This build_prompt function unconditionally appends "Think step
by step." to every prompt — even when the model is a reasoning
model (o3, Claude with thinking, etc.). On reasoning models, this
is at best wasted tokens and at worst hurts answer quality.
Fix the function to ONLY append the CoT prefix when the model is
NOT in REASONING_MODELS.
Expected output:
o3 cot? False
claude-haiku cot? True
The break is on line 5 — but read the whole snippet first.