Stably SDK — AI Rules (Playwright‑compatible)

Assumption: Full Playwright parity unless noted below.

Import

import { test, expect } from "@stablyai/playwright-test";

// Additional exports available:
import {
  defineConfig,       // Enhanced defineConfig with stably project support
  stablyReporter,     // Reporter for CI/cloud integration
  setApiKey,          // Programmatic API key configuration
  getDirname,         // ESM __dirname equivalent
  getFilename,        // ESM __filename equivalent
} from "@stablyai/playwright-test";

// Type imports for model selection
import type { AIModel } from "@stablyai/playwright-test";

// Email inbox testing
import { Inbox } from "@stablyai/email";

Install & Setup

npm install @playwright/test @stablyai/playwright-test
# Optional: install email testing package
npm install @stablyai/email
export STABLY_API_KEY=YOUR_KEY

import { setApiKey } from "@stablyai/playwright-test";
setApiKey("YOUR_KEY");

Environment Variables

Variable	Description	Default
STABLY_API_KEY	API key for authentication	(required)
STABLY_PROJECT_ID	Project ID for reporter	(required for reporter)
STABLY_API_URL	Custom API endpoint	https://api.stably.ai
STABLY_WS_URL	Custom WebSocket endpoint	wss://api.stably.ai/reporter

Both STABLY_API_KEY and STABLY_PROJECT_ID are shared across @stablyai/playwright-test and @stablyai/email.

Model Selection

All AI methods support an optional model parameter to specify which AI model to use:

import type { AIModel } from "@stablyai/playwright-test";

// Available models:
// - "openai/o4-mini" - OpenAI's efficient reasoning model
// - "google/gemini-3-pro-preview" - Google's most capable model
// - "google/gemini-3-flash-preview" - Google's fast, efficient model
// - Custom model strings are also supported

// Example usage with different methods:
await expect(page).aiAssert("Shows dashboard", { model: "openai/o4-mini" });
const data = await page.extract("Get heading", { model: "google/gemini-3-flash-preview" });
const { locator } = await page.getLocatorsByAI("the login button", { model: "google/gemini-3-pro-preview" });

If no model is specified, the backend default is used.

When to Use Stably SDK vs Playwright

Prioritization:

1. Test accuracy and stability are the most important factors - prioritize reliability over cost/speed.
2. Otherwise, use Playwright whenever possible since it's cheaper and faster.
3. For interactions: If the interaction will be hard to express as Playwright or will be too brittle that way (e.g., the scroll amount changes every time), then use agent.act(). Any canvas-related operations, or any drag/click operations that require coordinates, must use agent.act() (more semantic meaning, and less flaky).
4. For assertions: Use Playwright if it fulfills the purpose. But if the assertion is very visual-heavy, use Stably's aiAssert.
5. For email verification flows: Use @stablyai/email Inbox to receive and extract data from emails (OTP codes, magic links, confirmation emails).
6. Use Stably SDK methods if it helps your tests pass - when Playwright methods are insufficient or unreliable.

AI Assertions (intent‑based visuals)

Note: toMatchScreenshotPrompt is deprecated. Use aiAssert instead.

await expect(page).aiAssert(
  "Shows revenue trend chart and spotlight card",
  { timeout: 30_000 }
);
await expect(page.locator(".header"))
  .aiAssert("Nav with avatar and bell icon");

Signature: expect(page|locator).aiAssert(prompt: string, options?: ScreenshotOptions & { model?: AIModel })

• Use for dynamic UIs; keep prompts specific; scope with elements (using locators) when possible.
• Consider whether you need fullPage: true: Ask yourself if the assertion requires content beyond the visible viewport (e.g., long scrollable lists, full page layout checks). If only viewport content matters, omit fullPage: true — it's faster and cheaper. Use it only when you genuinely need to capture content outside the browser window's visible area.

AI Extraction (visual → data)

// Extract from entire page
const txt = await page.extract("List revenue, active users, and churn rate");

// Extract from specific element (more focused, better results)
const headerText = await page.locator(".header").extract("Get the username displayed");

Typed with Zod:

import { z } from "zod";
const Metrics = z.object({ revenue: z.string(), activeUsers: z.number(), churnRate: z.number() });
const m = await page.extract("Return revenue (currency), active users, churn %", { schema: Metrics });

// Also works on locators
const UserSchema = z.object({ name: z.string(), role: z.string() });
const userData = await page.locator(".user-panel").extract("Get user info", { schema: UserSchema });

Signatures:

• page.extract(prompt: string, options?: { model?: AIModel }): Promise<string>
• page.extract<T extends z.AnyZodObject>(prompt, { schema: T, model?: AIModel }): Promise<z.output<T>>
• locator.extract(prompt: string, options?: { model?: AIModel }): Promise<string>
• locator.extract<T extends z.AnyZodObject>(prompt, { schema: T, model?: AIModel }): Promise<z.output<T>>

AI Locator Finding (accessibility-based)

Use getLocatorsByAI to find elements using natural language based on the page's accessibility tree. Requires Playwright v1.54.1+.

// Find a single element
const { locator: loginBtn, count } = await page.getLocatorsByAI("the login button");
expect(count).toBe(1);
await loginBtn.click();

// Find multiple elements
const { locator: cards, count: cardCount } = await page.getLocatorsByAI("all product cards in the grid");
console.log(`Found ${cardCount} product cards`);
await expect(cards.first()).toBeVisible();

// With model selection
const { locator } = await page.getLocatorsByAI("the submit button", {
  model: "google/gemini-3-flash-preview"
});

Signature: page.getLocatorsByAI(prompt: string, options?: { model?: AIModel }): Promise<{ locator: Locator, count: number, reason: string }>

Returns:

• locator - Playwright Locator for the found elements (matches nothing if count is 0)
• count - Number of elements found
• reason - AI's explanation of what it found and why

Best Practices:

• Describe elements by their accessible properties (labels, roles, text) rather than visual attributes
• Use for elements that are hard to locate with traditional selectors
• Check the count to verify expected number of matches before interacting

AI Agent (autonomous workflows)

When Test Generation Uses agent.act() vs Raw Playwright

Stably's test generation agent intelligently chooses between raw Playwright code and agent.act():

• For repeatable, stable actions: The test generation agent will try to turn these into raw Playwright code when first generating tests. Raw Playwright is faster and more cost-effective.
• For dynamic or unstable pages: If the agent finds that the page changes frequently and it can't find stable Playwright selectors/code for an action, it will use agent.act() instead.

This means you may see a mix of both in generated tests—this is intentional and optimizes for reliability while keeping costs down where possible.

Using the Agent Fixture

Use the agent fixture to execute complex, human-like workflows:

test("complex workflow", async ({ agent, page }) => {
  await page.goto("/orders");
  await agent.act("Find the first pending order and mark it as shipped", { page });
});

// Or create manually
const agent = context.newAgent();
await agent.act("Your task here", { page, maxCycles: 10 }); // split into smaller steps if possible

Signature: agent.act(prompt: string, options: { page: Page, maxCycles?: number, model?: Model }): Promise<void>

• Throws Error if the agent fails to complete the task (check error.message for the AI's failure reason)
• Default maxCycles: 30
• Supported models: "anthropic/claude-sonnet-4-5-20250929" (default), "google/gemini-2.5-computer-use-preview-10-2025", or any custom model string

Passing Variables to Prompts

You can use template literals to pass variables into your prompts:

const duration = 24 * 7 * 60;
await agent.act(`Enter the duration of ${duration} seconds`, { page });

const username = "john.doe@example.com";
await agent.act(`Login with username ${username}`, { page });

Self-Contained Prompts

All prompts to Stably SDK AI methods (agent.act, aiAssert, extract) must be self-contained with all necessary information:

1. No implicit references to outside context - prompts cannot reference previous actions or state that the AI method doesn't have access to:

   • ❌ Bad: agent.act("Verify the field you just filled in the form is 4", { page })
   • ✅ Good: agent.act("Verify the 'timeout' field in the form has value 4", { page })
   • ❌ Bad: agent.act("Pick something that's not in the previous step", { page })
   • ✅ Good: const selectedItem = "Option A"; await agent.act(\Pick an option other than ${selectedItem}`, { page })`

2. Pass information between AI methods using explicit variables:

   // Extract data, then use it in next action
   const orderId = await page.extract("Get the order ID from the first row");
   await agent.act(`Cancel order with ID ${orderId}`, { page });

3. Include detailed instructions and domain knowledge to help the AI perform the task successfully:

   • ❌ Bad: agent.act("Fill in the form", { page })
   • ✅ Good: agent.act("Fill in the form with test data. On page 4 you might run into a popup asking for premium features - just click 'Skip' or 'Cancel' to ignore it", { page })

Optimizing Agent Performance

IMPORTANT: The fewer actions/cycles agent.act() needs to do, the better it performs. Offload work to Playwright code when possible:

1. If your prompt has work that could be done by Playwright code, use Playwright for that work, and only use agent.act() for actions that are hard for Playwright (canvas operations, dynamic decision making, etc.)
2. If your prompt has repetition (e.g., do it 5 times), calculations (e.g., type 24760 seconds), or other code-suitable tasks, use code for those, and only have agent.act() perform the agent-suitable part.
3. If your prompt has an if/else condition that can be expressed in code, use code for the condition, and only have agent.act() perform the agent-suitable part.

Examples:

• ❌ Bad: "Click the button 5 times"
• ✅ Good: "Click the button" (and include this in a loop that runs 5 times)
• ❌ Bad: "enter the duration of 24*7*60 seconds"
• ✅ Good: Calculate in code (const sum = 24*7*60), then use \enter the duration of ${sum} seconds``

Email Inbox Testing

The @stablyai/email package provides disposable email inboxes for testing email-dependent flows (OTP codes, verification links, order confirmations, etc.).

npm install @stablyai/email

Requires STABLY_API_KEY and STABLY_PROJECT_ID environment variables (same as @stablyai/playwright-test).

Creating an Inbox

import { Inbox } from "@stablyai/email";

const inbox = await Inbox.build({ suffix: `test-${Date.now()}` });
// Generates address like: "org+test-1706621234567@mail.stably.ai"

console.log(inbox.address);   // Full email address
console.log(inbox.suffix);    // Applied suffix
console.log(inbox.createdAt); // Inbox creation timestamp

Inbox.build() Parameters:

Parameter	Type	Description
suffix	string (optional)	Test isolation identifier appended to address
apiKey	string (optional)	Overrides STABLY_API_KEY env var
projectId	string (optional)	Overrides STABLY_PROJECT_ID env var

Waiting for Emails

const email = await inbox.waitForEmail({
  from: "noreply@example.com",
  subject: "verification",
  timeoutMs: 60_000,
});

inbox.waitForEmail() Parameters:

Parameter	Type	Default	Description
from	string	—	Sender address filter
subject	string	—	Subject line filter
subjectMatch	'contains' | 'exact'	'contains'	Subject matching mode
timeoutMs	number	120000	Max wait duration (ms)
pollIntervalMs	number	3000	Poll frequency (ms)

Returns an Email object or throws EmailTimeoutError. Only checks emails received after the inbox was created.

Email Object Properties

Property	Type	Description
id	string	Unique identifier
mailbox	string	Container (e.g., "INBOX")
from	{ address, name }	Sender info
to	array	Recipients
subject	string	Subject line
receivedAt	Date	Delivery timestamp
text	string?	Plain text body
html	array?	HTML body parts

AI Extraction from Emails

Use inbox.extractFromEmail() to extract structured data from emails using AI:

String extraction:

const { data: otp, reason } = await inbox.extractFromEmail({
  id: email.id,
  prompt: "Extract the 6-digit OTP code",
});

Structured extraction with Zod:

import { z } from "zod";

const { data } = await inbox.extractFromEmail({
  id: email.id,
  prompt: "Extract verification URL and expiration",
  schema: z.object({
    url: z.string().url(),
    expiresIn: z.string(),
  }),
});

inbox.extractFromEmail() Parameters:

• id (required): Email identifier
• prompt (required): Extraction instruction
• schema (optional): Zod schema for structured output

Returns { data, reason } or throws EmailExtractionError.

Listing & Retrieving Emails

// List emails with optional filters
const { emails, nextCursor } = await inbox.listEmails({
  from: "notifications@example.com",
  limit: 10,
});

// Get a specific email by ID
const email = await inbox.getEmail(emailId);

inbox.listEmails() Parameters:

Parameter	Type	Default	Description
from	string	—	Sender filter
subject	string	—	Subject filter
subjectMatch	'contains' | 'exact'	'contains'	Matching mode
limit	number	20 (max 100)	Result count
cursor	string	—	Pagination cursor
since	Date	—	Override default time filter
includeOlder	boolean	false	Include pre-creation emails

Cleanup

await inbox.deleteEmail(email.id);  // Delete single email
await inbox.deleteAllEmails();       // Delete all emails in this inbox

Only deletes inbox-scoped emails, not the entire org mailbox.

Playwright Fixture Pattern

Create a reusable inbox fixture for test isolation and automatic cleanup:

import { test as base } from "@stablyai/playwright-test";
import { Inbox } from "@stablyai/email";

const test = base.extend<{ inbox: Inbox }>({
  inbox: async ({}, use, testInfo) => {
    const inbox = await Inbox.build({ suffix: `test-${testInfo.testId}` });
    await use(inbox);
    await inbox.deleteAllEmails();
  },
});

test("OTP login flow", async ({ page, inbox }) => {
  await page.goto("/login");
  await page.getByLabel("Email").describe("Email input").fill(inbox.address);
  await page.getByRole("button", { name: "Send OTP" }).describe("Send OTP button").click();

  const email = await inbox.waitForEmail({
    subject: "verification code",
    timeoutMs: 60_000,
  });

  const { data: otp } = await inbox.extractFromEmail({
    id: email.id,
    prompt: "Extract the 6-digit OTP code",
  });

  await page.getByLabel("OTP").describe("OTP input").fill(otp);
  await page.getByRole("button", { name: "Verify" }).describe("Verify button").click();
  await expect(page).toHaveURL("/dashboard");
});

Email Testing Best Practices

• Always use unique suffixes (e.g., testInfo.testId or Date.now()) for parallel test isolation.
• Use the fixture pattern for automatic cleanup after each test.
• Keep timeoutMs reasonable — default is 120s, but 60s is usually sufficient.
• Prefer waitForEmail over polling with listEmails — it handles retry logic automatically.
• Use structured extraction with Zod schemas when you need typed data (URLs, codes, dates).
• Common extraction prompts:
  • "Extract the OTP or verification code"
  • "Extract the sign-in or verification URL"
  • "Extract the order number and delivery date"

CI Reporter / Cloud

npm install @stablyai/playwright-test

// playwright.config.ts
import { defineConfig, stablyReporter } from "@stablyai/playwright-test";

export default defineConfig({
  reporter: [
    ["list"],
    stablyReporter({
      apiKey: process.env.STABLY_API_KEY,
      projectId: process.env.STABLY_PROJECT_ID,
      // Optional: Scrub sensitive values from traces before upload
      sensitiveValues: ["secret-password", process.env.API_SECRET].filter(Boolean),
    }),
  ],
  use: {
    trace: "on", // Required for trace uploads
  },
});

Reporter Options:

• apiKey (required): Your Stably API key
• projectId (required): Your Stably project ID
• sensitiveValues (optional): Array of strings to scrub from trace files
• notificationConfigs (optional): Per-project notification settings for Slack/email

Notifications (Email/Slack)

Configure notifications per project via stably property in defineConfig:

import { defineConfig, stablyReporter } from "@stablyai/playwright-test";

export default defineConfig({
  reporter: [stablyReporter({ apiKey: "...", projectId: "..." })],
  projects: [
    {
      name: "smoke",
      stably: {
        notifications: {
          slack: {
            channelName: "#test-alerts",
            notifyOnStart: true,
            notifyOnResult: "failures-only", // "all" | "failures-only"
          },
          email: {
            to: ["team@example.com"],
            notifyOnResult: "all",
          },
        },
      },
    },
  ],
});

Commands

# Recommended for Stably reporter + auto-heal
npx stably test

# Still supported (requires your reporter/config to be set up)
npx playwright test
# All Playwright CLI flags still work (headed, ui, project, file filters…)

# When running tests for debugging/getting stacktraces:
npx playwright test --reporter=list  # disable HTML reporter, shows terminal output directly

ESM Utilities

For ESM projects needing __dirname or __filename equivalents:

import { getDirname, getFilename } from "@stablyai/playwright-test";

const __dirname = getDirname(import.meta.url);
const __filename = getFilename(import.meta.url);

// Use in tests for file paths
import path from "path";
await page.setInputFiles("input", path.join(__dirname, "fixtures", "file.pdf"));

Best Practices

• CRITICAL: All locators must use the .describe() method for readability in trace views and test reports. Example: page.getByRole('button', { name: 'Submit' }).describe('Submit button') or page.locator('table tbody tr').first().describe('First table row')
• Scope visual checks with locators; keep prompts specific with labels/units.
• Use toHaveScreenshot for stable pixel‑perfect UIs; aiAssert for dynamic UIs.
• Be deliberate with fullPage: true: Default to viewport-only screenshots. Only use fullPage: true when your assertion genuinely requires content beyond the visible viewport (e.g., verifying footer content on a long page, checking full scrollable lists). Viewport captures are faster and more cost-effective.

Troubleshooting

• Slow assertions → scope visuals; reduce viewport.
• Agent stops early → increase maxCycles or break task into smaller steps.

Minimal Template

import { test, expect } from "@stablyai/playwright-test";

test("AI‑enhanced dashboard", async ({ page, agent }) => {
  await page.goto("/dashboard");

  // Use agent for complex workflows
  await agent.act("Navigate to settings and enable notifications", { page });

  // Use AI assertions for dynamic content
  await expect(page).aiAssert(
    "Dashboard shows revenue chart (>= 6 months) and account spotlight card"
  );
});

---

Creating E2E Tests with Stably SDK

When creating end-to-end tests, follow these guidelines:

1. Understand Requirements

• Ask clarifying questions if the test scenario is unclear
• Identify the user flow, expected outcomes, and edge cases
• Determine which pages/components need testing

2. Choose the Right Tools

Use Playwright when:

• Simple, deterministic interactions (clicks, fills, selects)
• Static content that doesn't change
• Cost and speed are priorities

Use Stably SDK when:

• Visual assertions on dynamic UIs → aiAssert()
• Complex multi-step workflows → agent.act()
• Canvas interactions or coordinate-based operations → agent.act()
• Data extraction from UI → page.extract()
• Elements are hard to locate reliably → page.getLocatorsByAI()
• Email verification flows (OTP, magic links, confirmations) → Inbox from @stablyai/email

3. Structure the Test

import { test, expect } from "@stablyai/playwright-test";

test("descriptive test name", async ({ page, agent }) => {
  // 1. Setup & Navigation
  await page.goto("/your-page");

  // 2. Interactions (prefer Playwright, use agent for complex workflows)
  await page.getByRole('button', { name: 'Login' }).describe('Login button').click();

  // For complex workflows:
  await agent.act("Complete the multi-step checkout process", { page });

  // 3. Assertions (prefer Playwright, use AI assertions for dynamic content)
  await expect(page.getByText('Welcome')).toBeVisible();

  // For visual/dynamic assertions:
  await expect(page).aiAssert(
    "Dashboard shows revenue chart and user profile card"
  );
});

4. Best Practices

CRITICAL:

• Always use .describe() on locators for better traceability

  page.getByRole('button', { name: 'Submit' }).describe('Submit form button')

General:

• Write clear, descriptive test names
• Add comments explaining complex logic
• Use semantic locators (getByRole, getByLabel) over CSS selectors
• Keep prompts specific for AI assertions
• Scope visual checks with locators when possible
• Default to viewport screenshots (only use fullPage: true when needed)
• Handle async operations with proper waits
• Consider error states and edge cases

5. Example Patterns

Login Flow:

test("user can login successfully", async ({ page }) => {
  await page.goto("/login");
  await page.getByLabel('Email').describe('Email input').fill('user@example.com');
  await page.getByLabel('Password').describe('Password input').fill('password123');
  await page.getByRole('button', { name: 'Sign In' }).describe('Sign in button').click();

  await expect(page).toHaveURL('/dashboard');
  await expect(page.getByText('Welcome back')).toBeVisible();
});

Complex Workflow with AI:

test("complete checkout process", async ({ page, agent }) => {
  await page.goto("/products");

  // Use agent for complex multi-step process
  await agent.act(
    "Add 2 items to cart, proceed to checkout, and fill shipping details",
    { page }
  );

  // Verify outcome
  await expect(page).aiAssert(
    "Order confirmation page with order number and thank you message"
  );
});

Email Verification Flow:

import { Inbox } from "@stablyai/email";

test("signup with email verification", async ({ page }) => {
  const inbox = await Inbox.build({ suffix: `signup-${Date.now()}` });

  await page.goto("/signup");
  await page.getByLabel("Email").describe("Email input").fill(inbox.address);
  await page.getByLabel("Password").describe("Password input").fill("SecurePass123!");
  await page.getByRole("button", { name: "Sign Up" }).describe("Sign up button").click();

  // Wait for verification email
  const email = await inbox.waitForEmail({
    subject: "verify your email",
    timeoutMs: 60_000,
  });

  // Extract the verification link using AI
  const { data: verifyUrl } = await inbox.extractFromEmail({
    id: email.id,
    prompt: "Extract the email verification URL",
  });

  await page.goto(verifyUrl);
  await expect(page).toHaveURL("/welcome");

  // Cleanup
  await inbox.deleteAllEmails();
});

Visual Regression:

test("homepage matches design", async ({ page }) => {
  await page.goto("/");

  // AI-powered visual assertion for dynamic content
  await expect(page).aiAssert(
    "Hero section with call-to-action button, feature cards below, and navigation bar at top"
  );
});