yetone/magpie-releases v0.1.54 on GitHub

Fixes

Agents now know each model's context window. Pi used to treat every model through magpie as a 128K model, and Crush as a 200K one. A CLIProxyAPI's gpt-5.5, for example, actually holds 922K. magpie now takes each model's input limit from models.dev and writes it into Pi's contextWindow, Crush's context_window, and omp's and Codex's model lists. A proxy's variant ids (openai/gpt-5.4, gpt-5.4(high), glm-4.6:free) find their limit too, and a group uses the smallest limit among its members.
Provider changes now reach the agents' model lists. Some agents keep their own model list in a file: Codex, Pi, OpenCode, Crush, Command Code, omp, Hermes and dsh. magpie used to write those lists only when you picked a model, so a provider added later never appeared. They are now rewritten whenever a provider is saved, a vendor's model list is fetched again, or models.dev syncs, and once when magpie starts.
Codex signed in to ChatGPT refreshes its cached model list. Codex kept its model list cached under the backend's ETag, which magpie's changes never changed. That ETag now includes a tag of magpie's models, so the next Codex run fetches the list again.
A tool call from a ChatGPT account ends as a tool call. Anthropic clients were told end_turn under a tool_use block, so a client that reads stop_reason took the call for the final answer.

智能体现在知道每个模型的上下文窗口。 以前 Pi 把经过 magpie 的每个模型都当作 128K，Crush 都当作 200K。但比如 CLIProxyAPI 的 gpt-5.5 实际有 922K。magpie 现在从 models.dev 取每个模型的输入上限，写进 Pi 的 contextWindow、Crush 的 context_window，以及 omp 和 Codex 的模型列表。中转的变体 id（openai/gpt-5.4、gpt-5.4(high)、glm-4.6:free）也能查到上限；分组取成员里最小的那个上限。
Provider 的变更现在会同步到智能体的模型列表。 有些智能体把模型列表存在自己的文件里：Codex、Pi、OpenCode、Crush、Command Code、omp、Hermes 和 dsh。以前 magpie 只在你选模型时写这些列表，所以之后新加的 provider 不会出现。现在保存 provider、重新拉取厂商的模型列表、或者同步 models.dev 时都会重写这些列表，magpie 启动时也会写一次。
登录 ChatGPT 的 Codex 会刷新缓存的模型列表。 Codex 以前把模型列表缓存在后端的 ETag 下，而 magpie 这边的变更不会改变这个 ETag。现在 ETag 里带上了 magpie 模型的标记，所以下一次运行 Codex 会重新拉取列表。
ChatGPT 账号发起的工具调用现在会以工具调用结束。 以前 Anthropic 客户端在 tool_use 块下收到的是 end_turn，所以按 stop_reason 判断的客户端会把这次调用当成最终回答。

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>