streaming_proxy.h

#pragma once

#include <string>
#include <functional>
#include <nlohmann/json.hpp>
#include <httplib.h>
#include "utils/http_client.h"
#include "utils/aixlog.hpp"
#include <iomanip>

namespace lemon {

using json = nlohmann::json;

class StreamingProxy {
public:
    struct TelemetryData {
        int input_tokens = 0;
        int output_tokens = 0;
        double time_to_first_token = 0.0;
        double tokens_per_second = 0.0;

        void print() const {
            if (input_tokens > 0 || output_tokens > 0) {
                LOG(INFO, "Telemetry") << "=== Telemetry ===" << std::endl;
                LOG(INFO, "Telemetry") << "Input tokens:  " << input_tokens << std::endl;
                LOG(INFO, "Telemetry") << "Output tokens: " << output_tokens << std::endl;
                LOG(INFO, "Telemetry") << "TTFT (s):      " << std::fixed << std::setprecision(3)
                          << time_to_first_token << std::endl;
                LOG(INFO, "Telemetry") << "TPS:           " << std::fixed << std::setprecision(2)
                          << tokens_per_second << std::endl;
                LOG(INFO, "Telemetry") << "=================" << std::endl;
            }
        }
    };

    static void forward_sse_stream(
        const std::string& backend_url,
        const std::string& request_body,
        httplib::DataSink& sink,
        std::function<void(const TelemetryData&)> on_complete = nullptr,
        long timeout_seconds = 300,
        std::function<void()> on_chunk = nullptr
    );

    static void forward_byte_stream(
        const std::string& backend_url,
        const std::string& request_body,
        httplib::DataSink& sink,
        long timeout_seconds = 300,
        std::function<void()> on_chunk = nullptr
    );

    // Normalize streaming chat.completion.chunk SSE deltas for OpenAI API
    // compatibility. Applies:
    //   1. Injects `role: "assistant"` when null/missing on assistant deltas
    //   2. Injects `content: ""` alongside `reasoning_content` when absent
    static std::string normalize_chat_completion_chunk(const std::string& sse_chunk);

private:
    static TelemetryData parse_telemetry(const std::string& buffer);
};

} // namespace lemon