akashi

Provider-neutral core of the Akashi agent framework for Dart & Flutter. Pure Dart — no provider SDKs, no build step required.

You almost always pair this with a provider adapter such as akashi_google.

What's in here

• ToolLoopAgent — a streaming tool-loop agent. stream() is the primitive; run() collects over it, so streaming and buffered paths can never diverge.
• Agent<TDeps> — the agent contract (an interface, not just a class), so custom loops (durable, multi-agent) can drop in later without changing callers.
• tool<I, TDeps>(...) — define a tool from a description, a typed input Schema<I>, and an execute callback that receives typed input plus a typed ToolContext<TDeps> (dependency injection).
• Schema<T> — a zero-codegen runtime schema builder. Optional build_runner codegen (akashi_gen, later) emits the same type.
• Sealed unions — Message/Part and AgentEvent for exhaustive switch.
• StopCondition — composable stop conditions (stepCountIs, hasText, hasToolCall).

Example

import 'package:akashi/akashi.dart';

class Deps {
  Deps(this.weather);
  final WeatherApi weather;
}

final getWeather = tool<({String city}), Deps>(
  name: 'get_weather',
  description: 'Current weather for a city.',
  inputSchema: Schema.object(
    {'city': Schema.string(description: 'City name')},
    required: ['city'],
    fromJson: (j) => (city: j['city']! as String),
  ),
  execute: (input, ctx) async {
    final w = await ctx.deps.weather.current(input.city);
    return {'tempC': w.tempC, 'summary': w.summary};
  },
);

// Pair with a provider adapter to get a `LanguageModel`, then:
final agent = ToolLoopAgent<Deps>(
  model: model,
  instructions: 'Help users dress for the weather.',
  tools: [getWeather],
  stopWhen: [stepCountIs(8), hasText()],
);

final result = await agent.run('What should I wear in Bogotá today?',
    deps: Deps(weatherApi));
print(result.text);

Streaming, with an exhaustive switch:

await for (final event in agent.stream('Tell me a 3-sentence story.')) {
  switch (event) {
    case TextDelta(:final text):       stdout.write(text);
    case ToolCallReady(:final call):   print('\n[tool: ${call.toolName}]');
    case RunFinish(:final usage):      print('\n(${usage.totalTokens} tokens)');
    case ErrorEvent(:final error):     stderr.writeln('error: $error');
    default:                            break;
  }
}

Status

v0.3 — multi-agent orchestration, durable execution, and Flutter-native integration have shipped. See the repository roadmap for what's done (v0.1–v0.3) and what's planned (v0.4 multi-modal + Ollama, v0.5 agent-as-MCP-server).

License

MIT.