@type t() :: %LlmComposer.LlmResponse{
  cached_tokens: non_neg_integer() | nil,
  cost_info: LlmComposer.CostInfo.t() | nil,
  input_tokens: non_neg_integer() | nil,
  main_response: LlmComposer.Message.t() | nil,
  metadata: map(),
  output_tokens: non_neg_integer() | nil,
  previous_response: map() | nil,
  provider: provider(),
  provider_model: String.t() | nil,
  raw: any(),
  reasoning_tokens: non_neg_integer() | nil,
  response_id: String.t() | nil,
  status: :ok | :error,
  stream: nil | Enumerable.t()
}

Normalized response from any LLM provider.

• :status — :ok or :error.
• :main_response — the primary assistant Message.t().
• :input_tokens — number of input tokens consumed.
• :output_tokens — number of output tokens generated.
• :cached_tokens — number of tokens served from provider cache.
• :reasoning_tokens — tokens used for internal reasoning (where supported).
• :provider — atom identifying the provider (see provider/0 type).
• :provider_model — model identifier string as reported by the provider.
• :response_id — provider-assigned response ID, used for multi-turn continuations.
• :previous_response — raw previous response map, used by stateful APIs (e.g. OpenAI Responses).
• :stream — enumerable of LlmComposer.StreamChunk.t() when streaming is enabled, otherwise nil.
• :cost_info — CostInfo.t() with cost breakdown when :track_costs is enabled.
• :metadata — arbitrary map for provider-specific extra data.
• :raw — original decoded provider payload, useful for debugging.