Agent mode selected but model still receives Multitask Mode system_reminder

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Summary

Composer UI shows Agent mode, but the model is still injected with a Multitask Mode <system_reminder> that forces background subagent delegation for all tool use. This overrides project agent rules and normal Agent behavior even though Multitask is not selected.

Local state check (for triage)

On disk, composer.composerHeaders in
~/Library/Application Support/Cursor/User/globalStorage/state.vscdb
showed unifiedMode: agent for composers; none had unifiedMode: multitask. So the sticky Multitask reminder does not appear to match persisted UI mode — likely request-pipeline / in-memory injection, not the mode file the picker writes.

Impact

  • Breaks established Implementer-in-chat workflows (e.g. CLAUDE.md / AGENTS.md)
  • User cannot clear the injection from chat or repo rules; only new chat / restart seems reliable
  • Platform reminder outranks explicit user + project instructions

Environment

  • App: Cursor (macOS)
  • Mode UI: Agent
  • Model: Cursor Grok 4.5 High Fast
  • Happy to provide Cursor version / request IDs if you say where to find them.

Steps to Reproduce

Repro (approximate)

  1. Use Agent mode in a long-running chat.
  2. At some point use background/subagent Task tooling and/or discuss separate reviewer agents (without selecting Multitask in the mode menu).
  3. Continue in Agent mode — observe model still following Multitask delegation rules / admitting Multitask reminder is present.
  4. Confirm UI still shows Agent, not Multitask.

Expected Behavior

Expected

When the mode picker is set to Agent, the model should get Agent instructions only — not Multitask coordinator / “MUST DELEGATE” instructions.

Actual

Across many turns in the same chat:

  • Mode chip / dropdown: Agent (checkmark on Agent; Multitask only listed as an option)
  • Model context still included a system reminder roughly:
    “The user has engaged Multitask Mode… YOU MUST DELEGATE TO AN ASYNCHRONOUS SUBAGENT ANY TIME YOU NEED TO USE ANY TOOLS.”
  • Assistant then spawned background Task subagents for ordinary work (reads, commits, docs), contrary to repo workflow and user instructions.

Asking for a separate Reviewer session (normal multi-agent workflow via docs) did not intentionally enable Multitask. Switching the UI to Agent did not clear the injection for that conversation.

Operating System

MacOS

Version Information

Version: 3.9.16
VS Code Extension API: 1.105.1
Commit: 042b3c1a4c53f2c3808067f519fbfc67b72cad80
Date: 2026-06-27T06:41:01.941Z
Layout: editor
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Darwin arm64 25.5.0

For AI issues: which model did you use?

Grok 4.5

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey @Michael_Lewis, thanks for the detailed report.

I reproduced your flow on 3.13.25 (current stable; you’re on 3.9.16): the Multitask → Agent switch now injects an explicit “Multitask exited, proceed normally” reminder, and the model behaved correctly as Agent afterwards (read files inline, no subagent). So the transition itself looks improved on the current build.

Could you update to 3.13.25 and see if it still repros? If it does, a few things would help narrow it down:

  1. Roughly how many turns in when it hit, and had the chat been going long enough to hit summarization?
  2. How did you switch back to Agent: mode picker, the ✕ on the purple Multitask pill, or a slash command?
  3. Does it still repro on 3.13.25 with Auto, or only with Grok 4.5?
  4. Does your state.vscdb still show composer.composerHeaders.unifiedMode = "agent" on 3.13.25?
  5. A request ID from a bad turn if you can grab one (three-dot menu on the message → Copy Request ID).

If you’re stuck mid-chat on 3.9.16, clicking the small ✕ on the purple Multitask pill is stronger than picking Agent from the mode picker.