OpenCode overrides the samplers for Qwen models to the wrong values

This is invisible with llama.cpp or derivatives, but ninfer helpfully logs the sampler settings on each request and auto-configures the correct ones for the model. Basically, OpenCode will always send top-p=1.0 (which means there is no filtering of low-probability tokens except with top-k) instead of the correct 0.95 (thinking) or 0.80 (no-thinking). The commit that added this is supposedly for another fix and has no explanation for the change whatsoever: github.com/anomalyco/opencode/commit/0b132c032aae15a99907a5979f471c3b5bb2e3dc You can't easily fix this per/model provider as far as I can tell, though this works, but it will affect all models/providers. "agent": { "build": { "top_p": 0.95 } }