OpenAI Adapter

@accordkit/provider-openai decorates the official OpenAI JavaScript SDK. It emits normalized AccordKit events (message, tool_call, usage, tool_result, span) without changing how you call chat.completions.create or the newer beta namespaces.

Install

pnpm add @accordkit/tracer @accordkit/provider-openai openai

Wrap the client

src/openai.ts

import OpenAI from 'openai';
import { FileSink, Tracer } from '@accordkit/tracer';
import { withOpenAI } from '@accordkit/provider-openai';

const tracer = new Tracer({ sink: new FileSink() });

const client = withOpenAI(
  new OpenAI({ apiKey: process.env.OPENAI_API_KEY }),
  tracer,
);

withOpenAI returns a proxy that mirrors the SDK. Re-wrapping the same client is idempotent—the adapter remembers the proxy and returns it.

Emitted events

Event	When it fires	Key fields
message	Prompts (before the request) and assistant completions	role, content, model, requestId
tool_call	Assistant requests a function/tool	tool, parsed input, $ext.id
usage	OpenAI reports token accounting	inputTokens, outputTokens, $ext.totalTokens
tool_result	Request completes (success or error)	ok, latencyMs, output summary
span	Surrounds each API invocation	operation, durationMs, status, attrs.model

All events share the same trace context so downstream tooling can correlate them easily.

Options

Adapter options

withOpenAI(client, tracer, {
  enableResponsesApi: true,
  enableImagesApi: true,
  enableAudioApi: false,
  provider: 'openai',
  operationName: 'my-app.chat',
  emitPrompts: true,
  emitResponses: true,
  emitToolCalls: true,
  emitUsage: true,
  emitToolResults: true,
  emitSpan: true,
});

Option	Default	Effect
enableResponsesApi	false	Wrap the beta responses namespace so responses.create emits AccordKit events.
enableImagesApi	false	Instrument images.generate to emit tool_result and span events (no binary payloads).
enableAudioApi	false	Instrument audio.speech/transcriptions/translations.
provider	'openai'	Provider label attached to every emitted event.
operationName	'openai.chat.completions.create'	Name recorded on tool_result and span events.
emitPrompts	true	Emit message events for system/user prompts.
emitResponses	true	Emit message events for assistant replies.
emitToolCalls	true	Emit tool_call events for tool/function requests.
emitUsage	true	Emit usage events when the SDK returns usage.
emitToolResults	true	Emit tool_result events summarising success/error and latency.
emitSpan	true	Emit a span around every request.

Streaming behaviour

Streaming responses are detected automatically. The adapter waits for finalChatCompletion() (or an equivalent helper) and then emits the same completion, usage, tool result, and span events you get for non-streaming calls. Intermediate deltas are not emitted individually yet, so the stream API remains unchanged for your application.

Redacting sensitive fields

Use tracer middleware to scrub content before it leaves your process.

src/redaction.ts

import { HttpSink, Tracer, type TraceMiddleware } from '@accordkit/tracer';

const redactEmails: TraceMiddleware = (event) => {
  if (event.type === 'message' && typeof event.content === 'string') {
    event.content = event.content.replace(/\S+@\S+/g, '[redacted]');
  }
  return event;
};

const tracer = new Tracer({
  sink: new HttpSink({ endpoint: 'https://example.com/ingest' }),
  middlewares: [redactEmails],
});

Middleware runs on every emitted event, regardless of whether it originated from your code or the adapter.

Next steps

• Review the event model to see the exact JSONL shapes.
• Explore sinks to decide where events should be delivered.
• Check the examples/openai-examples project for runnable CLIs.