import { C as Capabilities, B as BaseLLMAdapter, L as LLMProviderConfig, b as LLMCompletionRequest, c as LLMCompletionResponse } from '../base-adapter-BQZiL4zs.cjs';

/**
 * BitNet Adapter
 *
 * Connects to a local bitnet.cpp inference server.
 * BitNet runs 1-bit quantized models that use ~10x less memory than
 * standard GGUF models and run faster on CPU — no GPU required.
 *
 * Setup (one-time):
 *   git clone --recursive https://github.com/microsoft/BitNet
 *   cd BitNet
 *   pip install -r requirements.txt
 *   python setup_env.py -md microsoft/bitnet-b1.58-2B-4T -q i2_s
 *
 * Run the server:
 *   python run_inference.py --serve --port 8080 --host 0.0.0.0
 *
 * The server exposes: POST http://localhost:8080/v1/chat/completions
 *
 * Why BitNet over standard GGUF?
 *   - ~500 MB model size vs ~4 GB for Mistral-7B Q4
 *   - Faster on CPU (1-bit arithmetic, no FP multiply)
 *   - Deployable on Railway, Fly.io, any server without GPU
 *   - See: https://github.com/microsoft/BitNet
 *
 * @version 2.0.0
 */

type BitNetAdapterConfig = Omit<LLMProviderConfig, 'apiKey'> & {
    apiKey?: string;
    model?: string;
};
declare const BITNET_MODELS: readonly ["microsoft/bitnet-b1.58-2B-4T", "1bitLLM/bitnet_b1_58-large"];
type BitNetModel = (typeof BITNET_MODELS)[number];
/**
 * Short aliases accepted by bitnet.cpp (maps to full HuggingFace model IDs)
 */
declare const BITNET_MODEL_ALIASES: Readonly<Record<string, BitNetModel>>;
/**
 * Capability manifest — BitNet 1-bit quantized models on local CPU
 * (~500MB model, no GPU required). Tiny model with limited capabilities;
 * keep declaration minimal and honest. Brains routing here should expect
 * the no-frills baseline: text in, text out, $0 marginal cost.
 *
 * Exported as a constant so the capability-aware router can read it
 * without instantiating the adapter — single source of truth per W.GOLD.006.
 */
declare const BITNET_CAPABILITIES: Capabilities;
declare class BitNetAdapter extends BaseLLMAdapter {
    readonly name: "bitnet";
    readonly models: readonly ["microsoft/bitnet-b1.58-2B-4T", "1bitLLM/bitnet_b1_58-large"];
    readonly defaultHoloScriptModel: string;
    readonly capabilities: Capabilities;
    private readonly localBaseURL;
    constructor(config?: BitNetAdapterConfig);
    protected getDefaultModel(): string;
    /**
     * Send a chat completion request to the bitnet.cpp server.
     * Uses the OpenAI-compatible /v1/chat/completions endpoint.
     */
    complete(request: LLMCompletionRequest, model?: string): Promise<LLMCompletionResponse>;
    /**
     * Returns the HoloScript-tuned system prompt for BitNet's smaller model context.
     * Overrides the base class version to use the compact prompt.
     */
    protected getHoloScriptSystemPrompt(): string;
    /**
     * Check if the local BitNet server is reachable.
     * Delegates to BaseLLMAdapter.healthCheckLocalServer — same /health →
     * /v1/models fallback, branded error with bitnet.cpp setup hint.
     */
    healthCheck(): Promise<{
        ok: boolean;
        latencyMs: number;
        error?: string;
    }>;
}

export { BITNET_CAPABILITIES, BITNET_MODELS, BITNET_MODEL_ALIASES, BitNetAdapter, type BitNetModel };