Self-healing vs static Playwright selectors, and the third option nobody mentions

Default to stable semantic locators (getByRole, getByLabel, getByText) because they survive most DOM churn without any extra machinery. Reserve self-healing for genuinely high-churn UI where selectors change every sprint, and never let a heal auto-pass unsupervised in CI: a selector that 'fixes itself' can mask a real regression. If you want to skip the dilemma entirely, store no selector and resolve the element fresh on every run.

A static selector is a string you freeze at write time, for example page.locator('.btn-primary'). When the class changes, the test throws. A self-healing selector starts from that same stored string, but when it fails the framework tries fallback strategies (other attributes, nearby text, a learned pattern) and rewrites the selector to whatever still matches. The heal keeps the test green, which is the point and also the danger.

Because a passing test is only useful if it fails when the product breaks. If a button's purpose changed but a heal re-pointed the selector to a different element that happens to match, the test passes while the user-facing flow is broken. Self-healing trades brittleness for the risk of false confidence. The mitigation is to surface every heal as a reviewable event, not to let it run silently in CI.

Assrt does not store a CSS or XPath selector at all. A scenario step is saved as natural language. In the source, a TestStep is { action, description } (src/core/types.ts), never a locator string. At run time the agent calls browser_snapshot to read the live accessibility tree, acts on a ref ID, and if a ref is stale it re-snapshots to get fresh refs (agent.ts, Selector Strategy section). There is no frozen locator to break, so there is nothing to heal.

No tool can promise that. Resolving from the accessibility tree by role and accessible name is closer to how a user finds an element than a CSS class is, so it tracks intent better than a static string. But a step described as 'click Submit' will still find a Submit-shaped element. You keep assertions doing the real verification work. The difference is that you are not also maintaining a separate brittle layer of selectors on top.

Assrt generates standard Playwright files you can inspect, modify, and run in any CI pipeline. It is fully open source with zero vendor lock-in, so the output lives in your repo and runs on your runners. That is the opposite of closed self-healing SaaS where the selector logic lives behind someone else's API.

When you control the markup and can add stable test ids, static getByTestId selectors are explicit, fast, and trivially debuggable. For a small, stable surface that rarely changes, a hand-written locator is perfectly fine and arguably clearer than any healing layer. The cost only shows up at scale, on high-churn UI, or on apps where you do not own the DOM.