magpie-releases Updates & Release Notes

Bug Fixes

• Fixed Gemini reasoning responses so they stream progressively instead of leaving the interface empty during thinking.
• Fixed an issue where a 1M-token model could be misidentified as a smaller 272K GPT model.
• Fixed detection of the goose command on PATH so it’s not confused with other Go programs.
• Fixed Qoder upstream errors being incorrectly reported as 502 Bad Gateway by requiring Qoder version 0.2.24 or later.

Improvements

• Optimized model catalog lookups by reusing cached results until the catalog changes.
• Reuse the compressed request decoder between requests to reduce overhead and improve performance.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

错误修复

• 修复了 Gemini 推理响应，使其逐步流式输出，而不是在思考期间让界面保持空白。
• 修复了一个问题：1M-token 模型可能被误识别为较小的 272K GPT 模型。
• 修复了对 PATH 中 goose 命令的检测，使其不会与其他 Go 程序混淆。
• 修复了 Qoder 上游错误被错误报告为 502 Bad Gateway 的问题，要求 Qoder 版本 0.2.24 或更高。

改进

• 优化了模型目录查找，通过复用缓存结果直到目录发生变化。
• 在请求之间复用压缩请求解码器，以减少开销并提高性能。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Features

Added support for portable Codex compaction through Magpie gateway routing.

Ported Codex rollouts now read the same through Magpie as the original compaction item.

Bug Fixes

Fixed Codex threads compacted with Magpie stopping when Magpie was quit or disconnected; they now continue as expected.

Improvements

Improved Codex summary consistency by giving each Magpie compaction item and its associated message a single owner.

Documentation

Added gateway-routing documentation explaining how Magpie’s Codex compaction is made portable.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

功能

新增通过 Magpie 网关路由支持可移植 Codex 压缩。

移植后的 Codex 推出内容现在通过 Magpie 读取时与原始压缩项一致。

错误修复

修复了使用 Magpie 压缩的 Codex 线程在 Magpie 退出或断开连接时停止的问题；现在它们会按预期继续运行。

改进

通过为每个 Magpie 压缩项及其关联消息指定单一所有者，提高了 Codex 摘要一致性。

文档

新增网关路由文档，说明 Magpie 的 Codex 压缩如何实现可移植性。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Bug Fixes

• Fixed routing so requests for agents omitted from the stage are listed instead of flown.
• Fixed session handling so DSH frames are read into a single correctly sized buffer.
• Fixed trace polling to read long lines once at their correct size.

Improvements

• Stopped off-screen pages from building content for each request (#2515, #2513).
• Made GUI tests more reliable by waiting for UI updates and scrolling rows before clicking (#2658).

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

Bug Fixes

• 修复了路由，使未包含在阶段中的代理请求被列出，而不是被直接发送。
• 修复了会话处理，使 DSH 帧被读入一个大小正确的单一缓冲区。
• 修复了跟踪轮询，使其按正确大小一次性读取长行。

Improvements

• 阻止屏幕外页面为每个请求构建内容 (#2515, #2513)。
• 通过等待 UI 更新并在点击前滚动行，使 GUI 测试更加可靠 (#2658)。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Features

• Settings for Codex and Claude Code now appear in each agent’s row on the Agents page instead of under Settings.
• Dialogs now grow out of the row that opened them and fold back into it.
• Added support for Ollama’s /api/chat and /api/generate endpoints for Zed’s Ollama provider.
• ClinePass now pins glm-5.3-flash through OpenRouter’s provider.only setting.

Bug Fixes

• Workbuddy users now get plugin 0.1.27, which fixes handling for models that only advertise supportsReasoning and prevents DeepSeek Harness from being incorrectly refused as error 11128.

Improvements

• Devin users now get plugin 0.1.21, which sends large chat requests gzipped to match the Devin CLI.
• Qoder and Qoder CN users now get plugin 0.2.23.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

功能

• Codex 和 Claude Code 的设置现在显示在 Agents 页面中每个智能体所在的行，而不是在“设置”下。
• 对话框现在会从打开它们的行中展开，再折叠回该行。
• 为 Zed 的 Ollama provider 添加了对 Ollama /api/chat 和 /api/generate 端点的支持。
• ClinePass 现在通过 OpenRouter 的 provider.only 设置固定使用 glm-5.3-flash。

错误修复

• Workbuddy 用户现在可获得 0.1.27 版插件，该插件修复了对仅声明 supportsReasoning 的模型的处理，并防止 DeepSeek Harness 被错误拒绝并报错误 11128。

改进

• Devin 用户现在可获得 0.1.21 版插件，该插件会对大型聊天请求进行 gzip 压缩，以与 Devin CLI 保持一致。
• Qoder 和 Qoder CN 用户现在可获得 0.2.23 版插件。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Improvements

• Improved performance when handling packed rollouts by gathering long lines once.
• Optimized handling of packed Codex rollouts with changed stamps: decompressed once at startup for both Sessions and Requests.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

Improvements

• 通过一次性收集长行，提升了处理打包 rollouts 的性能。
• 优化了对带有变更戳的打包 Codex rollouts 的处理：在启动时解压一次，供 Sessions 和 Requests 使用。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Bug Fixes

• Fixed incorrect accounting of image decoding when image metadata is located in different places.
• Fixed a bug where a Codex image's metadata could cause decoding to use more memory than declared.

Improvements

• Improved performance when reading packed session files by reusing decompression tools between reads.
• Codex list tags are now regenerated only when the data they depend on changes.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

Bug Fixes

• 修复了当图像元数据位于不同位置时，图像解码计量不正确的问题。
• 修复了 Codex 图像的元数据可能导致解码使用的内存超过声明值的问题。

Improvements

• 通过在多次读取之间复用解压缩工具，提升了读取打包会话文件时的性能。
• Codex 列表标签现在仅在它们所依赖的数据发生变化时才会重新生成。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Bug Fixes

Fixed a routing error that occurred when a request time wasn’t a valid date; such rows now render without a time.

Fixed Linux rendering on systems without a GPU by using non-accelerated compositing (#2827).

Fixed Windows command cleanup so ending a command context also ends its child processes (#2341).

Fixed a Linux alert issue where closing with × could open the main window over the desktop (#2513).

Fixed lightweight mode so it stays off when a user turns it off and applies correctly to users who never chose it.

Fixed handling of in-use answers while a library sync is in progress.

Performance Improvements

Reduced redundant catalog counting in the gateway; counts refresh only when the catalog changes.

Cached AutoVision and searcher provider rankings until the model list, provider, or settings change.

Skipped unnecessary agent wiring work when nothing changed.

Reused read buffers and avoided building all subagents during session scans.

Read subagent task files only when needed and once per change.

Cached the Providers page’s last list and Routing trace account names in the UI.

Built provider model lists once per load without copying every model.

Updated routing arcs and request rows only when needed rather than on every update (#1481).

Showed the previous Providers page list immediately while the updated list loads in the background.

Reloaded Codex logs only when their database or write-ahead log changed.

Reduced unnecessary lock usage when reading LAN sharing, listing accounts, and remembering logins.

Looked up provider group members once per request instead of on every expansion (#2502).

Avoided a slow process check on macOS so /api/state no longer waits seconds (#1962, #2341).

Cleaned stale plugin versions from Bun’s cache after installs (#1988).

For days with thousands of routing requests, render only visible rows and recalculate numbers only on width changes (#2524).

Reduced routing page redraws: frames load only while moving, and the live dot pulses only during active requests (#2827).

Stopped hidden tray panels from redrawing the Routing tab for every request (#2515).

Handled long session lines by gathering them once at their own size (#2515).

Only created lock folders when they were missing.

Improvements

Limited Claude Code processes based on available memory; when full, returns a 503 and fails over to the next provider (#2956).

Optimized Codex turn rewrites to avoid decoding unchanged items (#1532, #2354).

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

错误修复

修复了当请求时间不是有效日期时发生的路由错误；此类行现在渲染时不显示时间。

修复了在没有 GPU 的系统上通过使用非加速合成进行 Linux 渲染的问题 (#2827)。

修复了 Windows 命令清理问题，使结束命令上下文时也会结束其子进程 (#2341)。

修复了 Linux 警报问题：使用 × 关闭时可能会在桌面之上打开主窗口 (#2513)。

修复了轻量模式：用户关闭后保持关闭，并对从未选择过该模式的用户正确应用。

修复了在资料库同步进行期间处理正在使用的回答的问题。

性能改进

减少了网关中冗余的目录计数；计数仅在目录发生更改时刷新。

缓存了 AutoVision 和 searcher 提供程序排名，直到模型列表、提供程序或设置发生更改。

在没有任何变化时跳过了不必要的 agent 连接工作。

重用了读取缓冲区，并避免在会话扫描期间构建所有子代理。

仅在需要时读取子代理任务文件，且每次更改只读取一次。

在 UI 中缓存了 Providers 页面的上次列表和 Routing trace 账户名称。

每次加载时只构建一次提供程序模型列表，而无需复制每个模型。

仅在需要时更新路由弧线和请求行，而不是每次更新时都更新 (#1481)。

在后台加载更新列表时，立即显示之前的 Providers 页面列表。

仅当 Codex 日志的数据库或预写日志发生更改时才重新加载 Codex 日志。

减少了在读取局域网共享、列出账户和记住登录信息时不必要锁的使用。

每次请求只查找一次提供程序组成员，而不是每次展开时都查找 (#2502)。

避免了 macOS 上缓慢的进程检查，因此 /api/state 不再等待数秒 (#1962, #2341)。

安装后清理了 Bun 缓存中过时的插件版本 (#1988)。

对于有数千个路由请求的日期，仅渲染可见行，并且仅在宽度变化时重新计算数值 (#2524)。

减少了路由页面重绘：帧仅在移动时加载，实时点仅在活动请求期间闪烁 (#2827)。

阻止隐藏的托盘面板为每个请求重绘 Routing 标签页 (#2515)。

通过一次性按其自身大小收集来处理长会话行 (#2515)。

仅在锁文件夹缺失时才创建它们。

改进

根据可用内存限制 Claude Code 进程；当内存已满时，返回 503 并故障转移到下一个提供程序 (#2956)。

优化了 Codex 轮次重写，以避免解码未更改的项目 (#1532, #2354)。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Improvements

Requests are now routed to free models before paid models when both are available.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

改进

当免费模型和付费模型均可用时，请求现在会优先路由到免费模型。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Bug Fixes

Fixed a concurrency issue where a request waiting for room availability could end its wait before properly releasing its place.

Improvements

Improved Routing page performance: paths are now recalculated only when the path actually moves, reducing lag when opening the page.

Reduced memory usage by reading request bodies in place instead of copying them for every reader.

Chat requests are now decoded for silent assistant turns only when such a turn is present, avoiding unnecessary processing.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

Bug Fixes

修复了一个并发问题：等待房间可用性的请求可能会在正确释放其位置之前就结束等待。

Improvements

改进了 Routing 页面性能：现在仅当路径实际移动时才重新计算路径，减少了打开页面时的延迟。

通过就地读取请求体而不是为每个读取器复制来减少内存使用。

现在仅当存在静默助手轮次时，才会对聊天请求进行解码以用于静默助手轮次，避免不必要的处理。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Features

• Added a local privacy sidecar that runs OpenAI’s Privacy Filter for masking, with Vulkan build support.
• Improved masking of Chinese personal names when the model misses or overreaches.
• The local model can now catch personal information that rule-based masking can’t.
• Settings > Privacy can turn on the local model, show where it runs, and keep your choice after saving.
• Routing rules now list known agents and accept a typed User-Agent name.

Bug Fixes

• Fixed Grok Build and Grok App showing the wrong reason and reference.
• Fixed Codex on ChatGPT retrying an already-used turn on overload before using fallbacks.
• Fixed Codex on Copilot showing a relayed message more than once.
• Fixed retired Antigravity 2.5 Flash models not being named correctly.
• Fixed plugin GitHub lists disappearing when they couldn’t be read; they now show an error and a Try again option.
• Fixed Open at Login not warning when the Mac’s own login item switch is off.
• Fixed Windows update handling when antivirus holds a download: waits up to 30 seconds and avoids duplicate trial starts after retry.
• Fixed Magpie-written models keeping native web search enabled or re-enabling models you unticked.

Improvements

• Adjusted the reply limit display to use at most a quarter of the window, leaving room for compaction.
• Updated Ollama and Ollama Cloud decision models to classify routing groups through System One.
• Adjusted the core script budget to make room for the local privacy model.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

功能

• 新增了一个本地隐私 sidecar，用于运行 OpenAI 的 Privacy Filter 来脱敏，并支持 Vulkan 构建。
• 改进了模型漏检或过度屏蔽时对中文人名的脱敏处理。
• 本地模型现在可以捕获基于规则的脱敏无法捕获的个人信息。
• 设置 > 隐私可以开启本地模型、显示其运行位置，并在保存后保留你的选择。
• 路由规则现在会列出已知用户代理，并接受手动输入的 User-Agent 名称。

错误修复

• 修复了 Grok Build 和 Grok App 显示错误原因和引用的问题。
• 修复了 Codex on ChatGPT 在过载时会先重试已使用过的轮次，然后才使用回退的问题。
• 修复了 Codex on Copilot 重复显示转发消息的问题。
• 修复了已退役的 Antigravity 2.5 Flash 模型名称不正确的问题。
• 修复了插件的 GitHub 列表在无法读取时消失的问题；现在会显示错误并提供“重试”选项。
• 修复了当 Mac 自身的登录项开关处于关闭状态时，“登录时打开”不会发出警告的问题。
• 修复了防病毒软件占用下载文件时 Windows 更新处理的问题：现在最多等待 30 秒，并避免重试后出现重复的试用启动。
• 修复了由 Magpie 写入的模型会保持原生网络搜索启用，或重新启用你已取消勾选的模型的问题。

改进

• 调整了回复上限显示，使其最多占用窗口的四分之一，为压缩留出空间。
• 更新了 Ollama 和 Ollama Cloud 决策模型，以通过 System One 对路由组进行分类。
• 调整了核心脚本预算，为本地隐私模型留出空间。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Features

• Plugin pages can now add actions to the row menus for providers and agents.
• The status dot on the Plugins page now explains its meaning and links to plugins with available updates.
• Plugin panels can appear in Magpie pages, and plugins can replace built-in pages.
• Plugin titles now appear in the window's language.

Improvements

• Plugin panels now match the application's visual style.

Install

• Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.
• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

功能

• 插件页面现在可以向提供者和代理的行菜单添加操作。
• 插件页面上的状态点现在会解释其含义，并链接到有可用更新的插件。
• 插件面板可以出现在 Magpie 页面中，并且插件可以替换内置页面。
• 插件标题现在以窗口语言显示。

改进

• 插件面板现在与应用程序的视觉风格相匹配。

安装

• 从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Improvements

Updated provider middleware to communicate which provider a request is routed to.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

Improvements

更新了 provider 中间件，以告知请求被路由到哪个 provider。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Provider Changes

Retired the built-in providers for Command Code, Devin, Grok, Factory, ZCode, Xiaomi MiMo, WorkBuddy, Qoder, Cursor, Kiro, and Zed. They are now served by their plugins.

All built-in plugins are enabled by default, including on new installs.

Removed legacy code that only the retired built-ins used.

Bug Fixes

Failed moves for retired built-ins are retried after 10 minutes instead of 6 hours.

Retired built-ins still show their own account for the mover.

Unmoved accounts now get a Move option when a built-in provider is retired and its plugin is available.

Fixed the "leanest" compaction window label when choosing a window past the leanest band, and the recommended band is now offered.

Legacy compaction requests for another account now route to the correct account or are refused with an explanation.

ChatGPT turns now use a 2 MiB compression window instead of 8 MiB.

Improvements

Account fetching no longer waits for CLI input to identify the signed-in user; it fetches a list when no user is available.

Updated documentation to reflect the retired built-ins and their removal from the gateway and account pages.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

提供商变更

已停用以下产品的内置提供商：Command Code、Devin、Grok、Factory、ZCode、Xiaomi MiMo、WorkBuddy、Qoder、Cursor、Kiro 和 Zed。它们现在由各自的插件提供服务。

所有内置插件默认启用，全新安装时也是如此。

移除了仅供这些已停用内置提供商使用的旧代码。

错误修复

已停用内置提供商的迁移失败后，将在 10 分钟后重试，而不是 6 小时后。

已停用的内置提供商仍会在迁移工具中显示其自己的账户。

当某个内置提供商被停用且其插件可用时，未迁移的账户现在会显示“迁移”选项。

修复了在选择超出最精简区间（leanest band）的窗口时“leanest”压缩窗口标签的问题，现在会提供推荐区间。

针对其他账户的旧版压缩请求现在会路由到正确的账户，或者被拒绝并给出解释。

ChatGPT 轮次现在使用 2 MiB 的压缩窗口，而不是 8 MiB。

改进

获取账户信息不再等待 CLI 输入来识别已登录用户；当没有可用用户时，会改为获取列表。

更新了文档，以反映已停用的内置提供商以及它们从网关和账户页面中的移除。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Features

• Added DeepSeek Search API as a new search provider.

Bug Fixes

• Fixed display-name model matching so models like “GPT-6 Sol” are grouped correctly.
• Fixed routing for short follow-up messages by providing Jev with the previous user message and reply.
• Fixed auto mode review to avoid using another subscription’s Claude model and to route Claude Code’s selected model correctly.
• Fixed ChatGPT accounts’ own model list rejecting GPT Image.

Improvements

• Updated Cursor plugin to 0.2.19, Qoder and Qoder CN plugins to 0.2.18, and Factory plugin to 0.1.33.
• Improved privacy for Jev by masking turn details unless Jev is allowed to see them.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

Features

• 添加了 DeepSeek Search API 作为新的搜索提供方。

Bug Fixes

• 修复了显示名称模型匹配，使“GPT-6 Sol”等模型能正确分组。
• 修复了短后续消息的路由，向 Jev 提供上一条用户消息和回复。
• 修复了自动模式审查，避免使用其他订阅的 Claude 模型，并正确路由 Claude Code 选择的模型。
• 修复了 ChatGPT 账户自有模型列表拒绝 GPT Image 的问题。

Improvements

• 将 Cursor 插件更新至 0.2.19，将 Qoder 和 Qoder CN 插件更新至 0.2.18，并将 Factory 插件更新至 0.1.33。
• 改进 Jev 的隐私保护，除非允许 Jev 查看，否则屏蔽对话轮次详情。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。

Features

• Added a diagnostic command that records CPU activity for 30 seconds to help investigate high CPU usage.
• Added support for reaching Claude Code through a SOCKS proxy via magpie's HTTP bridge.
• CodeBuddy Code sessions now appear on the Sessions page and can be resumed or deleted.

Bug Fixes

• Fixed automatic model selection when a provider is switched off, so agents move to the remaining provider.
• Fixed tool and reasoning requests on chat-only relays so the group advances correctly.
• Fixed tools with dots in their names so they keep the client-declared name.
• Fixed the account message for natively compacted conversations to show the correct account.
• Fixed typing with an input method editor so the page no longer scrolls.
• Fixed editing patch lists that start with a YAML document marker so the marker is preserved.
• Fixed the title bar in narrow windows so tool icons remain visible and nothing overlaps the title.
• Fixed DSML messages written with pipe separators so they are executed as calls rather than left as text.
• Fixed fast mode for priority-served plugin models regardless of plugin.
• Fixed Codex native compaction so conversations stay on the original account only when enabled, and explain why when unavailable.
• Fixed custom tool rejections from a Responses upstream being returned as raw 400 errors; they are now translated properly.
• Fixed quota refresh indicators so they stay centered and don't jump between turns.
• Fixed Codex model listings to show the correct 272K context window instead of 922K.
• Fixed the gateway's caller key being lost when picking a model in Claude Code and Codex configurations.
• Fixed a Claude Code tier selected as main model showing as "Same as main model" instead of its own name.
• Fixed the Claude model chip to match the number shown in Claude Desktop's menu.

Improvements

• Model list now includes reasoning levels as plain names for Catwalk clients.

Install

Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.

• macOS (Apple Silicon): magpie-darwin-arm64.dmg
• macOS (Intel): magpie-darwin-amd64.dmg
• Windows: magpie-windows-amd64.exe (ARM: magpie-windows-arm64.exe)
• Linux: magpie-linux-amd64 / magpie-linux-arm64 (needs GTK 3 and WebKitGTK 4.1)
• Terminal only: magpie-cli-<os>-<arch>

功能

• 新增了一个诊断命令，可记录 30 秒的 CPU 活动，以帮助排查高 CPU 使用率问题。
• 新增支持通过 magpie 的 HTTP 桥接经由 SOCKS 代理访问 Claude Code。
• CodeBuddy Code 会话现在显示在 Sessions 页面，并且可以恢复或删除。

错误修复

• 修复了某个提供程序关闭时的自动模型选择问题，使代理转移到剩余的提供程序。
• 修复了仅聊天中继上的工具和推理请求，使分组正确推进。
• 修复了名称中带点的工具，使其保留客户端声明的名称。
• 修复了原生压缩对话的账户消息，使其显示正确的账户。
• 修复了使用输入法编辑器输入导致页面滚动的问题。
• 修复了编辑以 YAML 文档标记开头的补丁列表时标记被保留的问题。
• 修复了窄窗口中的标题栏，使工具图标保持可见且没有内容与标题重叠。
• 修复了使用竖线分隔符编写的 DSML 消息，使其作为调用执行而不是保留为文本。
• 修复了优先服务的插件模型的快速模式，使其不受具体插件影响。
• 修复了 Codex 原生压缩，使对话仅在启用时保留在原账户上，并在不可用时说明原因。
• 修复了来自 Responses 上游的自定义工具拒绝被返回为原始 400 错误的问题；现在它们会被正确转换。
• 修复了配额刷新指示器，使其保持居中，不会在轮次之间跳动。
• 修复了 Codex 模型列表，显示正确的 272K 上下文窗口，而不是 922K。
• 修复了在 Claude Code 和 Codex 配置中选择模型时网关的调用方密钥丢失的问题。
• 修复了被选为主模型的 Claude Code 层级显示为“Same as main model”而不是其自身名称的问题。
• 修复了 Claude 模型标签，使其与 Claude Desktop 菜单中显示的数字匹配。

改进

• 模型列表现在以纯名称形式为 Catwalk 客户端显示推理级别。

安装

从 usemagpie.ai 下载，或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。