Subagent started with unsupported model fails the whole session

Describe the bug

In a long running agent session on autopilot (~2 hours), with plenty use of subagents, the main agent was Opus 5, but started gemini-3.6-flash when building and executing tasks. Later during that same autopilot session gemini-3.6-flash was no longer supported. The main agent attempts to launch gemini-3.6-flash agent again, which results in this error:

Execution failed: CAPIError: 400 The requested model is not supported.

This breaks the entire session and the main agent stops. Instead the main agent should have seen that error, been able to react on it and try to find a different model it could use (e.g. Haiku, GPT Luna), and launch the subagent

Affected version

GitHub Copilot CLI 1.0.79.

Steps to reproduce the behavior

1. Have a session running where you instruct the main agent to launch subagents using a model that you are going to retract from the pool of models
2. On your side, change which models are supported to ban a model that you made the session use
3. See the subagent fail and the main agent fail together

Expected behavior

The main agent should have seen that error, been able to react on it and try to find a different model it could use (e.g. Haiku, GPT Luna), and launch the subagent

Additional context

No response

首先跟踪 CLI 如何启动 subagents 并传播模型执行错误。没有指定文件或测试，因此先定位 session 和 subagent 的错误处理入口；完成的标准是，不受支持的 subagent 模型失败不再终止 main session，并且可以针对另一个模型进行处理。