Skip to content

macOS 0.1.2 — Third-party LLMs on stock firmware

Pre-release
Pre-release

Choose a tag to compare

View all tags
@rudyll rudyll released this 22 Sep 08:53
macos-v0.1.2
195eed0

StackChan AI Server for macOS 0.1.2 — Third-party LLMs on stock firmware

更新内容 / Changes

  • 回合制管线（openai_compatible / tokenhub / openrouter）现在由服务端判断说话结束。
    出厂固件以 auto 模式监听、从不发送 listen:stop，此前双方互等，永远不会触发 STT。
    设备若自己发来 listen:stop，仍以先到者为准。新增 compatible_server_vad（默认开启）
    和 compatible_vad_silence_ms（默认 800 毫秒）。

  • 修复 Base URL 写成 .../v1 时请求打到 /v1/v1/... 的问题，STT、LLM、TTS 三段以及
    内置的 OpenRouter 默认值都受影响。现在带不带 /v1 都能正确解析。

  • Realtime 地址可配置（openai_base_url），可指向任何实现 OpenAI Realtime 协议的服务。
    第三方 realtime 端点未经验证。

  • 日志新增上行音频帧数/字节数统计和设备声明的监听模式，用来区分“没收到音频”和
    “收到音频但没触发回合”。没有产生音频的回合现在也会上报结束。

  • The turn-based pipeline (openai_compatible, tokenhub, openrouter) now detects
    the end of speech server-side. Firmware listening in auto mode never sends
    listen:stop, so both sides waited and no turn ever started. A device's own
    listen:stop still ends the turn when it arrives first. New compatible_server_vad
    (default on) and compatible_vad_silence_ms (default 800 ms).

  • Fixed base URLs written as .../v1 producing /v1/v1/... request paths across all
    three stages, including the built-in OpenRouter default. Both spellings now work.

  • Added a configurable openai_base_url for the realtime provider. Third-party
    realtime endpoints are untested.

  • Added incoming-audio counters and the announced listening mode to the log, and an
    end-of-turn report for turns that produce no audio.

下载与升级 / Download and upgrade

下载 StackChan-AI-Server-0.1.2-macos-universal.dmg；无需自行构建或安装 Docker。
安装包包含 Apple Silicon 和 Intel 两种架构，构建目标为 macOS 12+，静态打包 OPUS。
升级前用“活动监视器”退出本应用的 stackchan-server 进程，再把新应用拖入“应用程序”替换。
设置与对话历史保留在原来的 Application Support 目录；不要同时运行两个副本。
旧版 0.1.1 与 0.1.0 安装包继续保留，便于回退。

Download the universal DMG from Assets. No user build or Docker installation is needed.
Apple Silicon and Intel executables target macOS 12+, with OPUS included statically.
Stop the old app's stackchan-server processes in Activity Monitor before replacing
the app in Applications. Existing settings/history are retained; do not run two copies.
The 0.1.1 and 0.1.0 downloads remain available. Verify the matching checksum first:

shasum -a 256 -c StackChan-AI-Server-0.1.2-macos-universal.dmg.sha256

许可与源码 / License and corresponding source

这是本应用的第一个 AGPL-3.0-only 构建。Release 附带与该 tag 完全一致的源码包
stackchan-ai-server-macos-v0.1.2-source.tar.gz 及其校验文件；设置页面内也带有源码链接。
已发布的 macos-v0.1.1、macos-v0.1.0 与 HA v2.8.0-beta.3 保留原有 MIT 许可，未被重建或覆盖。

This is the first AGPL-3.0-only build of this app. The Release carries the
corresponding-source archive built from this exact tag
(stackchan-ai-server-macos-v0.1.2-source.tar.gz) with its checksum, and the settings
page carries a source link. Previously published artifacts — macos-v0.1.1,
macos-v0.1.0 and HA v2.8.0-beta.3 — keep their original MIT license and were
neither rebuilt nor overwritten. Go module and OPUS sources are not vendored in the
archive; their versions are pinned in go.mod/go.sum and macos/prepare-opus.sh,
which downloads a checksum-pinned OPUS release.

限制与验证 / Limitations and verification

仍是临时签名预览版，尚无 Developer ID 签名或 Apple 公证，也没有菜单栏控制和自动更新。
确认来源后，无法验证开发者的提示可参考
Apple 安装说明；不要全局关闭 Gatekeeper 或绕过恶意软件警告。
服务端静音判定只有单元测试覆盖，没有在 auto 模式的真机上验证过；欢迎回报日志。
本次不改变 HA 插件版本，插件侧对应的是 v2.8.0-beta.4。

Still an ad-hoc signed preview, not Developer ID signed or Apple notarized; no
menu-bar controls or automatic updates. Follow the Apple guidance above for
unidentified-developer warnings; never disable Gatekeeper globally or bypass malware
warnings. The server-side turn detection is covered by unit tests only — no physical
device running factory firmware in auto mode was available, so logs from real
hardware are welcome. This release does not change the HA add-on version; the matching
add-on release is v2.8.0-beta.4.

Assets 6
Loading