Skip to content

0.17.1 Alpha 1 — corrected release with bundled dictionary and plugin store preview

Latest

Choose a tag to compare

@Lucas1479 Lucas1479 released this 22 Oct 19:13
9547f39

0.17.1 Alpha 1 — bundled dictionary and plugin store preview

Release tracking and feedback #260 · 中文说明

This is the corrected follow-up to the earlier Alpha 0 preparation in #245, integrated through #258. It uses a new version and tag; Alpha 0's tag and existing bytes are unchanged.

Download and run

Download amadeus-0.17.1-alpha.1-source.zip below for the complete application source distribution, including the default Japanese dictionary. GitHub's automatically generated “Source code” archives do not include the external dictionary payload. Follow the README inside the full archive to install Python/Node dependencies and run the app. This is a source Alpha, not a self-contained desktop installer; model weights, character media and credentials are not included.

Recommended setup for new installations — retained from 0.17.0

Alpha 1 includes the local ASR/TTS and rendering optimizations introduced in 0.17.0. Install matching assets and verify the actual runtime paths below to benefit from them; upgrading code alone does not replace old explicit settings.

  1. Download amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip from the official resource center. This color-corrected pack includes video for 40 clips / 7,672 frames and the new key-point return transition. For Google Drive, download all 28 parts and the helper files, then run Assemble-Kurisu.ps1 to reconstruct and verify the full ZIP; do not install an individual part. From the project root, using your installed environment:

    uv run --locked --no-sync python tools\external_assets.py verify "C:\Downloads\amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip"
    uv run --locked --no-sync python tools\external_assets.py install "C:\Downloads\amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip"

    To replace an older character pack, stop the backend and close rendering windows, back up customized assets, then append --overwrite to the install command. Character media is separate from the bundled dictionary.

  2. In Settings → Graphics & performance, select Standard (60 FPS), Frame source: Automatic · video when available, and Frame cache budget: Automatic · video 384 MiB / static 2 GiB. Enable Sample animation textures and Reuse converted textures on disk. These are the fresh Standard-profile defaults; texture sampling defaults on at 60 FPS. Existing explicit settings remain authoritative, so check for an older static-source, fixed-budget or disabled-sampling choice.

  3. Restart the backend and reopen the character/wallpaper. Check Actual animation frame source for Video with static heads · 384 MiB. Selecting Automatic alone does not confirm video use. Missing/incompatible video or decoder support retains static frames, whose automatic budget is 2 GiB. The budget bounds retained frame data, not total application RAM/VRAM or decoder memory.

  4. For NVIDIA local voice optimizations, install the T4 local voice environment and Qwen/GPT-SoVITS assets following the installation guide. Windows local-cu124 remains the reference profile; RTX 50-series users should evaluate the separate local-cu128 candidate instead of treating cu124 as a Blackwell baseline. Keep Qwen device on auto, Qwen CUDA Graph enabled (QWEN3_ASR_CUDA_GRAPH=true), and local TTS inference on Automatic (ENABLE_CUDA_GRAPH=auto, TTS_T2S_FLASH_ATTN=auto). Qwen FlashAttention defaults on (QWEN3_ASR_FLASH_ATTN=true), but FlashAttention paths require a compatible optional extension and supported hardware; the app does not install that extension automatically. See extension compatibility and installation. Restart the backend after changes and check startup diagnostics for the selected path. Remote voice, CPU, MPS and ROCm configurations do not receive these NVIDIA CUDA optimizations.

The 0.17.0 performance comparisons retain their original workloads and limitations. They are historical measurements, not new Alpha 1 benchmarks or a promise of the same gain on every device.

Upgrading: extract into a new directory, preserve settings, personal data and separately installed assets, and settle active Work before replacing the runtime. Retain the complete optional-extra selection for your installation profile when syncing dependencies, rebuild Electron and restart the backend. Explicit voice/render settings and independently selected dictionaries remain authoritative.

Changes since 0.17.0

Compared with the previous public release, v0.17.0-alpha.0, this release includes the following features and follow-up fixes. View the full comparison.

  • Plugin store preview: browse the official catalog, inspect packages, install with progress/cancellation, check for updates manually and download through HTTPS mirrors. Withdrawal, persistent blocking/unblocking and damaged-state recovery share the existing installation lifecycle. Authors and maintainers also gain repeatable package checking and catalog publication tools. The official catalog is currently empty; remote safety updates require an explicit check. #256.
  • Desktop console and plugin management: unified controls across Chat, Settings and Plugins, a simpler composer, compact history/artifact access, inline chat rename fixes and approvals above the composer. Plugin cards expose identity, state and enable switches, with configuration and diagnostics in the existing details flow. Local manifest-v1 icons are supported; their official-catalog admission still awaits separate checker qualification. #255, #256.
  • Bundled, independently versioned Japanese dictionary: the full 2026.10.22.1-rc2 resource now ships in application source archives and packaged resources. Independent updates/rollback and personal user.csv retain precedence, including older format-1 selections. Fixes also cover personal-entry case/width matching and repeated full-vocabulary copies during loading. The context-reading experiment remains separate and default-off. #237, #250, #254.
  • Pi, MCP and approval control: Pi moves to 1.2.0 with Host-managed MCP tools/resources, explicit credential bindings and per-call approvals. Image metadata for legacy DeepSeek selections and transport cleanup after cancellation are repaired. Users can explicitly opt into persistent automatic approval, or grant revocable matching-call permission for the current Pi MCP run. Automatic approval defaults off and does not bypass Host permission boundaries. #221, #225, #226, #234.
  • Chat and subtitle preferences: Chat and ordinary Wallpaper subtitles can target Simplified Chinese or English; Chinese remains the default, while VN retains its own translation settings. Main Chat receives a configurable role temperature and prompt adjustments to reduce repeated openings and misplaced nickname objections, without changing control-query sampling. #239, #229.
  • Local voice reliability: GPT-SoVITS v3 references load through SoundFile without requiring TorchCodec, incomplete reference caches are rejected, and synthesis failures no longer masquerade as successful silence. Shared playback stop/initialization is serialized to prevent an active flag with no usable output stream. #238, #242.
  • Windows desktop and wallpaper recovery: preserve window position/size during desktop-layer reconciliation on mixed-DPI displays, retain the original failure during rollback, and improve recovery after wallpaper-host restarts using the actual receiving process identity. #220, #228.
  • Cross-platform installation and checks: Linux AEC builds use a disposable source copy and verify wrapper/binding consistency; macOS gains CPython 3.12.14 migration guidance and a real Tk/ImageTk image-transfer probe. Electron smoke tests now track the native process correctly and exercise the appropriate controls on each platform. #249, #253, #248.
  • Dependency maintenance: patched MCP 2.2.0, aiohttp/yarl, Pillow 13, certifi and OpenTelemetry dependencies. Qwen-ASR's Gradio demo dependency is separated from inference installation; local TTS no longer imports the Gradio/pandas UI stack through its shared decoder. These are targeted repairs, not a claim that every optional environment is vulnerability-free. #232, #243, #252, #256.
  • Setup and release delivery: updated installation/performance guidance, the Pi acknowledgement, and a new Alpha 1 identity with a complete dictionary-bearing source archive, manifest and checksums. This corrects the earlier Alpha 0 preparation without replacing its tag or files. #216, #231, #258.

Dictionary scope

The immutable rc2 resource contains all nine pinned members and 134,201 source rows. Preserved source rows have not all been individually reviewed or listened to; the rc2 designation and reviewed/unverified distinctions are retained. The context-reading model remains separate and disabled by default. Source/license notices are included with the data.

The independent dictionary ZIP is also attached for data-only updates and rollback. Its SHA-256 is a935b3adfded9022afe3b2384cebc2ef1e4023f54f48df1f9ded9ff45b896989.

Verification

  • The full tag-triggered Source Release workflow passed on 9547f39a, including public dictionary download, extracted-source installation, native bundled dictionary checks, packaged Windows resources and actual desktop launch.
  • 18/18 checks passed on candidate e26d204e; the merged commit has the identical Git tree.
  • Clean full-source build: 4,299 selected files, zero release gate errors or warnings. Final release regenerated from the merged commit, with every payload member unchanged except source-manifest commit metadata.
  • 41 native dictionary/lifecycle tests passed. Actual extracted-source and packaged-resource loading, independent update/rollback, upgrade precedence and personal data preservation were verified.
  • Actual packaged desktop: 15 desktop, 15 plugin and 18 fault/recovery checks passed.
  • Formal catalog-toolchain roundtrip: 15 desktop and 27 publication checks passed, including six local Git/HTTPS publications, real install/update, mirror download, withdrawal, block/unblock, restarts and simulated NUL-damaged settings recovery.

Fault injection does not claim a physical power-loss test, and native frontend checks do not claim exhaustive linguistic or human listening validation. See source-manifest.json and SHA256SUMS.txt for release identity and integrity.


中文说明

0.17.1 Alpha 1:随包词典与插件商店预览

本版本发布反馈与验收追踪 #260

这是早期 #245 中 Alpha 0 发布准备的更正与整合版本,经 #258 合并。Alpha 1 使用新的版本号和标签,旧 Alpha 0 标签及文件保持不变。当前已标记为 Latest,但仍是 Alpha 版本。

下载与运行

请下载 amadeus-0.17.1-alpha.1-source.zip,这是已随附默认日语词典的完整应用源码包。GitHub 自动生成的两个“Source code”归档不包含外置词典数据。解压后按包内 README 安装 Python/Node 依赖并启动。

本次提供源码 Alpha,不是自带完整运行环境的桌面安装器;模型权重、角色素材和凭据需另行准备。默认词典已经包含在完整源码包中,无需再下载一次。

新安装推荐配置与生效检查——延续 0.17.0 的性能优化

Alpha 1 已包含 0.17.0 引入的本地 ASR/TTS 与渲染优化。建议安装后按以下步骤配置并确认实际路径;只更新代码不会覆盖旧的显式设置。

  1. 从官网资源中心下载 amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip,包含重新校色的素材、40 组视频(7,672 帧)和新增 key-point 返回过渡动画。Google Drive 下载需取齐 28 个分卷及辅助文件,用 Assemble-Kurisu.ps1 合并并校验完整 ZIP,不能直接安装单个分卷。在项目根目录使用已安装的环境执行:

    uv run --locked --no-sync python tools\external_assets.py verify "C:\Downloads\amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip"
    uv run --locked --no-sync python tools\external_assets.py install "C:\Downloads\amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip"

    覆盖旧角色包前,先停止后端、关闭渲染窗口并备份自定义素材,再给 install 命令添加 --overwrite。角色素材与本次随包词典是不同资源。

  2. 设置 → 图形与性能:选择标准(60 FPS)、帧源选择:自动 · 可用时优先视频、帧缓存预算:自动 · 视频 384 MiB / 静态 2 GiB,开启启用动画纹理采样和复用磁盘中的已转码纹理。这些对应全新安装的标准档默认值,纹理采样在 60 FPS 下默认开启;旧用户的静态帧源、固定预算或关闭采样等显式设置仍会保留,需要核对。

  3. 重启后端并重新打开角色/壁纸,检查当前实际动画帧源是否显示视频与静态入口帧 · 384 MiB。只选择“自动”不代表视频已生效;缺少匹配视频或解码能力时继续使用静态帧,自动预算为 2 GiB。该预算限制保留的帧数据,不等于整个应用、解码器的总内存或显存上限。

  4. 使用本地 NVIDIA 语音优化时,按安装指南安装 T4 本地语音环境及 Qwen/GPT-SoVITS 资产。Windows local-cu124 仍是参考配置;RTX 50 系应评估独立的 local-cu128 候选,不能把 cu124 作为 Blackwell 的运行依据。保持 Qwen 设备为 auto、Qwen CUDA Graph 开启(QWEN3_ASR_CUDA_GRAPH=true),本地 TTS 推理模式为自动(ENABLE_CUDA_GRAPH=auto、TTS_T2S_FLASH_ATTN=auto)。Qwen FlashAttention 默认开启(QWEN3_ASR_FLASH_ATTN=true),但相关路径仍需兼容的可选扩展和硬件支持,应用不会自动安装扩展;参见兼容性与安装说明。修改后重启后端,通过启动诊断确认实际推理路径。远程语音、CPU、MPS 和 ROCm 配置不享有这些 NVIDIA CUDA 优化。

0.17.0 性能对比保留原始测试负载和适用边界;这些是历史测量,不是本轮重新跑出的 Alpha 1 基准,也不保证所有设备获得相同比例的提升。

升级用户:建议解压到新目录,保留设置、个人数据和独立安装的资产,待活动 Work 收尾后再替换运行环境。同步依赖时保留所选安装配置的完整 extras,重新构建 Electron 并重启后端。既有语音/渲染显式设置,以及明确选择的独立词典,仍优先生效。

自 0.17.0 起的更新

以下以此前公开发布的 v0.17.0-alpha.0 为基线,概括本版本包含的新功能与后续修复。查看两个版本之间的完整差异。

  • 插件商店预览:支持官方目录浏览、包详情、安装进度与取消、手动更新和 HTTPS 下载镜像;撤回、持久封禁/解封及损坏恢复复用现有安装流程。作者和维护者也有了可重复执行的包校验与目录发布工具。官方目录目前为空,远程安全状态需用户主动检查。#256。
  • 桌面控制台与插件管理:统一 Chat、Settings 和 Plugins 的控件,简化输入框与历史/产物入口,修复聊天内重命名,并把待审批项放到输入框上方。插件卡片集中展示身份、状态和启用开关,配置与诊断保留在详情流程中。支持本地 manifest-v1 图标;带图标包进入官方目录仍需校验器单独完成准入验证。#255、#256。
  • 完整日语词典随包与独立版本:完整 2026.10.22.1-rc2 已纳入应用源码包和打包资源。独立更新/回退与个人 user.csv 保持原有优先级,旧 format-1 独立版本也不会被应用升级覆盖;同时修复个人词条大小写/宽度匹配,以及加载时反复复制全量词表的问题。上下文读音实验仍独立提供、默认关闭。#237、#250、#254。
  • Pi、MCP 与审批控制:Pi 升至 1.2.0,接入 Host 管理的 MCP 工具/资源、明确的凭据绑定和逐次调用审批;修复旧 DeepSeek 模型选择下的图片元数据,以及任务取消后的连接清理。用户可主动开启持久自动审批,或仅在当前 Pi MCP 运行中授权参数一致的调用并随时撤销。自动审批默认关闭,也不会越过 Host 权限边界。#221、#225、#226、#234。
  • 聊天与字幕偏好:Chat 和普通壁纸字幕可选择简体中文或英文,默认仍为中文;VN 保留独立翻译设置。主聊天增加可配置的角色温度,并调整提示词以减少重复开场和不合语境的昵称抗议,不改变控制判断的采样设置。#239、#229。
  • 本地语音可靠性:GPT-SoVITS v3 改用 SoundFile 读取参考音频,无需依赖 TorchCodec;拒绝不完整的参考缓存,并让合成错误正常暴露,避免把静音当作成功。共享播放器的停止与初始化改为串行处理,避免出现“标记为播放中、实际却没有可用输出流”的状态。#238、#242。
  • Windows 桌面与壁纸恢复:混合 DPI 屏幕下调整桌面层级时保留窗口位置和大小;回退时保留原始错误,并依据实际接收进程的身份改进壁纸宿主重启后的恢复。#220、#228。
  • 跨平台安装与检查:Linux AEC 在临时源码副本中构建,并校验 wrapper/binding 一致性;macOS 补充 CPython 3.12.14 迁移指南及真实 Tk/ImageTk 图像传递探针。Electron 冒烟测试改为跟踪真正的原生进程,并按平台验证实际存在的控件。#249、#253、#248。
  • 依赖维护:更新至修复已知问题的 MCP 2.2.0、aiohttp/yarl、Pillow 13、certifi 和 OpenTelemetry 依赖;将 Qwen-ASR 的 Gradio 演示依赖与推理安装分离,本地 TTS 不再经共享解码器导入 Gradio/pandas UI 依赖。这些是具体修复,不代表所有可选环境都已不存在漏洞。#232、#243、#252、#256。
  • 安装指南与发布交付:完善安装和性能配置说明,补充 Pi 致谢,并以新的 Alpha 1 版本提供包含完整词典的源码包、清单和校验和。本轮更正旧 Alpha 0 发布准备,但不替换旧标签和文件。#216、#231、#258。

词典范围与限制

固定摘要的 rc2 资源包含九个文件和 134,201 行源数据。保留的源词条并非全部经过逐词审核或试听,因此继续使用 rc2 标识,并区分已审核和未审核内容。可选上下文读音模型仍独立提供、默认关闭;源数据与授权说明随词典附带。

Release 同时提供独立词典 ZIP,供只更新数据或回退使用。其 SHA-256 为 a935b3adfded9022afe3b2384cebc2ef1e4023f54f48df1f9ded9ff45b896989。

验证结果

  • 完整标签发布工作流在 9547f39a 上通过,覆盖公开词典下载、解压源码安装、原生词典加载、Windows 打包资源和实际桌面启动。
  • 候选 e26d204e 的 18/18 项检查通过,合并提交与该候选具有相同文件树。
  • 干净完整源码构建收录 4,299 个文件,发布检查无错误或警告;正式产物从合并提交重新生成,除源码清单的提交元数据外,其余包内文件与已验收候选一致。
  • 41 项原生词典与生命周期测试通过;验证实际源码/打包资源加载、独立更新与回退、升级优先级和个人数据保留。
  • 实际桌面包通过 15 项桌面、15 项插件和 18 项故障恢复检查。
  • 正式目录工具链闭环通过 15 项桌面和 27 项发布检查,覆盖六次本地 Git/HTTPS 发布、真实安装更新、镜像、撤回、封禁/解封、重启和设置文件 NUL 损坏恢复。

故障注入不等于真实断电试验,原生前端检查也不等于全量语言质量或人工试听验收。产物身份和完整性以 source-manifest.json 与 SHA256SUMS.txt 为准。