A published npm package (v1.0.0) that fixes broken Playwright selectors. It fingerprints elements while your test suite is green, and when a selector breaks, matches the element's new identity and suggests the fix — as a diff on your pull request. It never edits your source files.

Status: v1.0.0 · stable public API · security-reviewed · MIT

Quick links

The 30-second version

// before
import { test, expect } from "@playwright/test";
// after
import { test, expect } from "playwright-eir";

That's the whole integration. Run your suite green once (calibration), and from then on, broken selectors produce suggestion diffs with confidence scores instead of just red failures.

Measured results

Seed-42, 8-class mutation benchmark — reproducible with one command (pnpm bench:all --seed 42):

False-heal rate is 0.0% in every class — the number the project is actually optimized for: a wrong heal that silently passes is worse than an honest failure. The two ⚠️ rows are documented structural ceilings (explained in the repo's Known Limitations), not tuning gaps — and stating them plainly is part of the point.

How it works (one paragraph)

Every successful interaction captures a compact fingerprint of the element (tag, stable attributes, text, label association, ancestor context, sibling position, quantized geometry) into committed, git-diffable JSON. On failure, a triage funnel confirms the failure is genuinely locator drift, then six deterministic similarity scorers rank every candidate in the live DOM against the fingerprint. Two gates stand between a match and a heal: a confidence threshold and a decision margin (the gap to the runner-up — what stops near-duplicate elements from being confused). An optional heal-and-continue mode retries the action and then verifies the post-condition (did the route/DOM actually change the way it did during calibration) before accepting the heal. Results flow into a custom reporter and a GitHub Action that maintains a single auto-updating PR comment.

Design principles

• Suggest, never silently modify — user source files are never touched, in any mode.
• Query methods never heal — isVisible(), count() etc. are structurally incapable of triggering healing.
• Deterministic core — the matching engine is pure heuristics. An optional Gemini fallback exists, is structurally capped at suggestion strength, and ships disabled — because 74 real measured invocations showed zero accuracy benefit over the heuristics. The measurement is public.
• Honestly benchmarked — the results table, the tuning log (including failed iterations), the security review, and every design decision's reasoning are all in the repository.

Deep-dive documents (in the repo)