[Feature Request] Add user_message solver

[kaifronsdal]

I often find I want to add scafolding to an agent/prompt it in multiple stages. Currently this requires writing custom solvers to add the subsequent user messages (prompt_template only works for the first message). This is the usecase I'm thinking of is something like this

chain(
    system_message("You are a math expert."),
    user_message("How would you go about solving 3x^2+2x+1=10?"),
    generate(),
    user_message("Now solve for x."),
    generate(),
    user_message("What is your final answer for x?"),
    generate()
)

The implementation would be pretty simlar to system_message. Perhaps something like this

@solver
def user_message(template: str, **params: Any) -> Solver:
    """Solver which appends a user message to the conversation.

    User message template containing any number of optional `params`.
    for substitution. All values contained in sample `metadata` are also
    automatically included in the `params`.

    Args:
      template (str): Template for user message.
      **params (dict[str,Any]): Parameters to fill into the template.

    Returns:
      A solver that appends the parameterised user message.
    """
    # read template
    content = resource(template)

    async def solve(state: TaskState, generate: Generate) -> TaskState:
        kwargs = state.metadata | params
        state.messages.append(
            ChatMessageUser(content=content.format(**kwargs))
        )
        return state

    return solve

I could also see a usecase for wanting assistant_message and tool_message solvers as well that simply append the relevent message type. This is somwhat similar to sglang's model interface.

[kaifronsdal]

A minor update to the sugestion. Instead of using state.metadata to fill in the template, it would be more useful to be able to use store() to fill in the template.