metame-cli 1.5.10 → 1.5.12

package/scripts/ops-verifier.js

@@ -0,0 +1,197 @@
+'use strict';
+
+const { execSync } = require('child_process');
+const fs = require('fs');
+const path = require('path');
+
+/**
+ * MetaMe Ops verifier — hard gate based on syntax + test results.
+ *
+ * Phases:
+ *   diagnose  — workspace/ops/diagnosis.md must exist
+ *   plan      — workspace/ops/fix-plan.md must exist
+ *   fix       — all .js files pass syntax check + all tests pass
+ *   test      — same as fix (double-check after implementation)
+ */
+
+// ── Protected files: changes to these BLOCK the gate ─────────
+// These are load-bearing files. If broken, the entire system dies.
+// Agent can diagnose issues in these files, but fixes require human approval.
+
+// Scripts that must not be modified (basename match)
+const PROTECTED_SCRIPTS = [
+  'daemon.js',                    // main daemon process
+  'daemon-reactive-lifecycle.js', // perpetual loop engine
+  'daemon-team-dispatch.js',      // dispatch + prompt enrichment
+  'daemon-claude-engine.js',      // Claude API bridge
+  'daemon-bridges.js',            // Feishu/Telegram bridges
+  'daemon-remote-dispatch.js',    // cross-device dispatch
+  'ops-verifier.js',              // this file — agent must not weaken its own gate
+];
+
+// Repo-root files that must not be modified (relative path match)
+const PROTECTED_ROOT_FILES = [
+  'perpetual.yaml',               // manifest — controls which verifier runs
+  'perpetual.yml',
+  'CLAUDE.md',                    // agent identity — weakening it weakens all safety
+];
+
+/**
+ * Check if any protected file was modified (repo-internal only).
+ * daemon.yaml lives outside repo and has its own backup mechanism — not our concern.
+ * FAIL-CLOSED: if git is unavailable, returns ['__GIT_UNAVAILABLE__'].
+ */
+function checkProtectedFiles(cwd) {
+  try {
+    const diff = execSync('git diff --name-only HEAD', { cwd, timeout: 5000, encoding: 'utf8', stdio: 'pipe' }).trim();
+    const staged = execSync('git diff --cached --name-only', { cwd, timeout: 5000, encoding: 'utf8', stdio: 'pipe' }).trim();
+    const changed = [...new Set([...diff.split('\n'), ...staged.split('\n')])].filter(Boolean);
+
+    const violations = [];
+    for (const f of changed) {
+      if (PROTECTED_SCRIPTS.includes(path.basename(f))) violations.push(path.basename(f));
+    }
+    for (const f of changed) {
+      if (PROTECTED_ROOT_FILES.includes(f)) violations.push(f);
+    }
+    return violations;
+  } catch {
+    return ['__GIT_UNAVAILABLE__'];
+  }
+}
+
+const GATES = {
+  diagnose: {
+    check: (cwd) => fileExists(cwd, 'workspace/ops/diagnosis.md'),
+    failDetail: 'workspace/ops/diagnosis.md not found',
+    hints: ['Scan logs and produce a diagnosis report at workspace/ops/diagnosis.md'],
+  },
+  plan: {
+    check: (cwd) => fileExists(cwd, 'workspace/ops/fix-plan.md'),
+    failDetail: 'workspace/ops/fix-plan.md not found',
+    hints: ['Write a fix plan at workspace/ops/fix-plan.md before implementing'],
+  },
+  fix: {
+    check: (cwd) => syntaxAndTests(cwd),
+    failDetail: 'syntax check or tests failed',
+    hints: ['Ensure all scripts pass `node -c` and `node --test`'],
+  },
+  test: {
+    check: (cwd) => syntaxAndTests(cwd),
+    failDetail: 'tests failed on verification pass',
+    hints: ['All tests must pass before marking mission complete'],
+  },
+};
+
+function fileExists(cwd, relPath) {
+  try { return fs.statSync(path.join(cwd, relPath)).isFile(); } catch { return false; }
+}
+
+function syntaxAndTests(cwd) {
+  try {
+    // Find changed .js files (staged + unstaged) — only verify what was touched
+    let changedFiles = [];
+    try {
+      const diff = execSync('git diff --name-only HEAD -- scripts/', { cwd, timeout: 5000, encoding: 'utf8', stdio: 'pipe' }).trim();
+      const staged = execSync('git diff --cached --name-only -- scripts/', { cwd, timeout: 5000, encoding: 'utf8', stdio: 'pipe' }).trim();
+      changedFiles = [...new Set([...diff.split('\n'), ...staged.split('\n')])]
+        .filter(f => f.endsWith('.js'))
+        .map(f => path.basename(f));
+    } catch { /* git not available or not a repo — check all */ }
+
+    // Syntax check changed .js files (or all if git unavailable)
+    const srcFiles = changedFiles.length > 0
+      ? changedFiles.filter(f => !f.endsWith('.test.js'))
+      : fs.readdirSync(path.join(cwd, 'scripts')).filter(f => f.endsWith('.js') && !f.endsWith('.test.js'));
+
+    for (const f of srcFiles) {
+      execSync(`node -c scripts/${f}`, { cwd, timeout: 5000, stdio: 'pipe' });
+    }
+
+    // Run associated test files
+    const testFiles = changedFiles.length > 0
+      ? changedFiles.filter(f => f.endsWith('.test.js'))
+      : [];
+    // Also run tests for changed source files (e.g., foo.js → foo.test.js)
+    for (const f of srcFiles) {
+      const testName = f.replace('.js', '.test.js');
+      if (!testFiles.includes(testName)) {
+        const testPath = path.join(cwd, 'scripts', testName);
+        if (fs.existsSync(testPath)) testFiles.push(testName);
+      }
+    }
+
+    // If no specific tests identified, run the core lifecycle test as baseline
+    if (testFiles.length === 0) {
+      testFiles.push('daemon-reactive-lifecycle.test.js');
+    }
+
+    for (const f of [...new Set(testFiles)]) {
+      const testPath = path.join(cwd, 'scripts', f);
+      if (!fs.existsSync(testPath)) continue;
+      const out = execSync(`node --test scripts/${f}`, { cwd, timeout: 60000, encoding: 'utf8', stdio: 'pipe' });
+      if (out.includes('# fail') && !out.includes('# fail 0')) return false;
+    }
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+function verify(cwd, phase) {
+  const gate = GATES[phase];
+  if (!gate) {
+    return {
+      passed: false, phase: phase || '',
+      details: `unknown phase: "${phase || ''}"`,
+      artifacts: [], hints: ['Valid phases: diagnose, plan, fix, test'],
+    };
+  }
+
+  // Safety gate: block if protected files were touched
+  if (phase === 'fix' || phase === 'test') {
+    const violated = checkProtectedFiles(cwd);
+    if (violated.length > 0) {
+      return {
+        passed: false, phase,
+        details: `PROTECTED_FILE_VIOLATION: ${violated.join(', ')}`,
+        artifacts: [],
+        hints: [
+          `These files are daemon-critical and cannot be auto-merged: ${violated.join(', ')}`,
+          'Revert changes to protected files. Write the fix plan to workspace/ops/ and notify the user for manual review.',
+        ],
+      };
+    }
+
+  }
+
+  const passed = gate.check(cwd);
+
+  // Collect artifacts
+  const opsDir = path.join(cwd, 'workspace', 'ops');
+  let artifacts = [];
+  try {
+    if (fs.existsSync(opsDir)) {
+      artifacts = fs.readdirSync(opsDir)
+        .filter(f => f.endsWith('.md'))
+        .map(f => `workspace/ops/${f}`);
+    }
+  } catch { /* ok */ }
+
+  return {
+    passed, phase,
+    details: passed ? `phase "${phase}" gate passed` : gate.failDetail,
+    artifacts,
+    hints: passed ? [] : gate.hints,
+  };
+}
+
+// CLI entry
+if (require.main === module) {
+  const cwd = process.env.VERIFIER_CWD || process.cwd();
+  const phase = process.env.VERIFIER_PHASE || 'unknown';
+  const result = verify(cwd, phase);
+  process.stdout.write(JSON.stringify(result));
+}
+
+module.exports = { verify };

package/scripts/signal-capture.js

@@ -97,7 +97,7 @@ function withBufferLock(fn) {
       acquired = true;
       break;
     } catch (e) {
-      if (e.code !== 'EEXIST') throw e;
+      if (e.code !== 'EEXIST') return false; // non-EEXIST errors (EMFILE, EACCES) → skip lock gracefully
       try {
         const age = Date.now() - fs.statSync(LOCK_FILE).mtimeMs;
         if (age > LOCK_STALE_MS) {
@@ -159,12 +159,12 @@ process.stdin.on('end', () => {
     // Skip empty or very short messages
     // Chinese chars carry more info per char, so use weighted length
     const weightedLen = [...prompt].reduce((sum, ch) => sum + (ch.charCodeAt(0) > 0x2e80 ? 3 : 1), 0);
-    if (weightedLen < 15) {
+    if (!isMeta && weightedLen < 15) {
       process.exit(0);
     }
 
     // Hard cap to prevent giant prompt pastes from poisoning distill budget.
-    if (prompt.length > ABSOLUTE_MAX_CAPTURE_CHARS) {
+    if (!isMeta && prompt.length > ABSOLUTE_MAX_CAPTURE_CHARS) {
       process.exit(0);
     }
     if (prompt.length > MAX_CAPTURE_CHARS) {

package/scripts/skill-evolution.js

@@ -75,6 +75,7 @@ const DEFAULT_POLICY = {
   min_evidence_for_gap: 3,
   max_updates_per_analysis: 3,
   max_gaps_per_analysis: 2,
+  max_signals_per_analysis: 30,  // cap signals sent to Haiku per run
 
   // Workflow discovery
   workflow_discovery_interval: 2,   // every N cold-path cycles
@@ -200,6 +201,7 @@ function sanitizePolicy(input) {
     min_evidence_for_gap: clampInt(merged.min_evidence_for_gap, DEFAULT_POLICY.min_evidence_for_gap, 1, 20),
     max_updates_per_analysis: clampInt(merged.max_updates_per_analysis, DEFAULT_POLICY.max_updates_per_analysis, 1, 20),
     max_gaps_per_analysis: clampInt(merged.max_gaps_per_analysis, DEFAULT_POLICY.max_gaps_per_analysis, 1, 20),
+    max_signals_per_analysis: clampInt(merged.max_signals_per_analysis, DEFAULT_POLICY.max_signals_per_analysis, 5, 100),
     workflow_discovery_interval: clampInt(merged.workflow_discovery_interval, DEFAULT_POLICY.workflow_discovery_interval, 1, 100),
     min_signals_for_workflow: clampInt(merged.min_signals_for_workflow, DEFAULT_POLICY.min_signals_for_workflow, 1, 100),
     workflow_proposal_threshold: clampInt(merged.workflow_proposal_threshold, DEFAULT_POLICY.workflow_proposal_threshold, 2, 50),
@@ -610,8 +612,15 @@ async function distillSkills() {
   if (!content) return null;
 
   const lines = content.split('\n');
-  const signals = lines.map(l => { try { return JSON.parse(l); } catch { return null; } }).filter(Boolean);
-  if (signals.length < policy.min_signals_for_distill) return null;
+  const allSignals = lines.map(l => { try { return JSON.parse(l); } catch { return null; } }).filter(Boolean);
+  if (allSignals.length < policy.min_signals_for_distill) return null;
+
+  // Cap signals sent to Haiku to avoid prompt bloat / timeout.
+  // Keep most recent signals (higher relevance); overflow is still cleared.
+  const maxSignals = policy.max_signals_per_analysis || 30;
+  const signals = allSignals.length > maxSignals
+    ? allSignals.slice(-maxSignals)
+    : allSignals;
 
   // Get installed skills list
   const installedSkills = listInstalledSkills();

package/skills/agent-browser/SKILL.md

@@ -0,0 +1,153 @@
+---
+name: agent-browser
+description: AI Agent 专属无头浏览器 CLI（省 token 80%）。触发：网页自动化、打开网页、点击填表、截图抓快照、agent-browser。
+version: "1.0.0"
+tool: agent-browser
+install: npm install -g agent-browser && agent-browser install
+---
+
+# agent-browser — AI Agent 专属浏览器
+
+## 核心优势
+
+| 特性 | agent-browser | MCP playwright |
+|------|--------------|----------------|
+| 快照 token 消耗 | ~80% 更少 | 较多 |
+| 安装方式 | 全局 CLI | MCP Server |
+| 调用方式 | Bash 工具 | 专用 MCP 工具 |
+| 底层引擎 | Playwright | Playwright |
+| 状态持久化 | session 文件 | MCP 进程生命周期 |
+| Electron 桌面 App | ✅ 支持 | ❌ 不支持 |
+
+## 安装验证
+
+```bash
+agent-browser --version   # 应输出版本号（当前 0.15.2）
+```
+
+若未安装：
+```bash
+npm install -g agent-browser
+agent-browser install      # 安装 Chromium
+```
+
+## 核心工作流
+
+### 1. 打开页面 + 抓快照（核心循环）
+
+```bash
+agent-browser open <url>
+agent-browser snapshot          # 输出带 @ref 的可交互元素树
+```
+
+快照格式示例：
+```
+- heading "Example Domain" [ref=e1] [level=1]
+- link "Learn more" [ref=e2]
+```
+
+AI 直接用 `@e1`、`@e2` 引用元素，无需 CSS/XPath 选择器。
+
+### 2. 常用操作命令
+
+```bash
+# 导航
+agent-browser open <url>
+agent-browser back / forward / reload
+
+# 交互
+agent-browser click @e3
+agent-browser fill @e5 "搜索内容"
+agent-browser press Enter
+agent-browser hover @e7
+agent-browser select @e4 "选项值"
+agent-browser upload @e6 /path/to/file
+
+# 信息获取
+agent-browser get text @e2
+agent-browser get title
+agent-browser get url
+agent-browser is visible @e1
+agent-browser is enabled @e2
+
+# 视觉工具
+agent-browser screenshot /tmp/debug.png
+agent-browser screenshot --annotate /tmp/annotated.png   # 带元素标注
+
+# 等待
+agent-browser wait @e1            # 等元素出现
+agent-browser wait 2000           # 等 2 秒
+
+# 执行 JS
+agent-browser eval "document.title"
+```
+
+### 3. 语义查找（不用 @ref）
+
+```bash
+agent-browser find role button "提交"      # 按 ARIA role 找
+agent-browser find text "登录"             # 按文本找
+agent-browser find label "用户名"          # 按 label 找
+agent-browser find placeholder "请输入"    # 按 placeholder 找
+```
+
+### 4. 状态管理
+
+```bash
+# 保存登录状态
+agent-browser set session ~/.agent-sessions/github.json
+
+# 后续会话复用
+agent-browser open github.com --session ~/.agent-sessions/github.json
+```
+
+## 标准 Agent 任务模板
+
+```bash
+# Step 1: 打开目标页面
+agent-browser open <url>
+
+# Step 2: 抓快照，识别元素
+agent-browser snapshot
+# AI 分析快照，找到目标 @ref
+
+# Step 3: 执行操作
+agent-browser click @eN
+agent-browser fill @eM "内容"
+agent-browser press Enter
+
+# Step 4: 等待并验证结果
+agent-browser wait ".success-msg"
+agent-browser get text ".result"
+
+# Step 5: 截图留档（可选）
+agent-browser screenshot /tmp/result.png
+```
+
+## 与 MCP Playwright 协作策略
+
+**用 agent-browser 的场景：**
+- token 预算紧张的长流程任务
+- 需要控制 Electron 桌面 App（Discord、Figma、Notion 等）
+- 简单的网页自动化脚本
+
+**继续用 MCP playwright 的场景：**
+- 已有 playwright MCP 连接且不想切换
+- 需要精细的 Playwright API（network mock、tracing）
+- 现有工作流已稳定
+
+## 常见错误处理
+
+| 错误 | 处理 |
+|------|------|
+| `command not found` | 运行 `npm install -g agent-browser` |
+| 浏览器未安装 | 运行 `agent-browser install` |
+| 元素点击失败 | 先 `snapshot` 确认 @ref，改用 `find` 语义查找 |
+| 页面未加载完 | 在操作前加 `agent-browser wait 1500` 或等待特定元素 |
+
+## 进阶：连接现有浏览器
+
+```bash
+# 连接 CDP 调试端口（可复用已登录的 Chrome）
+agent-browser connect 9222
+```

package/skills/agent-reach/SKILL.md

@@ -0,0 +1,66 @@
+---
+name: agent-reach
+description: 免费多平台信息采集CLI（Twitter/YouTube/B站/小红书/抖音/Reddit/LinkedIn/GitHub/RSS/网页）。触发：agent reach、爬取、读网页、搜推特、搜B站。**WebFetch 失败时优先用此 skill**：凡遇到 JS 渲染页面（Twitter/X、Instagram、LinkedIn 等）或 WebFetch 返回空内容/JavaScript错误，无需重试 WebFetch，直接调用此 skill。
+---
+
+# Agent Reach
+
+免费开源的多平台信息采集 CLI，安装后直接调用上游工具。
+
+## 首次使用 / 诊断
+
+```bash
+agent-reach doctor        # 查看各平台状态
+agent-reach install --env=auto  # 自动安装依赖
+```
+
+## 详细命令参考
+
+完整的平台命令和配置指南在上游 SKILL.md：
+```bash
+cat /tmp/Agent-Reach/agent_reach/skill/SKILL.md
+```
+
+如果 /tmp 已清理，重新获取：
+```bash
+pip show agent-reach | grep Location
+# 然后读取 <location>/agent_reach/skill/SKILL.md
+```
+
+或直接查看：https://github.com/Panniantong/Agent-Reach/blob/main/agent_reach/skill/SKILL.md
+
+## 快速速查
+
+| 平台 | 工具 | 示例 |
+|------|------|------|
+| 任意网页 | Jina | `curl -s "https://r.jina.ai/URL"` |
+| Twitter | xreach | `xreach search "query" --json -n 10` |
+| YouTube | yt-dlp | `yt-dlp --dump-json "URL"` |
+| B站 | yt-dlp | `yt-dlp --dump-json "bilibili URL"` |
+| Reddit | curl | `curl -s "https://reddit.com/r/xxx/hot.json?limit=10"` |
+| GitHub | gh | `gh search repos "query" --sort stars` |
+| 小红书 | mcporter | `mcporter call 'xiaohongshu.search_feeds(...)'` |
+| 抖音 | mcporter | `mcporter call 'douyin.parse_douyin_video_info(...)'` |
+| LinkedIn | mcporter | `mcporter call 'linkedin.get_person_profile(...)'` |
+| RSS | feedparser | Python `feedparser.parse(url)` |
+
+## Cookie 配置（需登录的平台）
+
+```bash
+agent-reach configure --from-browser chrome    # 自动提取
+agent-reach configure twitter-cookies "auth_token=xxx; ct0=yyy"  # 手动
+```
+
+> 提醒用户使用专用小号，Cookie 登录有封号风险。
+
+
+<!-- METAME-EVOLUTION:START -->
+
+## User-Learned Best Practices & Constraints
+
+> **Auto-Generated Section**: Maintained by skill-evolution-manager. Do not edit manually.
+
+### Known Fixes & Workarounds
+- Repeatedly failing on Twitter/X links (signals #7, #15, #30 all tool_failure); improve X.com social media parsing robustness or add fallback extraction method for social platforms
+
+<!-- METAME-EVOLUTION:END -->

package/skills/agent-reach/evolution.json

@@ -0,0 +1,13 @@
+{
+  "last_updated": "2026-03-17T13:41:59.591Z",
+  "fixes": [
+    "Repeatedly failing on Twitter/X links (signals #7, #15, #30 all tool_failure); improve X.com social media parsing robustness or add fallback extraction method for social platforms"
+  ],
+  "insights_stats": {
+    "Repeatedly failing on Twitter/X links (signals #7, #15, #30 all tool_failure); improve X.com social media parsing robustness or add fallback extraction method for social platforms": {
+      "success_count": 3,
+      "fail_count": 0,
+      "last_applied_at": "2026-03-20T00:11:44.684Z"
+    }
+  }
+}

package/skills/deep-research/SKILL.md

@@ -0,0 +1,77 @@
+---
+name: deep-research
+description: AI深度研究助手。对任何主题进行迭代式深度研究，结合搜索引擎、网页抓取和大语言模型生成详细研究报告。触发词：深度研究、调研、research。
+version: 2.0.0
+created_at: 2026-01-29
+---
+
+# Deep Research 深度研究
+
+基于 [dzhng/deep-research](https://github.com/dzhng/deep-research) 思路，用 Claude Code 原生能力实现。
+
+## 研究流程
+
+收到研究请求后，按以下步骤执行：
+
+### 1. 分解查询（广度）
+根据主题生成 3-5 个不同角度的搜索查询：
+- 基础概念/定义
+- 最新进展/动态
+- 技术实现/原理
+- 应用场景/案例
+- 对比/竞品
+
+### 2. 迭代搜索（深度）
+对每个查询：
+1. `WebSearch` 获取搜索结果
+2. `WebFetch` 抓取 2-3 个高质量页面详情
+3. 提取关键信息，发现新问题
+4. 对新问题继续搜索（迭代 1-2 轮）
+
+### 2.5 NotebookLM 补充查询（可选）
+
+如果用户的 NotebookLM 中有与主题相关的笔记本，可作为**私有知识源**补充公开搜索的盲区：
+
+```bash
+# 搜索是否有相关笔记本
+python ~/.claude/skills/notebooklm-skill/scripts/run.py notebook_manager.py search --query "关键词"
+
+# 查询笔记本获取有引用的回答
+python ~/.claude/skills/notebooklm-skill/scripts/run.py ask_question.py --question "具体问题" --notebook-id ID
+```
+
+**使用时机**：
+- 话题涉及用户个人积累的领域知识（已上传到 NotebookLM）
+- 公开搜索结果不够深入，需要私有资料补充
+- 用户主动要求"查一下我的 NotebookLM"
+
+**不要**每次调研都自动查 NotebookLM，只在相关时使用。
+
+### 3. 综合报告
+整合所有信息，输出结构化报告：
+
+```markdown
+# [主题] 深度研究报告
+
+## 摘要
+[3-5 句核心发现]
+
+## 背景与定义
+## 技术原理/核心机制
+## 最新进展
+## 应用场景
+## 竞品对比（如适用）
+## 未来展望
+## 参考来源
+```
+
+## 使用示例
+
+用户说：
+- "深度研究一下 AI Agent"
+- "调研 MCP 协议的最新进展"
+- "research Claude Code 的竞品"
+
+## 输出位置
+
+报告写入：`/tmp/research_report.md`