magpie v0.1.1235

Changes

• gateway: a plugin's non-streaming Responses request is relayed streamed and given back whole, so Grok's x_search runs (plugins#112, Watchyourback)
• gui: an agent's model picker lists magpie's models before the ones it asks for itself (abledo on Discord)
• gui: a picker on a phone keeps to the room it was placed in, never off the window's top
• gateway: Claude Code through magpie on any model lets a relay be tested as Claude Code
• gui: a template picked from Routing's Templates menu opens filled in, in the editor and in view
• tune: a compaction window advised never makes you wait longer on compactions, and says the wait (HirotaNagai on X)
• agent: connecting an app that gets magpie's models as its provider, with none listed for it, says why instead of reporting success (Lumens on Discord)
• gui tests: Settings tabs are opened with Enter, not a click that waits on frames a loaded WebKit doesn't draw (#2150)
• gateway: a decision model's intent counts by its option probabilities, not its own confidence field (@VersunPan on X)
• provider: SiliconFlow offers its international site as a region beside China
• provider: a balance query's header takes the key as {token} too, and the key pasted as its token goes as a Bearer (lx_78654 on Discord)
• tui: w changes a custom provider's base URLs (luluxiu_1_4392 on Discord)
• provider: a sub2api or new-api account query left without its token asks the key's query instead (lx_78654 on Discord)
• tests: TestAgentGoneInTheRetryPause cancels once the last try is in its pause, not 100ms after the upstream counted it
• gateway, provider: send no prompt_cache_options to Codex, and ask again without it when an upstream refuses it (#2154)
• catalog: a relay channel's word before a known model is that model, so Codex lets it see images (#2157)
• routing: the hub's mode chip is written only when it changes, so a press held over the once-a-second redraw still clicks in WebKit (#2151)

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>

更新

• 网关：插件的非流式 Responses 请求改为流式转发、再整体返回，Grok 的 x_search 可以运行了（plugins#112，Watchyourback）
• 界面：智能体的模型选择器先列出 magpie 的模型，再列出它自己请求的模型（Discord 上的 abledo）
• 界面：手机上的选择器保持在它所在的区域内，不会超出窗口顶部
• 网关：任意模型下通过 magpie 运行的 Claude Code 都可以用来把中转站当作 Claude Code 测试
• 界面：从路由页「模板」菜单选择的模板，在编辑器和视图中都会以填好的状态打开
• 调优：建议的压缩窗口不会让你在压缩上等更久，并会写明需要等待的时间（X 上的 HirotaNagai）
• 智能体：连接一个以 magpie 的模型作为提供商、但没有为它列出任何模型的应用时，会说明原因，而不是报告成功（Discord 上的 Lumens）
• 界面测试：设置的标签页用 Enter 打开，不再用需要等待繁忙 WebKit 不绘制的帧的点击（#2150）
• 网关：决策模型的意图按选项概率计算，而不是它自己给出的置信度字段（X 上的 @VersunPan）
• 提供商：硅基流动（SiliconFlow）在中国区之外新增国际站区域
• 提供商：余额查询的请求头也接受 {token} 形式的密钥，作为 token 粘贴的密钥会以 Bearer 发送（Discord 上的 lx_78654）
• TUI：w 可以修改自定义提供商的基础 URL（Discord 上的 luluxiu_1_4392）
• 提供商：缺少 token 的 sub2api 或 new-api 账号查询，改用密钥的查询（Discord 上的 lx_78654）
• 测试：TestAgentGoneInTheRetryPause 在最后一次尝试进入停顿时才取消，而不是在上游计数后 100ms
• 网关、提供商：不再向 Codex 发送 prompt_cache_options，上游拒绝它时去掉后重新请求（#2154）
• 目录：中转渠道放在已知模型名前面的前缀仍被识别为该模型，Codex 可以让它看图片（#2157）
• 路由：中心的模式标签只在变化时才重写，在每秒一次的重绘期间按住也能在 WebKit 中点击（#2151）

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。

---

The same release, and the feed magpie updates from: https://github.com/yetone/magpie-releases/releases/tag/v0.1.1235