Method worker_loop

server/src/server/http_server.cpp:1804–3326

        const json body = json::parse(hr.body);
        req.raw_body = body;
        if (!parse_common_request_fields(fd, body, req)) return true;
        if (!parse_endpoint_request(
                hr.path, body, req, count_tokens_only)) return false;

        const std::vector<ChatMessage> chat_messages =
            normalize_chat_messages(req.messages, req.format, tool_memory_);
        // Reasoning must be applied BEFORE rendering: the template injects
        // the empty <think>\n\n</think>\n\n block when thinking is disabled.
        apply_request_reasoning(body, req);
        // Bandit: parse session_id from extra_body (opt-in adaptive keep_ratio).
        req.session_id = parse_session_id_from_body(body);
        if (!render_and_tokenize_request(fd, chat_messages, req)) return true;

        // count_tokens: short-circuit after tokenization. Skip generation
        // entirely — Anthropic's contract is just {"input_tokens": N}.
        if (count_tokens_only) {
            const json response = {
                {"input_tokens", (int) req.prompt_tokens.size()},
            };
            send_response(fd, 200, "application/json", response.dump() + "\n");
            return true;
        }
    } catch (const std::exception & e) {
        send_error(fd, 400, std::string("JSON parse error: ") + e.what());
        return true;
    }

    if (!validate_request_context(fd, req)) return true;
    log_parsed_request(req);
    enqueue_request_and_wait(fd, std::move(req));
    return true;
}

// ─── Worker thread ──────────────────────────────────────────────────────

void HttpServer::worker_loop() {
    while (true) {
        ServerJob * job = dequeue();
        if (!job) break;  // stopping

        int fd = job->fd;
        const auto & req = job->req;
        auto started_at = std::chrono::steady_clock::now();

        // Track live status for /status page. RAII guard ensures idle on all paths.
        std::string prompt_excerpt;
        if (!req.prompt_tokens.empty()) {
            // Decode first ~40 tokens as a prompt excerpt (cheap, bounded).
            const int excerpt_len = (std::min)((int)req.prompt_tokens.size(), 40);
            std::vector<int32_t> excerpt_toks(req.prompt_tokens.begin(),
                                               req.prompt_tokens.begin() + excerpt_len);
            prompt_excerpt = tokenizer_.decode(excerpt_toks);
            if (prompt_excerpt.size() > 200) prompt_excerpt.resize(200);
        }
        {
            ServerStatus::RequestInfo info;
            info.model = req.model;
            info.format = api_format_name(req.format);