Why Your OpenAI JSON Calls Randomly Fail with "could not parse JSON body" (And How to Fix It)

``You're calling OpenAI from Node.js. 99% of requests work. Then randomly:

BadRequestError: 400 could not parse JSON body

Or:

SyntaxError: Invalid JSON: EOF while parsing an object

Or, if you're using tool calling:

Invalid JSON in tool call arguments

You check your code. Nothing changed. You retry manually — it works. You deploy again. It breaks again.

What's actually happening

These errors are NOT your bugs. They're transient artifacts from:

1. Proxy / network corruption — the request body arrives at OpenAI partially mangled. OpenAI returns 400.
2. Streaming truncation — the SDK parses JSON before the stream signals completion. Especially with max_output_tokens set low.
3. Model-side tool call malformation — the model emits single quotes, trailing commas, or markdown code fences inside function.arguments.

Every one of these is intermittent. So you can't debug it with a stack trace. You just suffer.

The wrong fix

People usually do this:

try {
  const res = await client.chat.completions.create({...});
} catch (e) {
  const res = await client.chat.completions.create({...});
}

Problems:

• You retry errors that should NOT be retried (invalid API key, rate limit that needs backoff, context length exceeded).
• You don't sanitize tool call args, so malformed JSON still crashes downstream.
• You add 20 lines of retry boilerplate per call site.

The right fix: classify then act

Different errors need different handling:

Error	Correct action
Transient 400 (could not parse JSON body)	Retry with backoff + jitter
Truncated stream (EOF while parsing)	Retry, possibly with higher max_tokens
Malformed tool args	Sanitize inline, no retry needed
Invalid API key	Fail fast, do NOT retry
Rate limit	Retry with longer backoff

Doing this manually is ~150 lines of code you'll write badly. Or:

npm install @coder12-z/llm-shield

import { shield } from "@coder12-z/llm-shield";
import OpenAI from "openai";

const client = new OpenAI();

const safeCall = shield(async (prompt) =>
  client.chat.completions.create({
    model: "gpt-4o-mini",
    messages: [{ role: "user", content: prompt }],
    response_format: { type: "json_object" },
  }),
  {
    maxRetries: 3,
    onRetry: ({ attempt, kind }) => console.log(`retry #${attempt} (${kind})`),
  }
);

Three lines of config. Zero dependencies. Works with OpenAI, Anthropic, Gemini, or anything that throws Error objects with status and message.

What it doesn't do

• It doesn't repair arbitrary JSON strings — for that, use jsonrepair (https://www.npmjs.com/package/jsonrepair).
• It doesn't call LLMs for you — it wraps your existing calls.
• No API key. No network calls. No telemetry.

Links

• npm: https://www.npmjs.com/package/@coder12-z/llm-shield
• GitHub: https://github.com/zey-netizen/llm-shield
• Landing: https://zey-netizen.github.io/llm-shield/

MIT licensed. Feedback welcome.