GPT-5 and GPT-6 keep their reasoning through tool calls on Responses

Agno's OpenAIResponses._using_reasoning_model() now matches IDs starting with gpt-6 next to o3, o4-mini and gpt-5, and OpenRouterResponses matches openai/gpt-5 and openai/gpt-6 next to openai/o3 and openai/o4. That check decides how Agno sends the turn that returns a tool result.

For a reasoning model on OpenAI, Agno stores the response and continues from its previous_response_id, so the API still holds the reasoning item each tool call is tied to. OpenRouter is stateless, so Agno asks for reasoning.encrypted_content and replays that item with the call. When the check missed a model, the follow-up went out with neither and could fail with a missing reasoning-item error.

We built the request after one lookup tool call, with a stored response ID and an encrypted reasoning item on the assistant message, and printed the parameters each class sends.

# Before (real output)
OpenAIResponses(id='gpt-6-sol')._using_reasoning_model() -> False
OpenRouterResponses(id='openai/gpt-5.6-luna')._using_reasoning_model() -> False
OpenRouterResponses(id='openai/gpt-6-sol')._using_reasoning_model() -> False
OpenAIResponses gpt-6-sol follow-up: {'store': None, 'previous_response_id': None, 'include': None}
OpenRouterResponses openai/gpt-6-sol follow-up: {'store': False, 'previous_response_id': None, 'include': None}
 
# Now (real output)
OpenAIResponses(id='gpt-6-sol')._using_reasoning_model() -> True
OpenRouterResponses(id='openai/gpt-5.6-luna')._using_reasoning_model() -> True
OpenRouterResponses(id='openai/gpt-6-sol')._using_reasoning_model() -> True
OpenAIResponses gpt-6-sol follow-up: {'store': True, 'previous_response_id': 'resp_previous', 'include': None}
OpenRouterResponses openai/gpt-6-sol follow-up: {'store': False, 'previous_response_id': None, 'include': ['reasoning.encrypted_content']}

gpt-5 IDs on OpenAIResponses already matched and behave as before.

See the OpenRouter tool use cookbook, and learn more about OpenAI Responses and OpenRouter Responses in the documentation.