Had the same experience using it for a refactor of a 3k LOC monolith via the Pi harness and OpenRouter. After burning through $8 worth of tokens it left the code in a broken state, the "thoughts" were full of loops where it would edit the monolith, then refer back to the original file, not finding it and then overwriting its changes with "git checkout --"

It's probably bad harness. I had a similar bad experience with qwen max yesterday also through opencode.

In the past I tried Kimi thru Claude code I might try that again