+        e.timer = setTimeout(() => this.echoDue(), Math.max(0, endAt - this.ms()) + (this.deps.echoWaitMs ?? ECHO_WAIT_MS));
+    }
+    /** 热身等够了:孩子正在说(或说完了在等转写、按着)就等它判完;没开口记一笔,接着下一拍;连着 QUIET_BEATS 拍没开口停下交脑 */
+    echoDue() {
+        const e = this.echoing;
+        if (!e || this.ended)
+            return;
+        e.timer = null;
+        if (this.vadOn || this.vadAwait !== null || this.holding || this.echoHeld) {
+            e.overdue = true;
+            return;
+        }
+        this.noteEcho('silent');
+        if (this.quietBeats >= QUIET_BEATS) {
+            this.interrupt();
+            this.d.note({ e: 'release' });
+            this.quietBeats = 0;
+            this.appTurn('(没出声)', 'quiet');
+            return;
+        }
+        this.echoed();
+    }
+    /** 热身记一拍(摸底;录像里一条 {t:'echo'},只记不发) */
+    noteEcho(outcome, text) {
+        const e = this.echoing;
+        if (!e)
+            return;
+        const startMs = e.heardAt !== null && e.endAt !== null ? Math.max(0, e.heardAt - e.endAt) : undefined;
+        const beat = { word: e.echo.word, say: e.echo.say === true, outcome, ...(text !== undefined ? { text } : {}), ...(startMs !== undefined && outcome !== 'silent' ? { startMs } : {}), ...(text !== undefined && e.echo.word ? { hit: coverage(e.echo.word, text) } : {}) };
+        this.warm.push(beat);
+        this.quietBeats = outcome === 'silent' ? this.quietBeats + 1 : 0;
+        this.rec.out({ t: 'echo', ...beat });
+    }
+    /** 热身这拍过了:开闸,下一拍接着念;压低的恢复 */
+    echoed() {
+        if (this.echoing?.timer)
+            clearTimeout(this.echoing.timer);
+        this.echoing = null;
+        this.d.step({ e: 'echoed' });
+        this.resume();
+    }
+    /** 热身里按住说的那一下转写回来了:空的接着等;像跟读开闸;对老师说的同松手(打断、接住、交给脑) */
+    echoHeard(text) {
+        const body = this.echoHeld ?? {};
+        this.echoHeld = null;
+        if (!this.echoing)
+            return;
+        if (!text.trim()) {
+            if (this.echoing.overdue)
+                this.echoDue();
+            return;
+        }
+        const kind = echoKind(text, this.echoing.echo.word);
+        this.noteEcho(kind, text);
+        if (kind === 'said') {
+            this.echoed();
+            return;
+        }
+        this.send({ t: 'kid', text });
+        this.interrupt();
+        this.turn(this.d.asked && body.newThread !== true ? 'answer' : 'voice', body, text);
+    }
+    /** [[call]] 那句出齐了:它在孩子端放完时开始响铃、算打电话的起点;马上交脑写英文开场(铃声垫着,嘴不接) */
+    startPhone() {
+        if (this.phoneAt !== null)
+            return;
+        this.phoneAt = this.wall(this.d.end);
+        // 铃响够之前英文开场先不交(排队在 [[call]] 那句关着闸,rung 才开;孩子开口的暂停 / 放过开不了它)
+        this.timers.push(setTimeout(() => { if (!this.ended)
+            this.d.step({ e: 'rung' }); }, Math.max(0, this.d.end - this.ms()) + (this.deps.ringMs ?? RING_MS)));
+        this.appTurn('(打电话)', 'call');
+    }
+    /** 应用起的一轮(不是孩子说的):嘴不出声,接在这次通话的话题后面交给脑 */
+    appTurn(text, app) {
+        const seq = ++this.turnSeq;
+        this.caught = '';
+        this.caughtDone = true;
+        this.pendingVoice = null;
+        this.draft = null;
+        const body = { text, app, ...(this.turns.thread ? { thread: this.turns.thread } : {}), ...(this.ctxBody.device ? { device: this.ctxBody.device } : {}) };
+        void this.toBrain(seq, body);
+    }
+    /** 孩子端的一块声音:按着才推,一直听时一直推;嘴在现连时按着说的先攒着(最多一分钟),一直听的只留最近 10 秒 */
+    onAudio(pcm) {
+        if (this.ended)
+            return;
+        if (this.open)
+            this.lastAudioAt = Date.now();
+        if (!this.session) {
+            if (this.open && this.opened) {
+                this.rec.kid(pcm);
+                this.backlog.push({ pcm });
+                this.backlogBytes += pcm.length;
+                while (this.backlogBytes > MAX_OPEN_BACKLOG_BYTES) {
+                    const i = this.backlog.findIndex((x) => 'pcm' in x);
+                    if (i < 0)
+                        break;
+                    this.backlogBytes -= this.backlog[i].pcm.length;
+                    this.backlog.splice(i, 1);
+                }
+                void this.revive();
+                return;
+            }
+            if (this.backlogHolding && this.backlogBytes + pcm.length <= MAX_BACKLOG_BYTES) {
+                this.rec.kid(pcm);
+                this.backlog.push({ pcm });
+                this.backlogBytes += pcm.length;
+            }
+            return;
+        }
+        if (!(this.holding || this.mouthOpen) || !this.session.isReady)
+            return;
+        this.rec.kid(pcm);
+        this.pushAudio(pcm);
+    }
+    pushAudio(pcm) {
+        if (!(this.holding || this.mouthOpen) || !this.session?.isReady)
+            return;
+        if (this.holding)
+            this.heldBytes += pcm.length;
+        this.session.appendAudio(pcm);
+    }
+    /** 收:关嘴、落录像、告诉孩子端 */
+    async end(why) {
+        if (this.ended)
+            return;
+        this.ended = true;
+        for (const t of this.timers) {
+            clearTimeout(t);
+            clearInterval(t);
+        }
+        this.d.stop();
+        if (this.echoing?.timer)
+            clearTimeout(this.echoing.timer);
+        this.offEvents?.();
+        const s = this.session;
+        this.session = null;
+        s?.close('end');
+        this.send({ t: 'end', why });
+        // 字幕重放:挂的时候嘴还在出的那段,出了的那截也落;等这几轮都落完
+        if (this.said?.kind === 'read')
+            this.mouthEnd(this.said, false);
+        await this.takes.flush();
+        try {
+            const take = this.rec.finish();
+            if (take.reel.msgs.length > 2) {
+                const { ctx, tutor } = this.deps;
+                const at = this.rec.startedAt;
+                const dir = join(ctx.runner.ws.dirs.conversations, tutor, `${localDate(at)}.live`);
+                const name = [at.getHours(), at.getMinutes(), at.getSeconds()].map((n) => String(n).padStart(2, '0')).join('');
+                await mkdir(dir, { recursive: true });
+                await writeFile(join(dir, `${name}.wav`), wavOfPcm(take.callPcm, MIX_RATE));
+                await writeFile(join(dir, `${name}.reel.json`), JSON.stringify(take.reel));
+            }
+        }
+        catch (err) {
+            console.error('live: 录像落盘失败', err);
+        }
+        this.sock.close(1000, why);
+        this.deps.onFinished?.(why);
+    }
+    /** 孩子端断了 */
+    onSocketClose() { if (!this.ended)
+        void this.end('kid-gone'); }
+}
+const isObj = (v) => typeof v === 'object' && v !== null && !Array.isArray(v);
+/** 孩子端发来的图(data:image/jpeg;base64,…)→ 字节;不是 jpeg、太大(百炼一帧上限)回 null */
+const jpegOf = (v) => {
+    const m = typeof v === 'string' ? /^data:image\/jpeg;base64,([A-Za-z0-9+/=]+)$/.exec(v) : null;
+    return m && m[1].length <= MAX_IMAGE_B64 ? Buffer.from(m[1], 'base64') : null;
+};
+/** 这一条带着交的东西(交卡、纸上的画):没说清也要交给脑 */
+const handsIn = (body) => body.action === 'submit' || isObj(body.drawing);
+/** 录像里记的那条:去掉图,留一个 image: true */
+const withoutImages = (m) => {
+    if (typeof m.image === 'string')
+        return { ...m, image: true };
+    if (isObj(m.chat) && typeof m.chat.image === 'string')
+        return { ...m, chat: { ...m.chat, image: true } };
+    return m;
+};
+/** 孩子端交来的 ask:只留认得的字段 */
+const askOf = (v) => {
+    if (!isObj(v))
+        return {};
+    const colors = Array.isArray(v.colors) ? v.colors.filter((c) => isObj(c) && typeof c.name === 'string' && typeof c.strokes === 'number' && typeof c.where === 'string').slice(0, 6) : undefined;
+    return {
+        ...(typeof v.prompt === 'string' && v.prompt.trim() ? { prompt: v.prompt.trim().slice(0, 200) } : {}),
+        ...(typeof v.color === 'string' && v.color ? { color: v.color.slice(0, 8) } : {}),
+        ...(typeof v.ask === 'boolean' ? { ask: v.ask } : {}),
+        ...(typeof v.minutes === 'string' ? { minutes: v.minutes.slice(0, 20) } : {}),
+        ...(colors?.length ? { colors } : {}),
+        ...(typeof v.cleared === 'number' && v.cleared > 0 ? { cleared: Math.floor(v.cleared) } : {}),
+    };
+};
+/** 孩子这边的动作(嘴断了要攒着补的那几种) */
+const INPUTS = new Set(['hold', 'release', 'hush', 'send', 'staged']);