mohdel v0.102.0

Added

• xAI reasoning.effort support (grok-4.3+). xAI introduced a
  parametric reasoning.effort parameter on the Responses API
  (none / low / medium / high) — earlier xAI reasoning
  models had no such control, which is why mohdel previously
  skipped the field on xAI. The adapter now forwards reasoning: { effort } to xAI on the same path used for OpenAI gpt-5.x,
  including the literal 'none' to disable reasoning per xAI's
  documented contract.
• outputEffort: 'none' now reaches the wire. When a model
  spec declares 'none' in thinkingEffortLevels, the adapter
  emits the upstream-appropriate disable signal: OpenAI/xAI →
  reasoning: { effort: 'none' }; Cerebras zai →
  disable_reasoning: true (deprecated toggle, sunset
  2026-07-21); Fireworks zai → reasoning_effort: 'none'.
  Gated on spec.thinkingEffortLevels[effort] != null — models
  without 'none' in their catalog skip the block as before.

Changed

• temperature preserved when reasoning is disabled in the
  chat-completions builder (was unconditionally deleted inside the
  effort block).
• Deps: @anthropic-ai/sdk ^0.91.1 → ^0.95.1, @google/genai ^1.51.0 → ^2.0.0, openai ^6.35.0 → ^6.37.0,
  @opentelemetry/{exporter-trace-otlp-grpc,sdk-node} ^0.216.0 → ^0.217.0, lint-staged ^16.4.0 → ^17.0.3 (dev).
  @google/genai 2.0 breaking changes are scoped to the new
  Interactions API; the gemini adapter is unaffected.

Scope

• gemini and anthropic adapters still guard on effort !== 'none'; their upstream-specific disable shapes are deferred to
  keep this release tight to the four models declaring 'none'.

Tests

• New 'none' unit tests for openai/gpt-5.4, xai/grok-4.3+,
  cerebras/zai-glm-4.7, fireworks/zai-glm-5. The xAI test
  asserting the prior omit-reasoning behavior is updated to
  assert forwarding (now that grok-4.3 accepts the field). Full
  unit suite green.