Context overhead regression: ~20K extra tokens after 1.0.106 (desktop app 1.4.0) #5450

Describe the bug

Summary

After updating to GitHub Copilot desktop app 1.4.0 (CLI 1.0.106), the fixed context overhead per session grew by ~20K tokens overnight. With a 200K window and compaction at ~160K, usable working room dropped from ~112K to ~93K, causing noticeably faster auto-compaction. No changes were made to my MCP servers, skills, or instructions.

Affected version

GitHub Copilot CLI 1.0.95-3.

Steps to reproduce the behavior

No response

Expected behavior

No response

Environment

• Desktop app: 1.4.0 (previous: 1.1.35)
• Copilot CLI/runtime: 1.0.106-8 (previous: 1.0.102-0)
• Model: claude-opus-5.5, default context tier (200K)
• OS: macOS
• User MCP servers: none enabled (GitHub MCP default + additionalToolsets: copilot_automations, githubiq - unchanged since 2026-11-05)
• User skills/instructions: ~2K tokens total (unchanged)