v0.4.32

MusePi 0.4.32

Platform	下载
macOS (Apple Silicon)	MusePi-0.4.32-arm64.dmg
Windows	MusePi-0.4.32-setup.exe · MusePi-0.4.32-arm64-setup.exe
Linux x64	MusePi-0.4.32-x86_64.AppImage · MusePi-0.4.32-amd64.deb
Linux arm64	MusePi-0.4.32-arm64.AppImage · MusePi-0.4.32-arm64.deb
Android	MusePi-0.4.32-app-debug.apk (debug build, adb 安装)

[0.4.32] - 2026-09-17

Added

• 设计模式（预设 design，TUI 与 GUI 都可用）从「一条 persona」补成完整设计工作流：新增工作流、简报协议、产物契约、风格边界四个提示词区块——现在会先对齐简报（目标/平台/风格基准/参考/交付物）、先做结构与主导区域判断再谈视觉、产出物带 sidecar manifest 以便预览面板识别、需要落地实现代码时明确建议切回 work 模式。风格边界把设计规范接进了 agent 行为：圆角只用 --radius-* 阶梯、禁止字面 px，玻璃只用 --glass-* 四件套且只有背后有内容的悬浮层才允许。刻意不收窄工具白名单——设计要读代码库、看现有样式、改文件。
  • EN: The design mode (preset design, available in both TUI and GUI) grew from a single persona line into a complete design workflow: four new prompt sections cover workflow, brief protocol, artifact contract and style boundaries. It now aligns on a brief (goal / platform / style baseline / references / deliverable) before drawing, judges structure and the dominant region before picking colours, emits a sidecar manifest so the preview panel can recognise artifacts, and explicitly recommends switching back to work mode when implementation code is needed. The style-boundary section wires the design system into agent behaviour: radii come only from the --radius-* ladder (literal px banned), glass only from the --glass-* four-piece set, and only on floating layers that have content behind them. The tool whitelist is deliberately not narrowed — design needs to read the codebase, inspect existing styles and edit files.

Changed

• 内置预设模板现在带修订号，且只在「文件内容仍等于旧内置模板」时才升级——此前模板只在文件缺失时写入，内置模板改内容后老用户磁盘上的 preset 永远不会被替换（本次 design 就是这样被卡住的）；无条件覆盖又会抹掉用户改过的预设，所以改动过的文件一律保留，坏 JSON 也不动（交给 modes.validate 报告）。
  • EN: Built-in preset templates now carry a revision number and are upgraded only when the file still matches the old built-in template. Previously templates were written only when missing, so any content change to a built-in template never reached existing users (exactly what blocked the design upgrade here); overwriting unconditionally would instead destroy user-edited presets, so modified files are always left alone and malformed JSON is left untouched (reported by modes.validate).

Fixed

• 桌面口述（麦克风输入）四条产品级失败一次修掉：①「静音阈值」设置从未接到录音入口——所有入口只传麦克风 id，说完停 2 秒仍要等满 15 秒上限，现在在 voice 层统一读 stt.vadEndMs 并真正驱动 VAD 判停，且静音判停只在听到过语音后才武装（否则开口前的短暂停顿会被当成「说完了」）；②桌面录音用 AudioContext 默认采样率（Windows 常为 48 kHz）且不重采样，Parakeet 按 16 kHz 解码等于三倍速、常常返回空串、输入框永不回填——现在显式请求 16 kHz 并在引擎忽略时线性重采样（对齐 guest-client 本就正确的实现）；③再点麦克风会把整段缓冲丢弃不转写——现在改为提前收音并提交转写，急着停也不再白录；④麦克风经扬声器回放抬高噪声底、干扰判停——现在经零增益节点接线（图谱仍活着但无声音）；⑤16 kHz × 15 s 的裸 float JSON 仍可能撞上 daemon 4 MiB 请求上限（Request too large）——现在量化到 5 位小数（约 13 bit，对 16-bit 级别的 ASR 输入无感）。#23
  • EN: Four product-level failures of desktop dictation (microphone input) fixed at once: (1) the "silence threshold" setting was never wired into the recorder — every entry point passed only a microphone id, so speech followed by a 2 s pause still waited out the full 15 s cap; the voice layer now reads stt.vadEndMs itself and actually drives the VAD, which only arms after some speech was heard (otherwise the pause before you start talking ends the recording). (2) The desktop capture ran at the AudioContext's default rate (48 kHz on Windows) without resampling, so Parakeet — fed at a hardcoded 16 kHz — heard 3× speed and usually returned an empty string, leaving the input box unfilled; capture now requests 16 kHz explicitly and linearly resamples when an engine ignores it (matching the guest-client implementation that was already correct). (3) Pressing the mic a second time discarded the whole buffer without transcribing; it now finishes early and submits what was said. (4) The mic was played straight back out of the speakers, raising the noise floor against the VAD; the graph now runs through a zero-gain sink. (5) Even at 16 kHz, 15 s of raw-float JSON could still exceed the daemon's 4 MiB request cap ("Request too large"); the payload is now quantised to 5 decimals (~13 bits, inaudible for 16-bit-class ASR input). #23
• 移动端口述撞上 daemon 的 4 MiB 请求上限（Request too large）：上限提高到 16 MiB（与 collab relay 的 16 MiB 帧上限对齐，两个传输层不再各说各话），guest-client 的转写 payload 同步量化到 5 位小数——原先裸 float JSON 约 320 KB/s，十几秒就会失败，现在约 130 KB/s，手机上较长时间的口述也能正常转写。#23
  • EN: Mobile dictation hit the daemon's 4 MiB request cap ("Request too large"). The cap is now 16 MiB (aligned with the collab relay's 16 MiB frame limit, so the two transports finally agree), and the guest-client transcription payload is quantised to 5 decimals like the desktop — raw float JSON was ~320 KB/s and failed within seconds, now ~130 KB/s, so longer dictation on the phone transcribes normally. #23

Fixed

• Windows 上「启动 daemon broker 的那个进程」退出会连带杀死整个 broker，连另一个项目仍在使用的全局浏览器 relay 一起带走：broker 以 detached: false 被拉起，而 Bun 在 Windows 上会随父进程退出终止非 detached 子进程，于是「一个项目退出不得拆掉仍被他人持有的机器级单例」这条契约被打破（relay 端口消失，另一个项目的浏览器工具随即失效）。broker 现在显式声明必须比启动它的客户端活得更久；POSIX 行为不变（那里子进程本就会被 init 收养）。#24
  • EN: On Windows, the process that happened to start the daemon broker killed the whole broker when it exited — taking the machine-global browser relay with it even while another project was still using it. The broker was spawned detached: false, and Bun terminates non-detached children when their parent exits, which broke the contract that one project exiting must not tear down a singleton others still hold (the relay port vanished and the other project's browser tooling broke). The broker now declares that it must outlive the client that spawned it; POSIX behaviour is unchanged, since orphans are re-parented to init there. #24

Full Changelog: v0.4.31...v0.4.32

Full Changelog: v0.4.31...v0.4.32