Function recursiveStream

core/edit/recursiveStream.ts:24–121

(
  llm: ILLM,
  abortController: AbortController,
  type: StreamDiffLinesType,
  prompt: ChatMessage[] | string,
  prediction: Prediction | undefined,
  currentBuffer = "",
  isContinuation = false,
)

const RECURSIVE_PROMPT = `Continue EXACTLY where you left`;

export async function* recursiveStream(
  llm: ILLM,
  abortController: AbortController,
  type: StreamDiffLinesType,
  prompt: ChatMessage[] | string,
  prediction: Prediction | undefined,
  currentBuffer = "",
  isContinuation = false,
): AsyncGenerator<string | ChatMessage> {
  const maxTokens = llm.completionOptions?.maxTokens ?? DEFAULT_MAX_TOKENS;
  const safeTokens = maxTokens * INFINITE_STREAM_SAFETY;
  let totalTokens = 0;
  let buffer = currentBuffer;
  // let whiteSpaceAtEndOfBuffer = buffer.match(/\s*$/)?.[0] ?? ""; // attempts at fixing whitespace bug with recursive boundaries

  const injectApplyToken = type === "apply" && shouldInjectApplyToken(llm);
  if (typeof prompt === "string") {
    const finalPrompt = injectApplyToken ? prompt + APPLY_UNIQUE_TOKEN : prompt;

    const generator = llm.streamComplete(finalPrompt, abortController.signal, {
      raw: true,
      prediction: undefined,
      reasoning: false,
    });

    for await (const chunk of generator) {
      yield chunk;
      buffer += chunk;
      totalTokens += countTokens(chunk);

      if (totalTokens >= safeTokens) {
        throw new Error(
          "Token limit reached. File/range likely too large for this edit",
        );
        // const continuationPrompt = `${RECURSIVE_PROMPT}:\n\n${buffer}`;

        // await generator.return(DUD_PROMPT_LOG); // kill the previous generator

        // // TODO - Prediction capabilities lost because of partial input
        // yield* recursiveStream(
        //   llm,
        //   abortController,
        //   continuationPrompt,
        //   undefined,
        //   buffer,
        //   true,
        // ); // Recursively stream the continuation

        // return;
      }
    }
  } else {
    const promptMessages = injectApplyToken
      ? appendTokenToLastMessage(prompt, APPLY_UNIQUE_TOKEN)
      : prompt;

    const generator = llm.streamChat(promptMessages, abortController.signal, {
      raw: true,