Repository navigation
中文
概览
v0.4.6 在 v0.4.5 之后增加 OpenAI Realtime 流式语音合成,并修复聊天生成、消息编辑、Live2D 自定义模型激活和 Windows 安装包问题。
- 发布日期:
2026-11-03 - 上一版本:
v0.4.5 - 兼容性: Kokoro Engine
v0.4.6;现有0.4.x内容包兼容范围继续覆盖本版本。 - 隐私: Kokoro 不收集自动远程分析数据;反馈、可用性和外部集成证据均为主动提供或人工记录。
新增功能与体验改进
语音合成
- 新增独立的
openai_realtime语音服务商,可流式播放 Realtime WebSocket 返回的语音,并在设置中选择音色和朗读风格。 - 保留现有 OpenAI Speech API 服务商。Realtime 服务商默认关闭,不会替换当前默认语音服务商。
聊天与会话体验
- 改进待发送消息的编辑定位,按界面显示顺序匹配目标消息。
- 恢复聊天回复内容,并清理生成文本中泄漏的 Markdown 标记。
修复与可靠性改进
- 修复 Realtime 语音请求停止和取消后的连接生命周期,避免旧请求继续占用语音会话。
- 修复 GPT-SoVITS 提示词缺少语言设置的问题。
- 修复 Windows 上 release 优化导致聊天回合卡住的问题,并为 Windows 安装包捆绑 ONNX Runtime。
- 固定 Windows CI 使用的 Rust 工具链,降低安装包构建差异。
- 修复 Windows 路径规范化导致已导入的 Live2D 自定义模型回退到默认模型的问题。
- 消除 release 构建中仅用于 debug 日志的未使用变量警告。
其他更新
- 应用版本从
0.4.5升级到0.4.6,并同步更新 npm、Cargo、Tauri、Nix 和 lock 文件元数据。 - 保留现有 OpenAI Platform API、OpenAI-compatible、Anthropic、Ollama、llama.cpp 和 Codex Runtime 服务商路径。
- Windows、macOS 和 Linux 安装包由
v0.4.6标签对应的 GitHub Actions 工作流构建。
兼容性说明
- 内置角色和 MOD 包使用
>=0.3.1, <0.5.0的引擎兼容范围,覆盖v0.4.6;安装前仍会执行来源、版本、大小、checksum、权限和路径检查。 - 使用
openai_realtime需要配置OPENAI_API_KEY,并连接支持 Realtime WebSocket 的服务端。仅支持 Speech API 的网关无法使用此服务商,失败时不会自动切换到其他服务商。 - Realtime 语音由模型生成,实际措辞可能与输入文本不同。该服务商不使用麦克风,也不会生成或替换聊天回复。
- Codex Runtime 仍需要本机安装兼容版本的 Codex CLI,并继续遵循实验性 app-server 协议的兼容性限制。
- AstrBot 适配器仍要求 AstrBot
>=4.5.0和启用 Bearer 认证且可访问的 Kokoro Webhook;本地模拟测试不等同于真实频道或 marketplace 验收。
验证结果
rtk npm test:83 个测试文件、810 个测试通过。rtk npm run build:通过;Vite 报告已有的 dynamic-import 和 chunk 大小提示。rtk npm run check:ipc:172 个 IPC command names 通过。rtk cargo check --manifest-path src-tauri/Cargo.toml:通过;构建脚本报告本机已找到 ONNX Runtime。rtk cargo clippy --manifest-path src-tauri/Cargo.toml --lib -- -D warnings:通过;构建脚本报告本机已找到 ONNX Runtime。rtk cargo test --manifest-path src-tauri/Cargo.toml --no-run:通过,测试二进制编译完成;本次没有执行 Rust 测试。rtk cargo fmt --manifest-path src-tauri/Cargo.toml -- --check:未通过;检查报告仓库中多个 Rust 文件存在格式差异。rtk git diff --check:通过。- 外部 Registry、AstrBot marketplace 和真实频道验收未纳入本次发布验证;本地验证结果不代表外部发布状态。
完整变更
English
Overview
v0.4.6 adds OpenAI Realtime streaming speech synthesis and fixes chat generation, message editing, Windows Live2D model activation, and Windows installer reliability.
- Release date:
2026-11-03 - Previous release:
v0.4.5 - Compatibility: Kokoro Engine
v0.4.6; existing0.4.xcontent compatibility ranges continue to cover this release. - Privacy: Kokoro does not collect automatic remote analytics. Feedback, usability, and external-integration evidence is opt-in or manually recorded.
New Features and Experience Improvements
Speech synthesis
- Added a separate
openai_realtimeprovider that streams speech from a Realtime WebSocket and exposes voice and reading-style settings. - Kept the existing OpenAI Speech API provider. The Realtime provider is disabled by default and does not replace the current default provider.
Chat and conversation experience
- Improved pending-message editing by resolving the target against the visible message order.
- Restored chat replies and cleaned leaked Markdown markers from generated text.
Fixes and Reliability Improvements
- Fixed Realtime speech request cleanup after stop and cancellation so old requests do not retain a speech session.
- Fixed the missing language setting in GPT-SoVITS prompts.
- Fixed chat turns hanging with release optimization on Windows and bundled ONNX Runtime with Windows installers.
- Pinned the Rust toolchain used by Windows CI to reduce installer build differences.
- Fixed Windows path canonicalization that caused imported Live2D models to fall back to the default model.
- Removed release-only unused-variable warnings for values used by debug logging.
Other Updates
- Bumped the application from
0.4.5to0.4.6, synchronizing npm, Cargo, Tauri, Nix, and lock-file metadata. - Preserved the existing OpenAI Platform API, OpenAI-compatible, Anthropic, Ollama, llama.cpp, and Codex Runtime provider paths.
- Windows, macOS, and Linux installers are built by the GitHub Actions workflows for the
v0.4.6tag.
Compatibility Notes
- Bundled characters and MODs use an engine range of
>=0.3.1, <0.5.0, which coversv0.4.6; source, version, size, checksum, permission, and path checks still run before installation. - The
openai_realtimeprovider requiresOPENAI_API_KEYand a server that supports Realtime WebSockets. Speech API-only gateways cannot use this provider, and failures do not automatically fall back to another provider. - Realtime speech is generated by a model, so wording may vary from the supplied text. The provider does not use a microphone or generate or replace chat replies.
- Codex Runtime still requires a compatible Codex CLI installed locally and remains subject to the compatibility limits of the experimental app-server protocol.
- The AstrBot adapter still requires AstrBot
>=4.5.0and a reachable Kokoro Webhook with Bearer authentication enabled; local mocked tests do not prove a real channel or marketplace acceptance.
Verification
rtk npm test: 83 test files and 810 tests passed.rtk npm run build: passed; Vite reported existing dynamic-import and chunk-size notices.rtk npm run check:ipc: 172 IPC command names passed.rtk cargo check --manifest-path src-tauri/Cargo.toml: passed; the build script reported that local ONNX Runtime was found.rtk cargo clippy --manifest-path src-tauri/Cargo.toml --lib -- -D warnings: passed; the build script reported that local ONNX Runtime was found.rtk cargo test --manifest-path src-tauri/Cargo.toml --no-run: passed; test binaries compiled, but Rust tests were not executed in this release check.rtk cargo fmt --manifest-path src-tauri/Cargo.toml -- --check: failed because the check reported formatting differences across multiple Rust files in the repository.rtk git diff --check: passed.- External Registry, AstrBot marketplace, and real-channel acceptance were not part of this release validation; local verification does not imply external publication.