/**
 * Tier-based model routing.
 *
 * The bridge routes each MCP tool invocation to a model tier. Tiers are
 * size/latency buckets, not capability buckets — the tool decides which
 * tier it wants based on expected workload (short classify vs long summary).
 *
 * All tiers route through `MlxHttpBackend` to a local oMLX
 * (https://github.com/jundot/omlx) inference server. Single backend, single
 * inference engine, KV-cache persistence across requests, schema strictness
 * via OpenAI Structured Outputs (`response_format: { type: "json_schema",
 * strict: true }`).
 *
 * Default models (mlx-community on Hugging Face):
 * - Tier B: `Qwen3-4B-Instruct-2507-4bit` (~2.5 GB) — non-thinking variant.
 *   The "Instruct-2507" suffix matters: bare `Qwen3-4B`/`Qwen3-8B` are
 *   thinking models that emit `<think>...</think>` traces by default and
 *   burn the per-tool MAX_OUTPUT_TOKENS cap before producing real output.
 *   Instruct-2507 is the explicitly-non-thinking line.
 * - Tier C: `Qwen3-8B-4bit` (~5 GB) — thinking variant; bridge injects
 *   `/no_think` into user messages to disable reasoning at call time.
 *   Long-form summarize at numCtx=32 768.
 * - Tier D: `Qwen3-14B-4bit` (~8 GB) — also thinking; same `/no_think`
 *   injection. Hardest tasks (classify subtle, transform on dense input).
 *
 * All Apache-2.0. Download into `~/.omlx/models/<dirname>` via:
 *   npm run download-models
 * Start oMLX:
 *   brew services start jundot/omlx/omlx
 */
export type Tier = 'B' | 'C' | 'D' | 'V';
export interface TierConfig {
    /**
     * Base URL of the oMLX (or any OpenAI-compatible) inference server.
     * Default: `"http://127.0.0.1:8000"` (oMLX's default port).
     */
    mlxUrl?: string;
    /**
     * Model name sent in the OpenAI `model` field — must match a directory name
     * under `~/.omlx/models/`. oMLX serves multiple models from one process and
     * routes requests by this name.
     *
     * Example: `"Qwen3-4B-Instruct-2507-4bit"`
     */
    mlxModelName?: string;
    /**
     * Context window size in tokens.
     *
     * Informational for `MlxHttpBackend` — the server-side model's max context
     * is fixed at load time. Used for the chunker's safety margin and tool
     * `maxInputTokens`.
     *
     * Tier B → 8192   (fast 4B model)
     * Tier C → 32768  (8B model with longer context)
     * Tier D → 16384  (14B model; ~9-10 GB at this context size)
     * Tier V → 32768  (Qwen3-VL-4B vision model; image-bearing calls)
     */
    numCtx?: number;
    /**
     * How `MlxHttpBackend` suppresses the model's reasoning trace for this tier.
     *
     * - `'no_think'` (default): append `/no_think` to the user prompt. Works on
     *   Qwen3 thinking models (8B/14B); inert on the non-thinking Instruct-2507.
     * - `'chat_template'`: send `chat_template_kwargs: { enable_thinking: false }`
     *   in the request, leaving the prompt untouched. Required for Qwen3-VL /
     *   Qwen3.5 — `/no_think` does NOT disable thinking on those chat templates.
     *
     * Absent → `'no_think'` (preserves the exact B/C/D request contract captured
     * by `migration-snapshot.test.ts`).
     */
    thinkingMode?: 'no_think' | 'chat_template';
}
export interface BridgeConfig {
    tiers: Record<Tier, TierConfig>;
    /** Fallback tier when a tool has no explicit mapping. */
    defaultTier: Tier;
    /** Per-tool tier assignment. Absent keys fall back to defaultTier. */
    toolTierMap?: Record<string, Tier>;
}
export declare const DEFAULT_CONFIG: BridgeConfig;
export interface ResolveOptions {
    /** Override the model for a specific tier (e.g. via CLI flag). */
    tierOverrides?: Partial<Record<Tier, Partial<TierConfig>>>;
}
/**
 * Apply overrides on top of a base config. Leaves unspecified fields alone.
 */
export declare function withOverrides(base: BridgeConfig, opts?: ResolveOptions): BridgeConfig;
export declare function tierForTool(config: BridgeConfig, toolName: string): Tier;
export declare function modelForTool(config: BridgeConfig, toolName: string): TierConfig;
/**
 * Display/telemetry identifier for a tier. Prefers the explicit model name
 * when set; falls back to the URL. Always returns a non-empty string so
 * footers/_meta never carry undefined.
 */
export declare function tierModelLabel(tcfg: TierConfig): string;
//# sourceMappingURL=tiers.d.ts.map