Changed the value setting method for step.input and step.additional_input during step execution

[mahm]

With the current specification, even if the agent generates prompts for executing steps during the planning phase, as shown in the code below, None will be set if the client does not explicitly provide the step's body.

async def _plan(task: Task, step: Step) -> Step:
    planner = load_chat_planner(model)
    plans = planner.plan({"input": task.input}).steps
    for i, plan in enumerate(plans):
        is_last_step = i == len(plans) - 1
        await Agent.db.create_step(
            task_id=task.task_id,
            name=StepTypes.ACTION,
            input=plan.value,
            is_last=is_last_step
        )
    step.output = "\n".join([f"- {plan.value}" for s in plans])
    return step

I found it unnatural to always have to explicitly specify the prompt to be executed in the step from the client when calling the agent, so I created this pull request. If this change does not conflict with the design philosophy, please merge it.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Updated (UTC)
agent-protocol-docs	⬜️ Ignored (Inspect)	Visit Preview	Aug 19, 2023 0:38am

I missed the problem with user overriding the values

[jakubno]

@mahm, thanks for your input and sorry for late reply. Looking forward to hearing your opinion

[hackgoofer]

Hi, we are going to close this PR, this feels like a big enough change for RFC, mind submitting an issue about this?

We may be deprecating the "input" for the task. There have been some discussion on this for v2. :)
#70