v0.9.0-canary.20261119.29 — 2026-11-19
This Canary introduces a brand-new native agent runtime, queue/steer messaging while the assistant is busy, detachable background tasks, better prompt-cache stability, and an important CJK input fix on Windows.
⚠️ Upgrade Notes
- Removed
{Time}Prompt Placeholder: To keep prompt prefixes stable for provider caching, the{Time}placeholder has been removed (while{Date}remains supported). If any custom prompt still contains{Time}, update it to{Date}in Prompt Manager to clear the unknown placeholder warning. - Cloud Sync Compatibility: Conversations that use queued/steered messages or background task notifications require sync format generation 3. Other devices must update to
.29or later to sync those conversations. - Google Gemini Schemas Split: Legacy Google (Generate Content) (
/v1beta) is restored alongside the new Google (Interactions API) (/v1, used by the built-in Google preset). Review your custom assistant schema if you use a third-party Gemini proxy.
✨ Features & Improvements
- 🔥 Queue & Steer While Busy: You no longer need to wait for the current reply to finish before typing and sending follow-ups. Choose your default behavior in Settings → Chat Window → Default Send Behavior When Busy (Queue to run after the current turn finishes, or Steer to redirect at the next safe step), and press
Ctrl+Enter(Cmd+Enteron macOS) to send in the opposite mode on demand. A pending messages tray lets you switch items between Queue and Steer, withdraw them back into the input box to edit, delete them, or pause/resume after stopping generation. - 🔥 Background Terminal & Subagent Tasks: Long-running Execute In Terminal and Subagent tasks now automatically move to the background when you steer the conversation, keep streaming progress inline, appear in a Running in background tray with individual Stop controls, and wake the conversation automatically once finished.
- 🔥 Native Agent Runtime & Seamless Provider Switching: Rebuilt the core AI runtime directly on
Microsoft.Extensions.AI(Everywhere.Harness), removing Semantic Kernel and its runtime patches. Dialect-scoped reasoning signatures, encrypted reasoning blocks, and tool-call metadata are now preserved across OpenAI, OpenAI Responses, Anthropic, Google Gemini, Mistral, and Ollama—making it safe to switch models mid-conversation without breaking multi-turn tool replay. - Better Prompt Cache Hit Rates: Tool declarations and leading system instructions now stay byte-identical within an active conversation epoch, with mid-conversation changes (such as date rollovers, workspace changes, skills, or toggled tools) delivered as lightweight tail reminders instead of invalidating KV caches.
- Per-Chat Workspaces & Cleaner Foundation: Each new conversation now gets its own dedicated workspace directory instead of sharing a date folder, preventing tool files across same-day chats from colliding. Deferred startup initializers and upgraded
System.Reactiveto 7.0.0 (removing residual WPF/WinForms dependencies on Windows). - Refined Chat Timeline & Turn Actions: Simplified completed activity headers, anchored turn branching and regeneration cleanly to each turn, added subtle Queue / Steered conversation badges on busy-submitted messages, and polished message hover controls.
🐛 Bug Fixes
- Fixed an issue on Windows where certain CJK input methods could produce duplicate characters when committing text or toggling input modes (#277).
- Fixed the active conversation row flickering or losing selection in the history list when sending the first message, and fixed a crash when activating the chat window while model notifications are active.
- Fixed turn-level regeneration targeting when a turn spans multiple assistant steps.
- Fixed duplicate file references in file search and transfer tool previews, and normalized terminal command previews to a single line.
- Fixed broken
CONTRIBUTING.mdand Star History links in the repository READMEs (#489).
🎉 Contributors
- @xionglongztz made their first contribution in #489
本次 Canary 带来了全新的原生 Agent 运行时、回复期间的消息排队与即时引导(Queue / Steer)、可转入后台运行的工具任务、更高的提示词缓存命中率,以及 Windows 中文输入法重复上屏等重要修复。
⚠️ 升级提示
- 移除
{Time}提示词占位符: 为避免秒级时间戳破坏大模型提示词缓存,已移除{Time}占位符({Date}保持可用)。如果自定义提示词中仍包含{Time},请在提示词管理器中将其替换为{Date}。 - 云同步兼容性: 使用了排队/引导发送或后台任务通知的对话需要第 3 代同步格式,其他设备需升级至
.29及以上版本才能同步这类会话。 - Google Gemini 接口方案拆分: 现已将传统的 Google (Generate Content)(
/v1beta)与新的 Google (Interactions API)(/v1,现为内置 Google 预设默认)区分开来;若你使用第三方 Gemini 中转或自定义端点,可按实际接口选择对应方案。
✨ 新特性与优化
- 🔥 回复期间支持「排队」与「引导」: 助手正在思考或执行工具时,无需等待即可继续输入并提交消息。可在 设置 → 聊天窗口 → 忙碌时默认发送方式 中设为 排队(当前轮次结束后发送)或 引导(在下一个安全节点插入),也可用
Ctrl+Enter(macOS 为Cmd+Enter)临时以另一种方式发送。输入框上方的待处理消息托盘支持随时在排队与引导间切换、撤回编辑、删除,或在停止生成后继续/清空剩余消息。 - 🔥 后台终端与子助手任务: 当你在长耗时 在终端中执行 或 指派新助手 任务运行期间发送引导消息时,对应任务会自动转入后台继续运行并实时输出进度;输入框上方提供 后台运行中 任务列表与独立停止按钮,任务完成后会自动唤醒当前对话继续处理。
- 🔥 全新原生 Agent 运行时与跨模型续接: 基于
Microsoft.Extensions.AI重写底层运行时,彻底移除 Semantic Kernel 及相关运行时补丁。推理签名、加密推理块与工具调用元数据现已按协议来源完整持久化,支持在 OpenAI、OpenAI Responses、Anthropic、Google Gemini、Mistral 与 Ollama 之间无缝切换模型继续对话,不再轻易因签名缺失报错。 - 更高的提示词缓存命中率: 在同一活跃对话周期内保持工具声明与起始系统提示词字节完全一致,将日期变更、工作区切换、技能更新或工具开关等中途变化转为尾部轻量提醒,大幅减少 KV Cache 失效。
- 独立会话工作区与底层精简: 每个新对话现已拥有独立的按日期与会话标识隔离的工作区目录,避免同一天的不同对话互相干扰工具文件;优化启动项延迟初始化与退出生命周期,并将
System.Reactive升级至 7.0.0,移除 Windows 构建中残留的 WPF/WinForms 依赖。 - 对话呈现与交互优化: 简化已完成活动的摘要展示,统一轮次分支与重新生成锚点,为忙碌期间发送的消息显示轻量 排队 / 已引导对话 标记,并优化消息操作按钮的悬停显示效果。
🐛 Bug 修复
- 修复 Windows 下部分中日韩输入法在确认输入或切换中英文模式时可能重复输入字符的问题(#277)。
- 修复新建对话发送第一条消息时聊天历史列表条目闪烁或丢失选中状态的问题,以及存在模型通知时激活聊天窗口可能异常崩溃的问题。
- 修复单轮对话包含多个助手步骤时,点击重新生成定位不准确的问题。
- 修复文件搜索与文件传输工具预览中出现重复文件引用的问题,并将终端命令预览统一规范为单行显示。
- 修复 README 中
CONTRIBUTING.md与 Star History 链接错误的问题(#489)。
🎉 贡献者
- 感谢 @xionglongztz 在 #489 完成首次贡献,欢迎加入项目!
Full Changelog: v0.9.0-canary.20261116.28...v0.9.0-canary.20261119.29