py-llm7-code 2025.9.81220

LLM7-powered generator that creates a minimal single-function Python package and runner from a natural-language spec.

py_llm7_code

py_llm7_code generates a minimal single-function Python package from a natural-language specification using an LLM7 ChatLLM7 client.
It enforces one public function in __init__.py, performs safety/syntax/import checks, and runs a minimal demo (run.py).

Installation

# From PyPI
pip install py_llm7_code

# Or install directly from source
pip install "git+https://github.com/chigwell/py_llm7_code"

Usage

from langchain_llm7 import ChatLLM7
from py_llm7_code.orchestrator import generate_package_with_llm7

llm = ChatLLM7(base_url="https://api.llm7.io/v1")

spec = """
Create a simple package with exactly one function `greet()` that returns 'hello'.
No network calls. Pure stdlib.
"""

result = generate_package_with_llm7(
    llm=llm,
    spec_text=spec,
    package_name="greet_pkg",
    verbose=True,
)

assert result["success"], result.get("error_message")

print("Package dir:", result["package_dir"])
print("Package name:", result["package_name"])
print("Run stdout:", result["stdout"])  # e.g. {"result": "hello"}

Generated structure (example):

/tmp/llm7_genpkg_xxxxxx/
├─ greet_pkg/
│  └─ __init__.py      # contains exactly one function
└─ run.py              # imports and executes the function

Features

• One-function enforcement: rejects multiple functions/classes in __init__.py.
• Safety checks: blocks dangerous imports (subprocess, sockets, HTTP libs, etc.).
• Syntax & import validation: py_compile + import reflection.
• Sandboxed execution: timeouts and best-effort memory limits (POSIX resource).
• Optional venv & deps: create a temp venv and pip install requested packages.
• Retry loop: automatically retries generation with feedback.
• Allow-list for imports: restrict third-party imports to a provided list.
• Custom indexes: supports --index-url and --extra-index-url.

API

generate_package_with_llm7(
    llm: ChatLLM7,
    spec_text: str,
    *,
    package_name: str | None = None,
    max_retries: int = 10,
    exec_timeout_sec: int = 8,
    memory_limit_mb: int = 256,
    verbose: bool = False,
    pip_packages: list[str] | None = None,
    allowed_imports: list[str] | None = None,
    pip_index_url: str | None = None,
    pip_extra_index_urls: list[str] | None = None,
    pip_no_deps: bool = False,
) -> dict[str, any]

Key return fields:

• success: bool
• package_dir: str
• package_name: str
• init_py_code: str
• runner_code: str
• run_cmd: str
• stdout: str
• stderr: str
• attempt_count: int
• pip_packages_installed: list[str]
• venv_python: str

Notes

• Memory limits rely on POSIX resource and may be ignored on non-POSIX systems.
• By default, third-party imports are disallowed unless explicitly allow-listed via allowed_imports.
• Network/file-system side effects are restricted; the generator scans for disallowed patterns.

Contributing

Issues and PRs are welcome. Please use the issue tracker.

License

py_llm7_code is licensed under the MIT License.