Making DSPy reliable: self-correcting, schema-validated LLM outputs with automatic prompt optimization

I got tired of babysitting LLM prompts, so I built a small open source tool to stop.
The pattern that wore me down: every LLM call in a real agent needs a wrapper. Parse the JSON, catch the field that didn't come back, re-prompt, hope it works this time. And every time I swapped models, the prompt I'd spent an afternoon tuning would quietly stop working and I'd tune it again by hand. After enough of that, I stopped treating it as the cost of doing business and started treating it as a thing to fix.

So I made dspyer.

The idea is simple. You wrap an LLM step in a Pydantic schema. When the model returns something that doesn't fit, malformed JSON, a missing field, a citation it made up, dspyer tells the model exactly what was wrong and asks again until it conforms, or stops after however many retries you allow. It's one decorator on a normal typed function. No try/except, no parsing glue.

The part I actually care about is what that buys you. The step compiles down to a standard DSPy module, so instead of hand editing prompts you point a DSPy optimizer at a handful of examples and let it tune them, then save the result and load it in production. That was the whole reason I went down this road. I wanted my prompts to stop being something I maintain.

A few deliberate choices, since they're the part I'd want to know about if I were reading this:

• It doesn't care which model you run. OpenAI, Claude, Gemini, or a local Ollama model with no API key at all.
• It doesn't make you rewrite anything. If you're already on LangGraph, your deterministic and tool nodes stay plain Python. Only the reasoning nodes get wrapped, because those are the only ones a prompt optimizer can actually help.
• And it's honest about its limits. There is no auto convert your whole agent magic, and the README says so. The value is a small reliable runtime, not a framework you have to adopt.

It's early. 0.3.5, Apache 2.0, on PyPI. There's a quickstart that runs in about 30 seconds offline with no key if you just want to watch the self correction loop fire.

I'd genuinely rather hear where it breaks than hear that it's neat. If you build agents and any of this sounds familiar, I'd love for you to try it and tell me what's missing.

Link in the comments.