test

            _sendAll(combined.data(), combined.size());

        }

        // Read until we have full response header (\r\n\r\n) -- identical to

        // GetStream's HTTP/1.x header-read + streaming-mode detection below.

        // Read until we have full response header (\r\n\r\n) -- same structure

        // as GetStream's HTTP/1.x header-read, except the deadline honors

        // setTimeout() (_recvTimeoutSec) instead of a fixed 10s: a slow AI

        // backend (cold model load/swap) can take much longer than 10s to

        // send back even the response headers, well before any body/stream

        // data starts flowing.

        std::vector<char> raw;

        raw.reserve(16384);

        size_t header_end = std::string::npos;

        size_t search_from = 0;

        auto header_deadline = std::chrono::steady_clock::now() + std::chrono::seconds(10);

        auto header_deadline = std::chrono::steady_clock::now() + std::chrono::seconds(_recvTimeoutSec);

        for (;;) {

            if (raw.size() >= 4) {