Function compactChatHistory

extensions/cli/src/compaction.ts:53–167

(
  chatHistory: ChatHistoryItem[],
  model: ModelConfig,
  llmApi: BaseLlmApi,
  options?: CompactionOptions,
)

 * @returns The compacted history with compaction index
 */
export async function compactChatHistory(
  chatHistory: ChatHistoryItem[],
  model: ModelConfig,
  llmApi: BaseLlmApi,
  options?: CompactionOptions,
): Promise<CompactionResult> {
  const { callbacks, abortController, systemMessageTokens = 0 } = options || {};
  // Create a prompt to summarize the conversation
  const compactionPrompt: ChatHistoryItem = {
    message: {
      role: "user" as const,
      content: COMPACTION_PROMPT,
    },
    contextItems: [],
  };

  // Check if the history with compaction prompt is too long, prune if necessary
  let historyToUse = chatHistory;
  let historyForCompaction = [...historyToUse, compactionPrompt];

  const contextLimit = getModelContextLimit(model);
  const maxTokens = getModelMaxTokens(model);

  // Check if system message is already in the history to avoid double-counting
  const hasSystemMessageInHistory = chatHistory.some(
    (item) => item.message.role === "system",
  );

  // Account for system message (if not already in history) AND compaction prompt
  const systemMessageReservation = hasSystemMessageInHistory
    ? 0
    : systemMessageTokens;

  const availableForInput =
    contextLimit -
    maxTokens -
    systemMessageReservation -
    COMPACTION_PROMPT_TOKENS;

  // Check if we need to prune to fit within context
  while (
    countChatHistoryTokens(historyForCompaction, model) > availableForInput &&
    historyToUse.length > 0
  ) {
    logger.debug("Compaction history too long, pruning last message", {
      tokenCount: countChatHistoryTokens(historyForCompaction, model),
      availableForInput,
      historyLength: historyToUse.length,
    });
    const prunedHistory = pruneLastMessage(historyToUse);

    // Break if pruning didn't change the history (prevents infinite loop)
    if (prunedHistory.length === historyToUse.length) {
      logger.warn(
        "Cannot prune history further while maintaining valid conversation structure",
      );
      break;
    }