Skip to content
Navigation

Changelog

All notable changes to Orbiter are documented in this file.

All notable changes to Orbiter are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.1.0 - 2025-02-16

Initial release of Orbiter, a ground-up rewrite of the AWorld multi-agent framework.

Core Framework (orbiter-core)

  • Agent — Single Agent class replacing AWorld’s 5 agent types (LLMAgent, TaskLLMAgent, LoopLLMAgent, ParallelLLMAgent, SerialLLMAgent). Supports tools, handoffs, lifecycle hooks, structured output, and configurable max steps.
  • Tool — @tool decorator for turning functions into LLM-callable tools with auto-generated JSON schemas from type hints and docstrings. Tool ABC for complex tools. Sync functions auto-wrapped via asyncio.to_thread().
  • Swarm — Multi-agent orchestration with three modes: "workflow" (sequential pipeline via flow DSL), "handoff" (agent-driven delegation), and "team" (lead-worker pattern with auto-generated delegate tools).
  • Runner — Three entry points: run() (async), run.sync() (blocking), run.stream() (async generator). State tracking, loop detection, and retry logic built in.
  • Types — Typed message classes (UserMessage, AssistantMessage, SystemMessage, ToolResult) as frozen Pydantic models. RunResult, StreamEvent, Usage, AgentOutput.
  • Config — Pydantic v2 models: AgentConfig, ModelConfig, TaskConfig, RunConfig. parse_model_string() for "provider:model_name" format.
  • Registry — Generic Registry[T] with fail-fast duplicate detection. Pre-built agent_registry and tool_registry.
  • Events — EventBus for decoupled async pub/sub communication.
  • Hooks — HookManager with HookPoint enum (START, FINISHED, ERROR, PRE_LLM_CALL, POST_LLM_CALL, PRE_TOOL_CALL, POST_TOOL_CALL).
  • Internal — message_builder (correct-by-construction message ordering), output_parser (LLM response to AgentOutput bridging, structured output validation), call_runner (execution loop with state tracking and loop detection), state (RunState/RunNode lifecycle), graph (flow DSL parser, topological sort).

LLM Models (orbiter-models)

  • ModelProvider — Abstract base class with complete() and stream() methods.
  • OpenAIProvider — Full implementation for OpenAI’s chat completions API.
  • AnthropicProvider — Full implementation for Anthropic’s messages API.
  • Model Registry — get_provider("openai:gpt-4o") factory for building providers from model strings.
  • Types — ModelResponse, StreamChunk, ToolCallDelta, FinishReason, ModelError.

Package Structure

  • UV workspace monorepo with 13 packages under packages/
  • Namespace packages via pkgutil.extend_path() for unified orbiter.* import path
  • Zero heavy dependencies in orbiter-core (only pydantic)
  • Provider SDKs isolated in orbiter-models

Design Highlights

  • Async-first with single sync bridge (run.sync())
  • Parallel tool execution via asyncio.TaskGroup
  • Typed message union (Message = UserMessage | AssistantMessage | SystemMessage | ToolResult)
  • Frozen Pydantic models for all data types
  • Flow DSL for declarative multi-agent pipelines ("a >> b >> c", "a >> (b | c) >> d")
  • Structured exception hierarchy rooted at OrbiterError
  • Tool errors captured as ToolResult(error=...), not propagated
  • Exponential backoff retry for transient LLM errors

Planned Packages (stubs created)

  • orbiter-context — Context engine: hierarchical state, prompt neurons, processors, workspace, RAG
  • orbiter-memory — Short-term and long-term memory backends
  • orbiter-mcp — Model Context Protocol client and server
  • orbiter-sandbox — Local and Kubernetes sandboxed execution
  • orbiter-trace — OpenTelemetry-based tracing
  • orbiter-eval — Evaluation framework with scorers and reflection
  • orbiter-a2a — Agent-to-Agent protocol
  • orbiter-cli — Command-line interface
  • orbiter-server — HTTP server for agent serving
  • orbiter-train — Training: trajectory collection and trainers