atelier/sidecar/providers/openai.ts

import type { ImageAdapter, ImageEditRequest, ImageRequest, ImageUsage, ProviderResult, UploadedImage } from './types';

const GENERATIONS_ENDPOINT = 'https://api.openai.com/v1/images/generations';
const EDITS_ENDPOINT = 'https://api.openai.com/v1/images/edits';

/**
 * USD per million tokens, from OpenAI's published pricing.
 *
 * A table rather than a computed rate, and one that is allowed to be incomplete: a model missing here
 * yields a null cost, which is honest. Guessing a price for a model we do not know produces a number
 * that looks authoritative and sums into a total that is wrong. Every row also stores its raw `usage`,
 * so any cost here can be recomputed later if these figures move.
 */
const PRICING: Record<string, { textInput: number; imageInput: number; imageOutput: number }> = {
  'gpt-image-1': { textInput: 5, imageInput: 10, imageOutput: 40 },
  'gpt-image-1-mini': { textInput: 2, imageInput: 2.5, imageOutput: 8 },
  'gpt-image-1.5': { textInput: 5, imageInput: 8, imageOutput: 32 },
};

/** Only the parameters OpenAI accepts, and only those the caller actually set. */
function body(request: ImageRequest, n: number): Record<string, unknown> {
  const out: Record<string, unknown> = { prompt: request.prompt, model: request.model, n };

  for (const key of [
    'quality',
    'size',
    'response_format',
    'output_format',
    'output_compression',
    'background',
    'moderation',
    'style',
    'user',
  ] as const) {
    if (request[key] !== undefined) out[key] = request[key];
  }

  return out;
}

export const openaiAdapter: ImageAdapter = {
  name: 'openai',
  label: 'OpenAI',
  models: ['gpt-image-1.5', 'gpt-image-1', 'gpt-image-1-mini', 'dall-e-3', 'dall-e-2'],
  defaultModel: 'gpt-image-1',

  async generate(request, apiKey, signal): Promise<ProviderResult> {
    // One image per call, even when the caller asked for several. Each is its own row with its own
    // cost and duration, and one failure out of four should not discard the three that worked.
    const response = await fetch(GENERATIONS_ENDPOINT, {
      method: 'POST',
      headers: { 'content-type': 'application/json', authorization: `Bearer ${apiKey}` },
      body: JSON.stringify(body(request, 1)),
      signal,
    });

    return decodeImagesResponse(response, request.output_format, signal);
  },

  async edit(request, images, mask, apiKey, signal): Promise<ProviderResult> {
    // multipart/form-data, because that is what the endpoint takes. Built rather than translated:
    // this is a bridge, and every field goes across under the name OpenAI gave it.
    const form = new FormData();
    form.append('prompt', request.prompt);
    form.append('n', '1');

    // Repeated `image[]` — the wire form of an array of files, and what the API expects for the GPT
    // image models. One image still goes as `image[]`; the endpoint accepts either.
    for (const image of images) {
      form.append('image[]', new Blob([image.bytes as BlobPart], { type: image.type }), image.filename);
    }
    if (mask) form.append('mask', new Blob([mask.bytes as BlobPart], { type: mask.type }), mask.filename);

    for (const key of [
      'model',
      'quality',
      'size',
      'response_format',
      'output_format',
      'output_compression',
      'background',
      'input_fidelity',
      'user',
    ] as const) {
      const value = request[key];
      if (value !== undefined) form.append(key, String(value));
    }

    // No content-type header: fetch sets it, with the multipart boundary that only it knows.
    const response = await fetch(EDITS_ENDPOINT, {
      method: 'POST',
      headers: { authorization: `Bearer ${apiKey}` },
      body: form,
      signal,
    });

    return decodeImagesResponse(response, request.output_format, signal);
  },

  costUsd(model, usage) {
    const price = PRICING[model];
    if (!price || !usage) return null;

    const textTokens = usage.input_tokens_details?.text_tokens ?? usage.input_tokens ?? 0;
    const imageTokens = usage.input_tokens_details?.image_tokens ?? 0;
    const outputTokens = usage.output_tokens ?? 0;

    const total =
      (textTokens * price.textInput + imageTokens * price.imageInput + outputTokens * price.imageOutput) / 1_000_000;

    return total.toFixed(6);
  },
};

/**
 * Turn OpenAI's ImagesResponse into bytes, whichever endpoint produced it.
 *
 * Shared by generate and edit because the response shape is identical — only the request differs.
 */
async function decodeImagesResponse(
  response: Response,
  requestedFormat: string | undefined,
  signal?: AbortSignal,
): Promise<ProviderResult> {
  if (!response.ok) {
    const text = await response.text();
    // Surface OpenAI's own message where there is one — "content policy violation" is worth keeping.
    let detail = text.slice(0, 500);
    try {
      const parsed = JSON.parse(text) as { error?: { message?: string } };
      if (parsed.error?.message) detail = parsed.error.message;
    } catch {}
    throw new Error(`OpenAI ${response.status}: ${detail}`);
  }

  const payload = (await response.json()) as {
    data?: { b64_json?: string; url?: string; revised_prompt?: string }[];
    usage?: ImageUsage;
    background?: string;
    output_format?: string;
    size?: string;
    quality?: string;
  };

  const format = payload.output_format ?? requestedFormat ?? 'png';

  const images = await Promise.all(
    (payload.data ?? []).map(async (item) => {
      if (item.b64_json) {
        return { bytes: Buffer.from(item.b64_json, 'base64'), format, revisedPrompt: item.revised_prompt };
      }
      // dall-e-2 with the default response_format hands back a URL that expires in an hour. Fetch it
      // now: the whole point of Atelier is that the image outlives the request.
      if (item.url) {
        const bytes = new Uint8Array(await (await fetch(item.url, { signal })).arrayBuffer());
        return { bytes, format, revisedPrompt: item.revised_prompt };
      }
      throw new Error('OpenAI returned neither b64_json nor url');
    }),
  );

  return {
    images,
    usage: payload.usage,
    echoed: {
      background: payload.background,
      output_format: payload.output_format,
      size: payload.size,
      quality: payload.quality,
    },
  };
}