oturn 0.1.3

OTURN: Orchestrated Task Unified Reactive Nucleus

OTURN

OTURN = Orchestrated Task Unified Reactive Nucleus.

OTURN is a small async runtime for turn-based agent systems.

What It Provides

• Reducer-style stream state machine (agent_stream_state.py)
• Queue-based multi-subscriber pub/sub (subscribe(queue), unsubscribe(queue))
• Turn runtime with tool execution and approval/choice callbacks
• LiteLLM streaming adapter with normalized semantic events
• Session/history persistence via Context + SessionStore

Installation

cd oturn
uv sync
uv pip install -e .

Build:

uv build

Core Types

• Oturn: high-level agent runtime
• OTurnNucleus: generic reducer nucleus for non-LLM state machines
• Transition: reducer output (state, events, terminal)
• Context: conversation history + persistence wrapper
• BaseTool: tool protocol/base class

Oturn Event Stream

Published events include:

• assistant_delta
• assistant_reasoning_delta
• assistant_final
• tool_call_delta
• tool_call
• tool_output
• tool_exit
• approval_request
• choice_request
• message_sent
• run_aborted
• error
• debug_latency / tool_call_debug (when debug flags are enabled)

Minimal Oturn Example

import asyncio
from pathlib import Path

from oturn import (
    Oturn,
    OturnConfig,
    OturnProviderConfig,
    OturnModelConfig,
    OturnAgentConfig,
    Context,
    WeatherTool,
)


async def main() -> None:
    cfg = OturnConfig(
        model="<provider/model>",
        provider=OturnProviderConfig(api_key="<API_KEY>"),
        model_config=OturnModelConfig(max_output_tokens=4096, temperature=0.7),
        agent_config=OturnAgentConfig(max_steps=50, user_name="User"),
    )

    agent = Oturn(work_dir=Path.cwd(), config=cfg, tools=[WeatherTool])
    ctx = Context(Path.cwd() / ".oturn" / "sessions" / "demo.md", work_dir=Path.cwd())

    q: asyncio.Queue[dict] = asyncio.Queue()
    agent.subscribe(q)

    agent.enqueue_user_input_nowait("What's the weather in Tokyo?")
    task = asyncio.create_task(agent.run_queues(ctx))

    try:
        while True:
            ev = await q.get()
            print(ev)
    finally:
        task.cancel()


asyncio.run(main())

Tool Injection Model

OTURN uses explicit runtime tool injection. Pass tool classes when constructing Oturn:

agent = Oturn(work_dir=work_dir, config=cfg, tools=[WeatherTool])

Example tools in oturn.tools.examples:

• WeatherTool (get_weather)
• HostConfigTool (get_host_config)

Usage and Tool-Call Delta Accessors

Usage data is available via context accessors.

• get_last_usage(ctx) -> dict | None
• get_last_total_tokens(ctx) -> int | None
• get_tool_call_deltas() -> list[tool_call_delta]
• get_tool_call_delta(delta_id) -> tool_call_delta | None

Session/History

• Session persistence is handled by Context + SessionStore.
• Records are stored as markdown event/message documents (Chron-based serialization via chronml).
• Default metadata root name is .oturn (DEFAULT_META_DIRNAME).
• Session files are created under a date-partitioned global directory:
  • ~/.oturn/global_sessions/YYYYMMDD/<session_id>.md
• When work_dir is available, SessionStore also creates a local session link:
  • <work_dir>/.oturn/sessions/<session_id>.md -> global session file
• Persistence is incremental:
  • append_message(...) appends a single serialized record
  • append_compact(...) appends compaction markers the same way
• Context.restore() loads existing session markdown into in-memory history.
• Configurable knobs:
  • meta_dirname: rename .oturn namespace for local/global paths
  • global_sessions_root: override ~/.<meta_dirname>/global_sessions
• You can also import/export session markdown via:
  • Context.export_md_as_config(...)
  • Context.import_md_as_config(...)