Unable to spawn subagents with `mai-code-1-flash-picker` when the main agent model is `gpt-5.4` or `gpt-5.5` with `deferTools: never` config

### Describe the bug

# Summary

Spawning a subagent with the `mai-code-1-flash-picker` model appears to break only when `deferTools: never` is set for some MCP server configs.

At first this looked like a model-pairing issue between a main conversation running on `gpt-5.4` / `gpt-5.5` and a subagent using `mai-code-1-flash-picker`. After more testing, the more accurate trigger seems to be the newer `deferTools: never` MCP configuration rather than the subagent model alone.

# Environment

- GitHub Copilot CLI `1.0.64-1`
- Main model: reproduced from `gpt-5.4` / `gpt-5.5`
- Subagent model: `mai-code-1-flash-picker`
- Interface: GitHub Copilot CLI / agent tool flow
- Additional condition: `deferTools: never` enabled for some MCP configs

# Actual Result

When `deferTools: never` is enabled on affected MCP configs, spawning the subagent does not behave correctly. In my testing this showed up as a tool failure such as:

```text
✗ Execution failed: CAPIError: 400 Tool 'tool_search' is not supported with gpt-5. (Request ID: C57D:3931E0:32D4049:3A7FACE:6A367C5C)
```

In some cases the spawn path may also return no useful subagent output instead of a normal success/failure response.

### Steps to reproduce the behavior

1. Configure one or more MCP servers with `deferTools: never`.
2. Start a session using a main model such as `gpt-5.4` or `gpt-5.5`.
3. Try to spawn a subagent with `model: "mai-code-1-flash-picker"`.
4. Observe the failed or empty subagent execution behavior.

### Expected behavior

The subagent flow should either:

1. spawn successfully with `mai-code-1-flash-picker`, or
2. fail immediately with a clear validation or compatibility error that explains the unsupported configuration.

### Additional context

- Before isolating `deferTools: never`, this appeared to be a parent/subagent model compatibility issue.
- The newer finding is that the problematic behavior only occurs when `deferTools: never` is enabled for some MCP configs.
- That suggests the real issue is likely in tool resolution / tool exposure during subagent startup under this MCP configuration path.

thanks

Start by reproducing the GitHub Copilot CLI agent tool flow with an MCP server configured with `deferTools: never`, a `gpt-5.4` or `gpt-5.5` main model, and `mai-code-1-flash-picker` as the subagent. Trace the subagent spawn path around tool resolution and exposure; done means it succeeds or returns a clear compatibility or validation error instead of the reported tool failure or empty output.

Escrita pelo modelo de indexação a partir do texto da issue.