feat: Add OpenAI Batch API Instrumentation

[Luca Forstner (lforst)]

Adds resumable tracing for OpenAI Batch jobs without making OpenAI API calls. startOpenAIBatchTrace() starts a task span with pending LLM spans and returns instrumented create parameters; completeOpenAIBatchTrace() finishes those spans from caller-supplied batch files.

const inputFile = await client.files.create({ file, purpose: "batch" });

const params = await startOpenAIBatchTrace({
  inputFile,
  input: jsonl, // JSONL string or parsed records
  params: {
    endpoint: "/v1/chat/completions",
    completion_window: "24h",
  },
});

// Only user code calls OpenAI.
const batch = await client.batches.create(params);

Complete the trace from a poller or webhook handler. File requests start concurrently; supplied response bodies are streamed and consumed:

const batch = await client.batches.retrieve(batchId);

await completeOpenAIBatchTrace({
  batch,
  inputFile: client.files.content(batch.input_file_id),
  outputFile: batch.output_file_id
    ? client.files.content(batch.output_file_id)
    : undefined,
  errorFile: batch.error_file_id
    ? client.files.content(batch.error_file_id)
    : undefined,
});

openai.batch [task]
├── Chat Completion [llm]  input → output + tokens
├── Chat Completion [llm]  input → output + tokens
└── Chat Completion [llm]  input → error