@wenbin_wb/dsh-bridge 2.11.0 → 2.11.2

package/lib/platform/message-split.js

@@ -1,229 +1,229 @@
-// 出站消息分块与 SEND_FILE 指令解析（平台无关纯函数）
-// 自 conversation-bridge.js 拆出：按平台 maxMessageChars 分块、保留 fenced code block、
-// [SEND_FILE: ...] 显式指令提取与路径解析。
-import { statSync } from 'node:fs'
-import { isAbsolute, normalize, relative, resolve } from 'node:path'
-
-const FENCE_RE = /^```([^\n`]*)\s*$/
-
-function normalizeMarkdownBlocks(content) {
-  const lines = content.split('\n')
-  const out = []
-  let blankRun = 0
-  let inCode = false
-  for (const raw of lines) {
-    const line = raw.replace(/\s+$/, '')
-    if (FENCE_RE.test(line.trim())) {
-      inCode = !inCode
-      out.push(line)
-      blankRun = 0
-      continue
-    }
-    if (inCode) {
-      out.push(line)
-      continue
-    }
-    if (!line.trim()) {
-      blankRun += 1
-      if (blankRun <= 1) out.push('')
-      continue
-    }
-    blankRun = 0
-    out.push(line)
-  }
-  return out.join('\n').trim()
-}
-
-function splitMarkdownBlocks(content) {
-  const blocks = []
-  let current = []
-  let inCode = false
-  const flush = () => {
-    const block = current.join('\n').trim()
-    if (block) blocks.push(block)
-    current = []
-  }
-  for (const raw of content.split('\n')) {
-    const line = raw.replace(/\s+$/, '')
-    if (FENCE_RE.test(line.trim())) {
-      if (!inCode && current.length) flush()
-      current.push(line)
-      inCode = !inCode
-      if (!inCode) flush()
-      continue
-    }
-    if (inCode) {
-      current.push(line)
-      continue
-    }
-    if (!line.trim()) {
-      flush()
-      continue
-    }
-    current.push(line)
-  }
-  flush()
-  return blocks
-}
-
-function hardSplit(text, max) {
-  const chunks = []
-  let rest = text
-  while (rest.length > max) {
-    chunks.push(rest.slice(0, max))
-    rest = rest.slice(max)
-  }
-  if (rest) chunks.push(rest)
-  return chunks
-}
-
-function packBlocks(blocks, max) {
-  const units = []
-  let current = ''
-  for (const block of blocks) {
-    const candidate = current ? `${current}\n\n${block}` : block
-    if (candidate.length <= max) {
-      current = candidate
-      continue
-    }
-    if (current) units.push(current)
-    if (block.length <= max) {
-      current = block
-    } else {
-      units.push(...hardSplit(block, max))
-      current = ''
-    }
-  }
-  if (current) units.push(current)
-  return units
-}
-
-export function splitForIM(content, max = 2000) {
-  // 安全检查：防止畸形输入导致无限循环或崩溃
-  if (typeof content !== 'string' || content.length === 0) return []
-  if (content.length > 1_000_000) {
-    content = content.slice(0, 1_000_000) + '\n\n[已截断：内容过长]'
-  }
-  const normalized = normalizeMarkdownBlocks(content)
-  if (!normalized) return []
-  if (normalized.length <= max) return [normalized]
-  return packBlocks(splitMarkdownBlocks(normalized), max)
-}
-
-// 协议标记过滤：模型偶尔会把工具调用语法（如 <｜｜DSML｜｜tool_calls>…）当正文
-// 输出，这些是协议内容，转发到 IM 只会是乱码噪音。整块移除 tool_calls 段落，
-// 再剥掉残余的 DSML 标签；剩余正文照常发送。
-export function stripProtocolMarkup(text) {
-  if (!text || typeof text !== 'string' || !text.includes('DSML')) return text
-  let out = text.replace(/<[^<>]*DSML[^<>]*tool_calls[^<>]*>[\s\S]*?<[^<>]*DSML[^<>]*tool_calls[^<>]*>/g, '')
-  out = out.replace(/<[^<>]*DSML[^<>]*>/g, '')
-  return out
-}
-
-export function textOfAssistantMessage(message) {
-  const raw = (message.content ?? [])
-    .filter((block) => block?.type === 'text')
-    .map((block) => block.text)
-    .join('\n')
-  return stripProtocolMarkup(raw)
-}
-
-/**
- * 尝试将任意路径（绝对或相对当前工作区）解析为真实存在的本地文件绝对路径
- */
-export function resolveFilePath(rawPath, cwd = process.cwd()) {
-  if (typeof rawPath !== 'string') return null
-  let p = rawPath.trim()
-    .replace(/^["'`]|["'`]$/g, '')
-    .replace(/^file:\/\/\/?/, '')
-    .replace(/^[📁📄📦\s]+/, '')
-  if (!p) return null
-  // 排除 HTTP/HTTPS 网址
-  if (/^https?:\/\//i.test(p)) return null
-  const resolved = isAbsolute(p) ? normalize(p) : resolve(cwd, p)
-  try {
-    if (statSync(resolved).isFile()) {
-      return resolved
-    }
-  } catch {}
-  return null
-}
-
-/**
- * 判断解析后的文件路径是否允许经 [SEND_FILE] 发送给 IM。
- * 安全约束（防止模型被诱导后外发任意本地文件）：
- *   1. 必须位于 allowedRoots 中的某个根目录（含子目录）内 —— 默认仅会话 cwd；
- *   2. 路径任何一段不得命中敏感名单（.ssh/.gnupg/.aws/.git/.env/.credentials 等）。
- * @param {string} resolvedPath 已解析的绝对路径
- * @param {string|string[]} allowedRoots 允许的根目录（绝对路径）；默认 process.cwd()
- * @returns {boolean}
- */
-export function isPathAllowedForSend(resolvedPath, allowedRoots = process.cwd()) {
-  if (typeof resolvedPath !== 'string' || !resolvedPath) return false
-  const roots = Array.isArray(allowedRoots) ? allowedRoots : [allowedRoots]
-  if (roots.length === 0) return false
-
-  const normalized = resolve(resolvedPath)
-  const pathParts = normalized.split(/[\\/]/).filter(Boolean)
-  const SENSITIVE_PARTS = new Set([
-    '.ssh', '.gnupg', '.aws', '.azure', '.kube', '.git', '.svn', '.hg',
-    '.bash_history', '.zsh_history', '.profile', '.bash_profile', '.bashrc',
-    '.zshrc', '.netrc', '.env', '.npmrc', '.credentials', 'id_rsa', 'id_ed25519',
-    'id_ecdsa', 'id_dsa', 'shadow', 'passwd',
-  ])
-  for (const part of pathParts) {
-    if (SENSITIVE_PARTS.has(part.toLowerCase())) return false
-  }
-
-  for (const root of roots) {
-    if (typeof root !== 'string' || !root) continue
-    const normRoot = resolve(root)
-    if (normalized === normRoot) return true
-    // 用 relative 判断是否位于根内：越界时 rel 为 '..' 或以 '../' 开头
-    // （Windows 跨盘则 rel 是绝对路径，isAbsolute 拦截）。跨平台正确处理 \ 与 / 差异。
-    const rel = relative(normRoot, normalized)
-    if (rel && !rel.startsWith('..') && !isAbsolute(rel)) return true
-  }
-  return false
-}
-
-/**
- * 提取并过滤文本中的 [SEND_FILE: <path>] 显式发送指令
- * 由 AI 根据用户意图显式决定何时向用户发送文件附件，杜绝底层盲目扫描与误发。
- * @param {string} text - 原始助手回复文本
- * @param {string} cwd - 会话当前工作目录
- * @returns {{ cleanText: string, files: string[] }}
- */
-export function extractAndStripSendFileDirectives(text, cwd = process.cwd()) {
-  if (typeof text !== 'string' || !text.trim()) {
-    return { cleanText: text || '', files: [] }
-  }
-
-  const files = []
-  const directiveRegex = /\[(?:SEND_FILE|SEND-FILE|send_file|send-file|SEND_MEDIA|send_media):\s*[`"']?([^\]`"'\r\n]+?)[`"']?\s*\]/gi
-
-  let m
-  const re = new RegExp(directiveRegex)
-  while ((m = re.exec(text)) !== null) {
-    const rawPath = m[1].trim()
-    const resolved = resolveFilePath(rawPath, cwd)
-    if (resolved && !files.includes(resolved)) {
-      files.push(resolved)
-    }
-  }
-
-  // 从聊天正文中彻底剔除控制指令（保持 IM 聊天气泡的干净整洁）
-  const cleanText = text.replace(directiveRegex, '').replace(/\n{3,}/g, '\n\n').trim()
-
-  return { cleanText, files }
-}
-
-/**
- * 提取文本中的产物文件路径（基于显式指令）
- */
-export function extractFilePathsFromText(text, cwd = process.cwd()) {
-  return extractAndStripSendFileDirectives(text, cwd).files
-}
-
-// ---------------------------------------------------------------------------
-// digest 摘要
+// 出站消息分块与 SEND_FILE 指令解析（平台无关纯函数）
+// 自 conversation-bridge.js 拆出：按平台 maxMessageChars 分块、保留 fenced code block、
+// [SEND_FILE: ...] 显式指令提取与路径解析。
+import { statSync } from 'node:fs'
+import { isAbsolute, normalize, relative, resolve } from 'node:path'
+
+const FENCE_RE = /^```([^\n`]*)\s*$/
+
+function normalizeMarkdownBlocks(content) {
+  const lines = content.split('\n')
+  const out = []
+  let blankRun = 0
+  let inCode = false
+  for (const raw of lines) {
+    const line = raw.replace(/\s+$/, '')
+    if (FENCE_RE.test(line.trim())) {
+      inCode = !inCode
+      out.push(line)
+      blankRun = 0
+      continue
+    }
+    if (inCode) {
+      out.push(line)
+      continue
+    }
+    if (!line.trim()) {
+      blankRun += 1
+      if (blankRun <= 1) out.push('')
+      continue
+    }
+    blankRun = 0
+    out.push(line)
+  }
+  return out.join('\n').trim()
+}
+
+function splitMarkdownBlocks(content) {
+  const blocks = []
+  let current = []
+  let inCode = false
+  const flush = () => {
+    const block = current.join('\n').trim()
+    if (block) blocks.push(block)
+    current = []
+  }
+  for (const raw of content.split('\n')) {
+    const line = raw.replace(/\s+$/, '')
+    if (FENCE_RE.test(line.trim())) {
+      if (!inCode && current.length) flush()
+      current.push(line)
+      inCode = !inCode
+      if (!inCode) flush()
+      continue
+    }
+    if (inCode) {
+      current.push(line)
+      continue
+    }
+    if (!line.trim()) {
+      flush()
+      continue
+    }
+    current.push(line)
+  }
+  flush()
+  return blocks
+}
+
+function hardSplit(text, max) {
+  const chunks = []
+  let rest = text
+  while (rest.length > max) {
+    chunks.push(rest.slice(0, max))
+    rest = rest.slice(max)
+  }
+  if (rest) chunks.push(rest)
+  return chunks
+}
+
+function packBlocks(blocks, max) {
+  const units = []
+  let current = ''
+  for (const block of blocks) {
+    const candidate = current ? `${current}\n\n${block}` : block
+    if (candidate.length <= max) {
+      current = candidate
+      continue
+    }
+    if (current) units.push(current)
+    if (block.length <= max) {
+      current = block
+    } else {
+      units.push(...hardSplit(block, max))
+      current = ''
+    }
+  }
+  if (current) units.push(current)
+  return units
+}
+
+export function splitForIM(content, max = 2000) {
+  // 安全检查：防止畸形输入导致无限循环或崩溃
+  if (typeof content !== 'string' || content.length === 0) return []
+  if (content.length > 1_000_000) {
+    content = content.slice(0, 1_000_000) + '\n\n[已截断：内容过长]'
+  }
+  const normalized = normalizeMarkdownBlocks(content)
+  if (!normalized) return []
+  if (normalized.length <= max) return [normalized]
+  return packBlocks(splitMarkdownBlocks(normalized), max)
+}
+
+// 协议标记过滤：模型偶尔会把工具调用语法（如 <｜｜DSML｜｜tool_calls>…）当正文
+// 输出，这些是协议内容，转发到 IM 只会是乱码噪音。整块移除 tool_calls 段落，
+// 再剥掉残余的 DSML 标签；剩余正文照常发送。
+export function stripProtocolMarkup(text) {
+  if (!text || typeof text !== 'string' || !text.includes('DSML')) return text
+  let out = text.replace(/<[^<>]*DSML[^<>]*tool_calls[^<>]*>[\s\S]*?<[^<>]*DSML[^<>]*tool_calls[^<>]*>/g, '')
+  out = out.replace(/<[^<>]*DSML[^<>]*>/g, '')
+  return out
+}
+
+export function textOfAssistantMessage(message) {
+  const raw = (message.content ?? [])
+    .filter((block) => block?.type === 'text')
+    .map((block) => block.text)
+    .join('\n')
+  return stripProtocolMarkup(raw)
+}
+
+/**
+ * 尝试将任意路径（绝对或相对当前工作区）解析为真实存在的本地文件绝对路径
+ */
+export function resolveFilePath(rawPath, cwd = process.cwd()) {
+  if (typeof rawPath !== 'string') return null
+  let p = rawPath.trim()
+    .replace(/^["'`]|["'`]$/g, '')
+    .replace(/^file:\/\/\/?/, '')
+    .replace(/^[📁📄📦\s]+/, '')
+  if (!p) return null
+  // 排除 HTTP/HTTPS 网址
+  if (/^https?:\/\//i.test(p)) return null
+  const resolved = isAbsolute(p) ? normalize(p) : resolve(cwd, p)
+  try {
+    if (statSync(resolved).isFile()) {
+      return resolved
+    }
+  } catch {}
+  return null
+}
+
+/**
+ * 判断解析后的文件路径是否允许经 [SEND_FILE] 发送给 IM。
+ * 安全约束（防止模型被诱导后外发任意本地文件）：
+ *   1. 必须位于 allowedRoots 中的某个根目录（含子目录）内 —— 默认仅会话 cwd；
+ *   2. 路径任何一段不得命中敏感名单（.ssh/.gnupg/.aws/.git/.env/.credentials 等）。
+ * @param {string} resolvedPath 已解析的绝对路径
+ * @param {string|string[]} allowedRoots 允许的根目录（绝对路径）；默认 process.cwd()
+ * @returns {boolean}
+ */
+export function isPathAllowedForSend(resolvedPath, allowedRoots = process.cwd()) {
+  if (typeof resolvedPath !== 'string' || !resolvedPath) return false
+  const roots = Array.isArray(allowedRoots) ? allowedRoots : [allowedRoots]
+  if (roots.length === 0) return false
+
+  const normalized = resolve(resolvedPath)
+  const pathParts = normalized.split(/[\\/]/).filter(Boolean)
+  const SENSITIVE_PARTS = new Set([
+    '.ssh', '.gnupg', '.aws', '.azure', '.kube', '.git', '.svn', '.hg',
+    '.bash_history', '.zsh_history', '.profile', '.bash_profile', '.bashrc',
+    '.zshrc', '.netrc', '.env', '.npmrc', '.credentials', 'id_rsa', 'id_ed25519',
+    'id_ecdsa', 'id_dsa', 'shadow', 'passwd',
+  ])
+  for (const part of pathParts) {
+    if (SENSITIVE_PARTS.has(part.toLowerCase())) return false
+  }
+
+  for (const root of roots) {
+    if (typeof root !== 'string' || !root) continue
+    const normRoot = resolve(root)
+    if (normalized === normRoot) return true
+    // 用 relative 判断是否位于根内：越界时 rel 为 '..' 或以 '../' 开头
+    // （Windows 跨盘则 rel 是绝对路径，isAbsolute 拦截）。跨平台正确处理 \ 与 / 差异。
+    const rel = relative(normRoot, normalized)
+    if (rel && !rel.startsWith('..') && !isAbsolute(rel)) return true
+  }
+  return false
+}
+
+/**
+ * 提取并过滤文本中的 [SEND_FILE: <path>] 显式发送指令
+ * 由 AI 根据用户意图显式决定何时向用户发送文件附件，杜绝底层盲目扫描与误发。
+ * @param {string} text - 原始助手回复文本
+ * @param {string} cwd - 会话当前工作目录
+ * @returns {{ cleanText: string, files: string[] }}
+ */
+export function extractAndStripSendFileDirectives(text, cwd = process.cwd()) {
+  if (typeof text !== 'string' || !text.trim()) {
+    return { cleanText: text || '', files: [] }
+  }
+
+  const files = []
+  const directiveRegex = /\[(?:SEND_FILE|SEND-FILE|send_file|send-file|SEND_MEDIA|send_media):\s*[`"']?([^\]`"'\r\n]+?)[`"']?\s*\]/gi
+
+  let m
+  const re = new RegExp(directiveRegex)
+  while ((m = re.exec(text)) !== null) {
+    const rawPath = m[1].trim()
+    const resolved = resolveFilePath(rawPath, cwd)
+    if (resolved && !files.includes(resolved)) {
+      files.push(resolved)
+    }
+  }
+
+  // 从聊天正文中彻底剔除控制指令（保持 IM 聊天气泡的干净整洁）
+  const cleanText = text.replace(directiveRegex, '').replace(/\n{3,}/g, '\n\n').trim()
+
+  return { cleanText, files }
+}
+
+/**
+ * 提取文本中的产物文件路径（基于显式指令）
+ */
+export function extractFilePathsFromText(text, cwd = process.cwd()) {
+  return extractAndStripSendFileDirectives(text, cwd).files
+}
+
+// ---------------------------------------------------------------------------
+// digest 摘要

package/lib/platform/stream-slices.js

@@ -1,21 +1,21 @@
-// 流式打字机输出的公共切片工具
-//
-// 各平台网关（qq/feishu/telegram）的流式输出都是"增量前缀"模式：
-// 第 i 片内容 = 前 i 个基础分片的拼接（服务端逐片覆盖，消息逐渐变长）。
-// 基础分块器因平台而异（QQ markdown 感知分块 / 飞书换行窗口 / Telegram splitForIM），
-// 本模块只沉淀三者复制的"前缀累积"逻辑。
-
-/**
- * 把基础分片转成增量前缀切片。
- * @param {string[]} chunks 基础分片（非空内容应至少产生 1 片）
- * @returns {string[]} 第 i 个元素 = chunks[0..i] 拼接
- */
-export function cumulativeSlices(chunks) {
-  const slices = []
-  let acc = ''
-  for (const s of chunks) {
-    acc += s
-    slices.push(acc)
-  }
-  return slices
-}
+// 流式打字机输出的公共切片工具
+//
+// 各平台网关（qq/feishu/telegram）的流式输出都是"增量前缀"模式：
+// 第 i 片内容 = 前 i 个基础分片的拼接（服务端逐片覆盖，消息逐渐变长）。
+// 基础分块器因平台而异（QQ markdown 感知分块 / 飞书换行窗口 / Telegram splitForIM），
+// 本模块只沉淀三者复制的"前缀累积"逻辑。
+
+/**
+ * 把基础分片转成增量前缀切片。
+ * @param {string[]} chunks 基础分片（非空内容应至少产生 1 片）
+ * @returns {string[]} 第 i 个元素 = chunks[0..i] 拼接
+ */
+export function cumulativeSlices(chunks) {
+  const slices = []
+  let acc = ''
+  for (const s of chunks) {
+    acc += s
+    slices.push(acc)
+  }
+  return slices
+}

package/lib/restart-helper.mjs

@@ -9,7 +9,8 @@
 //
 // 用法（由 lib/index.js 组装 payload）：node restart-helper.mjs '<json>'
 import { spawn } from 'node:child_process';
-import { appendFileSync, openSync } from 'node:fs';
+import { appendFileSync, openSync, mkdirSync } from 'node:fs';
+import { dirname } from 'node:path';
 import { connect } from 'node:net';
 
 const payload = JSON.parse(process.argv[2] ?? '{}');
@@ -80,7 +81,10 @@ async function main() {
   }
 
   let stdio = 'ignore';
-  try { stdio = openSync(childLog, 'a'); } catch { /* 打不开就用 ignore */ }
+  try {
+    mkdirSync(dirname(childLog), { recursive: true });
+    stdio = openSync(childLog, 'a');
+  } catch { /* 打不开就用 ignore */ }
 
   // spawn 的失败是异步 'error' 事件（例如 execPath 不存在 → ENOENT），只 try/catch 抓不到；
   // 不挂监听会变成未捕获异常，而且父进程给助手设的是 stdio:'ignore' → 失败细节全部丢失。
@@ -90,7 +94,17 @@ async function main() {
     const settle = (value) => { if (!settled) { settled = true; resolve(value); } };
     let child;
     try {
-      child = spawn(execPath, argv, { cwd, env: process.env, detached: true, stdio: ['ignore', stdio, stdio] });
+      const childEnv = { ...process.env };
+      if (process.platform !== 'win32') {
+        const execDir = dirname(execPath);
+        const defaultPaths = [execDir, '/usr/local/bin', '/usr/bin', '/bin'].filter(Boolean);
+        const existingPath = childEnv.PATH || '';
+        const missingPaths = defaultPaths.filter((p) => !existingPath.split(':').includes(p));
+        if (missingPaths.length > 0) {
+          childEnv.PATH = existingPath ? `${existingPath}:${missingPaths.join(':')}` : missingPaths.join(':');
+        }
+      }
+      child = spawn(execPath, argv, { cwd, env: childEnv, detached: true, stdio: ['ignore', stdio, stdio] });
     } catch (err) {
       log(`派生新 dsh 失败（同步异常）：${err?.message ?? err}`);
       settle(null);