Copilot CLI loses conversation context quickly when using Claude Sonnet 4.6

### Describe the bug

I'm experiencing what appears to be excessive context loss in Copilot CLI when using Claude Sonnet 4.6.

Within a single session, after only a few exchanges, the model often forgets information, requirements, or instructions that were provided earlier in the conversation. This results in having to repeatedly restate context that was already discussed.

The model should retain awareness of previously discussed context throughout an active session and use that information when answering follow-up questions.

The model frequently loses track of earlier conversation details after only a handful of prompts and responds as if the context was never provided.

I first noticed this behavior around the time GitHub transitioned to AI Credits-based billing. I'm not sure whether the two are related, but the issue is noticeable enough that it impacts the usability of Copilot CLI for longer conversations.

### Affected version

v1.0.57

### Steps to reproduce the behavior

_No response_

### Expected behavior

_No response_

### Additional context

_No response_

No reproduction steps, files, or tests are provided. Start by reproducing the context loss in Copilot CLI v1.0.57 with Claude Sonnet 4.6 across several exchanges, then document the exact prompt sequence and confirm that context remains available throughout the session.

由索引模型根據 Issue 內容生成。