Custom subagent pinned to Grok 4.7 Extra High Fast runs on Opus or is rejected

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Custom subagents pinned to Grok 4.7 Extra High Fast either run on Opus, or are rejected before they start.

There are two failure modes:

  1. If the parent agent launches the subagent without an explicit model argument, and relies on the model pinned in ~/.cursor/agents/NAME.md, the subagent runs on Opus instead of Grok 4.7.

  2. If the parent agent passes the only Grok 4.7 slug that the Task tool lists as available, grok-4.7-xhigh-fast, the launch fails immediately with:

Error: AI Model Not Found
Invalid parameters for registry model: “grok-4.7”

The slug is advertised as a valid Task model in the same session, then rejected when it is used. The error quotes the base id “grok-4.7”, which looks like the client splits grok-4.7-xhigh-fast into a registry model plus parameters (effort=xhigh, fast=true) and the registry rejects that parameter set.

The same error happens after removing context=500k from the agent file, so this is not caused by that context setting.

Steps to Reproduce

  1. Create ~/.cursor/agents/implementer.md with:

name: implementer
model: grok-4.7[effort=xhigh,fast=true]

  1. From a parent agent, launch a Task subagent with subagent_type “implementer” and no model argument.
    Observed: the subagent model is Opus.

  2. Launch the same Task with model “grok-4.7-xhigh-fast”.
    Observed, immediately, no subagent output:

Error: AI Model Not Found
Invalid parameters for registry model: “grok-4.7”

  1. This explicit-slug failure was reproduced three times with the same arguments: twice while the agent file still contained context=500k, and once after that token was removed, leaving model: grok-4.7[effort=xhigh,fast=true].

A later launch that again omitted model (to retest the updated agent-file pin) was cancelled before it returned, so there is no second observation of the Opus fallback after the config edit.

I have not confirmed whether Grok 4.7 Extra High Fast works in a normal non-subagent chat. This report is only about the subagent / Task path.

Expected Behavior

Launching subagent_type implementer or reviewer should use the model pinned in that agent file: Grok 4.7, extra-high effort, fast.

If the parent must pass a model explicitly, model “grok-4.7-xhigh-fast” should start that same variant, because that slug is in the session’s allowed Task model list.

If that variant is not actually entitled, the Task allow-list should not offer the slug, and the error should name the rejected parameter (effort, fast, or something else) instead of only: Invalid parameters for registry model: “grok-4.7”.

A missing or unhonored pin should not silently fall back to Opus.

Operating System

MacOS

Version Information

IDE (read from the installed Cursor.app product.json, not from Help > About):
Version: 3.23.10
VSCode Version: 1.128.0
Commit: e5c99d77d20a6f86647a0a257b866dfe9b116980
Quality: stable
OS: macOS (darwin 27.0.0)

For AI issues: which model did you use?

Requested variant: Grok 4.7 Extra High Fast
Task slug offered in the session allow-list and passed explicitly: grok-4.7-xhigh-fast
Agent file pin: grok-4.7[effort=xhigh,fast=true]
Earlier pin, before the context setting was removed: grok-4.7[effort=xhigh,fast=true,context=500k]
A different allow-list entry, cursor-grok-4.6-xhigh-fast, was not the requested model.

For AI issues: add Request ID with privacy disabled

Not captured. The failure is returned as a Task tool error before the subagent starts, and no Request ID was shown in the parent chat.

Additional Information

What this rules out:

  • Not a bad prompt. The call fails before the subagent starts, and the error names the model id.
  • Not context=500k. The explicit slug still fails after that setting is removed.
  • Not “model missing from the picker.” grok-4.7-xhigh-fast is in the Task tool’s allowed list for the session. The registry then rejects the parameters attached to grok-4.7.

Agent files involved: ~/.cursor/agents/implementer.md and ~/.cursor/agents/reviewer.md.
Date observed: 2026-10-01.

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey @MakeSenseWorks,

  1. Pin runs on Opus: Grok 4.7’s effort option is reasoning_effort, not effort. Update both agent files to:
model: grok-4.7[reasoning_effort=xhigh,fast=true]
  1. “Invalid parameters” error: what Context does Grok 4.7 show in your model picker? If it’s 500K, switch it to 256K, start a new chat, and try again.

This isn’t intended behavior, and we’ve let the team know. Let me know what you see!

The Model parameters section looks out of date: Subagents | Cursor Docs

It says to set reasoning effort with effort, for example claude-opus-5[effort=high]. I configured a custom subagent that way, and invoking it failed with a parameter error.

Could that section be updated so the documented key matches what Cursor now accepts?