Subagents set to Grok 4.7 silently fall back to Auto (Composer); Grok 4.6 works

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Bugbot / Security Review Agent

Describe the Bug

Subagents that should run on Grok 4.7 silently run on Auto instead, which picks Composer.

It happens with both ways of choosing the subagent model:

  • the parent agent passes model: "grok-4.7-xhigh-fast" to the Task tool (this slug is in the available subagent model list), or
  • Grok 4.7 is selected in Settings → Models → Task Models → Explore Subagent Model.

There is no error. The subagent card in the chat shows “Auto”, and the subagent’s own system prompt says “You are an AI coding assistant, powered by Composer.” Auto and Composer 2.5 are both disabled in my model list.

With Grok 4.6 the exact same steps work: the subagent’s system prompt says “powered by Cursor Grok 4.6”. My workflow used Grok 4.6 subagents without problems and broke as soon as I switched them to Grok 4.7.

Test results (parent agent: Claude Opus 5.5; each subagent was asked to quote the sentence in its system prompt that names its model):

How the subagent model was chosen subagent_type Model it actually ran on
Task tool model: "grok-4.7-xhigh-fast" explore Composer (Auto)
Task tool model: "grok-4.7-xhigh-fast" generalPurpose Composer (Auto)
Setting “Grok 4.7 Extra High Fast” + Task model: "inherit" explore Composer (Auto)
Setting “Grok 4.7 Extra High” + Task model: "inherit" explore Composer (Auto)
Setting “Grok 4.7 High” + Task model: "inherit" explore Composer (Auto)
Task tool model: "cursor-grok-4.6-xhigh" explore Cursor Grok 4.6 (correct)
Setting “Grok 4.6 Extra High Fast” + Task model: "inherit" explore Cursor Grok 4.6 (correct)

The Grok 4.7 slug itself is accepted: passing an unknown slug (invalid-slug-probe) fails with “Invalid model selection”, and the allowed list in that error includes grok-4.7-xhigh-fast. So the Grok 4.7 request is accepted and then silently replaced by Auto.

Steps to Reproduce

  1. Settings → Models → Task Models → set “Explore Subagent Model” to a Grok 4.7 option (I tried Extra High Fast, Extra High and High).
  2. In an Agent chat, have the agent launch an explore subagent with the Task tool, either without model / with model: "inherit", or with model: "grok-4.7-xhigh-fast".
  3. In the subagent prompt, ask: “Quote verbatim the sentence in your system prompt that says which model powers you.”
  4. Result: the subagent card shows “Auto” and the subagent replies “You are an AI coding assistant, powered by Composer.”
  5. Change the setting to “Grok 4.6 Extra High Fast” (or pass model: "cursor-grok-4.6-xhigh") and repeat: the subagent replies “You are an AI coding assistant, powered by Cursor Grok 4.6.”

Expected Behavior

The subagent runs on the Grok 4.7 model I selected. If Grok 4.7 can’t be used for subagents, I’d expect a clear error (the Task call failing, or a warning in Settings) instead of a silent fallback to Auto/Composer, especially since Auto and Composer are disabled in my model list.

Operating System

Windows 10/11

Version Information

Version: 3.21.18 (system setup)
VS Code Extension API: 1.128.0
Commit: c4730f7d93d787d9ab120af715999f0345ee5bc0
Date: 2026-09-22T02:04:34.641Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 42.10.0
Chromium: 148.0.7778.280
Node.js: 24.18.1
V8: 14.8.178.38-electron.0
xterm.js: 6.1.0-beta.291
OS: Windows_NT x64 10.0.26200

For AI issues: which model did you use?

Parent agent: Claude Opus 5.5. Subagent: Grok 4.7 (Extra High Fast / Extra High / High).

Additional Information

  • Setup: Cursor desktop on Windows, connected via Remote-SSH to a Linux x64 server.
  • Started on Sep 22, 2026, when grok-4.7-xhigh-fast appeared in the subagent model list. Before that my subagents used cursor-grok-4.6-xhigh-fast and worked fine.
  • Changing the Explore Subagent Model setting applies immediately (switching it to Grok 4.6 took effect on the very next subagent), so the Grok 4.7 results are not caused by a stale setting.
  • Looking at the bundled cursor-agent-exec extension code, the subagent model resolution does pick grok-4.7-xhigh-fast and no forced subagent model policy is active, so the fallback to Auto seems to happen later, when the subagent request is actually sent or served.

Does this stop you from using Cursor

No - Cursor works, but with this issue

5

See Cursor Live Task Allow List does not include Grok 4.7

8

Hey @Doer, this is an issue we’re tracking.

Two things that work today:

  1. Pin your subagents to a Grok 4.6 option (Settings → Models → Task Models → Explore Subagent Model, or cursor-grok-4.6-xhigh in the Task call). They’ll stay on the model you pick.
  2. If you can run the work as a Cloud Agent, its subagents can already be pinned to Grok 4.7.

The same thing is being followed in Cursor Live Task Allow List does not include Grok 4.7, so I’ll post updates there.

9

Thanks Mohit. Do you have a rough ETA for when Grok 4.7 will be on the Task allow list? Even a ballpark (this week, next sprint, or no date yet) would help.

11

Hey @Doer,

Thanks for the report and the test matrix.

This has been addressed on our side and will be rolled out soon if it hasn’t been already. The change is server-side, so no app update is needed: subagents set to Grok 4.7 (through the Task tool’s model argument or the Task Models setting) should now run on Grok 4.7 instead of silently falling back to Auto.

Could you give it another try and let me know if a Grok 4.7 subagent still ends up on Auto?