Prompts

Prompts are reusable templates that guide LLM interactions. They can have arguments for dynamic content.

Registering a Prompt

server->RegisterPrompt(
    "code_review",
    PromptOptions{}.Description("Review code changes"),
    [](const std::string& name,
       const std::optional<JsonValue>& args) -> GetPromptResult {
        GetPromptResult result;

        std::string diff;
        if (args) {
            auto* v = args->Find("diff");
            if (v) diff = v->GetString();
        }
        PromptMessage msg;
        msg.role = "user";
        msg.content = TextContent{"text",
            "Review the following code changes:\n" + diff};
        result.messages.push_back(msg);

        return result;
    });

Prompt Arguments

Arguments can be declared directly on PromptOptions via the fluent .Arguments(...) setter:

PromptArgument arg;
arg.name = "diff";
arg.description = "The git diff to review";
arg.required = true;

server->RegisterPrompt(
    "code_review",
    PromptOptions{}.Description("Review code changes").Arguments({arg}),
    handler);

The arguments are propagated to the protocol-level Prompt struct and returned via prompts/list. For full control you can also set arguments on a Prompt struct directly:

Prompt p;
p.name = "code_review";
p.description = "Review code changes";
p.arguments = {arg};

::: note PromptOptions fields (description, title, icons, arguments) are propagated to the protocol-level Prompt struct returned via prompts/list. :::

Input Requests (input_required)

Prompt handlers may also return input_required (GetPromptResult::input_required) to start a multi-round input collection; the client retries the get carrying inputResponses + requestState (same MRTR loop as tool calls, with automatic request_state signing when request_state_key is configured).