v0.17.0-alpha.0 — Faster local voice and Work reliability

Amadeus 0.17.0 Alpha brings the local voice, rendering and Work improvements since 0.16.1. This is a source Alpha release. Desktop installers, model weights, character media, reference voices, Cubism Core and credentials are supplied separately.

• Local ASR/TTS: CUDA Graph ASR decoding, semantic-history/EOS improvements, shared nested KV storage and graph reuse (#202, #205, #206). Recorded results include a 75.17% median reduction in warm ASR backend latency across 140 fixed-clip pairs, and unique KV storage from 561 to 48 MiB in the measured default configuration. These are workload-specific results, not universal latency or total-VRAM claims.
• Rendering: matching video assets can use hybrid video/static frames, with static fallback and visible Graphics status (#200); ambient actions follow speech completion (#204). The color-corrected 2026.10.14-keypoint-return-v3 character pack is already published separately, with 40 video clips / 7,672 frames and a new key-point return transition animation. Old static-only packs remain supported.
• Work and MCP: stronger narration/completion evidence, permission synchronization, registered-file and AUIP-delivery separation, explicit MCP environment bindings, native approval responses and shutdown/cancellation fixes (#183, #186, #189, #197, #198, #208–#211). Unsupported Codex server requests no longer terminate the SDK reader.
• Audio output: one bounded PortAudio -9986 retry with a new output owner (#213, retaining #199's contribution). Other active capture owners keep their references; this is not a global reset or an AirPods reconnect qualification.
• Extensions: Plugins is hidden in both sidebar layouts; code and Settings → General → Manage plugins remain. The protocol and wider integrations are still experimental. Extension Host version comes from installed package metadata. GUI smoke assertion failures now correctly fail CI.

Release commit: 342354bff579cfd18035cb80b1b59582c78c44e3.

Validation: all 17 checks on final preparation commit 7dfdc29 passed. The tag's source-release workflow installed, built, packaged and launched the extracted sources. All 4,106 source files, their sizes/hashes and the embedded manifest were verified, and all three published assets were downloaded and checked. Local follow-up covered real SDK-reader regressions, 523 frontend tests, native GUI smokes and both successful/intentional-failure smoke exits. The detailed record separates historical benchmarks from newly executed tests.

Real-service acceptance exercised native Codex permissions, same-run steering, Work creation/amendment/restart and cancellation/retry, native surface handoffs, AUIP interaction, local TTS interruption/recovery and offline ASR Graph comparison. Reactor GUI plus local TTS passed 24/24 checks after correcting the sample's stability boundary. The generated-app control journey completed; a separate offline re-review passed 29/29 after fixing an inapplicable no-TTS speech assertion. Its raw report remains intact. These are not microphone/acoustic or every-device qualifications.

Known limitations accepted for this Alpha:

• Automatic Gomoku participation can stop with b2_candidate_not_available. The failed choice was not executed. Later gameplay completion does not establish consistent full-round success; original failures remain recorded.
• Deep Windows project/application paths can prevent AUIP SDK loading with ERR_FILE_NOT_FOUND / managed_sdk_unavailable. Observed failing SDK paths were 263–266 characters. Use a shallow workspace path; a byte-identical short-path copy and a fresh short-path control journey succeeded.
• Windows wallpaper input/recovery (#185), sequential exclusive MCP lifetime (#207), wider rendering and physical interruption/device qualification (#201/#103), and same-process macOS/AirPods reconnect remain open.
• Existing Pi and optional-model dependency advisories remain disclosed. This is not a vulnerability-free release, and the maintainer's release decision does not turn failed live runs into passes.

Upgrade: extract into a new directory, preserve settings/data and separately installed assets, and stop incoming turns 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. Existing explicit voice/render settings remain authoritative; stdio MCP values must use the documented explicit encrypted bindings.

Recommended setup for 0.17.0:

1. Download amadeus-character-kurisu-2026.10.14-keypoint-return-v3.zip from the official resource center. This color-corrected pack includes the video files and the new return transition. For Google Drive, download all 28 parts and the helper files, then run Assemble-Kurisu.ps1 to reconstruct and verify the 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 upgrade an older character pack, back up customized assets first, then append --overwrite to the install command.

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. Standard/automatic settings and disk reuse are the defaults; texture sampling defaults on at 60 FPS. Existing explicit settings are preserved, so upgrading alone may retain an older static/off setting.

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

4. For the local NVIDIA speech improvements, install the L4 local CUDA profile and the Qwen/GPT-SoVITS assets. 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 is enabled by default (QWEN3_ASR_FLASH_ATTN=true), but FlashAttention paths still need a separately installed compatible extension and supported hardware. Restart the backend after changes and check the actual selected path in startup diagnostics. Remote voice and CPU profiles do not receive these CUDA benefits.

This is a documentation clarification to the existing release. The tag, source ZIP, manifest and checksums are unchanged; the README inside the ZIP remains the original release snapshot. See the current English setup guide or Chinese setup guide for documentation updates.

Release notes · Acceptance · Performance comparisons · Changes since 0.16.1 · Preparation #214 · Tracking #215

---

中文：0.17.0 Alpha 汇集本地 ASR/TTS、视频供帧、Work 事实与权限同步、MCP 环境迁移以及退出/取消修复。已校色的 2026.10.14 视频角色包已独立发布，包含 40 组视频、7,672 帧，并新增 key point 返回过渡动画；旧静态包继续可用。插件侧栏隐藏，代码及设置管理保留，尚不代表插件商店成熟。

本次发布源码 ZIP、逐文件清单和 SHA-256 校验文件，不包含安装器、模型或角色媒体。最终准备提交 17 项 CI 全绿；tag 的源码工作流已验证从解压源码安装、构建、打包和启动，4,106 个文件及正式附件均完成哈希复核。优化数据保留各自基线和适用范围，不宣称通用端到端加速。

真实服务回归覆盖权限、运行中修改、重启续接、取消重试、AUIP 和本地语音。反应堆 GUI＋语音 24/24；完整生成控制链经修正无语音检查条件后的独立离线重评为 29/29，原始报告没有改写。

**本 Alpha 明确保留：**五子棋自动应对可能因候选不可用而停止；Windows 过深路径可能使 AUIP SDK 加载失败，建议浅层工作目录；壁纸 #185、独占 MCP #207、跨设备/真实打断、AirPods 重连和已有依赖告警仍未闭合。维护者已接受这些限制并授权发布，失败的真机结果不会改成通过。升级前停止活动任务、保留设置及外置资产，按完整安装 profile 同步依赖并重启。

0.17.0 推荐配置与生效检查：

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、视频文件及新增返回过渡动画。Google Drive 下载需取齐 28 个分卷及辅助文件，用 Assemble-Kurisu.ps1 合并、校验后再安装，不能直接安装单个分卷。在项目根目录使用已有环境执行：

   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"

   覆盖旧角色包前先备份自定义素材，再给 install 命令添加 --overwrite。

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。标准、自动帧源/预算和磁盘复用本就是默认配置，纹理采样在 60 FPS 下默认开启；旧用户保存过的静态帧或关闭选项会保留，需要自行核对。

3. 重启后端并重新打开角色/壁纸，查看当前实际动画帧源是否为视频与静态入口帧、预算是否为 384 MiB。只选了“自动”不代表已经使用视频；缺少匹配素材或解码能力时仍使用静态帧，自动预算为 2 GiB。帧预算限制的是保留的帧数据，不等于整个应用或解码器的总内存/显存。

4. 要使用本地 NVIDIA 语音优化，需安装 L4 本地 CUDA 环境及 Qwen/GPT-SoVITS 资产。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），但相关加速路径仍需另行安装兼容扩展且硬件支持。修改后重启后端，通过启动诊断确认实际路径；远程语音和 CPU 配置不享有这些 CUDA 优化。

本次仅补充现有发布的说明，tag、源码 ZIP、清单及校验和均不变，ZIP 内的 README 仍为发布时快照。后续文档更新请查看主线中文指南或英文指南。