dsh-speak 1.7.4 → 1.8.1

package/engine/speak.ps1

@@ -1,155 +1,214 @@
-# speak.ps1 — Harness-agnostic speech engine (Windows SAPI5 + NaturalVoiceSAPIAdapter)
-# ====================================================================================
-# Reads text (inline or from a UTF-8 file), cleans it for speech synthesis, and reads
-# it aloud through Windows SAPI5, preferring natural voices registered by
-# NaturalVoiceSAPIAdapter (https://github.com/gexgd0419/NaturalVoiceSAPIAdapter).
-#
-# This script knows NOTHING about any harness (DSH, Claude Code, ...). Any process
-# can call it:
-#
-#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speak.ps1 -Text "hello"
-#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speak.ps1 -File C:\tmp\msg.txt
-#
-# It is best-effort by design: it never throws, never blocks the caller for longer
-# than the utterance itself, and exits 0 even if something failed.
-#
-# Design notes (see docs/DESIGN.md for full rationale):
-#   * Markdown symbols, URLs and emoji are stripped before speaking — SAPI5 Speak()
-#     silently fails (produces no audio, no error) when it hits emoji/surrogates.
-#   * NaturalVoiceSAPIAdapter has a per-Speak character ceiling (~375-470 chars);
-#     beyond that it silently speaks nothing. Text longer than $MaxChars is replaced
-#     with $LongTextMessage instead.
-#   * Adapter-registered voices often have plain names ("Microsoft Xiaoxiao") that do
-#     not contain the word "Natural", so matching checks Name + Description.
-# ====================================================================================
-
-param(
-    [string]$Text = '',
-    [string]$File = '',
-    [int]$Volume = 50,
-    [int]$Rate = 1,
-    [int]$MaxChars = 300,
-    [string]$LongTextMessage = '本次播报内容较长，请自行阅读。',
-    [ValidateSet('message', 'heading')]
-    [string]$LongTextMode = 'message',
-    # 命令行参数一律是字符串（-File 模式不做类型转换），这里用 [string] 接收，
-    # 在脚本内部再转布尔，兼容 '1'/'0'/'true'/'True'/'yes'/'on' 等写法。
-    [string]$CleanMarkdownFormatting = 'true',
-    [string]$ReadInlineCode = 'true',
-    [ValidateSet('all', 'smart', 'replace')]
-    [string]$CodeBlocks = 'smart',
-    [int]$CodeBlockMaxChars = 300,
-    [string]$CodeBlockReplacementText = 'You can see the code in our history.',
-    # 手动重播完整朗读：跳过超长文本的 heading/message 截断，分段完整朗读
-    [string]$FullRead = '0'
-)
-
-$cleanMarkdown = $CleanMarkdownFormatting -in @('1', 'true', 'yes', 'on')
-$readInlineCode = $ReadInlineCode -in @('1', 'true', 'yes', 'on')
-# 注意：PowerShell 变量大小写不敏感，内部变量名不能与参数名仅差大小写
-# （曾用 $fullRead 导致自赋值污染参数 $FullRead，使 -not 判断失效）
-$fullReadMode = $FullRead -in @('1', 'true', 'yes', 'on')
-
-# ---------- input: pick text source ----------
-if ($File) {
-    if (-not (Test-Path $File)) { exit 0 }
-    $text = [System.IO.File]::ReadAllText($File, [System.Text.Encoding]::UTF8)
-} else {
-    $text = [string]$Text
-}
-if (-not $text -or -not $text.Trim()) { exit 0 }
-
-# ---------- length guard: adapter per-Speak ceiling ----------
-# 'message': fixed prompt. 'heading': speak the largest markdown heading instead
-# (fewest '#' wins, tie -> first; no heading -> first non-empty line; the
-# cleaned candidate is still subject to the ceiling below). FullRead 手动重播
-# 跳过该守卫（见文件底部"完整朗读"分支）。
-if (-not $fullReadMode -and $text.Length -gt $MaxChars -and $LongTextMode -eq 'heading') {
-    $candidate = ''
-    $bestLevel = 7
-    $firstNonEmpty = ''
-    # 代码块 fence 内的行跳过：其中的 "# 注释" 不是 markdown 标题，
-    # 否则长回复里的代码注释会被误当成标题只念注释
-    $inCodeBlock = $false
-    foreach ($line in ($text -split "`n")) {
-        if ($line -match '^\s*```') { $inCodeBlock = -not $inCodeBlock; continue }
-        if ($inCodeBlock) { continue }
-        if ($line -match '^\s*#{1,6}\s+') {
-            $level = ([regex]::Match($line, '^(\s*)(#+)')).Groups[2].Value.Length
-            if ($level -lt $bestLevel) {
-                $bestLevel = $level
-                $candidate = $line -replace '^\s*#+\s*', ''
-            }
-        } elseif (-not $firstNonEmpty -and $line.Trim()) {
-            $firstNonEmpty = $line
-        }
-    }
-    if (-not $candidate) { $candidate = $firstNonEmpty }
-    if ($candidate) { $text = $candidate }
-}
-
-# ---------- clean: Markdown -> natural speech text ----------
-if ($cleanMarkdown) {
-    $text = [regex]::Replace($text, '```[^\n]*\n?([\s\S]*?)```', {
-        param($match)
-        $code = $match.Groups[1].Value
-        if ($CodeBlocks -eq 'all' -or ($CodeBlocks -eq 'smart' -and $code.Length -le $CodeBlockMaxChars)) { return " $code " }
-        return " $CodeBlockReplacementText "
-    })
-    if ($readInlineCode) { $text = $text -replace '`([^`]*)`', '$1' } else { $text = $text -replace '`[^`]*`', ' ' }
-    $text = $text -replace '\[([^\]]*)\]\([^\)]*\)', '$1'
-    $text = $text -replace 'https?://\S+', ' '
-    $text = $text -replace '(?m)^\s{0,3}(?:#{1,6}\s+|[-*+]\s+|\d+[.)]\s+|>\s?)', ' '
-    $text = $text -replace '(\*\*|__|~~)(.*?)\1', '$2'
-    $text = $text -replace '[*_~]+', ''
-}
-# Keep all Unicode letters, including Portuguese accents; remove unsafe symbols.
-$text = [regex]::Replace($text, '[^\p{L}\p{N}一-龥　-〿＀-￯ - -~]', '')
-$text = $text -replace '\s+', ' '
-$text = $text.Trim()
-
-# ---------- final ceiling (also catches over-long heading candidates) ----------
-if (-not $fullReadMode -and $text.Length -gt $MaxChars) { $text = $LongTextMessage }
-
-# ---------- speak ----------
-Add-Type -AssemblyName System.Speech
-$synth = New-Object System.Speech.Synthesis.SpeechSynthesizer
-$synth.Volume = $Volume
-
-# prefer a zh natural voice (NaturalVoiceSAPIAdapter-registered), fall back to any zh
-$voices = $synth.GetInstalledVoices()
-$voice = $voices | Where-Object {
-    $_.VoiceInfo.Culture.Name -like 'zh*' -and
-    ($_.VoiceInfo.Name + ' ' + $_.VoiceInfo.Description) -match 'Natural|Online'
-} | Select-Object -First 1
-if (-not $voice) { $voice = $voices | Where-Object { $_.VoiceInfo.Culture.Name -like 'zh*' } | Select-Object -First 1 }
-if ($voice) { $synth.SelectVoice($voice.VoiceInfo.Name) }
-
-$synth.Rate = $Rate
-
-# ---------- 完整朗读（FullRead，手动重播） ----------
-# Windows SAPI 单次 Speak 有约 375-470 字上限，超长会静默失败；因此按句末
-# 标点切成不超过 450 字的段，逐段朗读（自动播报不经过这里，走上面的守卫）。
-$SPEAK_CHUNK = 400
-if ($fullReadMode -and $text.Length -gt $SPEAK_CHUNK) {
-    $parts = [regex]::Split($text, '(?<=[。！？；.!?;])')
-    $chunk = ''
-    foreach ($part in $parts) {
-        if ($part.Length -eq 0) { continue }
-        if ($chunk.Length + $part.Length -gt $SPEAK_CHUNK) {
-            if ($chunk) { $synth.Speak($chunk); $chunk = '' }
-            # 单段仍超长：硬切
-            while ($part.Length -gt $SPEAK_CHUNK) {
-                $synth.Speak($part.Substring(0, $SPEAK_CHUNK))
-                $part = $part.Substring($SPEAK_CHUNK)
-            }
-            $chunk = $part
-        } else {
-            $chunk += $part
-        }
-    }
-    if ($chunk) { $synth.Speak($chunk) }
-} else {
-    $synth.Speak($text)
-}
-exit 0
+# speak.ps1 — Harness-agnostic speech engine (Windows SAPI5 + NaturalVoiceSAPIAdapter)
+# ====================================================================================
+# Reads text (inline or from a UTF-8 file), cleans it for speech synthesis, and reads
+# it aloud through Windows SAPI5, preferring natural voices registered by
+# NaturalVoiceSAPIAdapter (https://github.com/gexgd0419/NaturalVoiceSAPIAdapter).
+#
+# This script knows NOTHING about any harness (DSH, Claude Code, ...). Any process
+# can call it:
+#
+#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speak.ps1 -Text "hello"
+#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speak.ps1 -File C:\tmp\msg.txt
+#
+# It is best-effort by design: it never throws, never blocks the caller for longer
+# than the utterance itself, and exits 0 even if something failed.
+#
+# Design notes (see docs/DESIGN.md for full rationale):
+#   * Markdown symbols, URLs and emoji are stripped before speaking — SAPI5 Speak()
+#     silently fails (produces no audio, no error) when it hits emoji/surrogates.
+#   * NaturalVoiceSAPIAdapter has a per-Speak character ceiling (~375-470 chars);
+#     beyond that it silently speaks nothing. Text longer than $MaxChars is replaced
+#     with $LongTextMessage instead.
+#   * Adapter-registered voices often have plain names ("Microsoft Xiaoxiao") that do
+#     not contain the word "Natural", so matching checks Name + Description.
+# ====================================================================================
+
+param(
+    [string]$Text = '',
+    [string]$File = '',
+    [int]$Volume = 50,
+    [int]$Rate = 1,
+    [int]$MaxChars = 300,
+    # 本脚本唯一的非 ASCII 代码字面量。丢了 UTF-8 BOM 时 PowerShell 5.1 会按 ANSI
+    # 代码页把它解码成乱码——DSH 插件总是显式传 -LongTextMessage，所以只影响手动
+    # CLI 调用；脚本逻辑（句末判定 / 字符过滤 / 分句）已全部改成纯 ASCII 源码。
+    [string]$LongTextMessage = '本次播报内容较长，请自行阅读。',
+    [ValidateSet('message', 'heading')]
+    [string]$LongTextMode = 'message',
+    # 命令行参数一律是字符串（-File 模式不做类型转换），这里用 [string] 接收，
+    # 在脚本内部再转布尔，兼容 '1'/'0'/'true'/'True'/'yes'/'on' 等写法。
+    [string]$CleanMarkdownFormatting = 'true',
+    [string]$ReadInlineCode = 'true',
+    [ValidateSet('all', 'smart', 'replace')]
+    [string]$CodeBlocks = 'smart',
+    [int]$CodeBlockMaxChars = 300,
+    [string]$CodeBlockReplacementText = 'You can see the code in our history.',
+    # 手动重播完整朗读：跳过超长文本的 heading/message 截断，分段完整朗读
+    [string]$FullRead = '0',
+    # 只把「将要朗读的文本」按 UTF-8 写到 stdout、完全不出声（调试清洗与长文
+    # 守卫用；正常调用无需传，插件不会传）
+    [string]$DryRun = '0'
+)
+
+$cleanMarkdown = $CleanMarkdownFormatting -in @('1', 'true', 'yes', 'on')
+$readInlineCode = $ReadInlineCode -in @('1', 'true', 'yes', 'on')
+$dryRunMode = $DryRun -in @('1', 'true', 'yes', 'on')
+# 注意：PowerShell 变量大小写不敏感，内部变量名不能与参数名仅差大小写
+# （曾用 $fullRead 导致自赋值污染参数 $FullRead，使 -not 判断失效）
+$fullReadMode = $FullRead -in @('1', 'true', 'yes', 'on')
+
+# ---------- input: pick text source ----------
+if ($File) {
+    if (-not (Test-Path $File)) { exit 0 }
+    $text = [System.IO.File]::ReadAllText($File, [System.Text.Encoding]::UTF8)
+} else {
+    $text = [string]$Text
+}
+if (-not $text -or -not $text.Trim()) { exit 0 }
+
+# ---------- length guard: adapter per-Speak ceiling ----------
+# 'message': fixed prompt. 'heading': speak the largest markdown heading instead
+# (fewest '#' wins, tie -> first; with NO heading, speak a coherent opening of
+# the text — see below; the cleaned candidate is still subject to the ceiling
+# below). FullRead 手动重播跳过该守卫（见文件底部"完整朗读"分支）。
+if (-not $fullReadMode -and $MaxChars -gt 0 -and $text.Length -gt $MaxChars -and $LongTextMode -eq 'heading') {
+    $candidate = ''
+    $bestLevel = 7
+    # 代码块 fence 内的行跳过：其中的 "# 注释" 不是 markdown 标题，
+    # 否则长回复里的代码注释会被误当成标题只念注释
+    $inCodeBlock = $false
+    foreach ($line in ($text -split "`n")) {
+        if ($line -match '^\s*```') { $inCodeBlock = -not $inCodeBlock; continue }
+        if ($inCodeBlock) { continue }
+        if ($line -match '^\s*#{1,6}\s+') {
+            $level = ([regex]::Match($line, '^(\s*)(#+)')).Groups[2].Value.Length
+            if ($level -lt $bestLevel) {
+                $bestLevel = $level
+                $candidate = $line -replace '^\s*#+\s*', ''
+            }
+        }
+    }
+    if ($candidate) {
+        # 有标题：只念最大的标题（"长回复只报标题"）
+        $text = $candidate
+    } else {
+        # 没有标题：**不能只念第一个非空行** —— 那会念出"……官方文档写明："这类
+        # 断头句然后静默停住，听感上就是"从第二行开始不念了"（1.8.0 修复）。
+        # 改为取开头 MaxChars 长度的窗口，并在窗口内最后一个句末标点处收尾。
+        # 中英双语判据：
+        #   * 全角 。！？； 与省略号 … 无条件算句末（中文标点不含歧义）；
+        #   * 半角 .!?; 只在后面跟空白、右引号/右括号时才算，否则 "0.1.2"、
+        #     "file.txt"、"e.g." 里的小数点/扩展名会被当成句子结尾；
+        #   * 半角标点不认"到窗口结尾"本身就结束：窗口末尾若是小数点，认了等于没
+        #     修剪。但会多读一位来判断（见下），所以"句号+空格"在窗口边缘照样成立。
+        # 收尾后若不足半个窗口，就保留整个窗口：一整句超长文本不该被砍成一个词。
+        #
+        # 标点类一律用 [char] 码位拼出来，让源码里**不出现非 ASCII 代码字面量**：
+        #   1) PowerShell 把 ’ ” ‘ “ 也当字符串引号，直接写进单引号串会提前截断
+        #      （报 Missing ')' in method call）；
+        #   2) 更要紧的是——脚本一旦丢掉 UTF-8 BOM，Windows PowerShell 5.1 会按
+        #      ANSI 代码页解码，代码里的中文标点会变乱码，句末判定**静默失效**。
+        #      代码保持纯 ASCII 后，丢 BOM 只会让中文注释变乱码，不影响行为。
+        #   。 ！ ？ ； … = 0x3002 0xFF01 0xFF1F 0xFF1B 0x2026
+        #   ） 】 」 』 = 0xFF09 0x3011 0x300D 0x300F
+        $fullWidthEnders = [string]([char]0x3002) + [char]0xFF01 + [char]0xFF1F + [char]0xFF1B + [char]0x2026
+        $closers = [string]([char]0x22) + [char]0x201D + [char]0x2019 + [char]0xFF09 + [char]0x3011 + [char]0x300D + [char]0x300F + [char]0x29 + '\' + [char]0x5D + [char]0x7D
+        $sentenceEndPattern = '(?:[' + $fullWidthEnders + ']|[.!?;](?=[\s' + $closers + ']))'
+        $window = $text.Substring(0, [Math]::Min($text.Length, $MaxChars))
+        # 多看一个字符再判定，但只在窗口内收尾：句号落在窗口最后一位时，它后面那个
+        # 空格在窗口之外，多看一位才能认出"句号+空格"确实是句末；而"句号+数字"
+        # （`Version 0.1.` 的第 300 位）依旧被拒。切点永远不超过窗口长度。
+        $scan = $text.Substring(0, [Math]::Min($text.Length, $MaxChars + 1))
+        # 取**落在窗口内**的最后一个句末标点：scan 多读的那一位会让最后一个匹配可能
+        # 落在窗口之外（窗口外正好是个全角句号时），那种匹配必须忽略——但也不能因此
+        # 丢掉窗口内更早的合法边界，所以逐个过滤而不是只看最后一个。
+        $cut = 0
+        foreach ($match in [regex]::Matches($scan, $sentenceEndPattern)) {
+            $end = $match.Index + $match.Length
+            if ($end -le $window.Length) { $cut = $end }
+        }
+        if ($cut -ge [Math]::Floor($window.Length / 2)) { $window = $window.Substring(0, $cut) }
+        $text = $window
+    }
+}
+
+# ---------- clean: Markdown -> natural speech text ----------
+if ($cleanMarkdown) {
+    $text = [regex]::Replace($text, '```[^\n]*\n?([\s\S]*?)```', {
+        param($match)
+        $code = $match.Groups[1].Value
+        if ($CodeBlocks -eq 'all' -or ($CodeBlocks -eq 'smart' -and $code.Length -le $CodeBlockMaxChars)) { return " $code " }
+        return " $CodeBlockReplacementText "
+    })
+    if ($readInlineCode) { $text = $text -replace '`([^`]*)`', '$1' } else { $text = $text -replace '`[^`]*`', ' ' }
+    $text = $text -replace '\[([^\]]*)\]\([^\)]*\)', '$1'
+    $text = $text -replace 'https?://\S+', ' '
+    $text = $text -replace '(?m)^\s{0,3}(?:#{1,6}\s+|[-*+]\s+|\d+[.)]\s+|>\s?)', ' '
+    $text = $text -replace '(\*\*|__|~~)(.*?)\1', '$2'
+    $text = $text -replace '[*_~]+', ''
+}
+# Keep all Unicode letters, including Portuguese accents; remove unsafe symbols.
+# 范围写成 \u 转义（纯 ASCII 源码，丢 BOM 也不会乱码）：
+#   \u4e00-\u9fa5 汉字、\u3000-\u303f 中文标点、\uff00-\uffef 全角、
+#   \u2000-\u206f 通用标点（含 … 和 —）、\u0020-\u007e ASCII 可打印。
+$text = [regex]::Replace($text, '[^\p{L}\p{N}\u4e00-\u9fa5\u3000-\u303f\uff00-\uffef\u2000-\u206f\u0020-\u007e]', '')
+$text = $text -replace '\s+', ' '
+$text = $text.Trim()
+
+# ---------- final ceiling (also catches over-long heading candidates) ----------
+if (-not $fullReadMode -and $text.Length -gt $MaxChars) { $text = $LongTextMessage }
+
+# ---------- dry run: expose the text that would be spoken, silently ----------
+# Maintainer aid: makes the cleaning pipeline and the long-text guard observable
+# without a speaker, so a truncated candidate can be diffed against its source.
+# Written as raw UTF-8 bytes so a redirected capture never depends on the
+# console code page.
+if ($dryRunMode) {
+    $bytes = [System.Text.Encoding]::UTF8.GetBytes($text)
+    [Console]::OpenStandardOutput().Write($bytes, 0, $bytes.Length)
+    exit 0
+}
+
+# ---------- speak ----------
+Add-Type -AssemblyName System.Speech
+$synth = New-Object System.Speech.Synthesis.SpeechSynthesizer
+$synth.Volume = $Volume
+
+# prefer a zh natural voice (NaturalVoiceSAPIAdapter-registered), fall back to any zh
+$voices = $synth.GetInstalledVoices()
+$voice = $voices | Where-Object {
+    $_.VoiceInfo.Culture.Name -like 'zh*' -and
+    ($_.VoiceInfo.Name + ' ' + $_.VoiceInfo.Description) -match 'Natural|Online'
+} | Select-Object -First 1
+if (-not $voice) { $voice = $voices | Where-Object { $_.VoiceInfo.Culture.Name -like 'zh*' } | Select-Object -First 1 }
+if ($voice) { $synth.SelectVoice($voice.VoiceInfo.Name) }
+
+$synth.Rate = $Rate
+
+# ---------- 完整朗读（FullRead，手动重播） ----------
+# Windows SAPI 单次 Speak 有约 375-470 字上限，超长会静默失败；因此按句末
+# 标点切成不超过 450 字的段，逐段朗读（自动播报不经过这里，走上面的守卫）。
+$SPEAK_CHUNK = 400
+if ($fullReadMode -and $text.Length -gt $SPEAK_CHUNK) {
+    # 分句标点同样用 \u 转义（纯 ASCII 源码）
+    $parts = [regex]::Split($text, '(?<=[\u3002\uff01\uff1f\uff1b.!?;])')
+    $chunk = ''
+    foreach ($part in $parts) {
+        if ($part.Length -eq 0) { continue }
+        if ($chunk.Length + $part.Length -gt $SPEAK_CHUNK) {
+            if ($chunk) { $synth.Speak($chunk); $chunk = '' }
+            # 单段仍超长：硬切
+            while ($part.Length -gt $SPEAK_CHUNK) {
+                $synth.Speak($part.Substring(0, $SPEAK_CHUNK))
+                $part = $part.Substring($SPEAK_CHUNK)
+            }
+            $chunk = $part
+        } else {
+            $chunk += $part
+        }
+    }
+    if ($chunk) { $synth.Speak($chunk) }
+} else {
+    $synth.Speak($text)
+}
+exit 0

package/engine/speak.sh

@@ -55,15 +55,45 @@ fi
 # `heading` is meaningful only when a positive long-text ceiling is configured.
 # -F (full read, manual replay) skips this guard entirely.
 if [ "$FULL_READ" != "1" ] && [ "$MAX_CHARS" -gt 0 ] && [ "${#TEXT}" -gt "$MAX_CHARS" ] && [ "$LONG_MODE" = "heading" ]; then
+  export DSH_SPEAK_MAX_CHARS="$MAX_CHARS"
   TEXT=$(printf '%s' "$TEXT" | /usr/bin/perl -CSD -e '
-    my $best = 7; my $cand = ""; my $first = ""; my $inCode = 0;
-    while (<STDIN>) {
-      if (/^\s*```/) { $inCode = !$inCode; next; }   # 跳过代码块内的 "# 注释"
+    my $text = do { local $/; <STDIN> };
+    my $best = 7; my $cand = ""; my $inCode = 0;
+    for my $line (split /\n/, $text, -1) {
+      $line =~ s/\r$//;
+      if ($line =~ /^\s*```/) { $inCode = !$inCode; next; }   # 跳过代码块内的 "# 注释"
       next if $inCode;
-      if (/^[ \t]*(\#{1,6})[ \t]+(.*)$/) { my $n = length($1); if ($n < $best) { $best = $n; $cand = $2; } }
-      elsif ($first eq "" && /\S/) { $first = $_; }
+      if ($line =~ /^[ \t]*(\#{1,6})[ \t]+(.*)$/) { my $n = length($1); if ($n < $best) { $best = $n; $cand = $2; } }
     }
-    print($cand eq "" ? $first : $cand);
+    if ($cand ne "") { print $cand; exit }
+    # 没有标题：**不能只念第一个非空行**（会念出"……写明："这种断头句然后静默
+    # 停住，听感上就是"从第二行开始不念了"）。改为取开头 MaxChars 窗口，并在窗口
+    # 内最后一个句末标点处收尾。中英双语判据与 speak.ps1 保持一致：
+    #   * 全角 。！？；… 无条件算句末；
+    #   * 半角 .!?; 只在后面跟空白、右引号/右括号时才算，否则
+    #     "0.1.2"、"file.txt"、"e.g." 里的小数点/扩展名会被当成句末；
+    #   * 半角标点不认"到窗口结尾"本身就结束：窗口末尾若是小数点，认了等于没修剪。
+    #     但会多读一位来判断，所以"句号+空格"在窗口边缘照样成立。
+    # 收尾后若不足半个窗口，就保留整个窗口，避免一整句超长文本被砍成一个词。
+    #
+    # 标点一律写成 \x{...} 码位转义，让这段 Perl 源码保持**纯 ASCII**：Perl 源码默认
+    # 按字节处理（这里没有 use utf8），直接写中文标点会被当成 Latin-1 单字节字符，
+    # 模式就再也匹配不到解码后的正文，表现为**完全不修剪**（1.8.0 实测踩到）。
+    # 码位与 speak.ps1 里 [char] 构造的集合一一对应：
+    #   。 ！ ？ ； …   = \x{3002} \x{FF01} \x{FF1F} \x{FF1B} \x{2026}
+    #   ） 】 」 』 = \x{FF09} \x{3011} \x{300D} \x{300F}
+    #   " ” ’ ) ] }    = " \x{201D} \x{2019} ) ] }
+    my $max = $ENV{DSH_SPEAK_MAX_CHARS} || 300;
+    my $window = length($text) > $max ? substr($text, 0, $max) : $text;
+    my $scan = length($text) > $max + 1 ? substr($text, 0, $max + 1) : $text;
+    my @ends;
+    while ($scan =~ /(?:[\x{3002}\x{FF01}\x{FF1F}\x{FF1B}\x{2026}]|[.!?;](?=[\s"\x{201D}\x{2019}\)\]\}\x{FF09}\x{3011}\x{300D}\x{300F}]))/g) { push @ends, pos($scan); }
+    # 只认落在窗口内的最后一个句末标点：多读的那一位可能让最后一个匹配落在窗口
+    # 之外，那种匹配要忽略，但不能因此丢掉窗口内更早的合法边界。
+    my $cut = 0;
+    foreach my $end (@ends) { $cut = $end if $end <= length($window); }
+    $window = substr($window, 0, $cut) if $cut >= int(length($window) / 2);
+    print $window;
   ')
 fi
 

package/engine/speech-prompt.ps1

@@ -1,22 +1,24 @@
-# speech-prompt.ps1 — Short prompt announcement (synchronous, blocking)
-# Use when a harness/agent needs the user's attention (a question, an approval
-# request). Reads the text through engine/speak.ps1 and waits until it finishes,
-# so the caller knows the announcement was actually spoken.
-#
-#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speech-prompt.ps1 -Text "请做出选择"
-
-param([string]$Text = '请做出选择')
-
-if (-not $Text) { exit 0 }
-
-$speak = Join-Path $PSScriptRoot 'speak.ps1'
-if (-not (Test-Path $speak)) { exit 0 }
-
-$tmp = Join-Path $env:TEMP ('speech-prompt-' + [guid]::NewGuid().ToString('N') + '.txt')
-try {
-    [System.IO.File]::WriteAllText($tmp, $Text, [System.Text.UTF8Encoding]::new($false))
-    & powershell.exe -NoProfile -ExecutionPolicy Bypass -File $speak -File $tmp
-    exit $LASTEXITCODE
-} finally {
-    if (Test-Path $tmp) { Remove-Item $tmp -Force -ErrorAction SilentlyContinue }
-}
+# speech-prompt.ps1 — Short prompt announcement (synchronous, blocking)
+# Use when a harness/agent needs the user's attention (a question, an approval
+# request). Reads the text through engine/speak.ps1 and waits until it finishes,
+# so the caller knows the announcement was actually spoken.
+#
+#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speech-prompt.ps1 -Text "请做出选择"
+
+# 唯一的非 ASCII 代码字面量：丢 UTF-8 BOM 时 PowerShell 5.1 会按 ANSI 代码页把它
+# 解码成乱码——调用方通常显式传 -Text，所以只影响不带参数的调用。
+param([string]$Text = '请做出选择')
+
+if (-not $Text) { exit 0 }
+
+$speak = Join-Path $PSScriptRoot 'speak.ps1'
+if (-not (Test-Path $speak)) { exit 0 }
+
+$tmp = Join-Path $env:TEMP ('speech-prompt-' + [guid]::NewGuid().ToString('N') + '.txt')
+try {
+    [System.IO.File]::WriteAllText($tmp, $Text, [System.Text.UTF8Encoding]::new($false))
+    & powershell.exe -NoProfile -ExecutionPolicy Bypass -File $speak -File $tmp
+    exit $LASTEXITCODE
+} finally {
+    if (Test-Path $tmp) { Remove-Item $tmp -Force -ErrorAction SilentlyContinue }
+}

package/engine/speech-summary.ps1

@@ -1,27 +1,27 @@
-# speech-summary.ps1 — Reply summary announcement (synchronous, blocking)
-# For harnesses with no "reply finished" event (e.g. DSH has no Stop hook): the
-# agent calls this at the end of its final reply.
-#
-#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speech-summary.ps1 -Text "总结文本"
-#
-# NOTE: keep this SYNCHRONOUS. An earlier version spawned the inner powershell
-# asynchronously with Start-Process; DSH's sandbox blocks nested sub-process
-# spawning, so it silently produced no audio. The synchronous call chain
-# (summary -> speak.ps1) is the reliable path (cost: caller waits for the
-# utterance to finish).
-
-param([string]$Text = '')
-
-if (-not $Text) { exit 0 }
-
-$speak = Join-Path $PSScriptRoot 'speak.ps1'
-if (-not (Test-Path $speak)) { exit 0 }
-
-$tmp = Join-Path $env:TEMP ('speech-summary-' + [guid]::NewGuid().ToString('N') + '.txt')
-try {
-    [System.IO.File]::WriteAllText($tmp, $Text, [System.Text.UTF8Encoding]::new($false))
-    & powershell.exe -NoProfile -ExecutionPolicy Bypass -File $speak -File $tmp
-    exit $LASTEXITCODE
-} finally {
-    if (Test-Path $tmp) { Remove-Item $tmp -Force -ErrorAction SilentlyContinue }
-}
+# speech-summary.ps1 — Reply summary announcement (synchronous, blocking)
+# For harnesses with no "reply finished" event (e.g. DSH has no Stop hook): the
+# agent calls this at the end of its final reply.
+#
+#   powershell.exe -NoProfile -ExecutionPolicy Bypass -File speech-summary.ps1 -Text "总结文本"
+#
+# NOTE: keep this SYNCHRONOUS. An earlier version spawned the inner powershell
+# asynchronously with Start-Process; DSH's sandbox blocks nested sub-process
+# spawning, so it silently produced no audio. The synchronous call chain
+# (summary -> speak.ps1) is the reliable path (cost: caller waits for the
+# utterance to finish).
+
+param([string]$Text = '')
+
+if (-not $Text) { exit 0 }
+
+$speak = Join-Path $PSScriptRoot 'speak.ps1'
+if (-not (Test-Path $speak)) { exit 0 }
+
+$tmp = Join-Path $env:TEMP ('speech-summary-' + [guid]::NewGuid().ToString('N') + '.txt')
+try {
+    [System.IO.File]::WriteAllText($tmp, $Text, [System.Text.UTF8Encoding]::new($false))
+    & powershell.exe -NoProfile -ExecutionPolicy Bypass -File $speak -File $tmp
+    exit $LASTEXITCODE
+} finally {
+    if (Test-Path $tmp) { Remove-Item $tmp -Force -ErrorAction SilentlyContinue }
+}