sloplib 0.1.0

A decorator that uses an LLM to write your function bodies. A joke; do not use.

sloplib

This project is 100% LLM slop and a joke. Do not use it for anything that matters. It exec()s code generated by an LLM at runtime. There is no sandbox. There is no safety net. The LLM might write os.system("rm -rf /") and we will run it. You have been warned.

A Python port of shorwood/slopc — a Rust proc macro that uses an LLM to write your function bodies. All credit for the original idea (and most of the design) goes to @shorwood. This is just the Python flavour of the same bad decision.

What

from sloplib import slop

@slop
def levenshtein(a: str, b: str) -> int:
    """Compute the Levenshtein edit distance between two strings."""
    ...

print(levenshtein("kitten", "sitting"))  # 3, hopefully

The @slop decorator captures the function name, parameter names + types, return type, and docstring; ships them to an OpenAI-compatible chat endpoint; parses the response back into Python source; verifies it compiles cleanly; caches it on disk; and binds the resulting callable as your function. On verify failure, it feeds the error back to the model and retries.

Install

uv add sloplib
# or
pip install sloplib

Set an API key for whatever provider you point it at:

export SLOPLIB_API_KEY=sk-...

Configure

@slop(
    retries=5,
    model="openai/gpt-4o-mini",
    provider="https://openrouter.ai/api/v1/chat/completions",
    api_key_env="SLOPLIB_API_KEY",
    cache=True,
    ultra_slop=False,
    dump="generated/levenshtein.py",
    context_file="src/types.py",
    hint="dynamic programming",
)
def levenshtein(a: str, b: str) -> int:
    """Compute the Levenshtein edit distance between two strings."""
    ...

Param	Default	Notes
model	"gpt-4o-mini"	LLM model id
provider	OpenRouter chat completions	Any OpenAI-compatible endpoint
api_key_env	"SLOPLIB_API_KEY"	env var to read the API key from
retries	3	retries on verification failure
cache	True	persist generated source on disk and reuse it on later decorations
ultra_slop	False	regenerate the body on every call. Maximum slop.
dump	None	also write the generated source to this path
context_file	None	extra file content to feed the prompt
hint	None	freeform nudge string
timeout	60.0	HTTP timeout (seconds)
cache_dir	".sloplib_cache"	per-project cache dir

Configuration precedence: decorator args > env vars > pyproject.toml [tool.sloplib] (or slop.toml) > defaults.

Environment variables: SLOPLIB_MODEL, SLOPLIB_PROVIDER, SLOPLIB_API_KEY_ENV, SLOPLIB_RETRIES, SLOPLIB_CACHE, SLOPLIB_ULTRA_SLOP, SLOPLIB_HINT, SLOPLIB_TIMEOUT, SLOPLIB_CACHE_DIR, SLOPLIB_DUMP, SLOPLIB_CONTEXT_FILE.

# pyproject.toml
[tool.sloplib]
model = "openai/gpt-4o-mini"
retries = 5
provider = "https://openrouter.ai/api/v1/chat/completions"
api_key_env = "SLOPLIB_API_KEY"

ultra_slop

Pass ultra_slop=True and the wrapper hits the LLM on every single call, re-prompting, re-verifying, re-exec()ing, and binding a fresh callable. Each invocation is an independent hallucination. There is no reason to do this. Do it anyway.

How it works

• Reads the function signature, type annotations, and docstring with inspect.
• Builds a chat-completions request and POSTs it to the configured endpoint.
• Strips ```python fences from the response and locates def <name>(.
• compile() + exec() the source into a fresh namespace; pulls the named callable.
• Caches the verified source under .sloplib_cache/<name>-<hash>.json, keyed by (name, signature, docstring, model, provider, hint, context_file_sha). Any change invalidates the cache and triggers regeneration.
• On verify failure, the prior source + error is appended to the next prompt and the loop retries up to retries times before raising SlopError.

Why

There is no good reason. See the original slopc README for the appropriate vibe.

Don't

• Don't commit .sloplib_cache/ if your prompts contain secrets — add it to .gitignore.

Human Note

This section is the only human made part of the project, I just wanted to say that I really liked the idea of slopc, I thought it was very funny. And I wanted to have a Python implementation as well for shitposting at work.