Imported from comet-ml/opik (.agents/skills/playwright-pom-discovery/SKILL.md). Install upstream with npx skills add comet-ml/opik --skill playwright-pom-discovery. Copyright stays with the author.

Playwright POM Discovery

This skill is the how of choosing selectors for a POM in the Opik E2E suite. You already know which POM you're building; this skill tells you how to figure out what's on the page, what selectors are stable, and how to verify your method actually works before checking it in.

Announce at start: "I'm using the playwright-pom-discovery skill to build the X page object."

When this skill applies

• You're touching anything under tests_end_to_end/e2e/pom/.
• You need to write a data-testid selector, a getByRole, or any other Playwright locator targeting the live Opik UI.
• You're adding a method to an existing POM that interacts with a new element you haven't seen before.

It does NOT apply to:

• Pure fixture work (no UI interaction).
• Matchers / type-only changes.

The procedure

digraph pom_discovery {
    rankdir=TB;
    "Identify target page + entity preconditions" [shape=box];
    "Seed required state (project, dataset, trace, etc.)?" [shape=diamond];
    "Create state via SDK / bridge before opening UI" [shape=box];
    "Open page via browser_navigate (with auth)" [shape=box];
    "browser_snapshot to get accessibility tree" [shape=box];
    "browser_evaluate to enumerate data-testids" [shape=box];
    "Pick selector preference (testid > role > label > css)" [shape=box];
    "Write POM method" [shape=box];
    "Verify via browser_generate_locator and test_run" [shape=box];
    "Method works?" [shape=diamond];
    "Selector unstable or missing?" [shape=diamond];
    "Flag for FE data-testid addition" [shape=box];
    "Commit POM method" [shape=box];

    "Identify target page + entity preconditions" -> "Seed required state (project, dataset, trace, etc.)?";
    "Seed required state (project, dataset, trace, etc.)?" -> "Create state via SDK / bridge before opening UI" [label="yes"];
    "Create state via SDK / bridge before opening UI" -> "Open page via browser_navigate (with auth)";
    "Seed required state (project, dataset, trace, etc.)?" -> "Open page via browser_navigate (with auth)" [label="no — stateless page"];
    "Open page via browser_navigate (with auth)" -> "browser_snapshot to get accessibility tree";
    "browser_snapshot to get accessibility tree" -> "browser_evaluate to enumerate data-testids";
    "browser_evaluate to enumerate data-testids" -> "Pick selector preference (testid > role > label > css)";
    "Pick selector preference (testid > role > label > css)" -> "Write POM method";
    "Write POM method" -> "Verify via browser_generate_locator and test_run";
    "Verify via browser_generate_locator and test_run" -> "Method works?";
    "Method works?" -> "Commit POM method" [label="yes"];
    "Method works?" -> "Selector unstable or missing?" [label="no"];
    "Selector unstable or missing?" -> "Flag for FE data-testid addition" [label="yes"];
    "Selector unstable or missing?" -> "browser_snapshot to get accessibility tree" [label="no, try different selector"];
    "Flag for FE data-testid addition" -> "Pick selector preference (testid > role > label > css)";
}

Step-by-step

1. Identify target page + preconditions

Before opening the browser, answer in writing:

• What page does the POM model? E.g., LogsPage models /<workspace>/projects/<projectId>/logs. Get the route from the FE source under apps/opik-frontend/src/v2/pages/ — each page has a directory matching its name.
• What entities must exist for the page to show real data? Examples:
  • LogsPage — needs a project AND at least one trace under it. An empty project shows only the empty state.
  • DatasetItemsPage — needs a dataset AND at least one item. Empty datasets show only the "Create item" CTA.
  • TestSuitesPage — needs at least one test suite to show row interactions.
  • OnlineEvaluationPage — works empty, but creating a rule shows the rule list.
• What auth state does the page need? The browser session needs the workspace authenticated. See the auth setup below.

If the page genuinely is stateless (e.g., an empty list page), skip seeding and go straight to step 3. If it's not, seed first — exploring an empty-state UI will lead you to write a POM that only ever sees the empty state.

2. Seed required state via SDK / bridge

Never click-create state through the UI just to populate the page you're exploring. Two reasons:

1. The page you're exploring may itself BE the UI-create flow you're trying to model — you'd be using the UI to test the UI.
2. UI-create is slower and flakier than SDK-create.

Instead, use the bridge or the TS SDK to seed before opening the browser. For the discovery phase, a one-off seed script is fine — you don't need to wire it into the test suite yet (the test's fixture handles that later).

Example seed for LogsPage discovery:

// scratch script, not committed
import { Opik } from 'opik';
const opik = new Opik({
  apiKey: process.env.OPIK_API_KEY,
  workspaceName: process.env.OPIK_WORKSPACE,
  apiUrl: process.env.OPIK_BASE_URL + '/api',
});

// project + a few traces with structure variety
const projectName = `discovery-logs-${Date.now()}`;
await opik.api.projects.createProject({ name: projectName });

import { track } from 'opik';
const decoratedFn = track({ name: 'discovery-trace', projectName }, async (input: string) => {
  return `output for ${input}`;
});
await decoratedFn('hello');
await decoratedFn('world');
await opik.flush();

console.log(`Discovery project: ${projectName}`);

Run it locally with staging or local-dev creds, note the project name, then navigate the UI to that project's logs page in the next step. Tear it down by hand at the end of the discovery session (backendClient.deleteProject if it's wired up, or curl otherwise) — discovery state should never accumulate.

Reusable seed patterns by page family:

Common gotchas:

• Trace ingestion is eventually consistent. After opik.flush(), the trace may take 1–3 seconds to appear in the Logs page. If you snapshot too fast, you'll see the empty state. Wait or refresh.
• Online scoring rules apply asynchronously — if you seed a trace and then snapshot the Online Evaluation page, scores may not have landed yet.
• Dataset items have a slight delay between insert and table-render — usually ms, but watch for it.

3. Open the page with an authed browser session

Auth setup once per discovery session:

The browser MCP needs an authenticated page. Against local OSS (http://localhost:5173, workspace default) there's no login wall — navigate straight to the page. Against an authenticated deployment, reuse the suite's storage state: global-setup mints .auth/user.json the first time you run a Playwright test, and the browser MCP can load it as its storageState. Don't script a full login flow during discovery — it's a distraction.

Arm the dialog handler before your first navigation (below). Several Opik pages guard against data loss with a native beforeunload "Leave site?" confirm — it fires the moment you navigate (or reload) with unsaved state: a staged draft on the Dataset Items page (useNavigationBlocker), a dirty form, an open editor. A native dialog is not in the accessibility snapshot, so browser_navigate just silently blocks on it until it times out — you won't see why. Do this before the navigation step below, not after. Guard against it two ways, both cheap:

1. Register an auto-accept handler at the start of the session, before your first navigation:

   mcp__Playwright__browser_handle_dialog(accept=true)
   

   This dismisses any beforeunload/confirm that appears so navigation never hangs. (If one has already blocked you, call the same tool to clear it, then carry on.)

2. Leave the page clean. Before navigating away, clear unsaved state through the UI the way a user would — commit or discard the draft, close the editor, reset the form. This is also what the POM method itself must do, so doing it in discovery validates that path. Don't rely on the auto-handler alone: a discarded draft leaves the page in a real, testable state; a force-dismissed dialog leaves stale draft state behind.

Standard discovery navigation:

mcp__Playwright__browser_navigate(url="http://localhost:5173/...")

Get the exact route from the FE source — apps/opik-frontend/src/v2/router.tsx for the route table, or the page directory under apps/opik-frontend/src/v2/pages/<PageName>/. Most data pages are project-scoped, e.g. /{workspace}/projects/{projectId}/datasets/ or /{workspace}/projects/{projectId}/logs. Confirm against the router rather than guessing — routes change.

4. Snapshot the accessibility tree

mcp__Playwright__browser_snapshot()

This returns the structured accessibility tree, not pixels. Each interactive element has:

• A role (button, textbox, link, combobox, row, etc.)
• An accessible name (the visible text or aria-label)
• A stable ref ID you can use with browser_click(ref="...") to interact

Why this is the right primitive: Playwright's getByRole(...) selectors map 1:1 to what the snapshot shows. If you see button "Create suite" in the snapshot, you can confidently write page.getByRole('button', { name: 'Create suite' }).

Read the snapshot before writing any selector. Don't guess. Don't grep the FE source for what you think the button is called — the rendered DOM is the only source of truth that matters.

5. Enumerate data-testids

The snapshot tells you what's interactive but not what's been explicitly marked stable by the FE team. For that:

mcp__Playwright__browser_evaluate(function="""() => {
  return Array.from(document.querySelectorAll('[data-testid]'))
    .map(e => ({
      testid: e.getAttribute('data-testid'),
      tag: e.tagName.toLowerCase(),
      text: (e.textContent || '').slice(0, 60).trim(),
      visible: e.offsetParent !== null,
    }))
    .filter(e => e.visible);
}""")

This returns every test id currently rendered on the page, with enough context to know what each one is. Test ids are the preferred selector — they're the FE team's contract for "this won't change." Use them when they exist.

6. Pick the right selector

In priority order:

1. data-testid — most stable. page.getByTestId('create-suite-button'). If a test id exists, use it.
2. getByRole(name) — stable across most refactors as long as the accessible name doesn't change. page.getByRole('button', { name: 'Create suite' }).
3. getByLabel for form inputs that have a label. page.getByLabel('Dataset name').
4. getByText — fragile if the text is dynamic or i18n'd. Use only for truly static labels.
5. CSS / XPath selectors — last resort. Use only when no test id and no accessible name exists, and leave a comment explaining why: // no test id; FE team to add — link to ticket.

The decision is binary at write time, not runtime. Pick one selector and commit to it — write deterministic selectors, not runtime-healing ones.

7. Use browser_generate_locator when you're unsure

If the accessibility tree shows three buttons with similar names, or the element has both a test id and a role and you're not sure which is canonical:

mcp__Playwright__browser_snapshot()  # to find the ref of the element
mcp__playwright-test__browser_generate_locator(ref="<the-ref>")

This returns the locator code Playwright itself would generate if you used codegen against this element. Trust that output — Playwright's locator selection logic is well-tuned for resilience.

If you don't have access to mcp__playwright-test__ tools, fall back to writing the selector manually based on the snapshot + test id list, then verify in step 8.

8. Verify by running

After writing the POM method, the verification loop is:

// scratch test, not committed
import { test, expect } from '@playwright/test';
import { LogsPage } from '../pom/logs.page';

test('discovery: LogsPage.filterByProject works', async ({ page }) => {
  await page.goto('http://localhost:5173/<workspace>/projects/<seed-project-id>/logs');
  const logs = new LogsPage(page);
  await logs.filterByProject('discovery-logs-...');
  expect(await logs.countTraces()).toBeGreaterThan(0);
});

Run it through:

mcp__playwright-test__test_run(testPath="path/to/scratch.spec.ts")

If it passes, the POM method works against the live page. If it fails, read the failure trace (Playwright's artifacts), don't just adjust selectors blindly. Common failures:

• Timeout waiting for element — selector is wrong. Re-snapshot and pick a different one.
• Element resolved but assertion failed — the POM method's logic is wrong, not the selector. Fix the method, not the selector.
• Multiple elements matched — selector isn't specific enough. Add a parent scope or use a more precise role.

9. When a stable selector doesn't exist

If after all of the above the only working selector is a CSS path like .MuiTable-root > tbody > tr:nth-child(2), stop and flag it. The missing-data-testid protocol:

• Default: add a data-testid to the FE component in the same change as your POM. Find the component under apps/opik-frontend/src/v2/pages/<Page>/... or its shared-component dependency, add data-testid="<descriptive-name>", then use that in the POM.
• Fallback: if blocked from touching the FE, use getByRole with explicit accessible name (survives most refactors).
• Last resort: structural CSS selector with a comment explaining why it's necessary and that a data-testid should be added.

The data-testid naming convention is kebab-case, descriptive, scoped to the page/component: dataset-items-table, create-suite-button, trace-row-{traceId}. Avoid generic names like submit-button that conflict across pages.

10. Commit and move on

Once the POM method works against the live UI:

• The POM file change goes in with the test that uses it.
• Any FE data-testid additions go in the same change (cross-package is fine; reviewers expect it for test-enablement work).
• Any scratch seed script and scratch test file are NOT committed. Delete them first.
• Tear down any state you created during discovery (backendClient.deleteProject(...) or curl -X DELETE ...).

Anti-patterns

These are red flags that mean you skipped a step:

Where this fits

This is the discovery sub-step of writing an E2E test. The writing-e2e-tests skill invokes it once it has scoped the test and analyzed the feature: you seed the page's state, explore the live UI, and come back with the selectors each POM method will use plus any data-testids the FE needs. The test and POM get written and run from there.