Comment by WilcoKruijer

I’ve seen people have success figuring out model reasoning by adding a required `reasoning` parameter to tool calls. Might be worth experimenting with this for the `write` tool call in a coding agent harness.