v0.2.56

[0.2.56] — 2026-10-02

Added

• Every engine call reports what it cost. roadmodel.usage records the
  token counts each provider returned for the last call: input, the part of
  it served from the prompt cache (and, on Anthropic, written to it), output,
  and reasoning. The hosted recommender meters a call from these instead of a
  fixed estimate. The estimate had drifted: the prompt is ~61k tokens now,
  not the 39.5k it assumed, and a warm prompt cache bills most of it at a
  tenth of the input price.

Fixed

• The gpt-6 line gets the reasoning cap. The OpenAI adapter capped
  reasoning.effort on gpt-5* models only, so GPT-6 Luna, Sol and Astra
  ran at their default effort: slower, dearer, and free to spend the whole
  output budget thinking. Every gpt-5+ model now gets the cap, at low
  where the model has no minimal rung.
• Gemini 3 gets a thinking level. Gemini 3 and later document discrete
  thinking levels, not the 2.5 token budget. thinking_budget now maps to a
  level on those models (0 → low, the floor every Gemini 3 model
  accepts); the 2.5 models keep the numeric budget.
• Claude engines cache the prompt and run at low effort. The Anthropic
  adapter sent the ~61k-token system prompt uncached on every call and could
  not lower reasoning. It now caches the system prompt, so a warm call reads
  it at a tenth of the input price, and maps thinking_budget to
  output_config.effort: low on the models that take effort (Sonnet 4.6+,
  Opus 4.5+, Fable). temperature is forwarded only to the models that
  still accept sampling (Haiku 4.5, Sonnet 4.6, Opus 4.6 and earlier);
  Sonnet 5 and Opus 4.7+ reject it with a 400.