Question: Does model: in agent frontmatter and task() calls control model selection? #2564
Description
Question
Does the model: property in .agent.md frontmatter and the model parameter in task() calls actually control which model is used?
Observed Behavior
Testing with Copilot CLI v1.0.19:
| Context | model: specified | Actual model |
|---|---|---|
Agent frontmatter model: claude-haiku-4.5 |
haiku 4.5 | Sonnet 4.6 (session default) |
Agent frontmatter model: claude-opus-4.6 |
opus 4.6 | Sonnet 4.6 (session default) |
task() parameter model: "claude-haiku-4.5" |
haiku 4.5 | Sonnet (per sub-agent self-report) |
All agents and sub-agents appear to run on the session default model regardless of explicit model configuration.
Expected Behavior
Per the custom agents configuration docs, model is a supported property. We expected it to route the agent to the specified model.
Questions
- Is
model:in frontmatter intended to be honored? Or is it overridden by auto-selection? - Does
modelintask()calls route sub-agents to different models? - What is the correct way for a plugin to request different model tiers for different agents (e.g., haiku for search, opus for architecture review)?
- Does the user need a specific Copilot plan tier for multi-model agent routing to work?
Context
Building a multi-agent orchestration plugin (28 agents) that assigns different model tiers based on task complexity. The model routing appears to have no effect.
Related: #2563 (tools: enforcement)