Function streamChatResponse

extensions/cli/src/stream/streamChatResponse.ts:423–591

(
  chatHistory: ChatHistoryItem[],
  model: ModelConfig,
  llmApi: BaseLlmApi,
  abortController: AbortController,
  callbacks?: StreamCallbacks,
  isCompacting = false,
)

// Main function that handles the conversation loop
// eslint-disable-next-line max-params
export async function streamChatResponse(
  chatHistory: ChatHistoryItem[],
  model: ModelConfig,
  llmApi: BaseLlmApi,
  abortController: AbortController,
  callbacks?: StreamCallbacks,
  isCompacting = false,
) {
  logger.debug("streamChatResponse called", {
    model,
    historyLength: chatHistory.length,
    hasCallbacks: !!callbacks,
  });

  const isHeadless = services.toolPermissions.isHeadless();

  let fullResponse = "";
  let finalResponse = "";
  let compactionOccurredThisTurn = false; // Track if compaction happened during this conversation turn

  while (true) {
    // If ChatHistoryService is available, refresh local chatHistory view
    chatHistory = refreshChatHistoryFromService(chatHistory, isCompacting);
    logger.debug("Starting conversation iteration");

    // Get system message once per iteration (can change based on tool permissions mode)
    const systemMessage = await services.systemMessage.getSystemMessage(
      services.toolPermissions.getState().currentMode,
    );

    // Recompute tools on each iteration to handle mode changes during streaming
    const rawTools = await getRequestTools(isHeadless);
    const tools = applyChatCompletionToolOverrides(
      rawTools,
      model.chatOptions?.toolOverrides,
    );

    // Pre-API auto-compaction checkpoint (now includes tools)
    const preCompactionResult = await handlePreApiCompaction(chatHistory, {
      model,
      llmApi,
      isCompacting,
      isHeadless,
      callbacks,
      systemMessage,
      tools,
    });
    chatHistory = preCompactionResult.chatHistory;
    if (preCompactionResult.wasCompacted) {
      compactionOccurredThisTurn = true;
    }

    logger.debug("Tools prepared", {
      toolCount: tools.length,
      toolNames: tools.map((t) => t.function.name),
    });

    // Get response from LLM