responses.Parameters.Reasoning holds model-specific reasoning settings. The selected adapter translates supported values into its provider’s request.
Given model and ctx:
Effort, Summary, and BudgetTokens do not have universal meanings or accepted values. BudgetTokens is used by adapters that support an explicit reasoning budget. Choose settings supported by the actual provider/model rather than treating the Go struct as a capability list.
Read returned reasoning
Reasoning items are separate from assistant text. A provider may return a summary, opaque/encrypted continuation content, or no reasoning output. A summary is not access to the model’s full private reasoning. When manually continuing a conversation, preserve returned reasoning items using the typed message conversion. RequestIncludableReasoningEncryptedContent only for providers that support it. Do not turn encrypted continuation content into user-visible text.
Higher effort can increase latency and token consumption. Bound total output and request lifetime separately, and inspect provider-reported usage when available.