Do you handle JSON numbers safely by default or do you require that people make their own deserializers for numbers that would lose precision when coerced into Python's float type? The most common mistake that I see JSON libraries make is using fixed precision floating point types somewhere in the process when handling numbers while JSON's number type specifies no such limitation, which then causes precision loss unless people catch the problem and do their own pre-serialization.

Hi HN - I’m the author of Jsonic.

I built it after repeatedly running into friction with Python’s built-in json module when working with classes, dataclasses, nested objects, and type hints.

Jsonic focuses on: - Zero-boilerplate serialization and deserialization - Strict type validation with clear errors - Natural support for dataclasses, enums, tuples, sets, nested objects etc. - Optional field exclusion (e.g. hiding sensitive data) - Extra features like transient fields definition, suport for __slots__ classes etc. - Clean interop with Pydantic models

The goal is to make JSON round-tripping feel Pythonic and predictable without writing to_dict() / from_dict() everywhere.

I’d really appreciate feedback on the API design and tradeoffs.

The degree of LLM writing here makes it hard to determine which parts of this are novel and which parts are derivations of existing popular libraries like Pydantic and msgspec.

I also don't think either Pydantic or msgspec struggles with any of the "gotcha" cases in the post. Both can understand enums, type tagging, literals, etc.

Were there any particular challenges when implementing your library? I have implemented my own serialization library [1] (with a focus on not allowing arbitrary code execution), but had skipped dataclasses for now, since they seemed difficult to get right. What was your experience?

[1] https://github.com/99991/safeserialize

Side note: I think that a warning in the README about arbitrary code execution for deserialization of untrusted inputs would be nice.

The article would benefit from a very clear and explicit section on pydantic model_dump_json() vs your tool. As that's the primary thing you're tool is likely competing against

Thanks for sharing, could you please comment on the performance aspect vis-a-vis json reader/writer provided by pydantic

Sorry to be a hater, but wouldn’t using Pydantic be better in almost every circumstance here?

Looks useful. Will try it out. Thanks for making it.