aicode-verify

aicode-verify

Catch AI-generated Python that looks right but breaks at runtime.

aicode-verify checks Python code without executing it. It catches hallucinated imports, wrong keyword arguments, missing required arguments, and common unsafe patterns such as eval, shell=True, hardcoded secrets, unsafe pickle usage, weak hashes, and SQL f-strings.

It is not a formatter. It is not a style linter.

It is a semantic safety net for code written by LLMs, coding agents, and hurried humans.

Why This Exists

AI-generated Python often fails in ways that style linters cannot see:

• imports that sound real but do not exist
• keyword arguments copied from old docs or imagined APIs
• helper functions called with missing arguments
• unsafe shell, SQL, pickle, or secret-handling snippets

aicode-verify focuses on that layer.

Install

pip install aicode-verify

For local development:

pip install -e ".[dev]"

Quickstart

aicode-verify .
aicode-verify src tests --fail-on high
aicode-verify script.py --format json
cat generated.py | aicode-verify --stdin

--fail-on accepts error, high, or medium.

Examples

Wrong keyword argument:

import math

math.sqrt(value=4)

Output:

AV101 ERROR line 3: math.sqrt() does not accept keyword argument 'value'
Suggestion: This callable does not accept keyword arguments

Hallucinated import:

from math import sqrtt

Output:

AV002 ERROR line 1: 'sqrtt' does not exist in 'math'
Suggestion: Did you mean 'sqrt'?

Alias-aware unsafe subprocess usage:

import subprocess as sp

sp.run(user_input, shell=True)

Output:

AV202 HIGH line 3: subprocess with shell=True enables shell injection
Suggestion: Pass a list of arguments instead: subprocess.run(['cmd', 'arg'])

SQL f-string:

query = f"SELECT * FROM users WHERE id = {user_id}"

Output:

AV204 HIGH line 1: SQL query built with f-string; possible injection surface
Suggestion: Use parameterized queries

Why Not Ruff, Mypy, Pyright, Bandit, or Pytest?

Use them. aicode-verify is designed to sit beside them.

Ruff keeps code clean. mypy and pyright reason about types. Bandit checks security smells. pytest proves behavior.

aicode-verify focuses on mistakes that are especially common in AI-generated Python: hallucinated imports, wrong function arguments, missing required parameters, and unsafe snippets.

Configuration

Add configuration to pyproject.toml:

[tool.aicode-verify]
fail-on = "high"
format = "text"
ignore = ["AV206"]
exclude = ["tests/fixtures/**", "migrations/**"]

CLI flags override config values:

aicode-verify . --config pyproject.toml
aicode-verify . --ignore AV203
aicode-verify --explain AV101

Ignore Comments

Suppress one finding on the next line:

# aicode-verify: ignore-next-line AV203
TEST_API_KEY = "fake-key-for-tests"

Suppress the next line for all rules:

# aicode-verify: ignore
eval(expr)

Inline ignores also work:

eval(expr)  # aicode-verify: ignore AV201

Rule Reference

Python API

from aicode_verify import format_report, verify

source = "import math\nmath.sqrt(value=4)\n"
findings = verify(source, ignore=["AV203"])
print(format_report(findings, source))

Pre-commit

Add this to .pre-commit-config.yaml:

repos:
  - repo: https://github.com/Github-Rajesh/AiCode-Verify
    rev: v0.2.0
    hooks:
      - id: aicode-verify
        args: ["--fail-on", "high"]

GitHub Actions

name: aicode-verify

on:
  pull_request:
  push:

jobs:
  verify-ai-code:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - run: pip install aicode-verify
      - run: aicode-verify . --fail-on high

Limitations

aicode-verify does not perform full type inference. Chained calls like df.groupby("x").agg(...) are only checked where the root object can be resolved statically. That keeps the tool fast and practical while leaving room for future stub-based inference.