Claude Desktop "Launch preview panel" leaks across chats — file edits in one chat hijack the preview in unrelated active chats

Summary

Claude Desktop appears to host a single shared Launch preview panel (the side panel that auto-renders files after Edit/Write tool calls). When an agent in chat A edits a file, the preview surfaces in whichever Claude Desktop chat or window the user currently has in focus — including unrelated chats B, C, etc. that the user is actively working in.

This is disruptive enough that the user repeatedly interrupted their session to call it out, including describing it as "totally inappropriate" when the preview popped up in a completely different conversation than the one driving the edits.

Where the hook surfaces

After every Edit or Write, the tool returns a <system-reminder> containing:

PostToolUse:Edit hook additional context: <path> is now visible in the Launch preview panel.

The system-reminder makes it sound like a per-chat side effect, but the rendered panel itself follows the user's foreground Claude Desktop window, not the chat that triggered the edit.

Steps to reproduce

1. Open two or more Claude Desktop chats in parallel.
2. In chat A, give an agent a task that involves repeated Edit operations on a renderable file (HTML, Markdown, image).
3. Switch focus to chat B and work there.
4. Whenever chat A's agent edits the file, chat B's Launch preview panel also displays the file — even though chat B has no relationship to that work.

Expected behavior

The preview panel should be scoped to the chat that triggered the edit. At minimum, edits made by an agent in chat A should never push a preview into chat B's UI while the user is actively focused there.

Why this matters

1. Disruption. The user has to dismiss / context-switch every time the preview leaks in.
2. Confidentiality risk. A file being edited as part of one project (e.g. a private codebase) gets rendered inside a chat that may be about an unrelated subject, in front of whatever the user is looking at.
3. Agent UX feedback loop. Agents instructed to "work independently" naturally end up editing files; if those edits keep poking the user in other chats, the user has no recourse beyond telling each agent to "minimize file edits" — but agents need to edit files to do work.

Workarounds the agent currently has

• Batch edits to the same file (e.g. avoid cache-buster bumps that touch HTML repeatedly).
• Move build-time concerns (cache busters) into server-rendered context so the template doesn't need to be touched per deploy.
• Avoid Edit-heavy iterative flows on files Claude Desktop renders previews for.

These help, but they're working around the real bug — they don't fix it.

Suggested fixes (no insider knowledge of the implementation)

1. Scope the Launch preview panel to the chat that triggered the edit, not the foreground chat.
2. Provide an explicit per-chat "show last edit in preview" toggle that defaults off when the chat is not foregrounded.
3. At minimum, suppress the preview push when the foreground chat ≠ the chat that issued the edit.

Environment

• Claude Desktop on macOS (Darwin 25.4.0)
• Multiple concurrent chats open in the same Claude Desktop window
• Agent issuing Edit / Write via claude-opus-4-7[1m]

Reported on behalf of a user who saw this happen repeatedly during a long session and asked me (the agent) to escalate.