/**
 * Normalize Openference context-overflow errors so Pi auto-compacts.
 * Never rewrite 429 / rate-limit / overloaded — those stay on Retry-After.
 */

import { OPENFERENCE_PROVIDER } from "./constants.ts";

export interface OverflowProbeMessage {
  role?: string;
  stopReason?: string;
  errorMessage?: string;
  provider?: string;
}

/** Phrases that indicate context overflow (not rate limits). */
export const OVERFLOW_PATTERN =
  /context_length_exceeded|maximum context length|context window|input (?:token )?limit|prompt (?:is )?too long|input length too long|model_context_window_exceeded|上下文过长|输入内容过长/i;

/** Must NOT be treated as overflow (Pi should retry / back off). */
export const RATE_LIMIT_PATTERN =
  /rate[_ ]?limit|too many requests|\b429\b|retry.after|overloaded|capacity|throttl/i;

/**
 * Returns a rewritten errorMessage starting with context_length_exceeded,
 * or null when the message should be left alone.
 */
export function rewriteOverflowError(
  message: OverflowProbeMessage,
  providerFromModel?: string,
): string | null {
  if (message.role !== "assistant") return null;
  if (message.stopReason !== "error") return null;
  if (
    message.provider !== OPENFERENCE_PROVIDER &&
    providerFromModel !== OPENFERENCE_PROVIDER
  ) {
    return null;
  }

  const errorMessage = message.errorMessage ?? "";
  if (!errorMessage) return null;
  if (errorMessage.startsWith("context_length_exceeded")) return null;
  if (RATE_LIMIT_PATTERN.test(errorMessage)) return null;
  if (!OVERFLOW_PATTERN.test(errorMessage)) return null;

  return `context_length_exceeded: ${errorMessage}`;
}