Where does the bug appear (feature/product)?
Cursor CLI
Describe the Bug
The Task tool’s model parameter only accepts "fast". Passing any specific model ID (e.g. claude-sonnet-4-6, claude-opus-4-6) returns:
Invalid model selection. Available models are: - fast
This contradicts the subagents documentation which states the model field supports three options: inherit, fast, or a specific model ID like claude-4-sonnet or claude-opus-4-6.
Custom subagent frontmatter with model: claude-sonnet-4-6 is also ignored — the tool schema enforces enum: ["fast"] at the parameter level, so the agent cannot even attempt to pass the value.
Steps to Reproduce
- Create a custom subagent at
~/.cursor/agents/test-model.md:
---
name: test-model
model: claude-sonnet-4-6
description: Test model selection
---
Say "hello, I am running on claude-sonnet-4-6" and stop.
-
In an agent conversation, ask the agent to dispatch a Task with
model: "claude-sonnet-4-6". -
The tool returns the error:
Invalid model selection. Available models are: - fast -
With
model: "fast", the subagent runs but self-reports as Sonnet 4.5, not the requested Sonnet 4.6. -
Without the
modelparameter, the subagent inherits the parent model — there is no way to target a middle-tier model.
Expected Behavior
The Task tool should accept specific model IDs as documented, e.g. claude-sonnet-4-6, claude-opus-4-6, matching the model configuration docs and the agent frontmatter spec.
Error output from Task tool:
Error: Invalid arguments:
model: Invalid model selection. Available models are:
- fast: Extremely fast, moderately intelligent model...
Tool schema (as presented to the agent):
"model": {
"description": "Optional model to use for this agent...",
"enum": ["fast"],
"type": "string"
}
Operating System
MacOS
Version Information
- Cursor Agent (CLI): 2026.03.30-a5d3e17
- Cursor IDE: 2.6.22 (c6285feaba0ad62603f7c22e72f0a170dc8415a0)
For AI issues: which model did you use?
claude-4.6-opus (parent agent), attempting to dispatch claude-sonnet-4-6 (subagent)
Additional Information
it prevents a key workflow. Orchestrator-pattern agents (documented in the subagents docs) rely on dispatching subagents on a middle-tier model for cost/capability balance. The only options today are fast (too weak for complex tasks) or omitting the param (inherits Opus — too expensive for routine work).
Does this stop you from using Cursor
Sometimes - I can sometimes use Cursor