Function createOpenAIResponse

test/harness/replayingCapiProxy.ts:1356–1407

(
  model: string,
  messages: NormalizedMessage[],
  responseStartIndex: number,
  workDir: string,
)

// Turns a normalized message back into a full OpenAI ChatCompletion that we can replay as a response
function createOpenAIResponse(
  model: string,
  messages: NormalizedMessage[],
  responseStartIndex: number,
  workDir: string,
): ChatCompletion {
  // Here we recreate the strange CAPI/productcode behavior of using multiple choices to represent
  // multiple assistant messages in a row. This is the inverse of the logic in transformOpenAIResponseChoice().
  // So, find all successive assistant messages starting from responseStartIndex.
  const choices: ChatCompletion.Choice[] = [];
  for (
    let index = 0;
    messages[responseStartIndex + index]?.role === "assistant";
    index++
  ) {
    const assistantMessage = messages[responseStartIndex + index];
    const toolCalls = assistantMessage.tool_calls?.map((tc, idx) => ({
      id: tc.id || `call_${idx}`,
      type: "function" as const,
      function: {
        name: expandToolName(tc.function?.name ?? ""),
        arguments: expandWorkDir(tc.function?.arguments, workDir, true) ?? "{}",
      },
    }));

    choices.push({
      index,
      message: {
        role: "assistant",
        content:
          expandWorkDir(assistantMessage.content, workDir, false) ?? null,
        refusal: assistantMessage.refusal ?? null,
        tool_calls: toolCalls,
      },
      finish_reason: toolCalls?.length ? "tool_calls" : "stop",
      logprobs: null,
    });
  }

  return {
    id: "cached-completion",
    object: "chat.completion",
    created: Math.floor(Date.now() / 1000),
    model,
    choices,
    usage: {
      prompt_tokens: 0,
      completion_tokens: 0,
      total_tokens: 0,
    },
  };
}

const STREAM_CHUNK_SIZE = 200;