Subagent Model Grade Selection

Goal

Let the model choose a user-defined model grade when spawning a regular subagent without exposing provider-specific model IDs in the Agent tool schema.

{
  "agents": {
    "modelGrades": {
      "small": "fast",
      "high": "qwen-max"
    },
    "allowedGrades": ["small", "high"]
  }
}

The Agent tool advertises model: "small" | "high" and resolves the selected grade immediately after loading the subagent configuration.

Resolution

The effective model selector uses this priority:

1. A non-built-in agent's explicit, non-inherit model
2. An allowed grade mapped by agents.modelGrades
3. The built-in Explore model setting
4. The inherited parent model

Unknown or disallowed grades are rejected. Forks reject the parameter because they must inherit the parent's model and prompt cache. Named team teammates also reject it because their backend model override accepts concrete model IDs rather than grade selectors.

Only configured, allowed grade names are included in the dynamic tool schema. Concrete model selectors remain private to user settings.

Verification

• Settings schema and CLI-to-core configuration forwarding
• Grade resolution, allowlist filtering, and custom-agent priority
• Dynamic Agent tool schema without concrete model IDs
• Regular foreground and background dispatch using the resolved model
• Fork validation