Choosing and checking a model

Nothing about specific models is built in. On first use, the harness probes each model: its context size, whether it can return JSON or call tools, whether its thinking can be turned off, and how fast it is. Values you set in a profile override the probe. Results are cached in .cache/probe/.

Practical advice:

• A coding model with 16K context or more, instruction-tuned, at a quantization that still follows JSON reliably.
• If a reasoning model spends its whole budget thinking, the harness retries once with thinking off. Setting thinking.control explicitly avoids the retry.
• Compare models with the evaluation: the same ten tasks, one table per profile mapping.