Skip to content

1.0.56

Choose a tag to compare

View all tags
@deskwand deskwand released this 01 Nov 03:56
· 31 commits to main since this release
1.0.56
c077c08

🎉 v1.0.56

📦 下载 / Downloads

平台 Platform 文件 File 大小 Size
macOS (Apple Silicon) DeskWand-1.0.56-mac-arm64.dmg 164 MB
Windows (x64) DeskWand-1.0.56-win-x64.exe 160 MB
Linux (AppImage) DeskWand-1.0.56-linux-x86_64.AppImage 221 MB
Linux (deb) DeskWand-1.0.56-linux-amd64.deb 149 MB

应用内自动更新走同一套文件源，无需手动下载。In-app auto-update uses the same file host — no manual download needed.

中文

✨ 新功能

  • 队友（Teammates）：绑定到某个会话的长期运行 Agent。它不是一次性任务 —— 它会被定期唤醒，也会自己安排唤醒（用 schedule_wakeup 工具规划后续动作），每次唤醒以一条 <agent_wakeup> 消息到达。你也可以直接在它所在的会话里跟它说话。
    • Agent 被告知要把长期状态（计划、已处理的事项、结果）写进工作目录里的笔记文件，因为早期上下文会被压缩 —— 这样它在唤醒之间不会「忘事」。
    • 侧栏里队友有头像，并且收在一个可折叠分组里，不会和普通会话混在一起。
  • 总览的上下文拆解更细：系统侧原本挤成一行，现在拆成三行 —— 系统提示词（去掉 <skills> 块）、工具定义、技能块。主进程在 SDK 会话上实测这三部分后以权重下发，渲染层仍负责拆分剩余部分，所以各行的和始终等于总输入。测量发生在当轮工具策略落定之后，因为语音轮次会收窄工具集并追加一段提示词，而 SDK 在两次运行之间会丢掉它。尚未测量过的会话保留原来的合并行与标签。
  • 输入区改版：布局更轻；工作区选择移到会话标题栏；展开按钮取消，编辑器自己长到 15 行 —— 上一版为了「打字时不位移」专门留的常驻槽位，现在连那个按钮和 44px 代价一起省掉了。

🐛 修复

  • 中转（relay）模型的思考等级被强制关闭：以前 reasoning 只从模型 id 的正则推断，而子 Agent 的 provider bridge 又硬编码成 false，于是 pi 把 thinking 钳到 off，界面却还显示你保存的偏好。现在分三层解析：每个模型的显式声明 → pi-ai 注册表条目 → id / 标签关键词。声明会被持久化，并在能力面板里给一个三态开关，你选什么就真的是什么。
  • 总览面板里原本并排的两条上下文进度条合并成一条，不再各说各话。

English

✨ New Features

  • Teammates: A long-running agent bound to one session. This is not a one-shot task — it is woken on a schedule, and by wakeups it schedules itself through the schedule_wakeup tool, each arriving as an <agent_wakeup> message. You can also just talk to it directly in its session.
    • The agent is instructed to keep long-term state — plans, what it has already handled, results — in note files in its working directory, because older context gets compacted. That is what stops it from "forgetting" between wakeups.
    • Teammates get avatars in the sidebar and collect in a collapsible group, instead of mixing in with ordinary sessions.
  • The Context Breakdown Gets Finer: The system side used to be one merged row; it is now three — the system prompt (with the <skills> block removed), tool definitions, and the skills block. Main measures the three on the SDK session and sends them as weights, while the renderer still splits the residual, so the rows always sum to total prompt input. Measurement happens after the turn's tool policy lands, because a voice turn narrows the tool set and appends a prompt section the SDK drops between runs. Sessions that have not been measured keep the merged row and its label.
  • The Composer Is Lightened: the workspace picker moves to the session header, and the expand button is gone — the editor grows to 15 lines on its own, retiring the always-rendered slot (and its 44px) that the previous version added to stop the shift while typing.

🐛 Bug Fixes

  • Reasoning Was Forced Off for Relay Models: reasoning was inferred from the model id regex alone, and the subagent provider bridge hardcoded false, so pi clamped the thinking level to off while the UI still showed your saved preference. It now resolves in three layers — an explicit per-model declaration, then the pi-ai registry entry, then id/label keywords. The declaration is persisted and exposed as a tri-state control in the capability panel, so what you choose is what runs.
  • The two context bars in the overview panel merge into one instead of disagreeing with each other.

101 files changed · +5911 −742
Full Changelog: 1.0.55...1.0.56

Assets 2
Loading