AI: AI conversation history leaks across projects (threads visible globally instead of scoped to active workspace)

[msageryd]

Summary

Description

AI assistant conversation threads created in one project appear in the Assistant panel history when working in other, unrelated projects. This causes cross-project context leakage and makes it easy to apply or resume a thread in the wrong workspace.

Expected behavior

• Assistant conversation history is scoped to the active project/workspace by default.

• Only threads created in the currently open project are shown, unless the user explicitly switches to a global or “All projects” view.

• Applying edits from a thread created in another project should be prevented or gated by a clear warning.

Actual behavior

• Threads created in Project A are listed and accessible when Project B is open.

• It’s possible to resume/edit from a Project A thread while in Project B, leading to irrelevant suggestions and potential unintended edits.

• History UI provides no clear indication that a thread belongs to a different project.

Impact

• High risk of context confusion and accidental edits in the wrong project.

• Productivity hit due to noise from unrelated threads.

• Potential privacy/information leakage: project-specific code snippets, paths, and screenshots are surfaced in other projects.

Steps to reproduce

1. Open Project A in Zed.

2. Create a new Assistant thread and interact (include file references or code from Project A).

3. Close Project A and open Project B (a different repo or workspace).

4. Open the Assistant panel.

5. Observe: the thread from Project A is visible and can be opened/applied within Project B.

Environment

• Zed version: 0.209.6

• OS: MacOS Sequoia 15.6

• AI setup:

  • External agent: Claude Code via API

Additional context

• This is not a feature request for per-project storage; it’s a functional bug: assistant history appears to be globally scoped with no project isolation in the UI.

• Even a temporary mitigation such as filtering by active project or tagging threads with their originating workspace would reduce risk.

• If there is a hidden setting or intended behavior to scope history by project, please point to docs—currently not discoverable.

Proposed mitigations (short-term)

• Default filter in Assistant panel: “This project.”

• Thread metadata badge: [Project name] [path] [last opened].

• Block or warn when applying edits from a thread whose originating project != active project.

Zed Version and System Specs

Zed: v0.209.6 (Zed)
OS: macOS 15.6
Memory: 32 GiB
Architecture: aarch64

[HactarCE]

This is working as intended for now. We'll be changing how some AI things work in the near future, so I'll pass this issue along to be reviewed as part of that.

[hallee9000]

How's the progress? I'm confused by this.