Function handleStreamChunk

web/src/llm-api/siliconflow.ts:500–572

({
  data,
  state,
  startTime,
  logger,
  userId,
  agentId,
  model,
}: {
  data: Record<string, unknown>
  state: StreamState
  startTime: Date
  logger: Logger
  userId: string
  agentId: string
  model: string
})

}

function handleStreamChunk({
  data,
  state,
  startTime,
  logger,
  userId,
  agentId,
  model,
}: {
  data: Record<string, unknown>
  state: StreamState
  startTime: Date
  logger: Logger
  userId: string
  agentId: string
  model: string
}): StreamState {
  const MAX_BUFFER_SIZE = 1 * 1024 * 1024

  if ('error' in data) {
    const errorData = data.error as Record<string, unknown>
    logger.error(
      {
        userId,
        agentId,
        model,
        errorCode: errorData?.code,
        errorType: errorData?.type,
        errorMessage: errorData?.message,
      },
      'Received error chunk in SiliconFlow stream',
    )
    return state
  }

  const choices = data.choices as Array<Record<string, unknown>> | undefined
  if (!choices?.length) {
    return state
  }
  const choice = choices[0]
  const delta = choice.delta as Record<string, unknown> | undefined

  const contentDelta = typeof delta?.content === 'string' ? delta.content : ''
  if (state.responseText.length < MAX_BUFFER_SIZE) {
    state.responseText += contentDelta
    if (state.responseText.length >= MAX_BUFFER_SIZE) {
      state.responseText =
        state.responseText.slice(0, MAX_BUFFER_SIZE) + '\n---[TRUNCATED]---'
      logger.warn({ userId, agentId, model }, 'Response text buffer truncated at 1MB')
    }
  }

  const reasoningDelta = typeof delta?.reasoning_content === 'string' ? delta.reasoning_content
    : typeof delta?.reasoning === 'string' ? delta.reasoning
    : ''

  // Track time to first token (TTFT) - set on first meaningful delta (content, reasoning, or tool_calls)
  const hasToolCallsDelta = delta?.tool_calls != null && (delta.tool_calls as unknown[])?.length > 0