Function endLLMRequestSpan

src/utils/telemetry/sessionTracing.ts:353–464

(
  span?: Span,
  metadata?: {
    inputTokens?: number
    outputTokens?: number
    cacheReadTokens?: number
    cacheCreationTokens?: number
    success?: boolean
    statusCode?: number
    error?: string
    attempt?: number
    modelResponse?: string
    /** Text output from the model (non-thinking content) */
    modelOutput?: string
    /** Thinking/reasoning output from the model */
    thinkingOutput?: string
    /** Whether the output included tool calls (look at tool spans for details) */
    hasToolCall?: boolean
    /** Time to first token in milliseconds */
    ttftMs?: number
    /** Time spent in pre-request setup before the successful attempt */
    requestSetupMs?: number
    /** Timestamps (Date.now()) of each attempt start — used to emit retry sub-spans */
    attemptStartTimes?: number[]
  },
)

 *   If not provided, falls back to finding the most recent llm_request span (legacy behavior).
 */
export function endLLMRequestSpan(
  span?: Span,
  metadata?: {
    inputTokens?: number
    outputTokens?: number
    cacheReadTokens?: number
    cacheCreationTokens?: number
    success?: boolean
    statusCode?: number
    error?: string
    attempt?: number
    modelResponse?: string
    /** Text output from the model (non-thinking content) */
    modelOutput?: string
    /** Thinking/reasoning output from the model */
    thinkingOutput?: string
    /** Whether the output included tool calls (look at tool spans for details) */
    hasToolCall?: boolean
    /** Time to first token in milliseconds */
    ttftMs?: number
    /** Time spent in pre-request setup before the successful attempt */
    requestSetupMs?: number
    /** Timestamps (Date.now()) of each attempt start — used to emit retry sub-spans */
    attemptStartTimes?: number[]
  },
): void {
  let llmSpanContext: SpanContext | undefined

  if (span) {
    // Use the provided span directly - this is the correct approach for parallel requests
    const spanId = getSpanId(span)
    llmSpanContext = activeSpans.get(spanId)?.deref()
  } else {
    // Legacy fallback: find the most recent llm_request span
    // WARNING: This can cause mismatched responses when multiple requests are in flight
    llmSpanContext = Array.from(activeSpans.values())
      .findLast(r => {
        const ctx = r.deref()
        return (
          ctx?.attributes['span.type'] === 'llm_request' ||
          ctx?.attributes['model']
        )
      })
      ?.deref()
  }

  if (!llmSpanContext) {
    // Span was already ended or never tracked
    return
  }

  const duration = Date.now() - llmSpanContext.startTime

  // End Perfetto span with full metadata
  if (llmSpanContext.perfettoSpanId) {
    endLLMRequestPerfettoSpan(llmSpanContext.perfettoSpanId, {
      ttftMs: metadata?.ttftMs,
      ttltMs: duration, // Time to last token is the total duration