v0.13.0

Minor Changes

• d58a55f: Derive reasoning levels from the endpoint's supported_efforts.

  thinkingLevelMap is now built by identity from each model's
  metadata.reasoning.supported_efforts (plus mandatory) via a single
  buildThinkingLevelMap helper, instead of hand-tuned per-family constants.
  Early-access models read the live metadata.reasoning block through the same
  helper. default_effort and effort_aliases are no longer consumed.

  Behavior changes for the public catalog:

  • DeepSeek V4 Flash: low is no longer exposed (the API aliases it to high;
    identity mapping drops it).
  • Kimi K3: off is now exposed (mandatory: false upstream).
  • Kimi K2.7 Code: high-only fallback map, unchanged in effect.