Tarseek API · Pre-production

Reasoning

Reasoning models may use hidden chain-of-thought. Tarseek should expose:

reasoning.enabled — on/off when the model allows.
reasoning.effort or today’s reasoning_effort: low | medium | high.
Usage split: completion_tokens_details.reasoning_tokens.
Whether reasoning text is returned, summarized, or omitted — must be per-model and documented on the capability card.

Partial

reasoning_effort is implemented. There is no reasoning object. Reasoning content visibility is provider-defined and not normalized. Reasoning tokens are not always present in usage.

Unresolved: hide all reasoning by default vs pass through summaries. Do not leak chain-of-thought from providers that forbid it.