tsdown-mono/packages/xllm/src/client/stream-with-tools.ts

import type {
  XChatWithToolsOptions,
  XClientOptions,
  XStreamEvent,
  XToolCall,
} from "../core/types";
import { stream } from "./stream";
import {
  buildAssistantMessage,
  buildRequestBase,
  createToolLoopConfig,
  executeToolCalls,
  throwIfMaxRoundsExceeded,
} from "./tool-loop-runner";

export const streamWithTools = async function* (
  options: XClientOptions,
  input: XChatWithToolsOptions,
  executors: Record<string, (args: unknown, toolCall: import("../core/types").XToolCall) => Promise<unknown> | unknown>,
): AsyncGenerator<XStreamEvent> {
  const config = createToolLoopConfig(input, options);
  const requestBase = buildRequestBase(input, true);

  let rounds = 0;
  let currentMessages = [...input.messages];

  while (rounds < config.maxRounds) {
    rounds += 1;
    const toolCalls: XToolCall[] = [];
    let assistantText = "";
    let assistantReasoning = "";

    for await (const event of stream(options, {
      ...requestBase,
      messages: currentMessages,
    })) {
      if (event.type === "text.delta") {
        assistantText += event.text;
      }
      if (event.type === "reasoning.delta") {
        assistantReasoning += event.text;
      }
      if (event.type === "tool_call.done") {
        toolCalls.push(event.toolCall);
      }
      yield event;
    }

    if (toolCalls.length === 0) {
      return;
    }

    const { toolMessages } = await executeToolCalls(config.provider, toolCalls, executors, config.strategy);

    const assistantMessage = buildAssistantMessage(assistantText, toolCalls, assistantReasoning);
    currentMessages = [...currentMessages, assistantMessage, ...toolMessages];
  }

  throwIfMaxRoundsExceeded(config);
};