fix(client): the voice reads a stored reply's words, not its untrusted fence or its delegation trace

changelog.d/1405.fixed.md

@@ -0,0 +1 @@
+Voice mode no longer reads a reply's framing aloud: the host header, a reply's delegation trace (`Chief → Product`) and the untrusted-fence markers a stored remote line carries are stripped along the same text chain a row preview reads, so `speechSynthesis` says the words and nothing else.

core/client.js

@@ -2532,6 +2532,9 @@ function dropKnownTrace(line2, names) {
   return /^[\s,，.。:：]*$/.test(rest) ? "" : rest;
 }
 var botNames = (roster) => [...roster?.bots ?? [], ...roster?.remoteBots ?? [], ...roster?.agents ?? []].map((entry) => entry?.name).filter((name) => typeof name === "string" && name !== "");
+function speechOf(text, names = []) {
+  return readableLine(text, names).replace(/```[\s\S]*?```/g, " ").replace(/\[([^\]]*)\]\([^)]*\)/g, "$1").replace(/[*_`#>|~-]+/g, " ").replace(/\s+/g, " ").trim();
+}
 var formatTokens = (count) => count >= 1e6 ? `${(count / 1e6).toFixed(1)}M` : count >= 1e3 ? `${(count / 1e3).toFixed(1)}K` : String(count);
 function usageOf(list, ids) {
   let input = 0;
@@ -14127,7 +14130,6 @@ function ExchangeDialog({ pair, roster, actions }) {
 
 // src/client/voice.js
 var import_react33 = require("react");
-var plainSpeech = (text) => stripHeader(text).replace(/```[\s\S]*?```/g, " ").replace(/\[([^\]]*)\]\([^)]*\)/g, "$1").replace(/[*_`#>|~-]+/g, " ").replace(/\s+/g, " ").trim();
 var VOICE_LABELS = {
   listening: "Listening",
   thinking: "Thinking",
@@ -14316,7 +14318,7 @@ function VoiceMode({ sessionId, roster, actions, useSessions, useSessionStatus }
     const [count, text] = [lastReply.slice(0, lastReply.indexOf("|")), lastReply.slice(lastReply.indexOf("|") + 1)];
     if (count === waiting.current || !text.trim()) return;
     waiting.current = null;
-    const spoken = plainSpeech(text);
+    const spoken = speechOf(text, botNames(roster));
     setSaid(spoken);
     setHeard("");
     if (!window.speechSynthesis || !spoken) {

core/src/client/text.js

@@ -332,6 +332,18 @@ function dropKnownTrace(line, names) {
 export const botNames = roster => [...(roster?.bots ?? []), ...(roster?.remoteBots ?? []), ...(roster?.agents ?? [])]
   .map(entry => entry?.name).filter(name => typeof name === 'string' && name !== '')
 
+// What voice mode reads aloud: a reply's words, read the one way a row reads a stored line — header,
+// fence and a known-name delegation trace dropped (#1395, #1408) — and then with the Markdown a speech
+// synthesiser would otherwise say out loud (a code fence, a link's target, emphasis marks).
+export function speechOf(text, names = []) {
+  return readableLine(text, names)
+    .replace(/```[\s\S]*?```/g, ' ')
+    .replace(/\[([^\]]*)\]\([^)]*\)/g, '$1')
+    .replace(/[*_`#>|~-]+/g, ' ')
+    .replace(/\s+/g, ' ')
+    .trim()
+}
+
 const formatTokens = count => (count >= 1e6 ? `${(count / 1e6).toFixed(1)}M` : count >= 1e3 ? `${(count / 1e3).toFixed(1)}K` : String(count))
 
 // Same buckets as the shell's usage pill: every prompt-side billing bucket plus output.

core/src/client/voice.js

@@ -1,5 +1,5 @@
 import { createElement as h, useEffect, useRef, useState } from 'react'
-import { stripHeader } from './text.js'
+import { botNames, speechOf } from './text.js'
 import { MicIcon, CloseIcon } from './icons.js'
 import { BotMark, RoomAvatar } from './mark.js'
 import { speechApi, useDictation } from './composer.js'
@@ -12,7 +12,6 @@ import { openCapture, speechCatalog, transcribeWave, turnDetector } from './spee
 // again. With DSH's Voice Input plugin on, its recognizer hears the user: one
 // recording ends after a pause and is transcribed whole. Otherwise the browser's
 // own speech recognition does. The browser's speech synthesis reads the answer.
-const plainSpeech = text => stripHeader(text).replace(/```[\s\S]*?```/g, ' ').replace(/\[([^\]]*)\]\([^)]*\)/g, '$1').replace(/[*_`#>|~-]+/g, ' ').replace(/\s+/g, ' ').trim()
 const VOICE_LABELS = {
   listening: 'Listening', thinking: 'Thinking', speaking: 'Speaking', paused: 'Paused',
   setup: 'Voice input needs its speech model', unsupported: 'Voice chat is not available in this browser',
@@ -184,7 +183,7 @@ export function VoiceMode({ sessionId, roster, actions, useSessions, useSessionS
     const [count, text] = [lastReply.slice(0, lastReply.indexOf('|')), lastReply.slice(lastReply.indexOf('|') + 1)]
     if (count === waiting.current || !text.trim()) return
     waiting.current = null
-    const spoken = plainSpeech(text)
+    const spoken = speechOf(text, botNames(roster))
     setSaid(spoken)
     setHeard('')
     if (!window.speechSynthesis || !spoken) { setPhase('listening'); return }

core/test/voice-speech.test.js

@@ -0,0 +1,82 @@
+import test from 'node:test'
+import assert from 'node:assert/strict'
+import { readFileSync } from 'node:fs'
+import { fileURLToPath } from 'node:url'
+import { build } from 'esbuild'
+import { fenceUntrusted } from '../src/host/delivery.js'
+
+// Voice mode reads a Bot's reply aloud, and a stored turn is framing plus words: the host's header,
+// a reply's delegation trace and an untrusted fence are routing, not content. The speech text is read
+// the one way a row reads a stored line — `readableLine(text, names)` in text.js — because
+// `speechSynthesis` repeats what it is handed: a fence mark reaches the user's ears as
+// `<untrusted message from="User"`, and a trace as "Chief → Product".
+
+// delivery.js:41 writes `headerFor(...)\n` and then the body, or `fenceUntrusted(...)` around it.
+const stored = (header, name, body) => `${header}\n${fenceUntrusted(name, body)}`
+
+const text = await build({
+  entryPoints: [fileURLToPath(new URL('../src/client/text.js', import.meta.url))],
+  bundle: true, format: 'esm', platform: 'node', write: false,
+})
+const { speechOf } = await import(`data:text/javascript;base64,${Buffer.from(text.outputFiles[0].text).toString('base64')}`)
+
+// The names a conversation's roster carries: a trace hop is a real Bot name, and only a hop that is
+// one is routing (#1408).
+const NAMES = ['Chief', 'Product', 'Dev Lead', 'Dev-1']
+
+test('a fenced stored reply is spoken as its words, never as its fence mark', () => {
+  assert.equal(typeof speechOf, 'function', 'text.js must export the read-side chain the voice speaks through')
+  const line = stored('[Message from User]', 'User', 'B踢后测4')
+  const spoken = speechOf(line, NAMES)
+  assert.equal(spoken, 'B踢后测4')
+  assert.equal(spoken.includes('<untrusted'), false, 'no fence mark is read aloud')
+  assert.equal(spoken.includes('from="'), false, 'no fence attribute is read aloud')
+  assert.equal(spoken.includes('untrusted-message'), false, 'no fence name is read aloud')
+})
+
+test('a reply\'s multi-line delegation trace is routing, so it is not read aloud', () => {
+  assert.equal(speechOf('Chief → Product\nB踢后测4', NAMES), 'B踢后测4')
+  assert.equal(speechOf('Chief → Product\nB踢后测4', NAMES).includes('Chief → Product'), false)
+  // A sentence that merely carries an arrow is not a trace, and is said in full.
+  assert.equal(speechOf('B踢后测4 → 已完成', NAMES), 'B踢后测4 → 已完成')
+})
+
+test('#1405: the outline flattens a trace onto the words, so a Latin sentence is not read with its routing', () => {
+  // The shape the outline stores — one line, trace and words sharing it, no newline after the
+  // trace (core/test/inbound-ui.test.js pins that same string). A rule that needs the trace to
+  // end at a line break never fires here, and `speechSynthesis` says the routing out loud.
+  assert.equal(speechOf('Chief → 助手 Hi! Message received.', ['Chief', '助手']), 'Hi! Message received.')
+})
+
+test('#1405: the same flat shape in front of a Chinese body is not read either', () => {
+  assert.equal(speechOf('Chief → Product B踢后测4', NAMES), 'B踢后测4')
+  assert.equal(speechOf('Chief → Dev Lead → Dev-1 点头', NAMES), '点头')
+})
+
+test('#1408: an ordinary arrow line keeps its words, as the row reads it', () => {
+  // The blind flatten this issue first used would read `Settings → Models 页面` as "页面";
+  // the shared name-aware reader keeps a line whose hops are not Bot names whole.
+  assert.equal(speechOf('Settings → Models 页面', NAMES), 'Settings → Models 页面')
+  assert.equal(speechOf('B踢后测4 → 已完成', NAMES), 'B踢后测4 → 已完成')
+  // With no names nothing that merely looks like a trace is removed: the voice passes the roster.
+  assert.equal(speechOf('Chief → 助手 Hi there'), 'Chief → 助手 Hi there')
+})
+
+test('a line that was never fenced is read byte for byte as it was', () => {
+  for (const line of ['B踢后测4', 'hello there', 'It costs $5.']) assert.equal(speechOf(line, NAMES), line)
+})
+
+test('the rest of what the voice says is unchanged: markdown is words', () => {
+  assert.equal(speechOf('**bold** and `code`', NAMES), 'bold and code')
+  assert.equal(speechOf('see [the issue](https://example.com/1)', NAMES), 'see the issue')
+  assert.equal(speechOf('```\nconst a = 1\n```', NAMES), '')
+  assert.equal(speechOf('  spaced\t\tout  ', NAMES), 'spaced out')
+})
+
+test('voice.js speaks through text.js, not through a chain of its own', () => {
+  const voice = readFileSync(new URL('../src/client/voice.js', import.meta.url), 'utf8')
+  assert.match(voice, /import \{[^}]*\bbotNames\b[^}]*\bspeechOf\b[^}]*\} from '\.\/text\.js'/, 'the speech text and the conversation names come from the read side')
+  assert.match(voice, /const spoken = speechOf\(text, botNames\(roster\)\)/, 'one chain: the voice hands the read side its names')
+  assert.equal(/plainSpeech/.test(voice), false, 'the voice keeps no second reading of its own')
+  assert.equal(/stripHeader\(text\)/.test(voice), false, 'the header, the trace and the fence are read once, not by the voice')
+})

patches/ds-bot/0001-core-fixes.patch

@@ -2379,7 +2379,7 @@ diff --git a/src/client/text.js b/src/client/text.js
  // parts) are not events in the chat, nor is a request handed back after its line was
  // already shown: deferred, resumed after a restart, or retried after a tool call
  // written as text.
-@@ -200,18 +200,72 @@ export function commRunOf(order, peers, turn, hiddenTurns = []) {
+@@ -200,18 +200,84 @@ export function commRunOf(order, peers, turn, hiddenTurns = []) {
  }
 
  // Latest visible line for a sidebar row; Bot-to-Bot exchange turns stay out of it.
@@ -2451,6 +2451,18 @@ diff --git a/src/client/text.js b/src/client/text.js
 +// A chat whose roster is not loaded still reads its text, just never as a trace.
 +export const botNames = roster => [...(roster?.bots ?? []), ...(roster?.remoteBots ?? []), ...(roster?.agents ?? [])]
 +  .map(entry => entry?.name).filter(name => typeof name === 'string' && name !== '')
++
++// What voice mode reads aloud: a reply's words, read the one way a row reads a stored line — header,
++// fence and a known-name delegation trace dropped (#1395, #1408) — and then with the Markdown a speech
++// synthesiser would otherwise say out loud (a code fence, a link's target, emphasis marks).
++export function speechOf(text, names = []) {
++  return readableLine(text, names)
++    .replace(/```[\s\S]*?```/g, ' ')
++    .replace(/\[([^\]]*)\]\([^)]*\)/g, '$1')
++    .replace(/[*_`#>|~-]+/g, ' ')
++    .replace(/\s+/g, ' ')
++    .trim()
++}
 +
  const formatTokens = count => (count >= 1e6 ? `${(count / 1e6).toFixed(1)}M` : count >= 1e3 ? `${(count / 1e3).toFixed(1)}K` : String(count))
 
@@ -2494,6 +2506,21 @@ diff --git a/src/client/text.js b/src/client/text.js
 diff --git a/src/client/voice.js b/src/client/voice.js
 --- a/src/client/voice.js
 +++ b/src/client/voice.js
+@@ -1,5 +1,5 @@
+ import { createElement as h, useEffect, useRef, useState } from 'react'
+-import { stripHeader } from './text.js'
++import { botNames, speechOf } from './text.js'
+ import { MicIcon, CloseIcon } from './icons.js'
+ import { BotMark, RoomAvatar } from './mark.js'
+ import { speechApi, useDictation } from './composer.js'
+@@ -12,7 +12,6 @@ import { openCapture, speechCatalog, transcribeWave, turnDetector } from './spee
+ // again. With DSH's Voice Input plugin on, its recognizer hears the user: one
+ // recording ends after a pause and is transcribed whole. Otherwise the browser's
+ // own speech recognition does. The browser's speech synthesis reads the answer.
+-const plainSpeech = text => stripHeader(text).replace(/```[\s\S]*?```/g, ' ').replace(/\[([^\]]*)\]\([^)]*\)/g, '$1').replace(/[*_`#>|~-]+/g, ' ').replace(/\s+/g, ' ').trim()
+ const VOICE_LABELS = {
+   listening: 'Listening', thinking: 'Thinking', speaking: 'Speaking', paused: 'Paused',
+   setup: 'Voice input needs its speech model', unsupported: 'Voice chat is not available in this browser',
 @@ -27,7 +27,7 @@ function useMicLevel(enabled) {
      let audio = null
      let frame = 0
@@ -2503,6 +2530,15 @@ diff --git a/src/client/voice.js b/src/client/voice.js
        if (stopped) { media.getTracks().forEach(track => track.stop()); return }
        stream = media
        audio = new AudioContext()
+@@ -184,7 +184,7 @@ export function VoiceMode({ sessionId, roster, actions, useSessions, useSessionS
+     const [count, text] = [lastReply.slice(0, lastReply.indexOf('|')), lastReply.slice(lastReply.indexOf('|') + 1)]
+     if (count === waiting.current || !text.trim()) return
+     waiting.current = null
+-    const spoken = plainSpeech(text)
++    const spoken = speechOf(text, botNames(roster))
+     setSaid(spoken)
+     setHeard('')
+     if (!window.speechSynthesis || !spoken) { setPhase('listening'); return }
 diff --git a/src/host/agents.js b/src/host/agents.js
 --- a/src/host/agents.js
 +++ b/src/host/agents.js

patches/ds-bot/0008-client-bundle.patch

@@ -1743,7 +1743,7 @@ diff --git a/client.js b/client.js
  var historyLoading = (snapshot) => snapshot?.openState === "cold" || snapshot?.openState === "loading";
  function toolArgs(root) {
    if (typeof root.args?.text === "function") {
-@@ -1597,17 +1597,43 @@ function commRunOf(order, peers, turn, hiddenTurns = []) {
+@@ -1597,17 +1597,46 @@ function commRunOf(order, peers, turn, hiddenTurns = []) {
    }
    return { count, peerIds };
  }
@@ -1787,6 +1787,9 @@ diff --git a/client.js b/client.js
 +  return /^[\s,，.。:：]*$/.test(rest) ? "" : rest;
 +}
 +var botNames = (roster) => [...roster?.bots ?? [], ...roster?.remoteBots ?? [], ...roster?.agents ?? []].map((entry) => entry?.name).filter((name) => typeof name === "string" && name !== "");
++function speechOf(text, names = []) {
++  return readableLine(text, names).replace(/```[\s\S]*?```/g, " ").replace(/\[([^\]]*)\]\([^)]*\)/g, "$1").replace(/[*_`#>|~-]+/g, " ").replace(/\s+/g, " ").trim();
++}
  var formatTokens = (count) => count >= 1e6 ? `${(count / 1e6).toFixed(1)}M` : count >= 1e3 ? `${(count / 1e3).toFixed(1)}K` : String(count);
  function usageOf(list, ids) {
    let input = 0;
@@ -16103,7 +16106,7 @@ diff --git a/client.js b/client.js
      let stop = false;
      let timer;
      let seen = null;
-@@ -8612,48 +8612,48 @@ function ExchangeDialog({ pair, roster, actions }) {
+@@ -8612,48 +8612,47 @@ function ExchangeDialog({ pair, roster, actions }) {
    const right = roster.byId[pair[1]];
    let lastDay = 0;
    const paneLeft = usePaneLeft();
@@ -16165,7 +16168,7 @@ diff --git a/client.js b/client.js
  // src/client/voice.js
 -var import_react22 = require("react");
 +var import_react33 = require("react");
- var plainSpeech = (text) => stripHeader(text).replace(/```[\s\S]*?```/g, " ").replace(/\[([^\]]*)\]\([^)]*\)/g, "$1").replace(/[*_`#>|~-]+/g, " ").replace(/\s+/g, " ").trim();
+-var plainSpeech = (text) => stripHeader(text).replace(/```[\s\S]*?```/g, " ").replace(/\[([^\]]*)\]\([^)]*\)/g, "$1").replace(/[*_`#>|~-]+/g, " ").replace(/\s+/g, " ").trim();
  var VOICE_LABELS = {
    listening: "Listening",
 @@ -8665,14 +8665,14 @@ var VOICE_LABELS = {
@@ -16251,7 +16254,7 @@ diff --git a/client.js b/client.js
      if (!speech || !catalog || !listen) return void 0;
      const own = turn.current;
      let capture = null;
-@@ -8838,7 +8838,7 @@ function VoiceMode({ sessionId, roster, actions, useSessions, useSessionStatus }
+@@ -8838,12 +8838,12 @@ function VoiceMode({ sessionId, roster, actions, useSessions, useSessionStatus }
        setServiceLevel(0);
      };
    }, [speech, catalog, listen, cycle]);
@@ -16260,6 +16263,12 @@ diff --git a/client.js b/client.js
      if (phase !== "thinking" || running || waiting.current === null) return;
      const [count, text] = [lastReply.slice(0, lastReply.indexOf("|")), lastReply.slice(lastReply.indexOf("|") + 1)];
      if (count === waiting.current || !text.trim()) return;
+     waiting.current = null;
+-    const spoken = plainSpeech(text);
++    const spoken = speechOf(text, botNames(roster));
+     setSaid(spoken);
+     setHeard("");
+     if (!window.speechSynthesis || !spoken) {
 @@ -8858,7 +8858,7 @@ function VoiceMode({ sessionId, roster, actions, useSessions, useSessionStatus }
      window.speechSynthesis.cancel();
      window.speechSynthesis.speak(utterance);