Allow specifying the model for the #runSubagent tool in Copilot Chat

[omarisw]

✨ Feature Description

The current implementation of the #runSubagent tool in Copilot Chat uses the model currently selected for the main chat session.

This request is to allow users to explicitly specify the model that should be used when executing a subagent via the #runSubagent tool, similar to how one can select the model for the primary chat.

💡 Why This is Needed (Use Case)

In complex development workflows, particularly involving custom-built Multi-Component Processing (MCP) tools or specialized agents, different models offer varying strengths in terms of context handling, speed, and token consumption.

For example:

• A larger, more context-aware model (e.g., Claude 4.5/5) might be needed for the main reasoning or issue-related tasks.
• A smaller, more efficient model (e.g., gpt-mini) could be used specifically for executing the subagent (MCP tool) to handle a specific, self-contained, or rapid task.

Current Problem: Running a large, "token-hungry" model for all subagent executions can be inefficient and costly, even when the subagent's task is simple.

Proposed Benefit: Providing model flexibility for subagents would allow developers to optimize their workflows for cost, efficiency, and performance by matching the right model to the right tool/task.

---

[phuzzie]

Multi-Component Processing (MCP)

Presumably this is intended to be to "Model Context Protocol (MCP)"?

[phuzzie]

This functionality is key for customers trying to optimize latency & cost and GitHub trying to optimize compute. Its implementation is a win-win. I'm hoping this issue might see some action.

The Claude Opus 4.5 System Card (pg. 23-25) does a fine job laying out the case.

This is going to become more of a need for Claude Opus 4.5 in a few days when the multiplier triples 😕 One expects model pricing will be ever evolving and being able to use more stably priced subagents will help smooth out such pricing changes on premier models.

[roblourens]

You can do this today, define a custom agent with whatever model, enable this setting "chat.customAgentInSubagent.enabled": true, tell the agent to use that agent by name. Note that this is experimental and might take some finesse to get it to successfully pick your custom agent. But if it does it will use that model.