feat(mcp): add compress option to browser_snapshot to collapse repeated ARIA nodes

[Josef-Le]

Closes #41395

Summary

Adds a compress?: boolean parameter to the browser_snapshot MCP tool. When set, a two-pass algorithm collapses repeated structural patterns in the ARIA snapshot YAML — keeping the first 10 occurrences of any pattern that appears more than 100 times — and emits a trailing note explaining how to enumerate the full list via browser_evaluate().

Motivation

On pages with large lists, data grids, or autocomplete menus, browser_snapshot can return thousands of lines. A GitHub issues page with 100 open issues produces ~1 800 lines; a spreadsheet with 500 rows produces ~6 000 lines. The vast majority is structurally identical — only the text and [ref=eN] differ. This wastes tokens and can cause the model to miss important elements buried in the noise.

Algorithm

Two-pass, O(n) time and space:

1. Pre-scan (safety gate): Normalise every line to a structural signature by stripping [ref=eN], accessible names, and bare numbers. Count (indent, signature) pairs. Only proceed if the maximum count exceeds FIRE_THRESHOLD = 100. This prevents false positives on diverse pages (dashboards, forms, settings pages).

2. Compression pass: Walk the YAML line by line; keep the first KEEP_N = 10 occurrences of any repeated pattern; collapse the rest along with their descendant subtrees. Lines matching interactive roles (button, input, link, checkbox, menuitem, …) are unconditionally kept.

3. Trailing note: Informs the model how many lines were removed and suggests browser_evaluate() for full enumeration.

Example

Page with 150 <li> elements:

# before: 152 lines
- list [ref=e2]:
  - listitem [ref=e3]: Item 1
  ... × 150

# after compress: true — 13 lines
- list [ref=e2]:
  - listitem [ref=e3]: Item 1
  ...
  - listitem [ref=e12]: Item 10

[playwright-compress: 140 repeated ARIA nodes collapsed — use browser_evaluate() to enumerate the full list]

Files changed

File	Change
packages/playwright-core/src/tools/backend/ariaCompression.ts	New — pure compression function, self-contained
packages/playwright-core/src/tools/backend/snapshot.ts	Add compress to browser_snapshot zod schema
packages/playwright-core/src/tools/backend/response.ts	Wire compress through setIncludeFullSnapshot()
tests/mcp/snapshot-compression.spec.ts	4 tests: happy path, small-list passthrough, interactive-element preservation, compress: false escape hatch

Design notes

• FIRE_THRESHOLD = 100: a typical sidebar/nav has <20 items; a real pagination list has >100. The gap makes false positives extremely unlikely.
• KEEP_N = 10: enough for the model to understand the list structure and extract the pattern for browser_evaluate().
• Explicit opt-in: the model decides when to use compression. It is never forced on, so existing behaviour is fully preserved.
• Interactive roles always kept: buttons, inputs, links, etc. carry distinct meaning even when visually identical in a list of search results. Collapsing them would silently hide actions.
• The compression function is a pure transform — no side effects, no I/O — so it can be unit-tested without a browser.

Testing

The 4 tests in tests/mcp/snapshot-compression.spec.ts cover:

1. compress: true collapses a 150-item list to the first 10 items
2. compress: false returns the full snapshot unchanged
3. compress: true does not fire on a 50-item list (below FIRE_THRESHOLD)
4. compress: true preserves all 150 buttons (interactive-role guard)

Algorithm correctness was also verified with a standalone Node.js harness (18 assertions, all pass) prior to integration.

---

This PR is opened as a draft pending maintainer triage of issue #41395.

[Josef-Le]

@microsoft-github-policy-service agree

[pavelfeldman]

Closing while discussion takes place in the issue.