Show HN: Repair-JSON-stream – Fix broken JSON from LLM streaming (1.7x faster)

[prxtenses]

I've been building LLM-powered apps and kept hitting the same problem: when you stream JSON from OpenAI/Anthropic, it arrives incomplete mid-generation.

{"message": "I'm currently generating your resp

JSON.parse dies. You either wait for the full response (slow) or try to parse incrementally (hard). I wrote a single-pass state machine that repairs broken JSON as chunks arrive.

Technical approach: - Zero external dependencies - everything from scratch - No regex (avoids ReDoS vulnerabilities) - O(n) single-pass processing - Stack-based context tracking - Character classification via bitmask lookup table - Works in Node.js, Deno, Bun, browsers, Cloudflare Workers

What it handles: - Truncated strings and unclosed brackets - Python constants (None, True, False) - Single quotes, trailing commas, unquoted keys - JSONP wrappers, MongoDB types (NumberLong) - LLM "thinking" blocks and markdown fences - String concatenation ("a" + "b")

The streaming benchmark shows 1.7x faster than jsonrepair - we avoid re-parsing the entire document on each chunk. 7KB minified. TypeScript-first with full type definitions.

Curious what edge cases others have hit - always looking to improve coverage.

[dmarwicke]

does this handle the thinking blocks that claude puts in? those break json.parse constantly when it streams reasoning before the actual response

[prxtenses]

Yes! We added that in v1.2.0 with the extract module

Supported thinking block patterns:

<thinking>...</thinking> (Claude) | <thought>...</thought> (DeepSeek) | <reasoning>...</reasoning> | <scratchpad>...</scratchpad> |

Also handles markdown code blocks and common LLM prose like "Here's the JSON:" or "Hope this helps!"