Broken Cache Management
Author: Metal-Gear-SolidCreated Sep 16, 2026Updated Sep 18, 2026
OMLX 0.7.0dev2
The moment the context window lands on 430K used, every single message becomes a full prefill reprocess from scratch.
It happens with Deepseek v4, Qwen 3.8 flash next, and GLM 5.3.
Using BYOK on VSCode Github Extension CoPilot
This has been the issue for at least 2 months and using the full 1M context is impossible, M# Ultra 512GB.
Source: jundot/omlx