/**
 * Estimate the number of tokens in `text` for the given model.
 * Heuristic only — for exact counts, use the provider's tokenizer.
 *
 * @example
 * ```ts
 * if (estimateTokens(prompt, 'gpt-4o') > 100_000) {
 *   throw new Error('prompt too long; consider chunking')
 * }
 * ```
 */
export declare function estimateTokens(text: string, model?: string): number;
/**
 * Estimate total tokens for a chat-completion request: sum of
 * every message's content plus a fixed per-message overhead
 * (matches the rough "+4 per message + 2 for the conversation"
 * heuristic OpenAI's docs publish).
 */
export declare function estimateMessageTokens(messages: Array<{ role: string, content: string | unknown }>, model?: string): number;
/**
 * Inspect `text` for common prompt-injection patterns. Returns
 * `{ ok, matched, cleaned }`. Apps decide what to do with the
 * result — reject the request (`if (!result.ok) throw...`), pass
 * the cleaned text on (`useText(result.cleaned)`), or just log
 * for audit while letting the original through.
 *
 * **Limits:** this is heuristic. Adversarial inputs can paraphrase
 * around any specific pattern. Use as a cheap first filter; for
 * real defense, isolate the user input from the system prompt
 * structurally (different roles, JSON-mode for the system layer)
 * and guard the output side too.
 *
 * @example
 * ```ts
 * const check = sanitizePrompt(userInput)
 * if (!check.ok) {
 *   log.warn('possible injection attempt', { patterns: check.matched })
 *   // option A: reject
 *   throw new HttpError(400, 'invalid input')
 *   // option B: pass cleaned
 *   await chat([{ role: 'user', content: check.cleaned }])
 * }
 * ```
 */
export declare function sanitizePrompt(text: string): SanitizeResult;
export declare interface SanitizeResult {
  ok: boolean
  matched: string[]
  cleaned: string
}