Mingbird v1.9.1

🇨🇳 中文

🌍 六个通用技能:不止写代码

• 此前内置的 11 个技能全是编码向(代码审查、重构、TDD……),日常电脑活没有对应手册。v1.9.1 加入六个通用技能:整理文件夹(归类/批量重命名/查重)、联网调研并写成报告(多查询检索+交叉验证)、长文速读(摘要+要点+行动项)、Word 文档生成与读取、Excel/CSV 表格处理、图片批量处理。技能仍是渐进式加载:模型平时只看到名字和一行描述,用到才载全文——出厂 prefill 一字未变。

🎤 语音输入:中英双语,开箱即用

• 旧版只内置中文模型,英文语音要自己在 Ollama 里拉 whisper。v1.9.1 把英文流式模型(en-20M)也打进安装包,麦克风旁新增 中/EN 切换,选择会记住。转写仍在纯 CPU 上本地完成,录音不出机器。

🐍 新工具 run_python:用户机器不用装 Python

• 办公/图片技能的执行底座:agent 生成的 Python 代码交给程序内置的解释器运行,pandas/numpy/PIL 随包可用——用户的电脑不需要装 Python,也不需要装 Office。安全模型与 run_bash 同款:危险命令硬拦、破坏性操作分级、工作目录沙箱、300 秒超时按进程树终止。

🧩 技能与 MCP 看得见了

• 技能窗口现在列出技能目录、自装方法(自己的 .md 放进 ~/.ollama_agent/skills/ 即成技能)、以及已配置的 MCP 服务器;skills/README.md 写明格式与查找顺序。

📦 安装包反而更小

• 内置语音模型改为只带实际加载的 int8 权重(此前打包里躺着一半没人读的 float32):多塞了一个英文模型之后,Windows 安装包反而从 187.4MB 降到 174.5MB。

472 项测试全绿(435 单测 + 26 集成 + 11 项 v1.9.1 新覆盖:run_python 分发与安全门、双语模型解析、int8 打包纪律、通用技能在位)。

📦 构建说明

Windows 两包从 v1.9.1 tag 源码构建,并在冻结产物上实测:依赖自检全 OK、--run-py 通道实跑通过、GUI 启动存活、stt 双模型仅 int8、17 个技能文件在位;Linux/macOS 三包由 CI 构建并含冒烟步骤;五个资产的 SHA-256 已与本地构建逐字节核对一致。从 v1.8.x 升级直接覆盖安装。完整变更见 CHANGELOG_zh.md。

🇬🇧 English

🌍 Six universal skills: beyond coding

• The 11 built-in skills were all coding-oriented (code review, refactoring, TDD...); everyday computer work had no playbook. v1.9.1 adds six universal skills: file organizing (sort / batch rename / dedup), web research with a written report (multi-query, cross-checked), long-document digest (summary + key points + action items), Word .docx generation and reading, Excel/CSV processing, and image batch processing. Skills stay progressive: the model only sees names and one-line descriptions until one is actually loaded — the factory prefill is unchanged byte for byte.

🎤 Voice input: Chinese & English, out of the box

• Previously only a Chinese model was bundled; English speech required pulling whisper into Ollama yourself. v1.9.1 bundles an English streaming model (en-20M) too, with a 中/EN toggle next to the mic (remembered). Transcription still runs locally on pure CPU; your audio never leaves the machine.

🐍 New run_python tool: no Python install needed on the user machine

• The execution backbone for the office/image skills: Python code the agent writes runs on the interpreter bundled inside the app, with pandas/numpy/PIL available — the user's computer needs neither Python nor Office. Same safety model as run_bash: danger patterns blocked outright, destructive operations classified, workspace sandbox, 300 s timeout with process-tree kill.

🧩 Skills & MCP made visible

• The Skills window now lists skill directories, how to install your own (drop a .md into ~/.ollama_agent/skills/), and the configured MCP servers; skills/README.md documents the format and lookup order.

📦 The installer actually got smaller

• Bundled voice models now ship only the int8 weights actually loaded (half the previous payload was float32 nobody read): after adding a second model, the Windows installer still shrank from 187.4 MB to 174.5 MB.

472 tests green (435 unit + 26 integration + 11 new v1.9.1 cases: run_python dispatch and safety gates, bilingual model resolution, int8-only packaging rule, universal skills presence).

📦 Build note

The two Windows packages were built from the v1.9.1 tag source and verified on the frozen binary: dependency selftest OK, the --run-py path exercised for real, GUI launch stays alive, both stt models int8-only, all 17 skill files present. The Linux/macOS packages are built by CI with a smoke step. All five asset SHA-256 digests were verified byte-for-byte against the local builds. Upgrading from v1.8.x is a plain overwrite install. Full changelog in CHANGELOG.md.

---

Install / 安装

• Windows: download Mingbird-v1.9.1-CN-Setup.exe (中文界面) or Mingbird-v1.9.1-EN-Setup.exe (English UI).
• Linux/macOS: experimental tarballs (mingbird-v1.9.1-linux-x64.tar.gz / -macos-arm64.tar.gz / -macos-x64.tar.gz), see install-unix.sh inside.

Mingbird v1.8.2

🇨🇳 中文

🔒 会话隔离(真实事故后的加固)

• 旧"清空对话"只清屏幕,下一条消息仍带着旧会话的完整历史继续——已废除,换为主操作区的 ＋ 新对话 按钮:会话、上下文、工作目录(回到默认)、计划面板全部复位。新对话不再继承旧对话的上下文与目录。
• 新对话首条消息强制清理工作目录遗留的状态文件;后续追问按"续跑"勾选决定(此前勾选被忽略)。
• 断网时要求联网搜索,任务开始即在会话里明示,不再让模型悄悄翻本地文件。

🖱️ 界面修复

• 流式重写:token 进缓冲、150ms 批量上屏——长思考不再长时间卡死界面。
• 会话区支持鼠标选中与复制(Ctrl+C)。

☁️ 云端模型选择

• 🔒 开关旁新增云端模型框:联网且已配置才可选;与本地模型框互斥(选一个清空另一个)。断网时禁用,云端通路架构级断死。? 给配置指引,配置即 config.json 的 cloud 段(README 有示例)。

461 项测试全绿(435 单测 + 26 集成)。

📦 构建说明(2026-09-23)

五个平台的安装包已于 2026-09-23 用同一份源码树重新构建并替换(CLI 传 --help 或无参数时改为打印用法,不再抛异常堆栈);版本号不变。Windows 两包已在冻结产物上实测(启动存活 + 全局错误日志零新增)并通过静默安装→启动→卸载验证,Linux/macOS 三包由 CI 构建含冒烟步骤。完整变更见 CHANGELOG_zh.md。

🇬🇧 English

🔒 Conversation isolation (hardening after a real incident)

• The old "Clear chat" only wiped the display while the next message silently continued the previous session's full history. Replaced by a prominent ＋ New Chat button that resets session, context, working directory (back to default) and the todo panel — a new conversation can no longer inherit an old one's context or folders.
• The first message of a new conversation force-clears stale state files; follow-ups honor the "resume" checkbox (previously ignored).
• Asking for web search while offline now produces an explicit transcript note at task start.

🖱️ UI fixes

• Streaming rewrite: tokens are buffered and rendered in 150 ms batches — long thinking no longer freezes the UI.
• The transcript is mouse-selectable and copyable (Ctrl+C).

☁️ Cloud model picker

• A cloud-model box next to the 🔒 toggle: selectable only when online and configured; mutually exclusive with the local-model box (one model per conversation). Offline disables it and the cloud path stays architecturally dead. ? shows guidance; config is a cloud section in config.json (example in README).

461 tests green (435 unit + 26 integration).

📦 Build note (2026-09-23)

All five platform packages were rebuilt and replaced on 2026-09-23 from the same source tree (CLI --help / no-arg now prints usage instead of a traceback); version number unchanged. The two Windows packages were verified on the frozen binary (stays alive, global error log clean) and through a silent install → launch → uninstall cycle; the Linux/macOS packages are built by CI, smoke step included. Full changelog in CHANGELOG.md.

---

Install / 安装

• Windows: download Mingbird-v1.8.2-CN-Setup.exe (中文界面) or Mingbird-v1.8.2-EN-Setup.exe (English UI).
• Linux/macOS: experimental tarballs (mingbird-v1.8.2-linux-x64.tar.gz / -macos-arm64.tar.gz / -macos-x64.tar.gz), see install-unix.sh inside.

Mingbird v1.8.1

🇨🇳 中文

🚑 修复:v1.8.0 桌面 GUI 启动即崩(v1.8.0 发行已撤回,功能说明并入本页)

• v1.8.0 全部发行包(Windows 安装包与 Linux/macOS tarball)构建自受影响源码:插入离线开关时新方法被误写进 _build_chat_page 内部,输入框/会话区/目录栏从未被创建,桌面 GUI 一启动即崩溃。请使用 v1.8.1 或更新版本。
• 修复已端到端实证:真实启动安装后的可执行文件(存活、全局错误日志零新增),459 项测试全绿。
• 顺带修复 CI:v1.8.0 的 macOS/Linux 包构建因 macos-13 runner 退役卡在队列 14 小时;本版起 x64 macOS 改用 macos-15-intel,Linux/macOS 包由 CI 真机构建并带"进程存活"冒烟。

🔒 一键断网/联网(offline mode)——把隐私做成架构承诺(v1.8.0 引入)

• 工具栏一键切换(🌐 联网 / 🔒 断网):断网=仅本地 Ollama 模型;联网内置工具(web_search / web_fetch 等)不装配进 prefill(模型看不到=不会调用);URL 型 MCP 服务器整个跳过;云端 provider 强制禁用。零出站可用 netstat 自行验证。
• CLI/环境变量:AGENT_OFFLINE=1 覆盖;offline_mode 持久化于 config.json。

☁️ 云端 provider(OpenAI 兼容端点,可选;v1.8.0 引入)

• config.json cloud 段:启用且非断网时,推理走 OpenAI 兼容 /chat/completions(本地优先不变,一键回退断网)。api_key 仅存本地配置。

🇬🇧 English

🚑 Fix: v1.8.0 desktop GUI crashed on launch (v1.8.0 withdrawn, its feature notes merged into this page)

• All v1.8.0 packages were built from affected source: the offline-toggle edit inserted a new method inside _build_chat_page, so the input box / transcript / directory bar were never constructed — the desktop GUI crashed immediately on launch. Use v1.8.1 or newer.
• The fix is verified end to end: the installed executable is actually launched (stays alive, zero new entries in the global error log), 459 tests green.
• CI also fixed: v1.8.0's macOS/Linux builds were stuck in queue for 14 h because the macos-13 runner was retired; x64 macOS now uses macos-15-intel, and Linux/macOS packages are built by CI with a binary-stays-alive smoke test.

🔒 One-click offline/online — privacy as an architectural commitment (introduced in v1.8.0)

• Toolbar toggle (🌐 online / 🔒 offline): offline = local Ollama models only; built-in web tools are not assembled into the prefill (the model cannot call what it cannot see); URL-based MCP servers are skipped entirely; the cloud provider is force-disabled. Zero outbound, verifiable with netstat.
• CLI/env: AGENT_OFFLINE=1 override; offline_mode persisted in config.json.

☁️ Cloud provider (OpenAI-compatible endpoint, optional; introduced in v1.8.0)

• config.json cloud section: when enabled and not offline, inference goes through the OpenAI-compatible /chat/completions. The api_key is stored only in the local config.

459 tests. Full changelog in CHANGELOG.md.

---

Install / 安装

• Windows: download Mingbird-v1.8.1-CN-Setup.exe (中文界面) or Mingbird-v1.8.1-EN-Setup.exe (English UI).
• Linux/macOS: experimental tarballs (mingbird-v1.8.1-linux-x64.tar.gz / -macos-arm64.tar.gz / -macos-x64.tar.gz), see install-unix.sh inside.

Mingbird v1.7.0

🇨🇳 中文

🎛️ 采样三态:移除隐藏默认,回归 Ollama 默认

鸣鸟此前默认强制 temperature=0、默认关闭思考——这是基准协议的残留,不是产品应有的行为。本版彻底修正:

• 温度三态:未设置 → 请求完全不带 temperature(由模型 manifest/Ollama 默认决定);显式设置(含 0)→ 按设置生效。
• 思考三态:未设置 → 不带 think 字段(Ollama 出厂默认);显式开/关 → 顶层 think 字段生效。
• 空思考轮防护:思考烧光生成预算导致回答为空时,思考如实展示,空回答进入防护梯(提醒→复位→优雅退出),不死循环。
• 设置面贯通:桌面 GUI 与 WebUI 的温度/思考设置端到端生效;修复旧版"关闭思考"勾选框接线反转且不持久化的存量 bug。
• 451 项测试全绿;新增 12 项采样三态载荷断言。

🇬🇧 English

🎛️ Three-state sampling: hidden defaults removed, back to Ollama defaults

Mingbird used to silently force temperature=0 and thinking-off — leftovers from the benchmark protocol, not proper product behavior. This release fixes both:

• Temperature, three states: unset → the request carries no temperature field (model manifest / Ollama default applies); explicitly set (including 0) → honored as-is.
• Thinking, three states: unset → no think field (Ollama factory default); explicitly on/off → top-level think field honored.
• Empty-thinking-turn guard: when thinking burns the generation budget and content comes back empty, thinking is surfaced honestly and the empty reply enters the protection ladder (nudge → reset → graceful exit) — no dead loops.
• Settings wired end-to-end: temperature & thinking settings in both desktop GUI and WebUI now actually take effect; fixes a legacy bug where the "disable thinking" checkbox was wired backwards and never persisted.
• 451 tests green; 12 new payload-level assertions for the three-state semantics.

---

Install / 安装

• Windows: download Mingbird-v1.7.0-CN-Setup.exe (中文) or Mingbird-v1.7.0-EN-Setup.exe (English)
• Linux (experimental): mingbird-v1.7.0-linux-x64.tar.gz → extract, sh install-unix.sh
• macOS Apple Silicon (experimental): mingbird-v1.7.0-macos-arm64.tar.gz → extract, sh install-unix.sh
• Requires Ollama installed on the same machine. Local-first: no cloud account needed.

Full changangelog / 完整变更: CHANGELOG.md

Mingbird v1.6.0

🇨🇳 中文

🛡️ 通用安全垫:小模型破坏性行为防护(四层)

面向 2-9B 本地小模型的实际失败模式,构建四层防护:

• 覆盖自毁防护:用短内容覆盖既有大文件被拒绝,需显式 replace=true;追加用 append_file,局部改用 edit_file。
• 可回滚删除:文件删除不再直接抹除,移入 .mingbird_trash/ 可恢复。
• 命令分级:卸载/环境变异/系统级破坏命令分级拒绝或询问;递归删除仅限工作目录内。
• **~ 路径拦截**:文件工具拒绝 ~ 开头路径,强制工作目录相对路径。

🧭 其他

• 基准消融框架(机制级消融独立目录,续跑互不污染)。

🇬🇧 English

🛡️ General safety net: four layers against destructive small-model behavior

Built for the real failure modes of local 2-9B models:

• Overwrite self-harm guard: overwriting an existing large file with much shorter content is refused unless replace=true; use append_file / edit_file for partial edits.
• Rollback-able deletion: deleted files go to .mingbird_trash/ instead of being wiped.
• Tiered command defense: package removal / environment mutation / system-destructive commands are refused or gated; recursive deletes restricted to the working directory.
• ~ path interception: file tools reject ~-prefixed paths in favor of working-directory-relative ones.

🧭 Misc

• Benchmark ablation framework (per-mechanism results in isolated directories).

---

Install / 安装

• Windows: download Mingbird-v1.6.0-CN-Setup.exe (中文) or Mingbird-v1.6.0-EN-Setup.exe (English)
• Linux (experimental): mingbird-v1.6.0-linux-x64.tar.gz → extract, sh install-unix.sh
• macOS Apple Silicon (experimental): mingbird-v1.6.0-macos-arm64.tar.gz → extract, sh install-unix.sh
• Requires Ollama installed on the same machine. Local-first: no cloud account needed.

Full changangelog / 完整变更: CHANGELOG.md

Mingbird v1.5.0

🇨🇳 中文

🛡️ harness 健壮性三连(72 格重跑实证驱动)

• read_file 爬行守卫 v2(字节预算制):预算耗尽后仍每次返回所请求区间,永不硬拒——v1 的次数硬拒在真实批次中把小模型逼成撞墙循环。
• 输出上限 2048 → 8192:单次工具调用可容纳完整中型代码文件;截断死亡签名会给"骨架+追加"的分块反馈。
• 空轮防护梯修复:纠正→硬复位→优雅退出梯队全部真实可达;未闭合 残骸正确计为空轮。

🧭 其他

• run_matrix 格级断点续跑;富版发布文档(中英 README + 基准证据包)入库。

🇬🇧 English

🛡️ Harness robustness trio (driven by a 72-cell re-run)

• read_file crawl guard v2 (byte-budget): once the budget is spent, each read still returns the requested window — never a hard refusal (v1's per-count refusal pushed small models into wall-banging loops in real batches).
• Output cap 2048 → 8192: one tool call can carry a complete mid-size code file; the truncation death-signature now gets a skeleton+append chunking hint.
• Empty-turn ladder fixed: nudge → hard reset → graceful exit all reachable; unclosed <think> debris counted as an empty turn.

🧭 Misc

• Cell-level resume in run_matrix; rich release docs (bilingual README + benchmark evidence pack).

---

Install / 安装

• Windows: download Mingbird-v1.5.0-CN-Setup.exe (中文) or Mingbird-v1.5.0-EN-Setup.exe (English)
• Linux (experimental): mingbird-v1.5.0-linux-x64.tar.gz → extract, sh install-unix.sh
• macOS Apple Silicon (experimental): mingbird-v1.5.0-macos-arm64.tar.gz → extract, sh install-unix.sh
• Requires Ollama installed on the same machine. Local-first: no cloud account needed.

Full changangelog / 完整变更: CHANGELOG.md

Mingbird v1.4.0

🇨🇳 中文

🎀 品牌定名:鸣鸟 / Mingbird

• 英文名、GUI 标题、安装器、文档统一以 Mingbird 发布;四注册表零冲突。
• 用户数据零迁移:配置目录不变;新增 MINGBIRD_HOME/MINGBIRD_DEPTH 环境变量。

交付自查门禁(finish 回读)

• 任务收尾时动态回注任务原文,要求模型对照题面自查格式/单位/子项后重新 finish(防长任务末尾语义漂移)。

🇬🇧 English

🎀 Branding: Mingbird

• Unified branding (GUI, installer, docs); zero conflicts across the four registries.
• Zero user-data migration; new MINGBIRD_HOME / MINGBIRD_DEPTH env vars.

Delivery self-check gate (finish read-back)

• On finish, the original task text is re-injected once so the model re-verifies format/units/sub-items before finishing (guards against late-task semantic drift).

---

Install / 安装

• Windows: download Mingbird-v1.4.0-CN-Setup.exe (中文) or Mingbird-v1.4.0-EN-Setup.exe (English)
• Linux (experimental): mingbird-v1.4.0-linux-x64.tar.gz → extract, sh install-unix.sh
• macOS Apple Silicon (experimental): mingbird-v1.4.0-macos-arm64.tar.gz → extract, sh install-unix.sh
• Requires Ollama installed on the same machine. Local-first: no cloud account needed.

Full changangelog / 完整变更: CHANGELOG.md