Skip to main content
Some providers expose a separate “thinking” channel that streams the model’s reasoning metadata alongside the normal message output. This lets you display or inspect the model’s inner monologue without sending it back to the model or to end users who just want the final answer. At the moment, OpenAI Responses is the only built-in provider that supports thinking metadata, but the generated metadata is provider-agnostic so additional providers can adopt it later without it requiring any changes to your code.

​
Enable Thinking

You can enabled thinking for the OpenAI Responses provider by setting the reasoningSummary parameter to OpenAIReasoningSummary.detailed. 
final agent = Agent(
  'openai-responses:gpt-5',
  chatModelOptions: const OpenAIResponsesChatModelOptions(
    reasoningSummary: OpenAIReasoningSummary.detailed,
  ),
);

final result = await agent.send('In one sentence, how does quicksort work?');
final thinking = result.metadata['thinking'] as String?;

if (thinking != null && thinking.isNotEmpty) {
  print('[[${thinking.trim()}]]');
}
print(result.output);
Key points:
  • The reasoning stream is emitted through ChatResult.metadata['thinking']
  • Thinking metadata doesn’t appear in ChatMessage.metadata
  • Metadata is not fed back to the model, so you control where (or if) it is displayed

​
Streaming Thinking

final history = <ChatMessage>[];
var stillThinking = true;
stdout.write('[[');

final thinkingBuffer = StringBuffer();
await for (final chunk in agent.sendStream(
  'In one sentence: how does quicksort work?',
)) {
  final thinking = chunk.metadata['thinking'] as String?;
  final hasThinking = thinking != null && thinking.isNotEmpty;
  final hasText = chunk.output.isNotEmpty;

  if (hasThinking) {
    thinkingBuffer.write(thinking);
    stdout.write(thinking);
  }

  if (hasText) {
    if (stillThinking) {
      stillThinking = false;
      stdout.writeln(']]\n');
    }
    stdout.write(chunk.output);
  }

  history.addAll(chunk.messages);
}

stdout.writeln('\n');
The stream delivers reasoning deltas incrementally, so you can render a live “thought bubble” while the model is working. Each chunk may include text output, thinking metadata, or both.

​
Examples