zuraffa_agent #

The agent engine of the Zuraffa ecosystem: missions, sessions, tool and MCP orchestration, an eval harness, and event-driven observability for building type-safe AI agents with Zuraffa clean-architecture conventions.

Getting started #

dependencies:
  zuraffa_agent: ^0.1.0

import 'package:zuraffa_agent/zuraffa_agent.dart';

The engine is organized as inside-out slices over typed ports:

• missions / engine — the mission runner and engine loop with a typed event bus.
• session storage — pluggable persistence (in-memory, JSONL, Hive) with compaction.
• llm — provider configuration over a swappable transport port.
• mcp — Model Context Protocol client plumbing and tool bridging.
• eval — scenario-driven evaluation with graders and reports.
• di — Zuraffa-style composition-root registration.

See the zuraffa package for the underlying UseCase/Result/DI foundations.

Development #

dart pub get
dart run build_runner build --delete-conflicting-outputs
dart test
dart analyze

License #

BSD-3-Clause. See LICENSE and NOTICE.

Configuring the engine #

ZuraffaConfig aggregates the six configuration sections the engine consumes (provider, agent spec, engine loop, MCP transport, stop policy, compaction strategy). Load it from a YAML document or an environment map, validate it at startup, and hand it to MissionRunner — an invalid configuration refuses to start before the first turn:

import 'package:zuraffa_agent/src/config/zuraffa_config_loader.dart';
import 'package:zuraffa_agent/src/engine/mission_runner.dart';

final config = ZuraffaConfigLoader.fromYaml(yamlDocument);
final issues = config.validate(); // List<ConfigIssue> — empty means runnable
final runner = MissionRunner(
  /* ...executor, dispatcher, stop policy, events... */
  config: config, // run() throws StateError listing every issue if invalid
);

Credentials never belong in the document — resolve them with a SecretResolver implementation (fromEnv / fromFile / fromVault; NullSecretResolver is the shipped stub).

Example document (each section is optional; validate() decides runnability):

provider:
  id: p1
  providerKind: openai
  baseUrl: https://llm.example.internal/api
  models: [internal/model]
  timeoutMs: 30000
agent_spec:
  id: spec-1
  name: research
  toolAllowlist: [search, read_file]
  systemPrompt: Research the topic.
engine_loop:
  id: loop-1
  sessionId: s1
  maxTurns: 8
  wallClockTimeoutMs: 60000
  repetitionThreshold: 5
mcp_transport:
  id: mcp-1
  transportType: sse
  endpoint: https://mcp.example.internal/sse
  authRequired: true
stop_policy:
  id: stop-1
  maxTurns: 16
  wallClockTimeoutMs: 0
  repetitionThreshold: 5
  enabled: true
compaction:
  id: comp-1
  sessionId: s1
  retainEntryIds: []
  summarizeEntryIds: []
  artifactRefs: []
  compactedAt: 0

Environment variables (all optional, prefix ZFA_): ZFA_PROVIDER_BASE_URL, ZFA_PROVIDER_PROVIDER_KIND, ZFA_PROVIDER_MODEL, ZFA_PROVIDER_TIMEOUT_MS, ZFA_AGENT_SPEC_ID, ZFA_AGENT_SPEC_NAME, ZFA_AGENT_SPEC_SYSTEM_PROMPT, ZFA_ENGINE_LOOP_MAX_TURNS, ZFA_ENGINE_LOOP_SESSION_ID, ZFA_ENGINE_LOOP_WALL_CLOCK_TIMEOUT_MS, ZFA_ENGINE_LOOP_REPETITION_THRESHOLD, ZFA_MCP_TRANSPORT_ENDPOINT, ZFA_MCP_TRANSPORT_TYPE, ZFA_STOP_POLICY_MAX_TURNS.