@xdfnet/ispeak 1.6.4 → 1.6.6

package/Docs/ARCHITECTURE.md

@@ -177,7 +177,7 @@ pending -> running -> delete
 ├── config.json      # API Key、音色配置
 ├── ispeak.sock      # Unix Socket
 ├── ispeak.log       # 日志（lumberjack 轮转）
-└── hook-speak.sh    # Claude/Codex Stop Hook
+└── hook-speak.sh    # Claude/Codex Hook
 
 ~/Library/LaunchAgents/
 └── com.iSpeak.plist # launchd 服务配置

package/Docs/HOOK_TEXT_EXTRACTION.md

@@ -0,0 +1,240 @@
+# Hook 文本提取链路
+
+本文记录 Claude Code / Codex CLI 在 Hook 中拿到“最后一条 assistant 回复”的实际方式。`hook-speak.sh` 的目标只做两件事：取最后一条 assistant 回复，发给 iSpeak socket。
+
+## 结论
+
+推荐优先级：
+
+1. **Codex `notify`**：从脚本第二个参数 `$2` 读取 JSON，取 `last-assistant-message`。
+2. **Claude / Codex Stop Hook**：从 stdin 读取 JSON，优先取 `last_assistant_message`。
+3. **明确 transcript**：如果没有直接字段，只读取 payload 里明确传入的 `transcript_path`。
+
+不扫描 `~/.codex/sessions`。没有 direct 字段也没有 `transcript_path` 时，本次不播报。
+
+## Codex CLI：notify
+
+当前本机版本：
+
+```text
+codex-cli 0.130.0
+```
+
+Codex CLI 的 `notify = [...]` 是 legacy notify 机制。官方源码里会把通知 JSON 追加成命令的最后一个 argv 参数，不写 stdin。
+
+配置示例：
+
+```toml
+notify = ["/Users/你的用户名/.config/iSpeak/hook-speak.sh", "codex"]
+```
+
+脚本实际收到：
+
+```bash
+$1 = "codex"
+$2 = '{"type":"agent-turn-complete",...,"last-assistant-message":"..."}'
+stdin = empty
+```
+
+核心字段：
+
+```json
+{
+  "type": "agent-turn-complete",
+  "thread-id": "...",
+  "turn-id": "...",
+  "cwd": "...",
+  "input-messages": ["..."],
+  "last-assistant-message": "最后一条 assistant 回复"
+}
+```
+
+所以 Codex `notify` 的正确读取方式是：
+
+```bash
+input="${2:-}"
+```
+
+然后解析：
+
+```js
+payload["last-assistant-message"]
+```
+
+源码依据：`codex-rs/hooks/src/legacy_notify.rs`。该文件把 `last_assistant_message` 序列化为 kebab-case 的 `last-assistant-message`，并在执行命令前 `command.arg(notify_payload)`。
+
+## Codex CLI：Stop Hook
+
+Codex 也支持 Claude 风格 Hook。Stop Hook 的输入 JSON 写入 stdin。
+
+配置示例：
+
+```json
+{
+  "hooks": {
+    "Stop": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "bash $HOME/.config/iSpeak/hook-speak.sh codex",
+            "timeout": 30
+          }
+        ]
+      }
+    ]
+  }
+}
+```
+
+脚本实际收到：
+
+```bash
+$1 = "codex"
+$2 = empty
+stdin = '{"hook_event_name":"Stop",...,"last_assistant_message":"..."}'
+```
+
+核心字段：
+
+```json
+{
+  "session_id": "...",
+  "turn_id": "...",
+  "transcript_path": "...",
+  "cwd": "...",
+  "hook_event_name": "Stop",
+  "model": "...",
+  "permission_mode": "bypassPermissions",
+  "stop_hook_active": false,
+  "last_assistant_message": "最后一条 assistant 回复"
+}
+```
+
+源码依据：
+
+- `codex-rs/hooks/src/events/stop.rs`：构造 `StopCommandInput`，包含 `last_assistant_message` 和 `transcript_path`。
+- `codex-rs/hooks/schema/generated/stop.command.input.schema.json`：Stop stdin schema。
+- `codex-rs/hooks/src/engine/command_runner.rs`：Hook 命令通过 stdin 接收 `input_json`。
+
+## Codex Transcript
+
+Codex 的 transcript/session 文件是 JSONL。实际 assistant 回复形态：
+
+```json
+{
+  "type": "response_item",
+  "payload": {
+    "type": "message",
+    "role": "assistant",
+    "content": [
+      {
+        "type": "output_text",
+        "text": "最后一条 assistant 回复"
+      }
+    ]
+  }
+}
+```
+
+提取规则：
+
+```js
+event.type === "response_item" &&
+event.payload?.type === "message" &&
+event.payload?.role === "assistant"
+```
+
+然后拼接：
+
+```js
+event.payload.content[].text
+```
+
+## Claude Code：Stop Hook
+
+Claude Code 官方 Stop Hook 通过 stdin 传 JSON，核心字段是：
+
+```json
+{
+  "session_id": "...",
+  "transcript_path": "...",
+  "hook_event_name": "Stop",
+  "stop_hook_active": false
+}
+```
+
+有些版本或场景可能直接提供：
+
+```json
+{
+  "last_assistant_message": "最后一条 assistant 回复"
+}
+```
+
+所以 Claude 的读取顺序是：
+
+1. `last_assistant_message`
+2. `message`
+3. `transcript_path`
+
+Claude transcript 常见 assistant 形态：
+
+```json
+{"role":"assistant","content":[{"type":"text","text":"..."}]}
+```
+
+或：
+
+```json
+{"message":{"role":"assistant","content":[{"type":"text","text":"..."}]}}
+```
+
+## 当前脚本策略
+
+`configs/hook-speak.sh` 当前入口：
+
+```bash
+input="${2:-}"
+if [[ -z "$input" ]]; then
+  input=$(cat)
+fi
+```
+
+含义：
+
+- Codex `notify`：读 `$2`
+- Claude / Codex Stop Hook：读 stdin
+
+Codex 文本字段优先级：
+
+```js
+payload["last-assistant-message"]
+payload.last_assistant_message
+payload.lastAssistantMessage
+payload.message
+payload.lastMessage
+payload.transcript_path
+payload.transcriptPath
+payload["transcript-path"]
+```
+
+Claude 文本字段优先级：
+
+```js
+payload.last_assistant_message
+payload.message
+payload.transcript_path
+payload.transcriptPath
+```
+
+## 为什么不能只读 stdin
+
+因为 Codex `notify` 不走 stdin。只读 stdin 会导致：
+
+```text
+TEXT_LEN: 0
+SPOKE: SKIP
+```
+
+正确做法是先读 `$2`，再读 stdin；不扫历史 session。

package/README.md

@@ -1,13 +1,13 @@
 # iSpeak
 
-![Version](https://img.shields.io/badge/version-1.6.3-blue)
+![Version](https://img.shields.io/badge/version-1.6.6-blue)
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![Go Version](https://img.shields.io/badge/Go-1.26-blue)](https://golang.org/dl/)
 ![Platform](https://img.shields.io/badge/platform-macOS-green)
 
 iSpeak 让 AI 编程助手开口说话。你写代码，它播结果——眼睛休息，耳朵来听。
 
-适合 Claude Code 或 Codex 常驻后台的开发者。AI 完成任务后自动播报；你发新消息时，旧播报立即中断，不花冤枉钱。
+适合 Claude Code 或 Codex 常驻后台的开发者。AI 完成任务后自动播报；你发新消息时，未开始的旧播报会被丢弃，不花冤枉钱。
 
 ## 效果示例
 
@@ -104,8 +104,7 @@ pending → running → delete
 - Markdown 链接：只保留链接标题，不播 URL
 - 绝对路径：简化为“路径”
 - 长 commit hash、UUID、长 ID：不播
-- 明显文件列表：如模型分片、代码文件列表、下载文件清单
-- 下载进度和终端噪声：百分比、速度、进度条、ANSI 控制符
+- 下载进度噪声：速度、ETA、预计剩余时间、ANSI 控制符
 
 保留优先级：结论、成功/失败状态、需要用户操作的下一步、关键错误原因。
 
@@ -131,7 +130,7 @@ ispeak-codex "消息"    # Codex 专属音色
 ```json
 {
   "apiKey": "你的火山引擎 API Key",
-  "endpoint": "https://openspeech.bytedance.com/api/v3/tts/unidirectional",
+  "endpoint": "https://openspeech.bytedance.com/api/v3/tts/unidirectional/sse",
   "defaultVoice": {
     "voice_type": "zh_female_mizai_uranus_bigtts",
     "resourceId": "seed-tts-2.0"
@@ -173,7 +172,13 @@ ispeak-codex "消息"    # Codex 专属音色
 
 ### Codex
 
-在 `~/.codex/hooks.json` 中添加 Stop Hook：
+推荐在 `~/.codex/config.toml` 中添加回合结束通知：
+
+```toml
+notify = ["bash", "/Users/你的用户名/.config/iSpeak/hook-speak.sh", "codex"]
+```
+
+如果你启用了 Codex hooks，也可以在 `~/.codex/hooks.json` 中添加 Stop Hook：
 
 ```json
 {

package/clean_text.go

@@ -0,0 +1,153 @@
+package main
+
+import (
+	"regexp"
+	"strings"
+)
+
+var (
+	markdownLinkRe     = regexp.MustCompile(`\[[^\]]+\]\(([^)]*)\)`)
+	absolutePathRe     = regexp.MustCompile(`/(?:Users|private|tmp|var|opt|usr|bin|sbin|etc|Library|Applications)/\S+`)
+	commitHashRe       = regexp.MustCompile(`\b[0-9a-f]{7,40}\b`)
+	uuidRe             = regexp.MustCompile(`\b[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}\b`)
+	urlRe              = regexp.MustCompile(`https?://\S+`)
+	ansiEscapeRe       = regexp.MustCompile(`\x1b\[[0-9;]*[A-Za-z]`)
+	multiSpaceRe       = regexp.MustCompile(`\s+`)
+	markdownListRe     = regexp.MustCompile(`^\s*(?:[-*+]\s+|\d+[.)]\s+)`)
+	htmlTagRe          = regexp.MustCompile(`<[^>]+>`)
+	codeFenceStartRe   = regexp.MustCompile("^```")
+	artifactStartRe    = regexp.MustCompile(`(?i)^<artifact\b`)
+	htmlDocumentLineRe = regexp.MustCompile(`(?i)^<!doctype html|^<html\b|^<head\b|^<body\b|^<style\b|^</`)
+	speedNoiseRe       = regexp.MustCompile(`(?i)\d+(?:\.\d+)?\s*(?:kb|mb|gb)/s`)
+	etaNoiseRe         = regexp.MustCompile(`(?i)\bETA\b|预计剩余|剩余时间`)
+)
+
+// 过滤格式符号，保留自然朗读文本。
+// 顺序很重要：先跳过跨行块结构，再跳过整行噪声，最后清理行内符号。
+func cleanText(text string) string {
+	var lines []string
+	rawLines := strings.Split(text, "\n")
+	inCodeBlock := false
+	inArtifact := false
+	inMarkdownTable := false
+	for i := 0; i < len(rawLines); i++ {
+		line := rawLines[i]
+		line = strings.TrimSpace(line)
+		if line == "" {
+			inMarkdownTable = false
+			continue
+		}
+		if codeFenceStartRe.MatchString(line) {
+			inCodeBlock = !inCodeBlock
+			continue
+		}
+		if inCodeBlock {
+			continue
+		}
+		if artifactStartRe.MatchString(line) {
+			inArtifact = !strings.Contains(strings.ToLower(line), "</artifact>")
+			continue
+		}
+		if inArtifact {
+			if strings.Contains(strings.ToLower(line), "</artifact>") {
+				inArtifact = false
+			}
+			continue
+		}
+		if isMarkdownTableSeparator(line) {
+			if len(lines) > 0 && isMarkdownTableRow(strings.TrimSpace(rawLines[i-1])) {
+				lines = lines[:len(lines)-1]
+			}
+			inMarkdownTable = true
+			continue
+		}
+		if inMarkdownTable {
+			if isMarkdownTableRow(line) {
+				continue
+			}
+			inMarkdownTable = false
+		}
+		if shouldSkipSpeechLine(line) {
+			continue
+		}
+
+		cleaned := cleanSpeechLine(line)
+		if cleaned != "" {
+			lines = append(lines, cleaned)
+		}
+	}
+	return strings.Join(lines, "，")
+}
+
+func shouldSkipSpeechLine(line string) bool {
+	if isMarkdownTableSeparator(line) {
+		return true
+	}
+	if strings.HasPrefix(line, "---") && strings.Count(line, "-") > 3 {
+		return true
+	}
+	if htmlDocumentLineRe.MatchString(line) {
+		return true
+	}
+	if isProgressNoiseLine(line) {
+		return true
+	}
+	if isMostlyTableRow(line) {
+		return true
+	}
+	return false
+}
+
+func isMarkdownTableSeparator(line string) bool {
+	line = strings.TrimSpace(line)
+	return strings.Contains(line, "|") && strings.Trim(line, "|-: ") == ""
+}
+
+func isMarkdownTableRow(line string) bool {
+	line = strings.TrimSpace(line)
+	return strings.Count(line, "|") >= 2
+}
+
+func cleanSpeechLine(line string) string {
+	// Markdown 链接必须在 URL 删除前处理，否则会丢掉链接标题。
+	line = ansiEscapeRe.ReplaceAllString(line, "")
+	line = markdownListRe.ReplaceAllString(line, "")
+	line = markdownLinkRe.ReplaceAllStringFunc(line, func(match string) string {
+		if end := strings.Index(match, "]"); end > 1 {
+			return match[1:end]
+		}
+		return ""
+	})
+	line = urlRe.ReplaceAllString(line, "")
+	line = absolutePathRe.ReplaceAllString(line, " 路径 ")
+	// UUID 必须在短 hash 前处理，避免先删短片段后破坏 UUID 识别。
+	line = uuidRe.ReplaceAllString(line, "")
+	line = commitHashRe.ReplaceAllString(line, "")
+	line = htmlTagRe.ReplaceAllString(line, "")
+	line = strings.NewReplacer(
+		"**", "",
+		"*", "",
+		"`", "",
+		"#", "",
+		">", "",
+		"✅", "",
+		"❌", "",
+		"✓", "",
+		"✗", "",
+		"→", "到",
+	).Replace(line)
+	line = strings.Trim(line, " \t-:|")
+	line = multiSpaceRe.ReplaceAllString(line, " ")
+	return strings.TrimSpace(line)
+}
+
+func isMostlyTableRow(line string) bool {
+	if !strings.Contains(line, "|") {
+		return false
+	}
+	return strings.Count(line, "|") >= 2 && len([]rune(line)) > 40
+}
+
+func isProgressNoiseLine(line string) bool {
+	return speedNoiseRe.MatchString(line) || etaNoiseRe.MatchString(line)
+}

package/configs/config.example.json

@@ -1,6 +1,6 @@
 {
   "apiKey": "your-api-key",
-  "endpoint": "https://openspeech.bytedance.com/api/v3/tts/unidirectional",
+  "endpoint": "https://openspeech.bytedance.com/api/v3/tts/unidirectional/sse",
   "defaultVoice": {
     "voice_type": "zh_female_mizai_uranus_bigtts",
     "resourceId": "seed-tts-2.0"

package/configs/hook-speak.sh

@@ -1,130 +1,144 @@
 #!/bin/bash
-# Stop Hook: 从 transcript 文件中提取本次会话所有 Claude 回复文本
-# iAgent 调用 Claude 时设 ISPEAK_SKIP=1，此时跳过（iAgent 自己播）
+# Claude Code / Codex 共用播报 Hook：
+# 只取最后一条 assistant 回复，加 `{source:<name>}` 前缀后发给 ispeakd。
 [[ "$ISPEAK_SKIP" == "1" ]] && exit 0
 
-# 来源参数: claude 或 codex
 SOURCE="${1:-claude}"
-
 SOCK="$HOME/.config/iSpeak/ispeak.sock"
 LOG="$HOME/.config/iSpeak/hook.log"
 
-input=$(cat)
-
-json_value() {
-  local key="$1"
-  if command -v node >/dev/null 2>&1; then
-    printf "%s" "$input" | node -e '
-      const key = process.argv[1];
-      let input = "";
-      process.stdin.setEncoding("utf8");
-      process.stdin.on("data", chunk => input += chunk);
-      process.stdin.on("end", () => {
-        try {
-          const value = JSON.parse(input)[key];
-          if (typeof value === "string") process.stdout.write(value);
-        } catch (_) {}
-      });
-    ' "$key"
-    return
-  fi
-
-  printf "%s" "$input" | sed -n "s/.*\"$key\"[[:space:]]*:[[:space:]]*\"\([^\"]*\)\".*/\1/p"
+# Codex `notify` 会把 JSON 作为最后一个参数传入；
+# Claude/Claude 风格 Stop Hook 会把 JSON 写到 stdin。
+input="${2:-}"
+if [[ -z "$input" ]]; then
+  input=$(cat)
+fi
+input_file=$(mktemp)
+trap 'rm -f "$input_file"' EXIT
+printf "%s" "$input" > "$input_file"
+
+text=$(SOURCE="$SOURCE" HOOK_INPUT_FILE="$input_file" node <<'NODE' 2>/dev/null
+const fs = require("fs");
+
+{
+  const input = readFile(process.env.HOOK_INPUT_FILE || "");
+  const payload = parseJSON(input) || {};
+  const source = process.env.SOURCE || "";
+  const text = source.startsWith("codex")
+    ? lastCodexAssistant(payload)
+    : lastClaudeAssistant(payload);
+
+  if (text) process.stdout.write(text);
 }
 
-extract_recent_assistant_text() {
-  local transcript="$1"
-  local cutoff="$2"
-
-  if command -v node >/dev/null 2>&1; then
-    node -e '
-      const fs = require("fs");
-      const file = process.argv[1];
-      const cutoff = Number(process.argv[2]);
-      const out = [];
-
-      function collectText(content) {
-        if (typeof content === "string") {
-          out.push(content);
-          return;
-        }
-        if (!Array.isArray(content)) return;
-        for (const item of content) {
-          if (item && typeof item.text === "string") out.push(item.text);
-        }
-      }
+function lastClaudeAssistant(payload) {
+  const direct = firstString(payload.last_assistant_message, payload.message);
+  if (direct) return direct;
 
-      for (const line of fs.readFileSync(file, "utf8").split(/\r?\n/)) {
-        if (!line.trim()) continue;
-        try {
-          const event = JSON.parse(line);
-          if (typeof event.timestamp === "number" && event.timestamp < cutoff) continue;
-          if (event.role === "assistant") collectText(event.content);
-          if (event.message && event.message.role === "assistant") collectText(event.message.content);
-        } catch (_) {}
-      }
-      process.stdout.write([...new Set(out.filter(Boolean))].join(" "));
-    ' "$transcript" "$cutoff" 2>/dev/null
-    return
-  fi
-
-  awk -v cutoff="$cutoff" '
-    {
-      if (match($0, /"timestamp"[[:space:]]*:[[:space:]]*[0-9]+/)) {
-        ts = substr($0, RSTART, RLENGTH)
-        gsub(/[^0-9]/, "", ts)
-        ts = int(ts)
-        if (ts < cutoff) next
-      }
+  const transcript = firstString(payload.transcript_path, payload.transcriptPath);
+  return transcript ? lastAssistantFromTranscript(transcript, "claude") : "";
+}
 
-      if (match($0, /"role"[[:space:]]*:[[:space:]]*"assistant"/)) {
-        if (match($0, /"content"[[:space:]]*:[[:space:]]*\[/)) {
-          gsub(/[^{]*\[/, "", $0)
-          gsub(/\].*/, "", $0)
-          while (match($0, /"text"[[:space:]]*:[[:space:]]*"[^"]*"/)) {
-            t = substr($0, RSTART, RLENGTH)
-            gsub(/"text"[[:space:]]*:[[:space:]]*"/, "", t)
-            gsub(/"$/, "", t)
-            if (t != "") print t
-            $0 = substr($0, RSTART + RLENGTH)
-          }
-        } else if (match($0, /"content"[[:space:]]*:[[:space:]]*"[^"]*"/)) {
-          t = substr($0, RSTART, RLENGTH)
-          gsub(/"content"[[:space:]]*:[[:space:]]*"/, "", t)
-          gsub(/"$/, "", t)
-          if (t != "") print t
-        }
-      }
-    }
-  ' "$transcript" 2>/dev/null | sort -u | tr '\n' ' '
+function lastCodexAssistant(payload) {
+  const direct = firstString(
+    payload["last-assistant-message"],
+    payload.last_assistant_message,
+    payload.lastAssistantMessage,
+    payload.message,
+    payload.lastMessage
+  );
+  if (direct) return direct;
+
+  const transcript = firstString(
+    payload.transcript_path,
+    payload.transcriptPath,
+    payload["transcript-path"]
+  );
+  return transcript ? lastAssistantFromTranscript(transcript, "codex") : "";
+}
+
+function readFile(file) {
+  try {
+    return fs.readFileSync(file, "utf8");
+  } catch {
+    return "";
+  }
 }
 
-# 从 stdin JSON 提取 transcript 路径和最后一条消息
-transcript=$(json_value "transcript_path")
-last_msg=$(json_value "last_assistant_message")
+function parseJSON(text) {
+  try {
+    return JSON.parse(text);
+  } catch {
+    return null;
+  }
+}
+
+function firstString(...values) {
+  for (const value of values) {
+    if (typeof value === "string" && value !== "") return value;
+  }
+  return "";
+}
+
+function collectText(content) {
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+  return content
+    .map(item => item && typeof item.text === "string" ? item.text : "")
+    .filter(Boolean)
+    .join(" ");
+}
 
-all_text="$last_msg"
+function lastAssistantFromTranscript(file, source) {
+  let data = "";
+  try {
+    data = fs.readFileSync(file, "utf8");
+  } catch {
+    return "";
+  }
+
+  let last = "";
+  for (const line of data.split(/\r?\n/)) {
+    if (!line.trim()) continue;
+    const event = parseJSON(line);
+    if (!event) continue;
+
+    if (source === "claude") {
+      if (event.role === "assistant") {
+        last = collectText(event.content) || last;
+      }
+      if (event.message && event.message.role === "assistant") {
+        last = collectText(event.message.content) || last;
+      }
+    }
 
-# 如果有 transcript 文件，提取最近 30 秒内的所有 assistant 消息
-if [[ -n "$transcript" && -f "$transcript" ]]; then
-  # 计算 30 秒前的时间戳（毫秒）
-  cutoff=$(($(date +%s) * 1000 - 30000))
+    if (source === "codex" &&
+      event.type === "response_item" &&
+      event.payload &&
+      event.payload.type === "message" &&
+      event.payload.role === "assistant"
+    ) {
+      last = collectText(event.payload.content) || last;
+    }
+  }
+  return last;
+}
 
-  # 优先用 JSON parser，Node 不存在时回退到简易 awk。
-  extra=$(extract_recent_assistant_text "$transcript" "$cutoff")
+NODE
+)
 
-  if [[ -n "$extra" ]]; then
-    all_text="$extra"
-  fi
+if [[ "$ISPEAK_HOOK_PRINT_TEXT" == "1" ]]; then
+  printf "%s" "$text"
+  exit 0
 fi
 
 echo "=== $(date) ===" >> "$LOG"
 echo "SOURCE: $SOURCE" >> "$LOG"
-echo "TEXT_LEN: ${#all_text}" >> "$LOG"
-echo "PREVIEW: ${all_text:0:150}" >> "$LOG"
+echo "TEXT_LEN: ${#text}" >> "$LOG"
+echo "PREVIEW: ${text:0:150}" >> "$LOG"
 
-if [[ -n "$all_text" && -S "$SOCK" ]]; then
-  printf "{source:%s}%s" "$SOURCE" "$all_text" | nc -U -w5 "$SOCK" 2>> "$LOG"
+if [[ -n "$text" && -S "$SOCK" ]]; then
+  printf "{source:%s}%s" "$SOURCE" "$text" | nc -U -w5 "$SOCK" 2>> "$LOG"
   echo "SPOKE: OK" >> "$LOG"
 else
   echo "SPOKE: SKIP" >> "$LOG"

package/main.go

@@ -17,7 +17,6 @@ import (
 	"os/exec"
 	"os/signal"
 	"path/filepath"
-	"regexp"
 	"strings"
 	"sync"
 	"syscall"
@@ -43,25 +42,6 @@ var tempDir string
 
 var errAlreadyRunning = errors.New("iSpeak already running")
 
-var (
-	markdownLinkRe     = regexp.MustCompile(`\[[^\]]+\]\(([^)]*)\)`)
-	absolutePathRe     = regexp.MustCompile(`/(?:Users|private|tmp|var|opt|usr|bin|sbin|etc|Library|Applications)/\S+`)
-	commitHashRe       = regexp.MustCompile(`\b[0-9a-f]{7,40}\b`)
-	uuidRe             = regexp.MustCompile(`\b[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}\b`)
-	urlRe              = regexp.MustCompile(`https?://\S+`)
-	ansiEscapeRe       = regexp.MustCompile(`\x1b\[[0-9;]*[A-Za-z]`)
-	multiSpaceRe       = regexp.MustCompile(`\s+`)
-	markdownListRe     = regexp.MustCompile(`^\s*(?:[-*+]\s+|\d+[.)]\s+)`)
-	htmlTagRe          = regexp.MustCompile(`<[^>]+>`)
-	codeFenceStartRe   = regexp.MustCompile("^```")
-	artifactStartRe    = regexp.MustCompile(`(?i)^<artifact\b`)
-	htmlDocumentLineRe = regexp.MustCompile(`(?i)^<!doctype html|^<html\b|^<head\b|^<body\b|^<style\b|^</`)
-	progressNoiseRe    = regexp.MustCompile(`(?i)(^\s*\d{1,3}%\s*$|\d{1,3}%.*\d+(?:\.\d+)?\s*(?:kb|mb|gb)/s|\bETA\b|^\s*[-=]{3,}\s*$)`)
-	speedNoiseRe       = regexp.MustCompile(`(?i)\d+(?:\.\d+)?\s*(?:kb|mb|gb)/s`)
-	etaNoiseRe         = regexp.MustCompile(`(?i)\bETA\b|预计剩余|剩余时间`)
-	fileListNoiseRe    = regexp.MustCompile(`(?i)\.(?:go|js|ts|tsx|jsx|json|md|yaml|yml|toml|sum|mod|lock|html|css|sh|plist|safetensors|mp3|wav|png|jpg|jpeg|pdf|docx)\b`)
-)
-
 type StreamPlayer interface {
 	Write(audio []byte) error
 	CloseAndWait() error
@@ -463,7 +443,7 @@ func loadConfig() Config {
 	// 回退到环境变量
 	return Config{
 		APIKey:   envOrDefault("IAGENT_TTS_API_KEY", ""),
-		Endpoint: envOrDefault("IAGENT_TTS_ENDPOINT", "https://openspeech.bytedance.com/api/v3/tts/unidirectional"),
+		Endpoint: envOrDefault("IAGENT_TTS_ENDPOINT", "https://openspeech.bytedance.com/api/v3/tts/unidirectional/sse"),
 	}
 }
 
@@ -590,8 +570,7 @@ func parseSSE(r io.Reader) ([]byte, error) {
 
 func parseSSEStream(r io.Reader, onAudio func([]byte) error) error {
 	audioChunks := 0
-	scanner := bufio.NewScanner(r)
-	scanner.Buffer(make([]byte, 256*1024), 256*1024)
+	reader := bufio.NewReaderSize(r, 64*1024)
 
 	var dataLines []string
 
@@ -608,40 +587,49 @@ func parseSSEStream(r io.Reader, onAudio func([]byte) error) error {
 		return err
 	}
 
-	for scanner.Scan() {
-		line := strings.TrimSpace(scanner.Text())
+	for {
+		rawLine, err := reader.ReadString('\n')
+		if err != nil && len(rawLine) == 0 {
+			if err == io.EOF {
+				break
+			}
+			return fmt.Errorf("read sse: %w", err)
+		}
+
+		line := strings.TrimSpace(rawLine)
 		if line == "" {
 			if err := flush(); err != nil {
 				return err
 			}
-			continue
-		}
-		if strings.HasPrefix(line, ":") || strings.HasPrefix(line, "event:") ||
+		} else if strings.HasPrefix(line, ":") || strings.HasPrefix(line, "event:") ||
 			strings.HasPrefix(line, "id:") || strings.HasPrefix(line, "retry:") {
-			continue
-		}
-		if strings.HasPrefix(line, "data:") {
+			// SSE metadata, ignored.
+		} else if strings.HasPrefix(line, "data:") {
 			dataLines = append(dataLines, strings.TrimPrefix(line, "data:"))
-			continue
-		}
-		// 非标准 JSON 直出
-		if err := flush(); err != nil {
-			return err
-		}
-		ok, err := processEvent(line, onAudio)
-		if ok {
-			audioChunks++
+		} else {
+			// 非标准 JSON 直出
+			if err := flush(); err != nil {
+				return err
+			}
+			ok, err := processEvent(line, onAudio)
+			if ok {
+				audioChunks++
+			}
+			if err != nil {
+				return err
+			}
 		}
+
 		if err != nil {
-			return err
+			if err == io.EOF {
+				break
+			}
+			return fmt.Errorf("read sse: %w", err)
 		}
 	}
 	if err := flush(); err != nil {
 		return err
 	}
-	if err := scanner.Err(); err != nil {
-		return fmt.Errorf("scan: %w", err)
-	}
 
 	if audioChunks == 0 {
 		return fmt.Errorf("no audio data")
@@ -661,6 +649,10 @@ func processEvent(payload string, onAudio func([]byte) error) (bool, error) {
 		return false, nil
 	}
 
+	if err := sseEventError(event); err != nil {
+		return false, err
+	}
+
 	if b64 := extractAudioBase64(event); b64 != "" {
 		data, err := base64.StdEncoding.DecodeString(b64)
 		if err != nil {
@@ -675,6 +667,35 @@ func processEvent(payload string, onAudio func([]byte) error) (bool, error) {
 	return false, nil
 }
 
+func sseEventError(event map[string]any) error {
+	codeValue, ok := event["code"]
+	if !ok {
+		return nil
+	}
+
+	var code int64
+	switch v := codeValue.(type) {
+	case float64:
+		code = int64(v)
+	case int:
+		code = int64(v)
+	case int64:
+		code = v
+	default:
+		return nil
+	}
+
+	if code == 0 || code == 20000000 {
+		return nil
+	}
+
+	message, _ := event["message"].(string)
+	if message == "" {
+		message = "unknown error"
+	}
+	return fmt.Errorf("tts sse error: code=%d message=%s", code, message)
+}
+
 func extractAudioBase64(event map[string]any) string {
 	for _, key := range []string{"data", "audio", "audio_data"} {
 		if v, ok := event[key].(string); ok && v != "" {
@@ -691,167 +712,6 @@ func extractAudioBase64(event map[string]any) string {
 	return ""
 }
 
-// 过滤格式符号，保留自然朗读文本。
-// 顺序很重要：先跳过跨行块结构，再跳过整行噪声，最后清理行内符号。
-func cleanText(text string) string {
-	var lines []string
-	rawLines := strings.Split(text, "\n")
-	inCodeBlock := false
-	inArtifact := false
-	inMarkdownTable := false
-	for i := 0; i < len(rawLines); i++ {
-		line := rawLines[i]
-		line = strings.TrimSpace(line)
-		if line == "" {
-			inMarkdownTable = false
-			continue
-		}
-		if codeFenceStartRe.MatchString(line) {
-			inCodeBlock = !inCodeBlock
-			continue
-		}
-		if inCodeBlock {
-			continue
-		}
-		if artifactStartRe.MatchString(line) {
-			inArtifact = !strings.Contains(strings.ToLower(line), "</artifact>")
-			continue
-		}
-		if inArtifact {
-			if strings.Contains(strings.ToLower(line), "</artifact>") {
-				inArtifact = false
-			}
-			continue
-		}
-		if isMarkdownTableSeparator(line) {
-			if len(lines) > 0 && isMarkdownTableRow(strings.TrimSpace(rawLines[i-1])) {
-				lines = lines[:len(lines)-1]
-			}
-			inMarkdownTable = true
-			continue
-		}
-		if inMarkdownTable {
-			if isMarkdownTableRow(line) {
-				continue
-			}
-			inMarkdownTable = false
-		}
-		if shouldSkipSpeechLine(line) {
-			continue
-		}
-
-		cleaned := cleanSpeechLine(line)
-		if cleaned != "" {
-			lines = append(lines, cleaned)
-		}
-	}
-	return strings.Join(lines, "，")
-}
-
-func shouldSkipSpeechLine(line string) bool {
-	if isMarkdownTableSeparator(line) {
-		return true
-	}
-	if strings.HasPrefix(line, "---") && strings.Count(line, "-") > 3 {
-		return true
-	}
-	if htmlDocumentLineRe.MatchString(line) {
-		return true
-	}
-	if isProgressNoiseLine(line) {
-		return true
-	}
-	if isMostlyTableRow(line) {
-		return true
-	}
-	if isMostlyFileListLine(line) {
-		return true
-	}
-	return false
-}
-
-func isMarkdownTableSeparator(line string) bool {
-	line = strings.TrimSpace(line)
-	return strings.Contains(line, "|") && strings.Trim(line, "|-: ") == ""
-}
-
-func isMarkdownTableRow(line string) bool {
-	line = strings.TrimSpace(line)
-	return strings.Count(line, "|") >= 2
-}
-
-func cleanSpeechLine(line string) string {
-	// Markdown 链接必须在 URL 删除前处理，否则会丢掉链接标题。
-	line = ansiEscapeRe.ReplaceAllString(line, "")
-	line = markdownListRe.ReplaceAllString(line, "")
-	line = markdownLinkRe.ReplaceAllStringFunc(line, func(match string) string {
-		if end := strings.Index(match, "]"); end > 1 {
-			return match[1:end]
-		}
-		return ""
-	})
-	line = urlRe.ReplaceAllString(line, "")
-	line = absolutePathRe.ReplaceAllString(line, " 路径 ")
-	// UUID 必须在短 hash 前处理，避免先删短片段后破坏 UUID 识别。
-	line = uuidRe.ReplaceAllString(line, "")
-	line = commitHashRe.ReplaceAllString(line, "")
-	line = htmlTagRe.ReplaceAllString(line, "")
-	line = strings.NewReplacer(
-		"**", "",
-		"*", "",
-		"`", "",
-		"#", "",
-		">", "",
-		"✅", "",
-		"❌", "",
-		"✓", "",
-		"✗", "",
-		"→", "到",
-	).Replace(line)
-	line = strings.Trim(line, " \t-:|")
-	line = multiSpaceRe.ReplaceAllString(line, " ")
-	return strings.TrimSpace(line)
-}
-
-func isMostlyTableRow(line string) bool {
-	if !strings.Contains(line, "|") {
-		return false
-	}
-	return strings.Count(line, "|") >= 2 && len([]rune(line)) > 40
-}
-
-func isProgressNoiseLine(line string) bool {
-	if !progressNoiseRe.MatchString(line) {
-		return false
-	}
-	if speedNoiseRe.MatchString(line) || etaNoiseRe.MatchString(line) {
-		return true
-	}
-	return !containsCJK(line)
-}
-
-func isMostlyFileListLine(line string) bool {
-	if !fileListNoiseRe.MatchString(line) {
-		return false
-	}
-	if containsCJK(line) {
-		return false
-	}
-	if strings.Contains(line, ".safetensors") {
-		return true
-	}
-	return strings.Count(line, ".") >= 2 || strings.Contains(line, "/") || strings.Contains(line, " - ")
-}
-
-func containsCJK(s string) bool {
-	for _, r := range s {
-		if r >= '\u4e00' && r <= '\u9fff' {
-			return true
-		}
-	}
-	return false
-}
-
 func main() {
 	log.SetFlags(log.Ltime | log.Lshortfile)
 

package/npm/postinstall.js

@@ -61,6 +61,43 @@ function copyIfMissing(src, dst, mode) {
   console.log(`配置文件已创建: ${dst}`);
 }
 
+function migrateDefaultEndpoint(configPath) {
+  if (!fs.existsSync(configPath)) {
+    return;
+  }
+  const oldEndpoint = "https://openspeech.bytedance.com/api/v3/tts/unidirectional";
+  const newEndpoint = "https://openspeech.bytedance.com/api/v3/tts/unidirectional/sse";
+  let config;
+  try {
+    config = JSON.parse(fs.readFileSync(configPath, "utf8"));
+  } catch (_) {
+    return;
+  }
+  if (config.endpoint !== oldEndpoint) {
+    return;
+  }
+  fs.copyFileSync(configPath, `${configPath}.bak`);
+  config.endpoint = newEndpoint;
+  fs.writeFileSync(configPath, `${JSON.stringify(config, null, 2)}\n`);
+  console.log(`配置 endpoint 已迁移到 SSE，旧配置备份: ${configPath}.bak`);
+}
+
+function installHook(src, dst) {
+  if (fs.existsSync(dst)) {
+    try {
+      if (fs.readFileSync(src, "utf8") !== fs.readFileSync(dst, "utf8")) {
+        fs.copyFileSync(dst, `${dst}.bak`);
+        console.log(`旧 Hook 已备份: ${dst}.bak`);
+      }
+    } catch (_) {
+      fs.copyFileSync(dst, `${dst}.bak`);
+      console.log(`旧 Hook 已备份: ${dst}.bak`);
+    }
+  }
+  copyExecutable(src, dst);
+  console.log(`Hook 脚本已安装: ${dst}`);
+}
+
 function sleep(ms) {
   Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms);
 }
@@ -106,8 +143,10 @@ function main() {
   symlinkForce(cliPath, path.join(binDir, "ispeak-claude"));
   symlinkForce(cliPath, path.join(binDir, "ispeak-codex"));
 
-  copyIfMissing(path.join(root, "configs", "config.example.json"), path.join(configDir, "config.json"));
-  copyIfMissing(path.join(root, "configs", "hook-speak.sh"), path.join(configDir, "hook-speak.sh"), 0o755);
+  const configPath = path.join(configDir, "config.json");
+  copyIfMissing(path.join(root, "configs", "config.example.json"), configPath);
+  migrateDefaultEndpoint(configPath);
+  installHook(path.join(root, "configs", "hook-speak.sh"), path.join(configDir, "hook-speak.sh"));
 
   const plist = fs
     .readFileSync(path.join(root, "configs", "com.iSpeak.plist"), "utf8")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xdfnet/ispeak",
-  "version": "1.6.4",
+  "version": "1.6.6",
   "description": "Local macOS TTS daemon for AI coding assistants, powered by Volcengine streaming TTS.",
   "license": "MIT",
   "homepage": "https://github.com/xdfnet/iSpeak#readme",
@@ -26,14 +26,16 @@
   "scripts": {
     "build": "go build -ldflags=\"-s -w\" -o build/ispeakd .",
     "test": "go test ./...",
+    "prepublishOnly": "make test",
     "postinstall": "node npm/postinstall.js",
     "pack:dry-run": "npm pack --dry-run"
   },
   "files": [
     "main.go",
+    "clean_text.go",
     "go.mod",
     "go.sum",
-    "scripts/",
+    "scripts/ispeak",
     "configs/",
     "npm/",
     "Docs/",

package/scripts/ispeak

@@ -2,7 +2,7 @@
 # ispeak — iSpeak 控制命令
 set -euo pipefail
 
-VERSION="1.6.3"
+VERSION="1.6.6"
 SOCK="$HOME/.config/iSpeak/ispeak.sock"
 PLIST="$HOME/Library/LaunchAgents/com.iSpeak.plist"
 CMD_NAME="$(basename "$0")"