Chat in the terminal

A workflow that asks the questions instead of answering them, driven from a terminal by agentdeck chat. No API key: a @workflow is your own Python, so nothing here reaches a model. Source.

.agentdeck/
└── workflows/shift_handover/workflow.py    # @workflow + ctx.ask(...)

The workflow

from agentdeck import WorkflowCtx, workflow


@workflow
async def shift_handover(ctx: WorkflowCtx, area: str) -> str:
    """Write a handover note for one area, asking the outgoing operator what to flag."""
    severity = await ctx.ask(
        f"how did {area} run this shift?",
        options=["quiet", "busy", "problems"],
    )
    if severity == "problems":
        detail = await ctx.ask("what should the next shift look at first?")
        return f"{area}: problems. First: {detail}"
    return f"{area}: {severity}, nothing to escalate."

Run it

uv pip install agentdeck-sdk
agentdeck chat shift_handover

> bay 4
? how did bay 4 run this shift?
  1) quiet
  2) busy
  3) problems
> 3
? what should the next shift look at first?
> the conveyor stalled twice
-- run.completed --
>

Ctrl-C cancels a run in flight, or leaves the prompt if nothing is running. Ctrl-D exits.

What to look at

ctx.ask suspends the run. Nothing polls and no prompt is held open: the run reaches WAITING_ANSWER, the terminal renders the question, and answering resumes the same run. Close the terminal mid-question and the run is still waiting.
The options are enforced, not suggested. An answer outside them is refused and the run stays answerable, which is why the terminal re-asks instead of failing.
agentdeck chat <target> takes any agent or workflow. Omit the name only when the deck holds exactly one target; a name that does not exist fails before the first prompt.
The terminal is a binding, not a special case. agentdeck chat is deck.serve(Terminal.stdio(target=...)), and the same run is reachable over Native.http() at the same time.

Next: Workflows · Terminal