mingdao-harness 0.2.4 → 0.2.5

package/src/tasks/worker.js

@@ -0,0 +1,126 @@
+// 后台任务 worker（Phase C C2）：独立进程执行一轮任务并写状态文件。
+// 由 cli.js 的内部入口 `mingdao run-worker <id> ...` 启动（startTask 拉起的子进程）。
+// 无交互：ask 权限降级 readonly；支持避峰顺延、MCP、自动标题、费用入账、完成通知与自动同步。
+import fs from 'node:fs';
+import path from 'node:path';
+import { loadConfig, ensureHome } from '../config.js';
+import { createProvider, resolveProviderConfig, helperProvider } from '../providers/index.js';
+import { startMcpServers } from '../mcp.js';
+import { patchTask } from '../tasks.js';
+import { notifyTaskDone } from '../notify.js';
+import { createAgent } from '../agent.js';
+import { createPermission } from '../permissions.js';
+import { buildSystemPrompt } from '../prompts.js';
+import { createSession, appendMessages, rewriteSession } from '../session.js';
+import { recordUsage } from '../cachestats.js';
+import { maybeAutoSync } from '../sync.js';
+import { isPeakHour, deferToOffpeak } from '../pricing.js';
+import { createIO } from '../ui.js';
+import { generateTitle, renameSessionFile, titleModel } from '../titles.js';
+
+/**
+ * 执行一轮后台任务：加载配置 → 建 agent → 跑回合 → 回写状态文件。
+ * @param {string} id 任务 id
+ * @param {string} question 任务内容
+ * @param {{permission?: any, model?: any, offpeak?: any}} opts
+ */
+export async function runWorkerTask(id, question, { permission, model, offpeak }) {
+  const home = ensureHome();
+  const finish = (/** @type {any} */ patch) => patchTask(home, id, patch);
+  let mcpFacade = null;
+  let cfg = null; // 提到 try 外：catch 分支也要读 cfg.notify
+  try {
+    cfg = loadConfig();
+    if (!cfg) throw new Error('未初始化配置，请先运行 mingdao init');
+    const modelName = model || cfg.model || 'deepseek-v4-flash';
+    const pc = resolveProviderConfig(cfg, modelName);
+    if (!pc.apiKey) throw new Error(`模型 ${modelName} 没有可用 API Key`);
+    const workingDir = process.cwd();
+    // 后台无交互：ask 权限降级为 readonly 并注明（需要写权限请 mingdao run --permission auto）
+    let perm = permission || cfg.permission || 'ask';
+    let note = '';
+    if (perm === 'ask') {
+      perm = 'readonly';
+      note = 'ask 权限下后台任务按只读执行';
+    }
+    // 避峰（评估 A2/Kimi P-1）：高峰时段（北京工作日 9:00–12:00、14:00–18:00）
+    // 自动顺延到最近闲时起点（12:00 / 18:00）执行，输入价省 50%
+    if (offpeak && isPeakHour(new Date())) {
+      const defer = deferToOffpeak(new Date());
+      finish({ note: `避峰等待至北京时间 ${defer.toISOString().slice(11, 16)}（闲时起执行，输入价省 50%）` });
+      await new Promise((r) => setTimeout(r, defer.getTime() - Date.now() + 2000));
+      finish({ note: '' });
+    }
+    const provider = await createProvider(cfg, modelName);
+    const io = createIO({ quiet: true });
+    const permissionObj = createPermission(perm, io);
+    /** @type {any} */
+    let mcpManager = null;
+    if (cfg.mcpServers && Object.keys(cfg.mcpServers).length) {
+      mcpFacade = {
+        toolSchemas: () => (mcpManager ? mcpManager.toolSchemas() : []),
+        call: (/** @type {any} */ n, /** @type {any} */ a) => (mcpManager ? mcpManager.call(n, a) : Promise.reject(new Error('MCP 未就绪'))),
+        isReadonly: (/** @type {any} */ n) => (mcpManager ? mcpManager.isReadonly(n) : false),
+        stop: () => {
+          if (mcpManager) mcpManager.stop();
+        },
+      };
+      startMcpServers(cfg.mcpServers, workingDir).then((m) => (mcpManager = m)).catch(() => {});
+    }
+    const sessionRef = /** @type {{name: any}} */ ({ name: null });
+    let persistedCount = 0;
+    const agent = createAgent({
+      provider,
+      permission: permissionObj,
+      io,
+      modelName,
+      workingDir,
+      cfg,
+      mcp: mcpFacade || undefined,
+      sessionRef,
+      onCompact: (/** @type {any} */ msgs) => {
+        rewriteSession(session.file, msgs);
+        persistedCount = msgs.length;
+      },
+    });
+    const session = createSession(home);
+    sessionRef.name = path.basename(session.file);
+    const messages = [
+      { role: 'system', content: buildSystemPrompt({ modelName, workingDir }) },
+      { role: 'user', content: question },
+    ];
+    appendMessages(session.file, messages);
+    persistedCount = messages.length;
+    const t0 = Date.now();
+    const res = await agent.runTurn(messages);
+    appendMessages(session.file, messages.slice(persistedCount));
+    if (cfg.autoTitle !== false && res.text) {
+      try {
+        const tModel = titleModel(cfg, modelName);
+        const title = await generateTitle(await helperProvider(cfg, tModel, provider), tModel, question);
+        if (title) renameSessionFile(fs, path, home, session, title);
+      } catch {}
+    }
+    const finalStatus = res.truncated ? 'failed' : res.aborted ? 'killed' : 'done';
+    recordUsage(res.perf?.usedModel || modelName, res.usage, /** @type {any} */ (res.perf));
+    finish({
+      status: finalStatus,
+      text: (res.text || '').slice(0, 2000),
+      usage: res.usage,
+      durationMs: Date.now() - t0,
+      session: path.basename(session.file),
+      note,
+    });
+    if (cfg.notify !== false && !process.env.MINGDAO_TASK_QUIET_NOTIFY) notifyTaskDone(question, finalStatus === 'killed' ? 'failed' : finalStatus);
+    try {
+      await maybeAutoSync();
+    } catch {}
+    process.exitCode = res.truncated ? 1 : 0;
+  } catch (/** @type {any} */ err) {
+    finish({ status: 'failed', error: String(err?.message || err) });
+    if (cfg?.notify !== false && !process.env.MINGDAO_TASK_QUIET_NOTIFY) notifyTaskDone(question, 'failed');
+    process.exitCode = 2;
+  } finally {
+    if (mcpFacade) mcpFacade.stop();
+  }
+}

package/src/tasks.js

@@ -11,11 +11,11 @@ import { atomicWriteFileSync, withFileLockSync } from './atomic-write.js';
 
 const CLI_PATH = fileURLToPath(new URL('./cli.js', import.meta.url));
 
-export function tasksDir(home) {
+export function tasksDir(/** @type {any} */ home) {
   return path.join(home, 'tasks');
 }
 
-export function listTasks(home) {
+export function listTasks(/** @type {any} */ home) {
   const dir = tasksDir(home);
   let files;
   try {
@@ -33,7 +33,7 @@ export function listTasks(home) {
   return out.sort((a, b) => b.startedAt - a.startedAt);
 }
 
-export function readTask(home, id) {
+export function readTask(/** @type {any} */ home, /** @type {any} */ id) {
   if (!isValidTaskId(id)) return null; // 防 id 路径穿越（id 直接拼进文件路径）
   try {
     return JSON.parse(fs.readFileSync(path.join(tasksDir(home), id + '.json'), 'utf8'));
@@ -42,18 +42,18 @@ export function readTask(home, id) {
   }
 }
 
-export function writeTask(home, task) {
+export function writeTask(/** @type {any} */ home, /** @type {any} */ task) {
   fs.mkdirSync(tasksDir(home), { recursive: true });
   // 原子写：先临时文件再改名，避免崩溃留下半截 JSON
   const target = path.join(tasksDir(home), task.id + '.json');
   atomicWriteFileSync(target, JSON.stringify(task, null, 2) + '\n'); // 质检 H4：tmp 名含 pid+随机
 }
 
-export function isValidTaskId(id) {
+export function isValidTaskId(/** @type {any} */ id) {
   return typeof id === 'string' && /^[a-z0-9]+$/.test(id) && id.length >= 4 && id.length <= 40;
 }
 
-export function startTask(home, question, { permission, model, cwd, offpeak, quietNotify } = /** @type {any} */ ({})) {
+export function startTask(/** @type {any} */ home, /** @type {any} */ question, { permission, model, cwd, offpeak, quietNotify } = /** @type {any} */ ({})) {
   const id = Date.now().toString(36) + Math.random().toString(36).slice(2, 6) + process.pid.toString(36);
   const task = {
     id,
@@ -83,16 +83,16 @@ export function startTask(home, question, { permission, model, cwd, offpeak, qui
   // 质检 M12：spawn 失败（ENOENT/ARG_MAX）必须有 error 监听，否则未捕获事件直接崩进程
   child.on('error', (err) => {
     try {
-      patchTask(home, id, { status: 'failed', error: `worker 启动失败：${err?.message || err}`, durationMs: Date.now() - task.startedAt });
+      patchTask(home, id, { status: 'failed', error: `worker 启动失败：${(/** @type {any} */ (err))?.message || err}`, durationMs: Date.now() - task.startedAt });
     } catch {}
   });
-  task.pid = child.pid;
+  task.pid = /** @type {any} */ (child.pid);
   writeTask(home, task);
   child.unref();
   return task;
 }
 
-export function patchTask(home, id, patch) {
+export function patchTask(/** @type {any} */ home, /** @type {any} */ id, /** @type {any} */ patch) {
   // 质检 H3：读-改-写加锁（worker finish 与 CLI kill 互斥，防 killed/done 互相覆盖）
   return withFileLockSync(path.join(tasksDir(home), '.lock'), () => {
     const t = readTask(home, id);
@@ -101,12 +101,12 @@ export function patchTask(home, id, patch) {
   });
 }
 
-export function killTask(home, id) {
+export function killTask(/** @type {any} */ home, /** @type {any} */ id) {
   if (!isValidTaskId(id)) return false;
   // 质检 H3：读-改-写加锁（与 worker 自身的状态写互斥）
   return withFileLockSync(path.join(tasksDir(home), '.lock'), () => killTaskInner(home, id));
 }
-function killTaskInner(home, id) {
+function killTaskInner(/** @type {any} */ home, /** @type {any} */ id) {
   const t = readTask(home, id);
   if (!t) return false;
   if (t.status === 'running' && t.pid) {
@@ -132,8 +132,8 @@ function killTaskInner(home, id) {
 
 const MARK = { running: '▶', done: '✓', failed: '✖', killed: '■' };
 
-export function formatTaskRow(t) {
-  const mark = MARK[t.status] || '?';
+export function formatTaskRow(/** @type {any} */ t) {
+  const mark = /** @type {any} */ (MARK)[t.status] || '?';
   const elapsed =
     t.status === 'running'
       ? relativeTime(t.startedAt).replace('前', '')

package/src/titles.js

@@ -4,21 +4,21 @@
 import { modelPreset } from './models.js';
 import { routingConfig } from './routing.js';
 
-export function titleModel(cfg, currentModel) {
+export function titleModel(/** @type {any} */ cfg, /** @type {any} */ currentModel) {
   const rc = routingConfig(cfg);
   if (rc) return rc.executor;
   // 路由关闭：辅助调用用当前模型（评估 P2-2：硬编码 flash 会在自定义网关 404 静默失败）
   return currentModel;
 }
 
-function cleanTitle(s) {
+function cleanTitle(/** @type {any} */ s) {
   return String(s ?? '')
     .trim()
     .replace(/["「」『』"'`*#\n\r]/g, '')
     .slice(0, 20);
 }
 
-export async function generateTitle(provider, model, firstUserText, { timeoutMs = 10000 } = {}) {
+export async function generateTitle(/** @type {any} */ provider, /** @type {any} */ model, /** @type {any} */ firstUserText, { timeoutMs = 10000 } = {}) {
   // 硬超时护栏（审计：「第二问无反应」根因）：标题生成发生在回合收尾阶段，若网关/网络层
   // 挂起（不抛错），回合永远无法 complete——前端停在生成态、后续发送全部静默。20s 兜底：
   // 超时放弃标题（会话仍以时间戳命名），绝不阻塞下一问。
@@ -65,13 +65,13 @@ export async function generateTitle(provider, model, firstUserText, { timeoutMs
   }
 }
 
-export function sanitizeTitle(s) {
+export function sanitizeTitle(/** @type {any} */ s) {
   const clean = String(s).replace(/[^\w\u4e00-\u9fa5 .-]/g, '_').slice(0, 40).trim();
   return clean || '会话';
 }
 
 // 重命名会话文件（与 /title 一致：确保存在、同名加随机后缀）
-export function renameSessionFile(fs, pathMod, home, session, title) {
+export function renameSessionFile(/** @type {any} */ fs, /** @type {any} */ pathMod, /** @type {any} */ home, /** @type {any} */ session, /** @type {any} */ title) {
   const safe = sanitizeTitle(title);
   let newFile = pathMod.join(home, 'sessions', safe + '.jsonl');
   try {

package/src/tokenizer.js

@@ -22,6 +22,7 @@ const SPLIT_RES = [
   /[!"#$%&'()*+,\-./:;<=>?@\[\\\]^_`{|}~][A-Za-z]+|[^\r\n\p{L}\p{P}\p{S}]?[\p{L}\p{M}]+| ?[\p{P}\p{S}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+/gu,
 ];
 
+/** @param {string} text */
 function pretokenize(text) {
   let pieces = [text];
   for (const re of SPLIT_RES) {
@@ -48,6 +49,7 @@ function pretokenize(text) {
 // 汉字会退化为逐字节计数（如「的」被计为 3 tokens 而非 1）。
 const BYTE_TO_UNICODE = (() => {
   const m = new Array(256);
+  /** @param {number} b */
   const self = (b) => (b >= 0x21 && b <= 0x7e) || (b >= 0xa1 && b <= 0xac) || (b >= 0xae);
   let n = 0;
   for (let b = 0; b < 256; b++) {
@@ -60,11 +62,14 @@ const BYTE_TO_UNICODE = (() => {
   return m;
 })();
 
+/** @param {string} s */
 function escapeRe(s) {
   return s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
 }
 
+/** @type {any} */
 let data = null;
+/** @type {any} */
 let loadError = null;
 
 function loadData() {
@@ -78,7 +83,7 @@ function loadData() {
       const [a, b] = parsed.merges[i];
       mergeRank.set(a + '\u0001' + b, i);
     }
-    const added = (parsed.added || []).filter(Boolean).sort((a, b) => b.length - a.length);
+    const added = (parsed.added || []).filter(Boolean).sort((/** @type {any} */ a, /** @type {any} */ b) => b.length - a.length);
     // 全部 added token 合并成单个正则（按长度降序排列 → 左起最长优先，与逐 token startsWith 语义一致），
     // 一次 matchAll 定位所有特殊 token，把 O(文本长 × 818 个 indexOf) 降到 O(n)
     const addedRe = added.length ? new RegExp(added.map(escapeRe).join('|'), 'gu') : null;
@@ -89,6 +94,7 @@ function loadData() {
   return data;
 }
 
+/** @param {any} modelName */
 export function isTokenizable(modelName) {
   if (typeof modelName !== 'string') return false;
   if (modelName.startsWith('deepseek')) return true;
@@ -124,15 +130,17 @@ const CJK_RANGES = [
   [0x3400, 0x4dbf], [0x4e00, 0x9fff], [0xf900, 0xfaff], // CJK 扩展/基本区/兼容
   [0x3040, 0x30ff], [0xac00, 0xd7af], // 假名 / 谚文
 ];
+/** @param {number} code */
 const isCjk = (code) => CJK_RANGES.some(([lo, hi]) => code >= lo && code <= hi);
 
+/** @param {any} text */
 export function heuristicTokens(text) {
   if (!text) return 0;
   let ascii = 0;
   let cjk = 0;
   let other = 0;
   for (const ch of String(text)) {
-    const code = ch.codePointAt(0);
+    const code = /** @type {number} */ (ch.codePointAt(0));
     if (code < 128) ascii += 1;
     else if (isCjk(code)) cjk += 1;
     else other += code > 0xffff ? 2 : 1; // 审计 B5：增补平面 emoji 按 2 token 保守计
@@ -144,6 +152,7 @@ export function heuristicTokens(text) {
 
 // 单个预分词片段的 BPE 计数（tiktoken 语义：优先合并 rank 最小的对，同 rank 取最左）。
 // 惰性最小堆实现（O(n log n)）：大文本（中文长文/工具输出）不再 O(n²) 全表扫描。
+/** @param {any} piece @param {any} d */
 function countPiece(piece, d) {
   const bytes = Buffer.from(piece, 'utf8');
   const syms = [];
@@ -151,17 +160,21 @@ function countPiece(piece, d) {
   if (syms.length <= 1) return syms.length;
 
   const alive = new Uint8Array(syms.length).fill(1);
+  /** @param {number} i */
   const leftOf = (i) => {
     for (let j = i - 1; j >= 0; j--) if (alive[j]) return j;
     return -1;
   };
+  /** @param {number} i */
   const rightOf = (i) => {
     for (let j = i + 1; j < syms.length; j++) if (alive[j]) return j;
     return -1;
   };
 
   // 最小堆（rank, leftIndex）；过期条目惰性丢弃/重推
+  /** @type {any[]} */
   const heap = [];
+  /** @param {number} rank @param {number} idx */
   const push = (rank, idx) => {
     heap.push([rank, idx]);
     let c = heap.length - 1;
@@ -182,6 +195,7 @@ function countPiece(piece, d) {
         const l = c * 2 + 1;
         const r = l + 1;
         let m = c;
+        /** @param {number} x */
         const better = (x) => heap[x][0] < heap[m][0] || (heap[x][0] === heap[m][0] && heap[x][1] < heap[m][1]);
         if (l < heap.length && better(l)) m = l;
         if (r < heap.length && better(r)) m = r;
@@ -238,6 +252,7 @@ const TOKEN_CACHE_MAX_ENTRIES = 512;
 const TOKEN_CACHE_MAX_LEN = 50000;
 const tokenCache = new Map();
 
+/** @param {string} s */
 function countDeepseek(s) {
   const d = loadData();
   if (!d) return heuristicTokens(s); // 词表缺失时优雅回退
@@ -254,12 +269,14 @@ function countDeepseek(s) {
   return total;
 }
 
+/** @param {any} piece @param {any} d */
 function countGap(piece, d) {
   let total = 0;
   for (const m of pretokenize(piece)) total += countPiece(m, d);
   return total;
 }
 
+/** @param {any} text @param {any} modelName */
 export function countTokens(text, modelName) {
   if (!text) return 0;
   const s = String(text);
@@ -279,9 +296,10 @@ export function countTokens(text, modelName) {
 }
 
 // 供上下文预算使用的计数器工厂
+/** @param {any} modelName */
 export function makeTokenCounter(modelName) {
   if (isTokenizable(modelName)) {
-    return (text) => countTokens(text, modelName);
+    return (/** @type {any} */ text) => countTokens(text, modelName);
   }
-  return (text) => heuristicTokens(text);
+  return (/** @type {any} */ text) => heuristicTokens(text);
 }

package/src/tools/bash.js

@@ -15,22 +15,22 @@ const MAX_TIMEOUT_SECONDS = 600;
 // 整体关闭（回到完全透传）。
 const SENSITIVE_ENV_PAIR = /(api[_-]?key|access[_-]?key|client[_-]?secret|private[_-]?key)/i;
 const SENSITIVE_ENV_SEGMENT = /(^|_)(token|secret|password|passwd|credential|authorization|auth)(_|$)/i;
-const isSensitiveEnv = (k) => SENSITIVE_ENV_PAIR.test(k) || SENSITIVE_ENV_SEGMENT.test(k);
+const isSensitiveEnv = (/** @type {any} */ k) => SENSITIVE_ENV_PAIR.test(k) || SENSITIVE_ENV_SEGMENT.test(k);
 
-function buildChildEnv(ctx, filterSensitive) {
+function buildChildEnv(/** @type {any} */ ctx, /** @type {any} */ filterSensitive) {
   if (!filterSensitive) return process.env;
   const keep = new Set((ctx?.cfg?.bashEnvKeep || []).map(String));
   const env = {};
   for (const [k, v] of Object.entries(process.env)) {
-    if (!isSensitiveEnv(k) || keep.has(k)) env[k] = v;
+    if (!isSensitiveEnv(k) || keep.has(k)) (/** @type {any} */ (env))[k] = v;
   }
   return env;
 }
 
-let sandboxSupport = null;
+let /** @type {any} */ sandboxSupport = null;
 
 export function detectSandbox() {
-  if (sandboxSupport !== null) return sandboxSupport;
+  if (/** @type {any} */ sandboxSupport !== null) return /** @type {any} */ sandboxSupport;
   sandboxSupport = 'none';
   if (process.platform === 'linux') {
     try {
@@ -46,12 +46,12 @@ export function detectSandbox() {
 // 输出折叠（审计 MiniMax §3.3-E / v0.1.48 P1-G）：模型回填的 bash 输出先折叠再截断——
 // 1) 剥离 ANSI 转义序列（CSI/OSC）；2) 连续重复行（>3 行相同）折叠为「首行 + 重复标记」。
 // npm install 类输出通常 30-50KB → 折叠后 5-10KB，单次工具回填省 60-70% prompt token。
-function stripAnsi(s) {
+function stripAnsi(/** @type {any} */ s) {
   return s
     .replace(/\x1b\[[0-9;?]*[ -/]*[@-~]/g, '')
     .replace(/\x1b\][^\x07\x1b]*(?:\x07|\x1b\\)/g, '');
 }
-function foldRepeats(s) {
+function foldRepeats(/** @type {any} */ s) {
   const lines = s.split('\n');
   const out = [];
   let i = 0;
@@ -69,12 +69,12 @@ function foldRepeats(s) {
   }
   return out.join('\n');
 }
-function tail(s, n) {
+function tail(/** @type {any} */ s, /** @type {any} */ n) {
   const folded = foldRepeats(stripAnsi(s));
   return folded.length > n ? `…[输出过长，已截断头部]\n${folded.slice(-n)}` : folded;
 }
 
-export function runBash(args, ctx) {
+export function runBash(/** @type {any} */ args, /** @type {any} */ ctx) {
   const command = String(args.command ?? '');
   if (!command.trim()) return { ok: false, error: 'command 参数为空。' };
   const timeoutSec = Math.min(Number(args.timeout) || 120, MAX_TIMEOUT_SECONDS);
@@ -125,13 +125,13 @@ export function runBash(args, ctx) {
     let done = false;
     let timedOut = false;
     // 输出增量截断：超长输出只保留尾部，避免内存无限累积
-    const cap = (s, d) => {
+    const cap = (/** @type {any} */ s, /** @type {any} */ d) => {
       const t = s + d;
       return t.length > MAX_OUTPUT * 2 ? t.slice(-MAX_OUTPUT * 2) : t;
     };
-    const killGroup = (sig) => {
+    const killGroup = (/** @type {any} */ sig) => {
       try {
-        process.kill(-child.pid, sig);
+        process.kill(-(/** @type {any} */ (child)).pid, sig);
       } catch {
         try {
           child.kill(sig);