v3.14.0

Added

• Cache- and reasoning-aware costs — total_cost now prices prompt-cache reads, cache writes, and reasoning/thinking tokens via RubyLLM::Cost (1.16) on top of the text input/output cost, instead of text-only pricing. The per-component breakdown is recorded in executions.metadata["cost_breakdown"]. No migration required
• tool_concurrency DSL — Run the tool calls in a single LLM response concurrently. Set per-agent (tool_concurrency :threads, :fibers, true, or false; inheritable) or globally via config.tool_concurrency. Mirrors RubyLLM 1.16's tool concurrency
• HTTP-level latency capture — The instrumentation middleware subscribes to RubyLLM 1.16's request.ruby_llm events and records real provider latency and request count into executions.metadata as llm_request_ms and llm_request_count (distinct from total pipeline duration; retries/fallbacks accumulate)
• New forwarded config knobs — bedrock_api_base, mistral_api_base, perplexity_api_base, vertexai_api_base, xai_api_base, faraday_adapter, deprecation_behavior, and tool_concurrency are forwarded to RubyLLM.config

Changed

• Bumped ruby_llm dependency to >= 1.16.0