/**
 * First-class local-LLM presets for Ollama, vLLM, and LM Studio.
 *
 * All three speak the OpenAI Chat Completions wire format at `/v1/chat/completions`
 * with SSE streaming, so the bulk of the work is a tuned `WireFormatConfig`. The
 * differences worth modeling:
 *
 *   - Ollama: no auth at all (no `Authorization` header). Supports
 *     `keep_alive` and `num_ctx` request fields. Historically omits `usage`
 *     on the final chunk, so we synthesize a `message_stop` via
 *     `finalizeStream` even when the upstream didn't emit one.
 *   - vLLM: auth is optional (off by default). The server may close the
 *     stream without a `data: [DONE]` sentinel — also handled via
 *     `finalizeStream`.
 *   - LM Studio: optional Bearer auth. Same stream shape as OpenAI, but
 *     `max_tokens` is the accepted field (we always use it via the
 *     `openai-compatible` family quirk).
 *
 * `createLocalLlmPreset` is the single source of truth — the three named
 * exports below are thin wrappers that pick the right defaults.
 */
import type { StopReason } from '@wrongstack/core/types';
export interface LocalLlmPresetOptions {
    /** Provider id used for logging and the registry (e.g. 'ollama'). */
    id: string;
    /** Default base URL when the user doesn't override. */
    defaultBaseUrl: string;
    /**
     * When true, the request is sent with no `Authorization` header. Use for
     * servers that reject any Authorization value (Ollama without auth).
     * When false/undefined, a `Bearer <apiKey>` header is sent (any non-empty
     * key is fine for servers that have auth disabled).
     */
    noAuth?: boolean | undefined;
    /**
     * Provider-specific request body extras. Keys here are merged into the
     * outgoing JSON body verbatim. Use for things like Ollama's `keep_alive`
     * or vLLM's `repetition_penalty`. Values that collide with canonical
     * fields (`model`, `messages`, `tools`, `stream`, `max_tokens`, …) are
     * dropped — canonical wins.
     */
    bodyExtras?: Record<string, unknown> | undefined;
    /** Default context window — surfaced via `capabilities.maxContext`. */
    maxContext?: number | undefined;
    /**
     * Whether the model advertises vision input. Local vision models are
     * rare but Ollama supports them via multimodal tags; default false.
     */
    vision?: boolean | undefined;
}
interface StreamingArgBuffer {
    chunks: string[];
    length: number;
}
interface LocalLlmStreamToolState {
    id?: string | undefined;
    name?: string | undefined;
    argBuf: StreamingArgBuffer;
    emittedStart: boolean;
    emittedChunkIndex: number;
}
interface LocalLlmStreamState {
    model: string;
    started: boolean;
    textOpen: boolean;
    thinkingOpen: boolean;
    toolByIndex: Map<number, LocalLlmStreamToolState>;
    usage: {
        input: number;
        output: number;
    };
    stopReason: StopReason;
    /** Tracks whether the upstream emitted a terminal `data: [DONE]` or `finish_reason`. */
    endedNaturally: boolean;
    finalEmitted: boolean;
}
export declare function createLocalLlmPreset(opts: LocalLlmPresetOptions): import("../wire-format.js").WireFormatConfig<LocalLlmStreamState>;
/**
 * Ollama — https://ollama.com
 *
 * Runs on `http://localhost:11434` by default and exposes an OpenAI-compatible
 * chat-completions endpoint at `/v1/chat/completions`. Ollama rejects any
 * `Authorization` header (it returns 400), so we send none. Ollama also
 * accepts the `keep_alive` body field (e.g. `"5m"`, `"-1"` for indefinite)
 * to control how long the model stays loaded in memory.
 */
export declare const ollamaWireFormat: import("../wire-format.js").WireFormatConfig<LocalLlmStreamState>;
/**
 * vLLM — https://docs.vllm.ai
 *
 * Default server URL is `http://localhost:8000/v1`. Auth is disabled by
 * default in vLLM; if enabled, it expects a `Bearer <key>` header. vLLM
 * passes the legacy `max_tokens` parameter through correctly and supports
 * OpenAI-style tool calls.
 */
export declare const vllmWireFormat: import("../wire-format.js").WireFormatConfig<LocalLlmStreamState>;
/**
 * LM Studio — https://lmstudio.ai
 *
 * Default server URL is `http://localhost:1234/v1`. LM Studio's local
 * server mirrors the OpenAI Chat Completions API exactly, with optional
 * Bearer auth.
 */
export declare const lmstudioWireFormat: import("../wire-format.js").WireFormatConfig<LocalLlmStreamState>;
export {};
//# sourceMappingURL=local-llm.d.ts.map