Works much worse since gpt 5.4 and codex performs remarkably better with the same model

### Describe the bug

Basically performs much worse on gpt 5.4 than it ever did on gpt 5.3-codex, additionally running on xhigh. For some reason codex performs way better with gpt 5.4 than copilot cli. Especially difficult is that it cannot even properly switch from planing to autopilot. But there are many other issues it is generally breaking off tasks often before finishing and so on.

### Affected version

GitHub Copilot CLI 1.0.11

### Steps to reproduce the behavior

_No response_

### Expected behavior

_No response_

### Additional context

_No response_

A issue menciona o GitHub Copilot CLI 1.0.11, mas não fornece arquivos, testes, pontos de entrada ou etapas de reprodução. Comece reproduzindo o comportamento relatado do GPT 5.4 e comparando-o com o GPT 5.3-codex e o Codex. O trabalho estará concluído quando a transição do planejamento para o autopilot e o comportamento de conclusão de tarefas forem identificados e verificados com um caso de teste concreto.

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