OpenAI adapter

@receipta/openai wraps the openai SDK (v5+, including current v6) so every chat completion emits a receipt.

How it works

No fork. The OpenAI SDK accepts a fetch constructor option and invokes it per HTTP attempt (including retries — verified firsthand). receipta injects a fetch that:

1. reads the request body (model, messages, params),
2. delegates to the real fetch,
3. clone()s the response and reads the clone (the original stays unconsumed so the SDK's own parser sees an intact body — non-interference),
4. builds a receipt with usage, model, and the x-request-id header,
5. appends it to the store,
6. returns the original response.

Receipt emission is wrapped in try/catch — it never throws into your call.

Usage

import OpenAI from 'openai';
import { withReceipts } from '@receipta/openai';
import { openStore, generateKeyPair } from '@receipta/core';

const store = await openStore('./receipts.log.receipta');
const signer = generateKeyPair();

const client = withReceipts(
  OpenAI,
  { apiKey: process.env.OPENAI_API_KEY! },
  {
    store,
    signer,
    actor: { type: 'service', id: 'my-app' },
    captureMode: 'full', // or "metadata_only" to omit content
  },
);

const res = await client.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Hello' }],
});

What gets captured

• provider: "openai"
• model: from the response body
• usage: prompt_tokens / completion_tokens (mapped to input_tokens / output_tokens)
• request_id: the x-request-id header
• attempt_index: best-effort, sourced from the Stainless SDK's x-stainless-retry-count request header (0 on the first attempt, incrementing on retry); omitted when the header is absent
• outcome: success for 2xx, error otherwise
• content: the request + response bodies (when captureMode: "full")
• content_commitments: HMAC-SHA256 over request/response (keyed, not bare digests)

Streaming

For streaming responses, the fetch wrapper buffers the cloned SSE stream and computes the output commitment over the final assembled message (not intermediate chunks).