"type":"message","message":{"id":"chatcmpl-...","type":"message","role":"assistant","model":"...","content":[{"type":"text","text":"..."}],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":10,"output_tokens":20}},"session_id":"..."} after dsh: body[0].message.content[0].text and okay. dsh test: pass model "harness" and reason="high", messages "ping", label is "harness" same fallback, Message and Response are round-trip. openai.bash: Copilot's vendor request (1b1c9f39, 349de64b) is a "message" string in json body right from /v1/responses; no response_format on the vendor request, copy /v1/chat/completions to /v1/responses. session: close blank responses made in a run when "request" fails. on 512MB CI, -run TestDshSession|TestVendorDshSession are blank. no sleeps. gofmt; vet; builds darwin+linux; go test -tags nogui -count=1 ./... pass.## Changelog
Improvements
- Restored automatic reasoning level detection for models hosted on Azure OpenAI, ensuring compatible models regain their full range of reasoning options.
- Updated Azure OpenAI recognition so hosted models no longer inherit incomplete capability lists from third-party directories.
- Improved group routing to ignore members without reasoning levels instead of removing reasoning options from the entire group.
- Enhanced reasoning level handling across related tools, including the command line interface, Codex catalog, and integrations like Pi and OpenCode.
- Simplified request handling for the harness integration by using standard message formats and removing unnecessary error fields.
Bug Fixes
- Fixed an issue where models hosted on Azure OpenAI lost their reasoning levels, affecting models such as Kimi K2.5, DeepSeek V3.2, DeepSeek R1, and Grok 4.20.
- Fixed group routing to preserve reasoning levels when group members include models without reasoning levels.
- Corrected request and response processing for the harness integration, ensuring messages round-trip correctly.
- Fixed vendor request handling for Copilot so response requests work properly without unsupported format fields.
- Addressed blank responses in sessions when a request fails, improving reliability under resource constraints.
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>
"type":"message","message":{"id":"chatcmpl-...","type":"message","role":"assistant","model":"...","content":[{"type":"text","text":"..."}],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":10,"output_tokens":20}},"session_id":"..."} 之后 dsh:body[0].message.content[0].text,且正常。dsh 测试:通过 model "harness",reason="high",messages "ping",label 为 "harness" 同样回退,Message 和 Response 可正常往返。openai.bash:Copilot 的供应商请求(1b1c9f39、349de64b)在 /v1/responses 的 JSON 正文中直接是 "message" 字符串;供应商请求上没有 response_format,将 /v1/chat/completions 复制到 /v1/responses。session:在 "request" 失败时关闭运行中产生的空白响应。在 512MB CI 上,-run TestDshSession|TestVendorDshSession 为空白。无 sleep。gofmt;vet;构建 darwin+linux;go test -tags nogui -count=1 ./... 通过。
Changelog
Improvements
- 恢复了对 Azure OpenAI 上托管模型的自动推理级别检测,确保兼容模型重新获得完整的推理选项。
- 更新了 Azure OpenAI 的识别机制,使托管模型不再从第三方目录继承不完整的能力列表。
- 改进了分组路由,忽略没有推理级别的成员,而不是移除整个分组的推理选项。
- 增强了相关工具中的推理级别处理,包括命令行界面、Codex 目录以及 Pi 和 OpenCode 等集成。
- 通过使用标准消息格式并移除不必要的错误字段,简化了 harness 集成的请求处理。
Bug Fixes
- 修复了托管在 Azure OpenAI 上的模型丢失推理级别的问题,受影响模型包括 Kimi K2.5、DeepSeek V3.2、DeepSeek R1 和 Grok 4.20。
- 修复了分组路由,在组成员包含没有推理级别的模型时保留推理级别。
- 修正了 harness 集成的请求和响应处理,确保消息正确往返。
- 修复了 Copilot 的供应商请求处理,使响应请求在没有不支持的格式字段时正常工作。
- 解决了请求失败时会话中出现空白响应的问题,提高了资源受限条件下的可靠性。
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。