Function start

web/src/llm-api/openai.ts:438–687

(controller)

  const stream = new ReadableStream({
    async start(controller) {
      const decoder = new TextDecoder()
      let buffer = ''

      controller.enqueue(
        new TextEncoder().encode(`: connected ${new Date().toISOString()}\n`),
      )

      heartbeatInterval = setInterval(() => {
        if (!clientDisconnected) {
          try {
            controller.enqueue(
              new TextEncoder().encode(
                `: heartbeat ${new Date().toISOString()}\n\n`,
              ),
            )
          } catch {
            // client disconnected
          }
        }
      }, 30000)

      try {
        let done = false
        while (!done) {
          const result = await reader.read()
          done = result.done
          const value = result.value

          if (done) {
            break
          }

          buffer += decoder.decode(value, { stream: true })
          let lineEnd = buffer.indexOf('\n')

          while (lineEnd !== -1) {
            const line = buffer.slice(0, lineEnd + 1)
            buffer = buffer.slice(lineEnd + 1)

            let billedCredits: number | undefined

            if (line.startsWith('data: ')) {
              const raw = line.slice('data: '.length).trim()
              if (raw !== '[DONE]') {
                try {
                  const obj = JSON.parse(raw)
                  const delta = obj.choices?.[0]?.delta

                  // Track time to first token (TTFT) - set on first meaningful delta (content, reasoning, or tool_calls)
                  const hasContentDelta = delta?.content && responseText.length === 0
                  const hasReasoningDelta = delta?.reasoning && reasoningText.length === 0
                  const hasToolCallsDelta = delta?.tool_calls && delta.tool_calls.length > 0
                  if (ttftMs === null && (hasContentDelta || hasReasoningDelta || hasToolCallsDelta)) {
                    ttftMs = Date.now() - startTime.getTime()
                  }

                  if (delta?.content && responseText.length < MAX_BUFFER_SIZE) {