Changelog

Highlights per release. Full notes live in RELEASE.md.

v0.3.0 — 2026-06-02 (current)

• 🤝 Multi-agent orchestration: handoff, supervisor, debate, parallel sampling, human-in-the-loop.
• 🧱 Sandboxed AgentSkill execution (docker / e2b / firejail) + LLM_AGENT mode + URI skill resolver.
• 📊 Observability: ExecutionTrace, Gantt / heatmap / animated playback, ChainVisualizer.
• 🧬 Chain generation (ChainBuilder.from_description) + evolutionary search (ChainEvolver).
• 🤖 Native AnthropicClient, record/replay cassettes, cost estimation, dataset evaluation.
• 🔌 MCP graduated from experimental; pause/resume, cancellation, streaming, lossless serialization.

v0.2.0 — 2026-04-15

• 🔧 Breaking: LLM client refactoring; removed mmar-llm-mapi.
• ✨ 74 new tests, an examples runner.
• 🐛 6 critical fixes (conditional steps, serialization, input mapping).
• 📝 Documentation improvements throughout.

v0.1.0

• 🚨 Typed step description classes (StepDescription deprecated).
• 📊 Structured logging with configurable levels.
• 🔍 Error-traceback preservation; ⏱️ chain-level timeout; 🔄 per-step retry.
• 🌐 OpenAI-compatible API support; 🎛️ per-step LLM config.
• ⚡ Execution modes (FAST / SELF_CRITIC); 🔀 chain-level RE-PLAN policy; 📊 evaluation metrics.

See also

• Migration: legacy → typed steps