Skip to main content

​
Migrating from 1.x to 2.0.0

​
Exposing dartantic_interface directly from dartantic_ai

It’s no longer necessary to manually include the dartantic_interface package. 
// OLD - had to import both packages
import 'package:dartantic_ai/dartantic_ai.dart';
import 'package:dartantic_interface/dartantic_interface.dart';

// NEW - one import does it all
import 'package:dartantic_ai/dartantic_ai.dart';

​
Provider Factory Registry

Provider lookup has been moved from the Providers class to Agent static methods. Providers are now created via factory functions not cached instances. 
// OLD
final provider = Providers.get('openai');
final allProviders = Providers.all;
Providers.providerMap['custom'] = MyProvider();
final provider2 = Providers.openai;

// NEW
final provider = Agent.getProvider('openai');
final allProviders = Agent.allProviders;
Agent.providerFactories['custom'] = MyProvider.new;
final provider2 = OpenAIProvider();

​
Moved OpenAI-compat providers to example (except OpenRouter)

Removed the following intrinsic providers from dartantic to the openai_compat.dart example:
  • google-openai
  • together
  • ollama-openai
The openrouter OpenAI-compatible provider remains as an intrinsic provider.

​
Simplified Thinking API

Extended thinking (chain-of-thought reasoning) is now a first-class feature in Dartantic with a simplified, unified API across all providers that support thinking: 
// OLD
final agent = Agent(
  'openai-responses:gpt5',
  chatModelOptions: OpenAIResponsesChatModelOptions(
    reasoningSummary: OpenAIReasoningSummary.detailed,
  ),
);
final thinking = result.metadata['thinking'] as String?;

// NEW
final agent = Agent('openai-responses:gpt5', enableThinking: true);
final thinking = result.thinking;
  • Provider-specific fine-tuning options remain for advanced use cases:
    • GoogleChatModelOptions.thinkingBudgetTokens
    • AnthropicChatOptions.thinkingBudgetTokens
    • OpenAIResponsesChatModelOptions.reasoningSummary

​
Removed ProviderCaps

The ProviderCaps type was removed from the provider implementation and moved to a helper function in the tests. 
// OLD
final visionProviders = Providers.allWith({ProviderCaps.chatVision});

// NEW
// use Provider.listModels() and choose via ModelInfo instead

​
Migrating from 0.9.7 to 1.0.0

This release went through a major refactoring to support many more providers for both chat and embeddings, as well as to support more features. The API is different in a lot of tiny ways, but the core concepts remain largely the same. This is a list of changes (some of them breaking) to help you migrate. And as long as this list is, I’m sure I’ve missed some. If you find something that’s not here, please let me know!

​
Dynamic => Static Provider factories

Provider access has moved to Agent static methods: 
// OLD (0.9.7)
final provider = OpenAiProvider();
final providerFactory = Agent.providers['google'];
final providerFactoryByAlias = Agent.providers['gemini'];

// NEW (2.0.0)
final provider1 = Agent.createProvider('openai');
final provider2 = Agent.createProvider('google');
final provider3 = Agent.createProvider('gemini');
If you’d like to extend the list of providers dynamically at runtime, you can use the providerFactories map on the Agent class: 
Agent.providerFactories['my-provider'] = MyProvider.new;

​
Agent.runXxx => Agent.sendXxx

The Agent.runXxx methods have been renamed for consistency with chat models and the new Chat class: 
// OLD
final result = await agent.run('Hello');
final typedResult = await agent.runFor<T>('Hello', outputSchema: schema);
await for (final chunk in agent.runStream('Hello')) {...}

// NEW
final result = await agent.send('Hello');
final typedResult = await agent.sendFor<T>('Hello', outputSchema: schema);
await for (final chunk in agent.sendStream('Hello')) {...}
Also, when you’re sending a prompt to the agent, instead of passing a list of messages via the messages parameter, you can pass it via the history parameter: 
// OLD
final result = await agent.run('Hello', messages: messages);

// NEW
final result = await agent.send('Hello', history: history);
The subtle difference is that the history is a list of previous messages before the prompt + optional attachments, which forms the new message. Love it or don’t, but it made sense to me at the time…

​
Agent.provider => Agent.forProvider

The Agent.provider constructor has been renamed to Agent.forProvider for clarity: 
// OLD
final agent = Agent.provider(OpenAiProvider());

// NEW
final agent = Agent.forProvider(Agent.createProvider('anthropic'));

​
Message => ChatMessage

The Message type has been renamed to ChatMessage for consistency with chat models: 
// OLD
var messages = <Message>[];
final response = await agent.run('Hello', messages: messages);
messages = response.messages.toList();

// NEW
var history = <ChatMessage>[];
final response = await agent.send('Hello', history: history);
history.addAll(response.messages);

​
toSchema => JsonSchema.create

The toSchema method has been dropped in favor of the built-in JsonSchema.create method for simplicity: 
// OLD
final schema = <String, dynamic>{
  'type': 'object',
  'properties': {
    'town': {'type': 'string'},
    'country': {'type': 'string'},
  },
  'required': ['town', 'country'],
}.toSchema();

// NEW
final schema = JsonSchema.create({
  'type': 'object',
  'properties': {
    'town': {'type': 'string', 'description': 'Name of the town'},
    'country': {'type': 'string', 'description': 'Name of the country'},
  },
  'required': ['town', 'country'],
});

​
systemPrompt + Message.system() => ChatMessage.system()

The systemPrompt parameter has been removed from Agent and model constructors. It was confusing to have both a system prompt and a system message, so I’ve simplified the implementation to use just an optional ChatMessage.system() instead. In practice, you’ll want to keep the system message in the history anyway, so think of this as a “pit of success” thing: 
// OLD
final agent = Agent(
  'openai',
  systemPrompt: 'You are a helpful assistant.',
);
final result = await agent.send('Hello');

// NEW
final agent = Agent('openai');
final result = await agent.send(
  'Hello',
  history: [
    const ChatMessage.system('You are a helpful assistant.'),
  ],
);

​
Agent chat and streaming

The agent now streams new messages as they’re created along with the output: 
final agent = Agent('openai');
final history = <ChatMessage>[];
await for (final chunk in agent.sendStream('Hello', history: history)) {
  // collect text and messages as they're created
  print(chunk.output);
  history.addAll(chunk.messages);
}
If you’d prefer not to collect and track the message history manually, you can use the Chat class to collect messages for you: 
final chat = Chat(Agent('openai'));
await for (final chunk in chat.sendStream('Hello')) {
  print(chunk.output);
}

// chat.history is a list of ChatMessage objects

​
DataPart.file(File) => DataPart.fromFile(XFile)

The DataPart.file constructor has been replaced with DataPart.fromFile to support cross-platform file handling, i.e. the web: 
// OLD
import 'dart:io';

final part = await DataPart.file(File('bio.txt'));

// NEW
import 'package:cross_file/cross_file.dart';

final file = XFile.fromData(
  await File('bio.txt').readAsBytes(), 
  path: 'bio.txt',
);
final part = await DataPart.fromFile(file);

​
Model String Format Enhanced

The model string format has been enhanced to support chat, embeddings and other model names using custom relative URI. This was important to be able to specify the model for chat and embeddings separately: 
// OLD
Agent('openai');
Agent('openai:gpt-4o');
Agent('openai/gpt-4o');

// NEW - all of the above still work plus:
Agent('openai?chat=gpt-4o&embeddings=text-embedding-3-large');

​
Agent.embedXxx

The agent gets new Agent.embedXxx methods for creating embeddings for documents and queries: 
final agent = Agent('openai');
final embedding = await agent.embedQuery('Hello world');
final results = await agent.embedDocuments(['Text 1', 'Text 2']);
final similarity = EmbeddingsModel.cosineSimilarity(e1, e2);
Also, the cosineSimilarity method has been moved to the EmbeddingsModel.

​
Automatic Retry

The agent now supports automatic retry for rate limits and failures: 
final agent = Agent('openai');
final result = await agent.send('Hello!'); // Automatically retries on 429

​
Agent<TOutput>(outputSchema) => sendForXxx<TOutput>(outputSchema)

Instead of putting the output schema on the Agent class, it’s now on the sendForXxx method: 
// OLD
final agent = Agent<Map<String, dynamic>>('openai', outputSchema: ...);
final result = await agent.send('Hello');

// NEW
final agent = Agent('openai');
final result = await agent.sendFor<Map<String, dynamic>>('Hello', outputSchema: ...);
This allows you to be more flexible from message to message.

​
AgentResponse to ChatResult<MyType>

The AgentResponse type has been renamed to ChatResult.

​
DotPrompt Support Removed

The dependency on the dotprompt_dart package has been removed from dartantic_ai. However, you can still use the DotPrompt class to parse .prompt files: 
import 'package:dotprompt_dart/dotprompt_dart.dart';

final dotPrompt = DotPrompt(...);
final prompt = dotPrompt.render();
final agent = Agent(dotPrompt.frontMatter.model!);
await agent.send(prompt);

​
Tool Calls with Typed Output

The Agent.sendForXxx method now supports specifying the output type of the tool call: 
final provider = Agent.createProvider('openai');
// Note: ProviderCaps removed - capabilities are now model-specific

// tools
final agent = Agent.forProvider(
  provider,
  tools: [currentDateTimeTool, temperatureTool, recipeLookupTool],
);

// typed output
final result = await agent.sendFor<TimeAndTemperature>(
  'What is the time and temperature in Portland, OR?',
  outputSchema: TimeAndTemperature.schema,
  outputFromJson: TimeAndTemperature.fromJson,
);

// magic!
print('time: ${result.output.time}');
print('temperature: ${result.output.temperature}');
Unfortunately, not all providers support this feature. You can check the provider’s capabilities to see if it does.

​
ChatMessage Part Helpers

The ChatMessage class has been enhanced with helpers for extracting specific types of parts from a list: 
final message = ChatMessage.system('You are a helpful assistant.');
final text = message.text; // "You are a helpful assistant."
final toolCalls = message.toolCalls; // []
final toolResults = message.toolResults; // []

​
Usage Tracking

The agent now supports usage tracking: 
final result = await agent.send('Hello');
print('Tokens used: ${result.usage.totalTokens}');

​
Logging

The agent now supports logging: 
Agent.loggingOptions = const LoggingOptions(level: LogLevel.ALL);

​
Migrating from 0.9.6 to 0.9.7

​
Breaking Change: DataPart.file → DataPart.stream

The DataPart.file constructor has been removed and replaced with DataPart.stream for file and image attachments. This change was made to enable compatibility with web and WASM platforms, where direct file access is not available.

​
What you need to do

  • Replace all usages of DataPart.file(File(...)) with await DataPart.stream(file.openRead(), name: file.path).
  • Always provide a name argument for best results (especially for web/WASM).

​
Before (0.9.6 and earlier)

final file = File('bio.txt');
final part = await DataPart.file(file);

​
After (0.9.7 and later)

final file = File('bio.txt');
final part = await DataPart.stream(file.openRead(), name: file.path);

​
Why?

  • This approach works on all Dart platforms, including web and WASM, by using streams instead of direct file APIs.
  • It also makes it easier to support file uploads from sources other than the local filesystem.