ClaudeView 1.7.67

The window follows the viewport, and the wheel follows the cursor

Hover-to-scroll actually works now

Scrolling followed focus, not the cursor: you had to click a panel before its wheel worked, then click back. Windows' "scroll inactive windows on hover" doesn't help here, because that setting only picks a top-level window — inside one window the wheel still goes to whichever child HWND holds focus, and the transcript is its own child HWND.

Two separate problems had to be solved:

• A WPF PreviewMouseWheel handler never fires while the cursor is over the chat. Airspace means the child HWND swallows mouse-move, Mouse.DirectlyOver goes stale, and the routed event is raised on nothing — it was hit-testing for exactly the case it couldn't see. Routing now happens at the WM_MOUSEWHEEL level, off the message's own screen point.
• The mirror case can't be handled in the host at all: once you click the transcript, focus sits on a WebView2 HWND owned by the browser process, so those wheel messages are delivered to a different process entirely. The page now catches its own wheel, and when the pointer isn't inside it, hands the delta back for the host to apply to whatever is genuinely under the cursor.

Every combination of focus and hover now scrolls the panel you're pointing at.

The window re-fits itself when the desktop resizes

Sign in over RDP from a laptop, then later from a larger machine, and a maximized window kept the size it was given on the old client — Windows never re-asks for the maximized bounds — so it had to be toggled two or three times by hand. ClaudeView now watches display-settings changes, WM_DISPLAYCHANGE and RDP connect/unlock, settles for a moment while the taskbar re-lays out, and re-fits only when the window really no longer matches the work area. A restored window that no longer fits is pulled back into view; one you deliberately sized and placed is left alone, and a window sitting in the tray is never pulled back onto the screen.

Background agent rows show which model each agent is running

Agents don't inherit the session's model, and Eco's explorer runs a cheaper one on purpose — but every row looked like the session's model. Each row now carries a quiet tag with the model that agent actually resolved to, so a fan-out shows what it's really costing.

The session you're working in appears in Active

Between turns, the session you were sitting in and typing into scored as idle, so the one project you were actually working in was the one missing from the Active tab. It now always lists, at a tier low enough that it can never push a session that needs an answer down the list.

Fixes

• Subagent token usage no longer overwrites the session's own context reading. Whenever an agent was the last thing to speak, the context ring reported the agent's few-thousand-token context as the session total — collapsing precisely when a fan-out had pushed the real number up.
• Maximized re-fit compares position as well as size, so a work area that moves without resizing (taskbar relocated, monitors re-arranged) is caught too.
• Agent SDK 0.3.263 to 0.3.267: prompt-cache stability fixes.