Local subagents ignore explicit model selection and always run Composer 2.5 Fast

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Custom local subagents ignore the explicitly requested model. When a subagent is launched with model claude-opus-5-thinking-high, Cursor runs it as Composer 2.5 Fast instead. The UI labels the run as Composer, and the subagent quoted its injected system prompt as: “You are Composer, a language model trained by Cursor.” This occurs even when the parent chat is running Claude Opus 5 Thinking High.

Steps to Reproduce

  1. In Cursor IDE, select Claude Opus 5 Thinking High for the parent chat.
  2. Launch a custom local subagent (for example, a general-purpose persona judge) with the explicit model parameter claude-opus-5-thinking-high.
  3. Open the subagent run in the Cursor UI.
  4. Observe that the model is shown as Composer 2.5 Fast.
  5. Ask the subagent to quote its injected system prompt; it reports: “You are Composer, a language model trained by Cursor.”
  6. Restart Cursor and repeat. The result is unchanged.

We reproduced this across multiple Alice, Engineering, and Design judge runs, with explicit model selection and with the parent session on Opus.

Expected Behavior

The local subagent should run the explicitly requested model. If that model is unavailable because of plan limits, admin policy, or another constraint, Cursor should show a clear error or confirmation before using a different model. It should not silently replace the requested model with Composer.

Operating System

MacOS

Version Information

Cursor 3.17.19
Commit/distro: d5c0e77a0214208f36b56d42e8e787de88d02ea4
OS: Darwin arm64 25.5.0

For AI issues: which model did you use?

Requested subagent: Claude Opus 5 Thinking High (claude-opus-5-thinking-high)
Parent chat: Claude Opus 5 Thinking High
Actual subagent: Composer 2.5 Fast

Additional Information

Scope: custom local IDE subagents, not cloud agents. The explicit model argument was passed on every launch. The Cursor UI consistently displayed Composer 2.5 Fast, matching the injected system prompt, so this does not appear to be only a UI-label issue. Restarting Cursor did not change the behavior.

Impact: persona reviews that require a capable model silently run on a different model, invalidating results. The workaround is to create separate top-level Opus chats for each judge, which is slow and requires continuous manual supervision. If local subagents intentionally cannot use the selected model, the product should disclose that limitation before launch.

Does this stop you from using Cursor

No - Cursor works, but with this issue

6

Hey @dsharvit, You’re right, those subagents really are running Composer 2.5 Fast, not just a mislabel.

On request-based (legacy) pricing plans, subagents run on Composer whenever Max Mode is off - turn on Max Mode in the parent chat’s model picker (with Opus 5 Thinking High selected) and they’ll run Opus instead. One caveat: Max Mode bills at token/usage rates, so watch cost on high-volume judge runs. (Your separate top-level Opus chats work for the same reason.)

Agreed that swapping the model silently with no notice is confusing - I’ve passed that along. I’ll follow up here if there’s an update.

Details: Subagents | Cursor Docs