feat: allow the agent tool to choose the sub-agent model tier per call

[UnderLotus]

Thanks for all the work on Crush.

I wanted to suggest a small follow-up feature around the agent tool.

Right now, the sub-agent seems to use the task agent’s configured model tier (large or small) as a default. It would be really useful if the main agent could optionally override that on a per-call basis when invoking the agent tool.

For example, something like:

{
  \"prompt\": \"Search the codebase for where config is loaded\",
  \"model\": \"small\"
}

or

{
  \"prompt\": \"Do a deeper architectural analysis of the session/agent flow\",
  \"model\": \"large\"
}

with the current task-agent configuration still acting as the fallback when model is omitted.

The main use case is letting the primary agent choose a cheaper/faster sub-agent for lightweight retrieval work, while still being able to use the larger model for more complex reasoning when needed.

This seems like it could fit naturally with the existing large / small model-tier design, without requiring full arbitrary per-call model selection.

Thanks for considering it.