Files
metona-ollama-desktop/README.md
T
thzxx 3fb293c618
CI / verify (push) Successful in 1m53s
v0.17.2: 上下文一致性根治 + 消息差量持久化 + 压缩摘要修复 + 安全层测试补课
- P0: 发送路径用户消息重复注入根治(history-builder 纯模块 + 单测,连带修复 maxCount 截断保留最旧消息缺陷);/undo、/retry 消息删除差量落库(新增 db:getMessageIds/deleteMessages 四层通道);/compress 摘要 role:user + 可折叠卡片渲染 + 旧 system 行读取归一化
- P1: memory search 默认 limit=8;工具缓存键/去重改稳定序列化;run_command 超时联动主进程杀子进程;记忆访问统计写回纳入写入锁;搜索自动抓取单页限幅 8k;Token 趋势采样移出 calculateContextStats 并记录裁剪后值;空白 assistant 幽灵消息跳过入库(新迭代/中止两路径)
- P2: 新增 tool-security(18 用例,平台自适应)与 history-builder(11 用例)测试;帮助/README/DEVELOPMENT 文案与代码事实对齐;vendor 失效 sourcemap 与 .npmrc 弃用配置清理;备份导入携带 attachments 修复
- 版本号升级 0.17.2(5 文件白名单);typecheck 零错误 / 301 测试通过 / 构建通过
2026-09-08 16:32:28 +08:00

26 KiB
Raw Blame History

Metona Ollama Desktop

Metona Ollama Desktop

🤖 本地 AI 桌面客户端 · Local AI Desktop Client

中文 English

version electron typescript license platform

所有 AI 推理在本地完成,数据不离开本机。
All AI inference runs locally — data never leaves your machine.


🇨🇳 中文

特性

功能 说明
🤖 ReAct Agent Loop 始终开启的唯一对话模式。8 状态机(INIT→THINKING→PARSING→EXECUTING→OBSERVING→REFLECTING→COMPRESSING→TERMINATED),最大 85 轮(可配置),智能重试(永久错误立即返回、瞬态错误指数退避),工具去重,智能路径依赖串行化,看门狗超时保护
🛡️ 提示词加固 参考数据标记(<<<REFERENCE_DATA_START>>>+ 工具结果仅为数据非指令的安全规则注入
📋 Plan Mode 开关切换,AI 首先生成执行计划(Markdown 渲染确认弹窗),批准后按步骤追踪执行,plan_track 工具标记完成状态,支持断点续传(中止后可恢复未完成计划)
🔧 33 个内置工具 文件系统(14个,含 diff) · 命令执行 · 联网搜索 · 浏览器控制(9个) · Git · 记忆 · 会话 · 子代理 · 系统工具 · Plan Mode 追踪
🧠 智能记忆系统 三类记忆(fact / preference / rule),存储于工作空间 MEMORY.md 文件,受路径保护仅 memory 工具可访问,写入前安全扫描,容量 500 条,对话结束自动提取,严格格式校验,访问统计随条目持久化(TTL 衰减保护)
📋 自定义文件 SOUL.md(人格,不可压缩)+ AGENT.md(行为准则,内置 fallback)+ USER.md(用户画像,仅工作空间读取,不存在则不注入)
🌐 MCP 协议扩展 JSON-RPC 2.0 over stdio,动态工具发现,Shadowing 防护
🔍 联网搜索(双模式) SearXNG 元搜索引擎 JSON API70+引擎聚合)/ 四引擎 HTML 解析(Bing+百度+搜狗+360),双模式可切换;web_fetch 支持反爬+UA切换+浏览器回退
🌏 浏览器控制 打开网页 · 截图 · 执行 JS · 提取内容 · 点击 · 输入 · 滚动 · 关闭
🖥️ 工作空间面板 终端(增量流式输出)+ 文件浏览器,命令安全检查
🔢 上下文长度手动控制 设置面板下拉选择(8K–1M 共 8 档),默认 128K,模型栏显示当前配置值,下拉框中显示每个模型自身的上下文长度
🗜️ 智能上下文管理 滑动窗口 + Token 自动校准 + 消息重要性评分 + LLM 结构化 JSON 压缩,智能触发(120 条增量压缩 + 300 条硬上限)
⏱️ 智能超时保护 Agent Loop 看门狗(可配,默认 30min)+ 流式总超时(可配,默认 300s)+ 工具 HTTP/MCP 超时可配
🪝 Hook 系统 4 阶段生命周期钩子(pre_tool / post_tool / post_iteration / pre_completion),内置安全检查(SecurityCheck)与文件写入去重(FileWriteDedup
👥 子代理委派 spawn_task 工具,独立上下文 + 超时保护;权限分级(readonly / limited_write / full_write)且只降不升(AI 请求的权限封顶于用户设置),写类工具与主 Agent 共用确认管线,路径沙箱覆盖全部文件类工具
📈 Token 仪表盘 全局 + 会话统计,消耗趋势柱状图,2 秒刷新,输入/输出分色
📋 系统提示词卡片 每条 AI 回复顶部折叠卡片,点击查看实际发送给模型的完整上下文
🎨 暖色调 UI 奶白 #FAF7F2 + 珊瑚橙 #E8734A,长时间使用不疲劳,流式渲染优化
🔔 系统托盘 原生托盘集成
🔐 备份加密 会话备份文件 AES-256-GCM 编码(密钥内置于应用内,用于防误读与基本混淆,不构成针对攻击者的安全边界,敏感数据请自行额外加密)
🧪 可观测性 执行轨迹缓冲批量写入 SQLite + Agent Metrics 仪表盘(JSON / Prometheus 文本双格式导出)

🔧 工具清单

📁 文件系统(14 个)
工具 功能
read_file 读取文件(文本/binary,2000行默认)
write_file 写入/追加文件(支持base64二进制)
list_directory 列出目录内容
search_files 搜索文件(正则/通配符)
create_directory 创建目录
delete_file 删除文件/目录(返回大小)
move_file 移动/重命名文件(跨盘自动回退 copy+delete
copy_file 复制文件
edit_file 编辑文件(支持正则替换)
tree 目录树结构(默认5层深度)
download_file 下载文件
read_multiple_files 批量读取(默认不截断,可按文件限制字符数)
compress 压缩文件/目录
diff 文件差异比对(file_vs_file / file_vs_content / file_vs_git_head 三种模式,unified diff 输出)
命令执行(1 个)
工具 功能
run_command 执行 shell 命令,实时流式输出,支持自动/需确认/禁用三种模式
🔍 联网搜索(2 个)
工具 功能
web_search SearXNG JSON API / 四引擎并行,默认 30 条结果
web_fetch 网页内容抓取,自动重试(指数退避),移动端UA切换,SPA页面自动升级到浏览器渲染
🌐 浏览器控制(9 个)
工具 功能
browser_open 打开网页
browser_screenshot 页面截图
browser_evaluate 执行 JavaScript
browser_extract 提取页面内容
browser_click 点击元素
browser_type 输入文本
browser_scroll 页面滚动
browser_wait 等待元素/延时
browser_close 关闭浏览器
📦 Git1 个,17 个子命令)
子命令 功能
init clone add commit push pull 基础操作
diff log status branch checkout 查看与切换
merge stash reset tag remote 高级操作
🧠 记忆管理(1 个)
工具 功能
memory 统一记忆管理(6 个 actionsearch 搜索 / add 添加 / replace 替换 / remove 删除 / remove_batch 批量删除 / read_all 读取全部)。工作空间 MEMORY.md 受路径保护,仅此工具可访问
📊 会话 / 子代理(3 个)
工具 功能
session_list 列出历史会话
session_read 读取会话内容
spawn_task 委派子代理(独立上下文 + 超时保护)
🕐 系统工具(1 个)
工具 功能
calculator 安全数学计算(+ - * / ** % (),递归下降解析器)
📋 Plan Mode1 个,仅 Plan 模式激活)
工具 功能
plan_track Plan 模式执行进度追踪,标记步骤完成,自动统计剩余

🏗️ 架构

用户消息 → 扫描工作空间 SOUL.md(不可压缩层)→ AGENT.md → USER.md(工作空间,不存在则跳过)
              ↓
         记忆检索 (MEMORY.md 关键词搜索) → 上下文注入
              ↓
         Agent Engine (8 状态机 ReAct Loop, ≤85 轮, 智能重试, 路径依赖串行化, 看门狗+分级超时)
              ↓
         提示词加固(参考数据标记 + 工具结果安全规则)
              ↓
         Ollama API (流式响应,num_ctx 用户可配 8K1M 共 8 档)
              ↓
         Tool Registry (33 内置 + MCP 动态 + Plan Mode plan_track)
              ↓
         Hook 系统 (pre/post tool/iteration/completion) → 观察结果 → 反思 → 循环 / 最终回答

🧩 UI 组件(原生 DOM

main.ts (入口)
  ├── header.ts              # 顶部导航栏
  ├── model-bar.ts           # 模型选择栏
  ├── chat-area.ts           # 聊天消息区域
  ├── input-area.ts          # 输入框 + 文件上传 + Plan Mode 开关
  ├── workspace-panel.ts     # 终端 + 文件浏览器 + 工具卡片
  ├── settings-modal.ts      # 设置面板(含看门狗超时配置)
  ├── history-modal.ts       # 会话历史
  ├── memory-modal.ts        # 记忆管理
  ├── tools-modal.ts         # 工具列表
  ├── token-dashboard.ts     # Token 消耗仪表盘
  ├── tool-confirm-modal.ts  # 工具执行确认
  ├── prompt-modal.ts        # 系统提示词查看 + Plan 确认弹窗
  ├── toast.ts               # Toast 通知
  └── lightbox.ts            # 图片灯箱

🗄️ 数据库

SQLite (sql.js WASM 内存库)6 张表;写入采用 防抖批量落盘(写操作合并为 300ms 一次全库快照 + temp 文件原子替换 + 退出强制刷盘),Schema 使用 PRAGMA user_version 顺序迁移管理:

用途
sessions 会话(parent_id 父子关系)
messages 消息(外键级联删除)
tool_calls 工具调用记录
settings 设置(JSON 序列化)
traces ReAct 执行轨迹
tool_audit 工具执行审计日志

🔒 安全机制

层级 措施
📁 文件系统 checkPathAllowed() — 路径黑名单(系统目录 + 用户敏感目录);系统目录为硬红线(工作空间/白名单也不可豁免)
🧾 身份文件保护 MEMORY.md 全工具禁写禁读(仅 memory 工具);SOUL.md / AGENT.md / USER.md 可读不可写(防提示注入改写人格实现持久化劫持,仅用户手动编辑)
命令执行 checkCommandAllowed() — 命令黑名单(含 POSIX + Windows+ 三种模式
🔀 参数注入防护 git 工具用户可控参数禁止 - 开头(防 --upload-pack 类注入);git add 强制 -- 分隔
🌐 SSRF 防护 net-guard — web_fetch / download_file / browser_open 拦截环回/内网/链路本地地址(域名 DNS 解析后校验真实 IP);browser_open 仅允许 http/https 协议
🖥️ 前端渲染 HTML 净化器(白名单标签 + URI 协议检查);toast/HTML 导出使用 textContent/escapeHtml 防 XSS
🔒 Electron webSecurity: true(同源策略开启,Ollama API 通过 webRequest CORS 允许清单精确放行)+ contextIsolation: true + IPC 白名单 + IPC fs 路径验证;Agent 浏览器使用 memory: 内存分区(退出清空 cookie/storage
👥 子代理权限 权限分级(readonly / limited_write / full_write只降不升(AI 请求封顶于用户设置);写类工具与主 Agent 共用确认管线;路径沙箱覆盖全部文件类工具
🔐 备份编码 AES-256-GCM(密钥内置,防误读混淆用途,非安全边界)
🧠 记忆安全 写入前 Prompt Injection / 敏感信息检测,多层级去重过滤
🌐 MCP 安全 Shadowing 防护 + 双下划线分隔防歧义;tools/list 支持分页(nextCursor
🌐 网络安全 web_fetch 流式体积限制(10MB+ 无 content-length 时防 OOM;搜索可达性预检只取响应头(Range: bytes=0-0
进程安全 Windows taskkill 强制终止 + 工作空间路径大小写不敏感校验;工作空间不可设为系统目录

🚀 快速开始

git clone https://gitee.com/thzxx/metona-ollama-desktop.git
cd metona-ollama-desktop
npm config set registry https://registry.npmmirror.com
npm install
npm start

📦 构建 Windows 安装包

# 环境:Ubuntu 24.04 + Node.js v22 + Wine 9.0+
ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/ npm run dist

产出:release/Metona Ollama Setup v0.17.2.exe

🛠️ 常用命令

npm start              # 构建并运行
npm run dev:renderer   # Vite watch(渲染进程)
npm run dev:main       # tsc watch(主进程)
npm run build          # 完整构建
npm run typecheck      # 双 tsconfig 类型检查
npm test               # Vitest 单元测试
npm run dist           # 构建 Windows 安装包

💻 系统要求

  • 🖥️ Windows 10/1164 位)
  • 🤖 Ollama 本地运行中(默认 http://localhost:11434
  • 💾 建议 8GB+ 内存

📄 许可证

MIT


🇬🇧 English

Features

Feature Description
🤖 ReAct Agent Loop Always-on, only chat mode. 8-state machine, up to 85 iterations (configurable), smart retry (permanent errors return immediately, transient errors use exponential backoff), tool dedup, path-aware dependency serialization, watchdog + tiered timeout protection
🛡️ Prompt Hardening Reference data markers (<<<REFERENCE_DATA_START>>>) + tool results are data-not-instructions safety rules
📋 Plan Mode Toggle switch. AI first generates an execution plan (Markdown-rendered confirmation dialog), then tracks step-by-step execution, supports resume (interrupted plans restore on next session)
🔧 33 Built-in Tools File system (14, incl. diff) · Command · Web search · Browser · Git · Memory · Sessions · Sub-agent · System · Plan Mode tracking
🧠 Smart Memory System Three types (fact / preference / rule), stored in workspace MEMORY.md file, path-protected — only the memory tool can access it, pre-write security scan, 500 capacity, auto-extract on conversation end, strict format validation, access stats persisted per entry (TTL decay protection)
📋 Custom Files SOUL.md (persona, never compressed) + AGENT.md (behavior rules, built-in fallback) + USER.md (user profile, workspace only, skipped if absent)
🌐 MCP Protocol Extension JSON-RPC 2.0 over stdio, dynamic tool discovery, Shadowing protection
🔍 Web Search (dual-mode) SearXNG meta-search JSON API (70+ engines) / quad-engine HTML parsing (Bing+Baidu+Sogou+360), switchable; web_fetch with auto-retry+mobile UA+SPA browser fallback
🌏 Browser Control Open pages · Screenshot · JS execution · Content extraction · Click · Type · Scroll · Close
🖥️ Workspace Panel Terminal (incremental streaming) + file browser, command security checks
🔢 Manual Context Length Settings dropdown (8 tiers, 8K1M), default 128K. Model bar shows configured value, dropdown shows per-model context length
🗜️ Smart Context Manager Sliding window + Token auto-calibration + message importance scoring + LLM structured JSON compression, smart triggers (120 msg incremental + 300 msg hard limit)
⏱️ Smart Timeout Protection Agent Loop watchdog (configurable, default 30min) + stream total timeout (configurable, default 300s) + configurable HTTP/MCP timeouts
🪝 Hook System 4-phase lifecycle hooks (pre_tool / post_tool / post_iteration / pre_completion), built-in SecurityCheck + FileWriteDedup
👥 Sub-agent Delegation spawn_task tool, isolated context + timeout protection; permission tiers (readonly / limited_write / full_write) that can only be downgraded (capped by user setting); write tools share the main agent's confirmation pipeline; path sandbox covers all file tools
📈 Token Dashboard Global + session stats, consumption trend bar chart, 2s refresh, input/output color-coded
📋 System Prompt Card Collapsible card atop each AI reply — click to inspect the full context sent to the model
🎨 Warm-tone UI Cream #FAF7F2 + coral #E8734A, fatigue-free for extended use, streaming render optimized
🔔 System Tray Native tray integration
🔐 Backup Encoding AES-256-GCM encoded session backups (key embedded in app — obfuscation against casual reading, not a security boundary)
🧪 Observability Buffered trace persistence to SQLite + Agent Metrics dashboard with JSON / Prometheus text export

🔧 Tool List

📁 File System (14)
Tool Function
read_file Read file (text/binary, 2000 lines default)
write_file Write/append file (base64 binary support)
list_directory List directory contents
search_files Search files (regex/wildcard)
create_directory Create directory
delete_file Delete file/dir (returns size)
move_file Move/rename (cross-device fallback to copy+delete)
copy_file Copy file
edit_file Edit file (regex support)
tree Directory tree (default 5 levels)
download_file Download file
read_multiple_files Batch read (no truncation by default, per-file char limit optional)
compress Compress file/directory
diff Compare files (file_vs_file / file_vs_content / file_vs_git_head, unified diff output)
Command Execution (1)
Tool Function
run_command Execute shell command, real-time streaming, auto/confirm/disabled modes
🔍 Web Search (2)
Tool Function
web_search SearXNG JSON API / quad-engine parallel, 30 results default
web_fetch Fetch web page content, auto-retry with backoff, mobile UA toggle, SPA auto-upgrade to browser rendering
🌐 Browser Control (9)
Tool Function
browser_open Open web page
browser_screenshot Page screenshot
browser_evaluate Execute JavaScript
browser_extract Extract page content
browser_click Click element
browser_type Type text
browser_scroll Scroll page
browser_close Close browser
📦 Git (1 tool, 17 subcommands)
Subcommands Function
init clone add commit push pull Basic operations
diff log status branch checkout View & switch
merge stash reset tag remote Advanced operations
🧠 Memory Management (1)
Tool Function
memory Unified memory management (6 actions: search / add / replace / remove / remove_batch / read_all). Workspace MEMORY.md is path-protected, only this tool can access it
📊 Sessions / Sub-agent (3)
Tool Function
session_list List session history
session_read Read session content
spawn_task Delegate sub-agent (isolated context + timeout)
🕐 System Tools (1)
Tool Function
calculator Safe math evaluation (+ - * / ** % (), recursive descent parser)
📋 Plan Mode (1, only active in Plan Mode)
Tool Function
plan_track Plan Mode execution progress tracking, mark steps complete, auto tally remaining

🏗️ Architecture

User message → workspace SOUL.md (never compressed) → AGENT.md → USER.md (workspace, skipped if absent)
              ↓
         Memory Retrieval (MEMORY.md keyword search) → Context Injection
              ↓
         Agent Engine (8-state ReAct Loop, ≤85 iter, smart retry, path-aware serialization, watchdog+tiered timeout)
              ↓
         Prompt Hardening (reference data markers + tool result safety rules)
              ↓
         Ollama API (Streaming Response, num_ctx user-configurable 8 tiers, 8K1M)
              ↓
         Tool Registry (33 Built-in + MCP Dynamic + Plan Mode plan_track)
              ↓
         Hook System (pre/post tool/iteration/completion) → Observation → Reflection → Loop / Final Answer

🧩 UI Components (Native DOM)

main.ts (Entry)
  ├── header.ts              # Top navigation bar
  ├── model-bar.ts           # Model selector
  ├── chat-area.ts           # Chat message area
  ├── input-area.ts          # Input box + file upload + Plan Mode toggle
  ├── workspace-panel.ts     # Terminal + file browser + tool cards
  ├── settings-modal.ts      # Settings panel (with watchdog timeout config)
  ├── history-modal.ts       # Session history
  ├── memory-modal.ts        # Memory management
  ├── tools-modal.ts         # Tool list
  ├── token-dashboard.ts     # Token consumption dashboard
  ├── tool-confirm-modal.ts  # Tool execution confirmation
  ├── prompt-modal.ts        # System prompt viewer + Plan confirm dialog
  ├── toast.ts               # Toast notifications
  └── lightbox.ts            # Image lightbox

🗄️ Database

SQLite (sql.js WASM in-memory), 6 tables; writes use debounced batched persistence (writes coalesced into one full snapshot every 300ms + atomic temp-file rename + forced flush on quit); schema managed via PRAGMA user_version sequential migrations:

Table Purpose
sessions Sessions (parent_id hierarchy)
messages Messages (cascade delete)
tool_calls Tool call records
settings Settings (JSON serialization)
traces ReAct execution traces
tool_audit Tool execution audit log

🔒 Security

Layer Measure
📁 File System checkPathAllowed() — path blacklist (system + user-sensitive dirs); system dirs are a hard red line (not exempted even by workspace/allowlist)
🧾 Identity File Protection MEMORY.md blocked from all tools (memory tool only); SOUL.md / AGENT.md / USER.md read-only for tools (prevents prompt-injection persona hijacking; user-editable only)
Command Execution checkCommandAllowed() — command blacklist (POSIX + Windows) + three modes
🔀 Argument Injection Guard git user-controllable args must not start with - (blocks --upload-pack style injection); git add forces -- separator
🌐 SSRF Guard net-guard — web_fetch / download_file / browser_open block loopback/private/link-local addresses (DNS-resolved IP checked); browser_open allows http/https only
🖥️ Frontend Rendering HTML sanitizer (whitelist tags + URI protocol check); toast/HTML export use textContent/escapeHtml against XSS
🔒 Electron webSecurity: true (same-origin enforced; Ollama API precisely allowed via webRequest CORS allowlist) + contextIsolation: true + IPC whitelist + IPC fs path validation; agent browser uses memory: in-memory partition (cookies/storage wiped on exit)
👥 Sub-agent Permissions Permission tiers that can only be downgraded (capped by user setting); write tools share the main agent's confirmation pipeline; path sandbox covers all file tools
🔐 Backup Encoding AES-256-GCM (embedded key — obfuscation, not a security boundary)
🧠 Memory Security Pre-write prompt injection / sensitive info detection, multi-layer dedup filtering
🌐 MCP Security Shadowing protection + double-underscore delimiter disambiguation; paginated tools/list (nextCursor)
🌐 Network Security web_fetch streaming size limit (10MB) + OOM prevention without content-length; search reachability precheck fetches headers only (Range: bytes=0-0)
Process Security Windows taskkill forced termination + case-insensitive workspace path validation; workspace cannot be set to a system directory

🚀 Quick Start

git clone https://gitee.com/thzxx/metona-ollama-desktop.git
cd metona-ollama-desktop
npm config set registry https://registry.npmmirror.com
npm install
npm start

📦 Build Windows Installer

# Environment: Ubuntu 24.04 + Node.js v22 + Wine 9.0+
ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/ npm run dist

Output: release/Metona Ollama Setup v0.17.2.exe

🛠️ Common Commands

npm start              # Build and run
npm run dev:renderer   # Vite watch (renderer)
npm run dev:main       # tsc watch (main process)
npm run build          # Full build
npm run typecheck      # Type-check both tsconfigs
npm test               # Vitest unit tests
npm run dist           # Build Windows installer

💻 System Requirements

  • 🖥️ Windows 10/11 (64-bit)
  • 🤖 Ollama running locally (default http://localhost:11434)
  • 💾 8GB+ RAM recommended

📄 License

MIT


Made with ❤️ by thzxx