diff --git a/mobile-uni/src/pages/user/sop/index.vue b/mobile-uni/src/pages/user/sop/index.vue index 624ed15c..5ceed312 100644 --- a/mobile-uni/src/pages/user/sop/index.vue +++ b/mobile-uni/src/pages/user/sop/index.vue @@ -108,7 +108,7 @@ {{ voiceLabel(msg) }} - @@ -231,7 +231,7 @@ - + 收起全文 ▲ @@ -497,6 +497,9 @@ interface MasterMsg { voiceDuration: number | null; unread: boolean; transcriptOpen: boolean; + /** Set once the reader toggles the transcript, so preparing voice never + * collapses text the reader chose to keep open. */ + transcriptPinned: boolean; } interface SystemMsg { @@ -993,7 +996,7 @@ const sendQuestion = async ( } retireEarlierMemoryCandidate(nextResult.memoryCandidate?.id); const spoken = (nextResult.answer || '').trim(); - const shouldPreloadVoice = Boolean(spoken) && !isNativeAppRuntime(); + const shouldPreloadVoice = Boolean(spoken); replaceMsg(pending.id, { id: pending.id, role: 'master', @@ -1008,7 +1011,10 @@ const sendQuestion = async ( voiceStatus: shouldPreloadVoice ? 'preparing' : 'unavailable', voiceDuration: null, unread: Boolean(spoken), - transcriptOpen: false + // Keep the answer readable while the voice is still synthesizing; it + // collapses on its own once the voice bubble can show a duration. + transcriptOpen: true, + transcriptPinned: false }); if (shouldPreloadVoice) void prepareVoice(pending.id, spoken); if (!nextResult.answer && !nextResult.snippets?.length && !nextResult.memoryCandidate) { @@ -1155,22 +1161,47 @@ const findMasterMsg = (id: number): MasterMsg | null => { return found && found.role === 'master' ? found : null; }; +/** Upper bound on the `preparing` voice bubble. Remote TTS has been seen to + * hang well past its own timeout; after this the reply degrades to the plain + * text bubble instead of showing an endless placeholder duration. */ +const VOICE_PREPARE_TIMEOUT_MS = 20000; + const spokenTextOf = (msg: MasterMsg) => (msg.result.answer || '').trim(); /** Pre-synthesize the reply audio so the voice bubble can show a duration * before the first tap, WeChat-style. Falls back to a plain text bubble * when remote TTS is unavailable. */ const prepareVoice = async (id: number, spoken: string) => { - const source = await speechPlayback.preload(spoken); + const source = await Promise.race([ + speechPlayback.preload(spoken), + new Promise((resolve) => { + setTimeout(() => resolve(null), VOICE_PREPARE_TIMEOUT_MS); + }) + ]); const msg = findMasterMsg(id); if (!msg) return; if (!source) { + // No voice to offer: leave the text bubble and its play button in place. msg.voiceStatus = 'unavailable'; msg.unread = false; return; } msg.voiceDuration = await measureAudioDuration(source); msg.voiceStatus = 'ready'; + // Deliberately not collapsing here: synthesis often finishes in 2-4s, which + // would pull the answer away mid-sentence. The transcript collapses when the + // reader starts the voice instead. +}; + +/** Pin on any manual toggle so a later `ready` voice keeps the reader's choice. */ +const toggleTranscript = (msg: MasterMsg) => { + msg.transcriptOpen = !msg.transcriptOpen; + msg.transcriptPinned = true; +}; + +const collapseTranscript = (msg: MasterMsg) => { + msg.transcriptOpen = false; + msg.transcriptPinned = true; }; const isVoicePlaying = (msg: MasterMsg) => @@ -1188,6 +1219,9 @@ const playVoice = (msg: MasterMsg) => { return; } msg.unread = false; + // Listening replaces reading, so fold the transcript away unless the reader + // explicitly asked to keep it open. + if (!msg.transcriptPinned) msg.transcriptOpen = false; void speechPlayback.toggle(`answer-${msg.id}`, spokenTextOf(msg)); }; diff --git a/mobile-uni/tests/agent.test.mjs b/mobile-uni/tests/agent.test.mjs index 21f115f6..c4a78148 100644 --- a/mobile-uni/tests/agent.test.mjs +++ b/mobile-uni/tests/agent.test.mjs @@ -38,12 +38,31 @@ test('Agent 响应按结构化状态渲染来源、澄清和确认动作', async assert.doesNotMatch(page, /result\.answer\.(?:includes|match|startsWith)/); }); -test('原生 App 立即显示 Agent 文字回答，语音仅按需生成', async () => { +test('回复语音优先，但文字在合成期间可读且不覆盖读者选择', async () => { const page = await source('../src/pages/user/sop/index.vue'); - assert.match(page, /const shouldPreloadVoice = Boolean\(spoken\) && !isNativeAppRuntime\(\)/); - assert.match(page, /voiceStatus: shouldPreloadVoice \? 'preparing' : 'unavailable'/); - assert.match(page, /if \(shouldPreloadVoice\) void prepareVoice\(pending\.id, spoken\)/); + // Voice-first applies on every platform, not just H5. + assert.match(page, /const shouldPreloadVoice = Boolean\(spoken\);/); + assert.doesNotMatch(page, /shouldPreloadVoice = [^;]*isNativeAppRuntime/); + + // Text is readable while the voice is still preparing. + assert.match(page, /transcriptOpen: true/); + + // Finishing synthesis must not yank the text away; only starting playback folds it. + assert.match( + page, + /msg\.unread = false;\s*(?:\/\/[^\n]*\n\s*)*if \(!msg\.transcriptPinned\) msg\.transcriptOpen = false;\s*void speechPlayback\.toggle/ + ); + assert.doesNotMatch( + page, + /msg\.voiceStatus = 'ready';\s*if \(!msg\.transcriptPinned\)/ + ); + assert.match(page, /const toggleTranscript = [^}]*transcriptPinned = true;/s); + assert.match(page, /const collapseTranscript = [^}]*transcriptPinned = true;/s); + + // A hanging synthesis degrades to the plain text bubble instead of an endless placeholder. + assert.match(page, /VOICE_PREPARE_TIMEOUT_MS/); + assert.match(page, /Promise\.race\(\[\s*speechPlayback\.preload\(spoken\)/); }); test('Agent 确认卡通过不透明 draftId 调用动作端点', async () => {