Agent integration with Bub main

Dohnuts provides one decision tool through Bub. The agent extra pins Bub main at 9bf70488e22a44eeb875a9d55d9a7d352b82e381. Integration uses its Python SDK to provide one batch decision tool and persistent session tapes.

Follow the runtime and checkpoint setup first.

pdm install --check --prod -G agent

from pathlib import Path

from dohnuts.bub_agent import create_agent
from dohnuts.predictor import Predictor

predictor = Predictor.from_checkpoint("PsiACE/Dohnuts-0.1.0-0.8B")
framework, agent = create_agent(
    predictor, workspace=Path.cwd(), tape_directory=Path("runs/agent-tapes")
)


async def decide():
    command = ',dohnuts.decide state=\'{"message":"Please refund this invoice."}\' '
    command += 'questions=\'{"refund":{"type":"noul","instructions":"Is a refund requested?"}}\''
    async with framework.running():
        stream = await agent.run_stream(session_id="billing", prompt=command)
        return [event async for event in stream]

The tool is named dohnuts.decide; its model-facing alias is dohnuts_decide. Supply independent questions together against one state. Its optional image_path argument names an image within the application’s workspace. The returned probabilities retain the predictor’s meanings. Calls are serialized per predictor instance, while each call batches its questions.

An application may configure Bub’s planner/provider and call run_stream with a natural-language prompt. The application owns the framework lifetime and decides how to act on the returned distributions.

The explicit comma-command calls the decision tool without an external planner. Planner quality and provider cost depend on the application’s provider settings.