Opencode harness suck or is it just me?
retained text
I tested opencode and codex official. Codex models would comply and follow directions while models, both proprietary and opensource, on Opencode kept intervening and rejecting requests, adding safety caveats and being VERY overly cautious with bio and nuclear stuff. I didn't know harnesses have such a big influence. I used the same model side-by-side a few times to ease out variance, and then switched to a different model, and the result was the same. Models tested, all on highest thinking: Sol 6.1 Astra GLM 5.3 Muse Spark 1.3 Qwen 3.8 Max
document fetched 2026-10-13 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1x3ybdl/opencode_harness_suck_or_is_it_just_me