Tuning Qwen 3.8 27B and OMP as a coding agent on 2× 3090s

A user successfully tuned Qwen 3.8 27B and Oh My Pi (OMP) as a coding agent on two 3090 GPUs, reducing the average latency per turn from 28 seconds to 7 seconds. This significant performance improvement was achieved by adjusting OMP settings, including: setting clear effort levels for each role, setting thinking_token_budget to 7500, increasing maxTokens from 8k to 32k to prevent file write truncation, directing tool output over 10 KB to files, and configuring up to 4 sub-agents with appendOnlyContext enabled.