jsonmend (JavaScript)

Mends the JSON your LLM almost wrote.

JavaScript port of jsonmend — same engine design, same repair semantics, verified against the same conformance corpus (485/485). The Python and JS implementations repair identically, so a model output parses the same on both sides of your stack.

Zero dependencies. Node ≥ 14, browsers, workers.

npm install jsonmend

Usage

import { repairJson, loads, Mender } from "jsonmend";

repairJson("{'name': 'John', age: 31");
// '{"name": "John", "age": 31}'

loads('```json\n{"ok": true,}\n```');
// { ok: true }

Streaming (true incremental)

const m = new Mender();
for await (const chunk of llmStream) {
  const partial = m.feed(chunk);   // best-effort value, O(new bytes)
  render(partial);                  // e.g. { answer: "The capital of Fr" }
}
const value = m.close();

Each feed() parses only the new bytes — re-parsing the whole buffer per chunk (what you have to do with a batch repairer) is O(n²) in total and hundreds of times slower on long outputs.

Note: V8 re-flattens the growing buffer string on access, so feeding a very large output one character at a time is super-linear on the JS side. Feed in reasonable chunks (whole SSE/tool-call deltas, or ≥256 B) and it stays fast. A prefix-discarding buffer that makes any feed size O(new bytes) is planned. (The Python package is already amortised O(1) per character.)

API

• repairJson(text, options?) → repaired JSON string. Options: { returnObjects, skipJsonParse, strict }.
• loads(text, options?) → parsed value (uses a JSON.parse fast path for valid input).
• mend(text, options?) → parsed value, always through the repair machine.
• new Mender() → feed(chunk), value, close().

Integers beyond Number.MAX_SAFE_INTEGER are preserved as BigInt and serialized with exact digits. Output is always valid RFC 8259 JSON (NaN/Infinity serialize as null) and always UTF-8 encodable (lone surrogates are replaced).

What it fixes

Truncated objects/arrays/strings/numbers/literals · markdown fences · single/smart quotes · unescaped inner quotes · bare keys/values · Python literals · comments · trailing/missing commas · mismatched brackets · NDJSON · string concatenation · JSONP/MongoDB wrappers · tuples · ellipsis · BOM · broken escapes · 100k-deep nesting (no recursion).

See the main README for the full story, benchmarks and the conformance corpus.

License

MIT. The conformance corpus is CC0.