Test Generation

The 8-Stage Pipeline

Each page snapshot goes through a structured pipeline — not a single prompt:

1. Crawl — visit pages, capture DOM snapshots
2. Filter — remove noise from interactive elements
3. Classify — identify page intent (AUTH, CHECKOUT, SEARCH, CRUD, NAVIGATION, CONTENT)
4. Plan — two-phase PLAN → GENERATE split avoids token truncation
5. Generate — writes focused Playwright tests per page intent
6. Deduplicate — removes redundant tests across the batch
7. Enhance — strengthens assertions for better coverage
8. Validate — rejects malformed or placeholder output, including raw-CSS visibility/text assertions (see below)

Validator — quality gate for assertions

The validator rejects tests that chain toBeVisible(), toContainText(...), or toHaveText(...) off a raw CSS page.locator():

// ❌ Rejected — bypasses self-healing, breaks on class renames or empty states
await expect(page.locator('.todo-count')).toContainText('0 items left');

// ✅ Accepted — semantic locator, resilient
await expect(page.getByText('0 items left')).toBeVisible();

// ✅ Accepted — self-healing waterfall
await safeExpect(page, expect, '0 items left');

Count, state, attribute, class, and CSS matchers on page.locator() are intentionally allowed — toHaveCount, toBeHidden, toHaveAttribute, toHaveClass, toHaveCSS. Those don't have a safe-helper equivalent and the generation prompt explicitly recommends page.locator() for them.

Advanced Playwright capability coverage

The pipeline preserves advanced Playwright primitives end-to-end so complex flows are not downgraded to basic click/fill between generation, enhancement, validation, and self-healing stages:

• Browser/context lifecycle — browser.newContext, context.newPage, storageState, addCookies, grantPermissions, setExtraHTTPHeaders.
• Frames & Shadow DOM — frameLocator() for iframe-scoped interactions and assertions.
• Network interception/mocking — page.route, route.fulfill/continue/fallback/abort, routeFromHAR, unroute.
• API request contexts — request.newContext plus get/post/put/patch/delete/fetch/dispose for hybrid UI+API tests.
• Rich interactions — dblclick, hover, check/uncheck, dragAndDrop, locator.dragTo, setInputFiles, press.
• Diagnostics — tracing.start/stop/startChunk/stopChunk, screenshots, testInfo.attach, expect.soft.
• Test runner structure — test.describe/describe.parallel/describe.configure, beforeEach/afterEach/beforeAll/afterAll, test.step, retries/timeouts.
• Emulation — setViewportSize, emulateMedia, geolocation/locale/timezone via context options.

Tests that use these primitives are detected as "advanced scenarios" and are skipped by the assertion enhancer so its generic templates can't break bespoke orchestration. The feedback loop also classifies advanced failures into dedicated categories (NETWORK_MOCK_FAIL, FRAME_FAIL, API_ASSERTION_FAIL) with targeted regeneration instructions instead of falling through to UNKNOWN.

Generate from Description

Skip crawling entirely — open Create Tests, write a plain-English scenario, and Sentri generates the steps and Playwright code. Watch AI output arrive token by token via LLM streaming.

Test Dials

Configure generation behaviour before hitting Generate:

• Strategy: happy path, edge cases, comprehensive, exploratory, regression
• Workflow: E2E, component isolation, multi-role persona, first-time user
• Quality checks: accessibility, security, performance, data integrity
• Output format: verbose, concise, Gherkin
• Test count and language

Presets like "Smoke Test" and "BDD Blueprint" auto-fill multiple dials. Config is validated server-side to prevent prompt injection.