Selector Drift Recovery

<objective>
A redesign shipped and 23 tests now fail with `TimeoutError: locator.* exceeded` — the DOM moved, not the product. This skill closes that maintenance loop: it diffs the old DOM against the new one, regenerates the broken selectors role-first against the new build, validates the whole suite, and ships the diff as one reviewable PR with a confidence score and screenshot per change. The trigger is an event (a refactor merged), not a flake. The output is a PR a human signs off on — not a silent runtime auto-heal.

This is the bulk, offline, batch counterpart to test-reliability. They share the multi-attribute locator + confidence scoring primitives but run in opposite directions: test-reliability heals one selector at runtime behind a guarded threshold; selector-drift-recovery regenerates N selectors offline against the new DOM and bundles them into a PR. If you are doing the second workflow inside test-reliability, switch here. </objective>

Quick Route

---

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it and skip anything answered there. It identifies your E2E framework, selector strategy, and known fragile areas. Then:

1. What triggered the drift? A planned refactor (Storybook can show the new DOM before merge), a shipped redesign (new DOM is in main), a dependency upgrade, or a Tailwind/CSS migration? The trigger decides whether you run pre-emptively or react to CI failures.
2. What is the blast radius? A single component, a page, or the whole app? Single component: scope recovery to the test files that touch it. Global: budget half a day to a day, and decide which tests should be rewritten rather than re-selected.
3. What is your current selector strategy? If selectors are mostly data-testid and the refactor preserved testids, recovery is trivial. If they are CSS-class or XPath based, expect 30–60% to need a new strategy, not just a new locator.
4. Is there a passing baseline? You need the old DOM somewhere: a previous CI trace artifact, a deployed staging build, a Storybook story, or git history of the components. No old-DOM reference means this degrades to "rewrite tests."
5. Are the broken locators inline or wrapped in a Page Object? The JSON reporter's error.location points at the failing line. For inline locators that is the locator itself; for POM-wrapped locators it points at the POM helper, not the test. Know this before you trust the auto-extracted line numbers (see Failure Modes).
6. Who reviews the resulting PR? Confidence-scored updates need a human signoff. Decide upfront whether the PR goes to the test author, the engineer who did the refactor, or the QA lead.

---

Core Principles

1. Recovery is event-driven, not failure-driven. Run this when a refactor is planned or just shipped, not when one test goes flaky. One broken test → test-reliability. Ten or more from the same event → this skill.

2. Old DOM, new DOM, mapped pair. The whole skill rests on a snapshot before the refactor and one after. Everything else is bookkeeping around that pair. Capture both as aria snapshots (await page.locator('body').ariaSnapshot()), not raw HTML — a role-tree diff is exactly the signal role-first recovery needs and ignores the cosmetic churn (class renames, wrapper divs) that a raw-HTML diff drowns in. If you cannot produce both snapshots, fix that first.

3. Role-first replacement, every time. Even when the old test used CSS, the regenerated selector should prefer getByRole + accessible name, then getByLabel for form fields, then getByTestId when the refactor added one. The recovery PR is your chance to ratchet the average selector stability score up (0–5 rubric below, shared with test-reliability).

4. Disambiguate by region scoping, not layout selectors. When role+name is ambiguous (two "Submit" buttons), narrow with getByRole('region', { name }).getByRole('button', …) or getByRole(...).filter({ hasText }). Do not reach for :near() / :right-of() — see the Avoid note. A score-3 candidate is one where region scoping has been applied and the locator now matches exactly one element.

5. One PR, grouped by file, with per-change evidence. Reviewers cannot eyeball 47 selector changes spread across 30 commits. Bundle into one PR, group hunks by test file, attach a confidence score + DOM screenshot per change.

6. The suite must pass before merge, and dead tests get deleted. A regenerated selector that doesn't run is a worse version of the original problem; the skill ends with green CI, not a generated diff. And if the refactor removed a feature, prune its tests — do not regenerate selectors for elements that no longer exist.

Avoid: Playwright layout selectors :near(), :right-of(), :left-of(), :above(), :below() as disambiguators — officially deprecated and "may be removed," because a 1px layout shift changes the match (Playwright docs, 2026). They also contradict the role-first thesis. Use region scoping / getByRole().filter() instead.

---

Workflow

Six phases, each gated by a check before the next.

Phase 1 — Snapshot the old DOM

You need a snapshot of every page or component the affected tests touch, in its pre-refactor state. Sources, in preference order:

1. Last green CI trace artifact. Most teams save Playwright traces on failure (trace: 'on-first-retry'). Download the green-run artifact, open a trace with npx playwright show-trace traces/checkout.zip, select an action, and read the per-action DOM snapshot panel for each surface. (The viewer no longer has a "Copy HTML at this step" menu item; you read the snapshot panel or, for a programmatic dump, replay with page.content() / ariaSnapshot().)
2. Storybook at a pre-refactor commit. git checkout <PRE_REFACTOR_SHA>, start Storybook, and dump each story with a tiny page.content() / ariaSnapshot() script.
3. A staging build still on the old version. Navigate the same flows and snapshot.
4. Git history of the components. Reconstructable but the most expensive — render in isolation.

Output: .drift-recovery/old/<page-or-component>.aria.yml (and .html if you also need raw markup) per affected unit.

Gate: You can answer "what did this page look like when the tests last passed?" from a snapshot file, not from memory.

Phase 2 — Snapshot the new DOM

Run the same surfaces in the post-refactor build — a Vercel/Netlify preview deploy is ideal, or a local dev server / the PR branch in CI. Wait for hydration (await page.waitForLoadState('networkidle')) before snapshotting, or SSR pages give you the pre-hydration tree and you miss client-rendered elements.

Output: .drift-recovery/new/<page-or-component>.aria.yml matching the old set.

Gate: Every old snapshot has a matching new one. If a route 404s in the new build, that flow was deleted — mark its tests for the deletion pile in Phase 6.

Phase 3 — Identify broken selectors and infer intent

For each test file, run against the new build with the JSON reporter, then parse it. Capture, per failure: file, line, old locator string, error type (timeout vs assertion), and inferred intent. Group the results by test file.

• Error classification: a drift failure is TimeoutError: locator.* exceeded. Distinguish it from an assertion failure (expect(...).toBe) so you don't try to re-select a locator that resolved fine but failed a value check.
• Inferred intent is mandatory and not in the reporter. Read the surrounding test code — what action is taken on the locator, what assertion follows — and store a short intent string ("submit the order", "read the order total"). The candidate generator keys off intent, so this step is load-bearing, not commentary.
• Page route is also not in the reporter. Map each locator to the snapshot it should resolve against (which .drift-recovery/new/*.aria.yml) so Phase 4 can load the right new DOM.

The result is a per-file table:

See references/recovery-scripts.md for identify-drift.ts, which produces exactly these rows (with the intent/route fields populated, not stubbed).

Gate: Every broken locator has an inferred intent and a page route. If you cannot infer intent, ask the test author or read the original PR — do not guess.

Phase 4 — Generate replacement candidates

For each row, generate candidates against the new DOM snapshot and score each on the 0–5 rubric (shared with test-reliability). Strategy ladder, best first:

1. New data-testid added by the refactor team — the most stable choice they made. Score 5.
2. getByRole + accessible name, unique on the page. Score 4.
3. getByLabel for a form field, when the intent is an input and a label exists. Score 4 (use over a bare role when the field has no name otherwise).
4. Role + name, region-scoped to a single match. If role+name alone returns >1 element, wrap it — getByRole('region', { name }).getByRole(role, { name }) or .filter({ hasText }) — and confirm the scoped locator now matches exactly one. Only score 3 after scoping makes it unambiguous.
5. Visible text only (getByText). Score 2 — fragile to copy changes.
6. CSS class on the changed structure. Score 1 — usually still broken.
7. No safe replacement. Score 0 — flag for human.

A candidate is score 3 only if scoping already resolved it to a single element. A still-ambiguous multi-match (count > 1, "needs scoping") is not a 3 — it is unfinished, and must not be auto-applied.

Output: .drift-recovery/candidates.json with { file, line, oldLocator, selector, score, rationale, screenshotPath } per change. See references/recovery-scripts.md for generate-candidates.ts.

Gate: Every row has a candidate scored ≥ 3 (and confirmed unique), or is flagged for human review. Score-0/1/2 are never auto-applied.