Background compaction consumes 1 premium request (VS Code does not)

[HA-Roy]

Describe the bug

Background compaction in Copilot CLI incorrectly consumes 1 premium request each time it triggers. The VS Code Copilot Chat extension performs the same compaction/summarization operation without consuming
a premium request.

Observations

• The VS Code extension handles compaction/summarization without consuming premium requests. It uses a caller-controlled boolean to mark requests as agent-initiated.
• The CLI's compaction appears to be incorrectly classified as a user-initiated request rather than an agent/background operation.
• Session naming (also a background operation in CLI) does NOT consume a premium request — only compaction does.
• This means long sessions with premium models consume significantly more premium requests than expected due to repeated compaction cycles.

Affected version

Copilot CLI v1.0.4 and v1.0.5

Steps to reproduce the behavior

1. Start a Copilot CLI session with a premium model (e.g., Claude Sonnet 4)
2. Have a long conversation until context reaches ~80% utilization
3. Send another message — this triggers background compaction
4. Check the premium request counter on the GitHub billing dashboard

Expected behavior

Background compaction should be free, consistent with VS Code behavior.

Additional context

No response