import type { CompletionRun, CompletionStats, ToolCall } from '@qvac/sdk';
export type OpenAiFinishReason = 'stop' | 'length' | 'tool_calls';
export interface DrainedCompletion {
    text: string;
    /** Concatenated `thinkingDelta` text; empty when the SDK captured no reasoning. */
    thinking: string;
    toolCalls: ToolCall[];
    stats: CompletionStats | undefined;
    /**
     * Terminal reason from the SDK `completionDone` event (`eos` / `length` /
     * `stopSequence` / `cancelled`), or undefined if the stream ended without
     * one. `error` and `cancelled` are never present here — `drainCompletion`
     * throws on both (502 for error, `InferenceCancelledError` for cancelled).
     */
    stopReason: string | undefined;
    /**
     * OpenAI `usage.completion_tokens`: prefers SDK `stats.emittedTokens`
     * (addon-streamed pieces), then `stats.generatedTokens` (decode count),
     * then a whitespace word count.
     */
    completionTokens: number;
    /** OpenAI `finish_reason`: `tool_calls` wins, then `length` on truncation, else `stop`. */
    finishReason: OpenAiFinishReason;
}
/**
 * Single-pass consumer of an SDK completion run, shared by every
 * chat-category route (chat / completions / responses). Draining
 * `result.events` once yields content text, tool calls, stats and the
 * terminal `stopReason` together, so the OpenAI `finish_reason` and token
 * accounting are derived in one place instead of drifting per route.
 *
 * Pass `onToken` to stream content deltas as they arrive (SSE paths); omit
 * it for blocking responses. Pass `onThinking` to stream reasoning deltas the
 * same way — only produced when the caller enabled `captureThinking` on the
 * SDK request.
 */
export declare function drainCompletion(result: CompletionRun, onToken?: (token: string) => void, onThinking?: (token: string) => void): Promise<DrainedCompletion>;
/**
 * OpenAI `usage.completion_tokens` for a drained run.
 *
 * Prefer SDK `emittedTokens` (non-empty addon stream pieces) over
 * `generatedTokens` (`llama_perf` `n_eval`), which can equal the predict /
 * `max_tokens` budget when fewer tokens were streamed. Normalized
 * `contentDelta` / `thinkingDelta` event counts are not used — those are
 * chunk boundaries, not tokenizer tokens.
 */
export declare function completionTokensFromStats(text: string, stats: CompletionStats | undefined): number;
//# sourceMappingURL=completion-result.d.ts.map