Fix "Premature close" error on large image responses (gpt-image-2)

[csulit]

Problem

Generating icons with larger models (notably gpt-image-2) fails deterministically:

Invalid response body while trying to fetch https://api.openai.com/v1/images/generations: Premature close

The request reaches OpenAI and the image is generated (and billed), but the CLI errors while reading the response body, so nothing is saved. Possibly the same failure as #15.

Root cause

openai@4.x ships its own HTTP stack — bundled node-fetch + agentkeepalive — selected at runtime in node_modules/openai/_shims/node-runtime.js. Reading a large base64 image payload (~1 MB) over that keep-alive socket throws Premature close before the body is fully consumed. This happens during body read, outside the SDK's maxRetries, so its built-in retries never fire. Node's built-in fetch (undici) reads the exact same responses without issue.

Fix

Pass a fetch to the OpenAI constructor that forwards to globalThis.fetch, so requests go through undici instead of the bundled node-fetch:

new OpenAI({
  apiKey,
  ...(typeof globalThis.fetch === "function"
    ? { fetch: (...args: Parameters<typeof globalThis.fetch>) => globalThis.fetch(...args) }
    : {}),
});

Guarded on typeof globalThis.fetch === "function"; engines already requires Node >= 18, where global fetch is always present.

Note: this coexists cleanly with #25 (both just add a property to the same constructor options). Upgrading to openai v5/v6 would also fix this natively (v5 dropped node-fetch for built-in fetch), but that raises the Node floor to 20 — this change is the low-risk fix for the shipping v4 line.

Verification

Same account/key/prompt, model gpt-image-2, 1024x1024:

Transport	Result
raw curl to the endpoint	OK
plain undici fetch() (no SDK)	OK
SDK default (bundled node-fetch)	FAIL (2/2)
SDK with fetch override (undici)	OK (3/3)

End-to-end confirmed: snapai icon --model gpt-image-2 now writes the file instead of throwing.

Tests

The test script pointed at jest with no config and no test files. This PR wires up ts-jest in ESM mode and adds coverage for the client construction/transport path (the fetch override and the missing-API-key error). Tests mock the openai module and ConfigService — no network calls, no real keys.

[cursor]

Thanks for the thorough investigation and writeup, @csulit! I reviewed this in depth and the fix checks out.

Verified:

• The root-cause analysis is accurate: openai@4.x bundles node-fetch v2 + agentkeepalive (confirmed in node_modules/openai/_shims/node-runtime.js), and the fetch: globalThis.fetch override routes requests through Node's built-in undici instead, bypassing both.
• Both new jest tests pass, and tsc builds cleanly.
• End-to-end: ran the CLI against a local server returning a ~2 MB payload — the request went through the override and the image saved successfully.
• The override composes cleanly with the custom base URL feature from Add baseURL configuration for OpenAI API #25/Support custom OpenAI base URL (completes #25) #35 (both are independent constructor options).

Two things needed before this can merge:

1. Merge conflicts with main — Fix CLI error double-wrapping, config read side effect, and stale flag docs #34 and Support custom OpenAI base URL (completes #25) #35 landed after this branch was cut, touching the same spots (new OpenAI({...}) in src/services/openai.ts, plus package.json and the lockfile). The resolution is straightforward and I've verified it builds and passes tests:

   • Keep both the baseURL: baseURL || undefined line from main and this PR's fetch override in the constructor
   • In package.json: keep main's sharp dependency and removed test:icon-batch script, take this PR's "test": "NODE_OPTIONS=--experimental-vm-modules jest"
   • Regenerate pnpm-lock.yaml

2. Lint failure — pnpm lint fails on the new test file:

   src/services/openai.test.ts
     12:36  error  '_options' is defined but never used  @typescript-eslint/no-unused-vars
   

   The ESLint config doesn't exempt underscore-prefixed args, so just drop the parameter: jest.fn(() => ({ images: { generate: generateMock } })).

Non-blocking notes: the ts-jest diagnostics exclusion for openai.ts is a reasonable, well-documented workaround (tsc still typechecks it), and agreed that the openai v5/v6 upgrade is the long-term fix but not worth raising the Node floor for right now.

Happy to merge once the branch is updated against main with the lint fix in.