/**
 * kosha-discovery — Tool-dialect and structured-output inference.
 *
 * Tool calling and structured output are two surfaces where every frontier
 * provider has settled on a *different* JSON shape. A consumer that wants
 * to route a single workload across multiple providers has to pick the
 * right adapter per model, and the registry is the natural place to tag
 * that. These helpers return best-effort hints inferred from the origin
 * provider plus the model ID — downstream callers treat them as defaults
 * and override with direct API probes when precision matters.
 * @module
 */
import type { StructuredOutputMode, ToolDialect } from "./types.js";
/**
 * Infer the tool-calling dialect a model natively speaks.
 *
 * The return value describes the wire format a consumer must emit to
 * invoke tools on this model. Returns `undefined` when the provider is
 * unknown and the model ID gives no signal — the caller should then
 * fall back to `"openai-tools"` (the de-facto compatibility dialect).
 *
 * Mapping summary:
 * - OpenAI GPT-4.1 / GPT-5 / o-series     → `"openai-responses"` (new API)
 * - OpenAI GPT-4 / GPT-4o / GPT-3.5       → `"openai-tools"`
 * - Anthropic Claude 3+                   → `"anthropic-tools"`
 * - Google Gemini 1.5+                    → `"gemini-functions"`
 * - Cohere Command R / R+                 → `"cohere-tools"`
 * - Mistral Large / Small                 → `"mistral-tools"`
 * - Meta Llama 3.1+ (instruct variants)   → `"llama3-tools"`
 * - Embedding / image / audio-only models → `"none"`
 *
 * NOTE on serving-layer proxies: this helper infers the model's *native*
 * dialect from its origin provider. Managed serving layers such as Groq,
 * Together, Fireworks, and OpenRouter expose open-weight models behind
 * an OpenAI-compatible tools API regardless of the underlying family —
 * so if you are calling a Llama 3.1 model through Groq, you should use
 * `"openai-tools"` on the wire, not the `"llama3-tools"` hint returned
 * here. Callers that care about the serving-layer wire format should
 * check `ModelCard.provider` (the serving slug) first and fall back to
 * this inference only for direct-origin routes. A future revision may
 * take `servingProvider` as a second argument.
 *
 * @param originProvider - Model creator slug (e.g. `"anthropic"`).
 * @param modelId        - Provider-canonical model ID.
 * @returns Best-effort dialect tag, or `undefined` when no safe guess exists.
 */
export declare function inferToolDialect(originProvider: string | undefined, modelId: string): ToolDialect | undefined;
/**
 * Infer the structured-output modes a model supports.
 *
 * Returns an ordered list with the most precise mode first. An empty
 * array means no structured-output enforcement is available and the
 * caller must fall back to prompt-level coercion.
 *
 * Precision ranking: `json-schema` > `response-schema` > `grammar`
 * > `json-mode` > `response-format` > `tool-choice` > `xml`.
 *
 * @param originProvider - Model creator slug.
 * @param modelId        - Provider-canonical model ID.
 * @returns Ordered list of supported modes (empty when unknown).
 */
export declare function inferStructuredOutputModes(originProvider: string | undefined, modelId: string): StructuredOutputMode[];
/**
 * Best-effort inference for whether a model supports parallel tool calls
 * (multiple tool invocations emitted in a single assistant turn).
 *
 * Frontier models (GPT-4o, Claude 3.5+, Gemini 1.5+) all support this;
 * older tool-capable models (GPT-3.5, early Claude 3) emit tool calls
 * serially. Returns `undefined` when the model does not appear to support
 * tool calling at all.
 *
 * @param originProvider - Model creator slug.
 * @param modelId        - Provider-canonical model ID.
 * @returns `true`, `false`, or `undefined` when not tool-capable.
 */
export declare function inferParallelToolCalls(originProvider: string | undefined, modelId: string): boolean | undefined;
//# sourceMappingURL=model-features.d.ts.map