Kohaku-Lab / KohakuTerrarium

What it does

• Agent-level abstraction. The six-module creature model is the first-class concept. A new agent shape is "write a config and maybe a few custom modules," not "rebuild the runtime."
• A real Python API. Agent.build, typed TurnResult turns with timeouts that actually cancel, streaming typed events, @kt.tool to turn any function into an agent tool, direct LLM-instance injection, strict-by-default errors instead of silent f
• Built-in session persistence and resume. The engine mints and owns session files (session= / Terrarium(session_dir=)); resume hours later with kt resume or Terrarium.resume. SessionReader replays any finished run offline.
• Searchable session history. Every event is indexed. kt search and the search_memory tool let you (and the agent) look up past work.
• Non-blocking context compaction. Long-running agents keep working while context compacts in the background.