GPT-5.6 Terra delegates to Opus subagent

Describe the bug

Asked Copilot to create a plan (from the IntelliJ plugin that is backed by the Copilot CLI now). Configured the model to gpt-5.6-terra. In the billing I noticed there was quite a big amount of credits spent on Opus (a model I rarely select so that triggered me). After looking at lots of sessions couldn't find myself using Opus.

Digging deeper I found one refactor that ran for quite some time. Started with a plan and then went to implementation as well. In the logs as presented in IntelliJ it's impossible to tell which models were used, I've installed the CLI and resumed the session there (this makes I'm not sure which CLI version was used initially). That history is better and contains opus references (so it wasn't me selecting the wrong model 😌 ).

But now I'm trying to understand why it selected opus for the job, thats a way more expensive model than the one I had selected for the job.

Learning the CLI now .. looking at the config I'd expect the code-review agent to inherit from the parent task (that would be Terra

Subagent         Origin    Model                              Overridden  Status

❯ explore          built-in  claude-haiku-4.5 (default model)   No          On
  task             built-in  claude-haiku-4.5 (default model)   No          On
  general-purpose  built-in  inherit (default behavior)         No          On
  rubber-duck      built-in  complementary (different family)   No          On
  code-review      built-in  inherit (default behavior)         No          On

Asked Terra what happened there and that scares me a bit,

There is no recorded rationale for why it selected Opus. It was an autonomous model choice made while producing the tool call, despite no prompt or instruction requiring it.

Is it really true that the model (Terra) selected Opus for the sub agents and ignored the Copilot settings to inherit the parent model? If my understanding is correct it is at least funny that it selected a model from a competitor 🙈

Affected version

"copilotVersion": "unknown", (from the logs) IntelliJ plugin version 1.14.2-261-macos-arm64 (assume the a cli is shipped with that as I didn't manually install before)

Steps to reproduce the behavior

No response

Expected behavior

I'd expect the sub agent to use the configured model, inherit Terra in this case.

Additional context

Had the Copilot cli log analyzed by Luna:

opus.jsonl is a copilot CLI log file where it turned out to have used Opus as the model voor some actions. Can you explain where it used Opus voor and also why because the chosen model by me as a user was GPT Terra

Opus (claude-opus-4.8) was used only for delegated subagents:

Planning: 16:52–16:54
General-purpose implementation/reasoning: 16:55–17:10, 17:10–17:48, and 18:05–18:27
Codebase exploration: 17:41–17:43
Code review: 17:48–18:05 and 18:54–19:00
There were 7 Opus subagent invocations. The primary assistant’s messages and normal tool calls remained on Terra.

Why this happened: the selected model applies to the main agent, while delegated subagents are routed independently. The CLI assigned different models to subagents, including Terra, Opus, and Sonnet. The log does not contain a reason or fallback event explaining the individual choices, and it does not show that you explicitly selected Opus. The most likely explanation is automatic subagent model routing or agent-specific defaults, especially for planning, exploration, and code-review agents.

首先使用 gpt-5.6-terra 重现委派，并结合 subagent 配置表检查 CLI 会话日志。将已配置的继承行为与规划、探索、实现和代码审查所记录的模型进行比较；当路由得到修正，或模型选择行为得到清晰记录时，即视为完成。