max-heal 0.2.3

Playwright wrapper with LLM-powered auto-heal and flaky auto-fix

MaxHeal

Playwright wrapper with LLM-powered auto-heal and flaky auto-fix for Python.

Features

• 🔧 Auto-heal broken selectors at runtime using an LLM (via OpenRouter)
• 🔁 Flaky guard — smart retry decorator for unstable async tests
• 🪶 Zero friction — drop-in wrapper around your existing Playwright page
• ⚡ Selector cache — healed selectors are reused within a session (no duplicate LLM calls)

Python Quick Start (Zero-Setup)

Install via pip:

pip install max-heal

Then, simply configure MaxHeal globally before your tests run (e.g., in a conftest.py or application setup file). MaxHeal will automatically intercept all playwright.sync_api.expect and playwright.async_api.expect calls across your entire framework.

import pytest
from playwright.sync_api import Page
from max_heal import MaxHealConfig, create_maxheal_page

# 1. Global AI Configuration
MAXHEAL_CONFIG = MaxHealConfig(
    api_key="sk-or-your-api-key",
    model="openai/gpt-4o-mini",
    max_retries=3
)

# 2. Wrap the global Playwright page fixture natively
@pytest.fixture
def page(page: Page):
    # This automatically activates the Global AssertPatch for this page
    return create_maxheal_page(page, MAXHEAL_CONFIG)

# 3. Write native tests normally — NO CODE CHANGES REQUIRED!
def test_login(page: Page):
    page.goto("https://example.com/login")
    
    # If the DOM changes and `#btn-signin` breaks, MaxHeal instantly freezes,
    # queries the LLM, injects the new selector, and retries the assert!
    from playwright.sync_api import expect
    expect(page.locator("#btn-signin")).to_be_visible(timeout=5000)
    page.locator("#btn-signin").click()

Providing "Intent" Context (V2)

MaxHeal works best when it knows what the user is trying to accomplish. You can inject this intent directly into the AI's brain using the built-in max_step context manager.

from max_heal import max_step

def test_login(page):
    with max_step("User fills out the login form with admin credentials"):
        page.fill("#user", "admin")
        page.fill("#pass", "secret")
        
    with max_step("User clicks the submit button"):
        page.click(".btn-primary")

If a selector fails anywhere inside a max_step block, that exact description is routed to the LLM to help exactly pinpoint the missing button or field!

Inline Intents

You can also pass intent= directly to actions for laser-focused precision:

page.click("#submit", intent="The login button on the top right corner")

Native Allure Integration

If your framework already uses allure.step, you don't even need to rewrite your code to use max_step. MaxHeal ships with a native Allure plugin that automatically parses your existing Allure steps and feeds them to the LLM.

Simply enable it in your MaxHealConfig:

MAXHEAL_CONFIG = MaxHealConfig(
    api_key="...",
    use_allure=True # Automatically syncs allure.step() into MaxHeal's LLM context!
)

Configuration

Option	Python default	Description
api_key	""	OpenRouter API key
model	openai/gpt-4o-mini	Chat model (any OpenRouter model)
base_url	https://openrouter.ai/...	OpenRouter base URL
max_retries	3	Heal attempts per selector failure
heal_enabled	True	Toggle auto-healing globally
timeout	30.0	HTTP timeout for LLM calls (seconds)
use_allure	False	Hook allure.step into LLM context

Concurrency & Thread-Safety (pytest-xdist)

MaxHeal is 100% thread-safe and async-safe. All intents, steps, and global_context mutations are securely bound using Python's native ContextVars. This means you can run hundreds of UI tests concurrently using pytest -n 16 or asyncio.gather(), and the LLM contexts will never leak or cross-pollinate between test workers!

How It Works

Test Action -> Timeout Error / Strict Mode Violation
        │
        ▼
  Snapshot DOM (aria tree + HTML)
        │
        ▼
  Prompt LLM (selector + error + DOM + Context)
        │
        ▼
  Parse healed selector
        │
        ▼
  Retry Native Assertion directly with new selector
        │
        ▼
  Cache result to prevent duplicate LLM calls

License

MIT