Browser e2e

<!-- Provenance: SYNTHESIZED (not re-authored from one external repo) for a Multica squad from research/2026-06-27-ai-agents-for-qa-and-testing.md §5D + findings #7/#8, which rest on WebArena (arXiv:2307.13854), VisualWebArena (2401.13649), SeeAct (2401.01614), Online-Mind2Web (2504.01382), WebTestPilot (2602.11724), and Practical Limits of Autonomous Test Repair (2605.01471). Wired to the target repo's real surfaces: the playwright MCP already mounted on the QA seat and the project's own UI bring-up harness. Subordinated to the QA card's scenario contract (qa.md — scenario shape, PASS/FAIL, two artifacts, smallest-repro) and eval-harness (determinism split, pass@k/pass^k). -->

Browser E2E (grounding + locators + escalate-don't-weaken)

Your scenario list, PASS/FAIL-per-scenario, two-artifacts, and smallest-repro live eager in your QA card (qa.md); your non-determinism handling (assert a predicate over k runs, pass@k vs pass^k) lives in eval-harness. This skill adds only the missing piece: how to drive a browser reliably so a UI verdict is grounded, not guessed. Expect a real human-vs-agent gap (web agents sit ~30–50 pts below humans, and grounding — intent→correct element — is the wall, not reasoning) → keep the human / strong-judge gate; never claim autonomous full-coverage regression.

When this mode applies

The target repo's web surface ({{UI_SURFACE}}) — NOT native-channel rendering. Drive it with the playwright MCP already on your runtime.

• Mode-1 mock-LLM (DEFAULT). UI behavior over the harness (routing, state render, ACL-gated views, form→request plumbing). Deterministic → assert exactly.
  • Bring-up (REQUIRED). Use {{E2E_HARNESS}}'s dedicated UI bring-up target — the base stack (gateway + mocks) is a separate lane from the UI stack on purpose, so a UI boot failure never gates the non-UI suite. Drive the browser at whatever URL the bring-up reports, and tear down with the matching teardown target.
  • Identity (REQUIRED for any RBAC assertion). Dashboard identity is resolved by whatever mechanism {{UI_SURFACE}} names (e.g. a proxy-injected header, a session cookie, a bearer token). A headless browser carries no real identity unless you set it explicitly: seed the target role/identity via the admin-side mechanism your bring-up exposes, then attach that identity to every navigation + XHR the same way the mechanism expects. Without this, "viewer sees X but not Y" — and even a plain authenticated-render assertion — silently passes against the WRONG identity. Do NOT build a browser RBAC deny-matrix if the repo already has one: check whether {{UI_SURFACE}} names an existing RBAC policy test first — a static/fail-closed test there is stricter than a browser can be.
• Mode-2 real-model only when the UI's correctness depends on model reasoning (e.g. an agent reply rendered in the UI): {{E2E_HARNESS}}'s live/real-dependency variant. Real cost, non-deterministic → defer to eval-harness (predicate over k runs, pass^k=100% on user-facing paths). Keep these suites small.
• This is distinct from native-channel rendering (a native messaging surface's own client), which stays un-automatable — a browser does not test a native channel.

Grounding — feed BOTH channels, never screenshot-only

Give the model the accessibility tree / DOM snapshot AND a screenshot for each decision. Screenshot-only grounding collapses on real tasks (≈5–6% on hard sets); the DOM gives stable identity, the screenshot gives layout/visual state. When marking elements for the model, use a Set-of-Marks overlay (numbered boxes) so "click element 7" is unambiguous. Assert on observable DOM/state (text content, aria, a mock-outbound field, a row count), never on pixels.

Locators — tiered, role-first; LLM only on the break

Resolve every target in this order, stopping at the first that's stable: get_by_role (name/role) → data-testid → ARIA (aria-label/role) → CSS → visible text. Role/test-id locators survive restyles; text/CSS are last resorts. Only invoke the LLM to re-ground a locator that actually broke — cache the healed selector for the rest of the run so you don't re-pay per step.

Self-heal that ESCALATES, never weakens (the load-bearing rule)

A bounded heal loop is allowed for a broken locator (the page changed, the intent didn't). It is NOT allowed to make a red test green by changing what the test checks. If convergence would require weakening an assertion, broadening a predicate, deleting a step, or dropping a scenario — STOP and escalate, do not weaken. That is the documented reward-hack anti-pattern (an enterprise self-healing study hit 70% "convergence" but only 10% first-attempt success, with 38% producing no test, via assertion-weakening + test-deletion). This also binds via the constitution's TEST INTEGRITY clause.

• Heal attempts count against your 2-cycle FAIL budget (your only fuse).
• Escalate through the existing paths: a code verdict → {{LEAD_SEAT_NAME}} mention; an environment blocker → the human-MEMBER mention + the one-line unblock command. Invent no new channel (anti-loop).

Bounds for a headless run

• Step cap 15–50 per scenario — your runtime has no per-run turn fuse, so cap explicitly and report partial state if you hit it.
• No-clarifying-question contract. You are headless; there is no human to answer mid-run. Never pause to ask — make the safe assumption and note it, or escalate via the MEMBER mention.

Oracle inference

Derive the expected result from the NL spec + the symbolized GUI elements (the WebTestPilot pattern): name the elements, state each scenario's expected post-condition as a binary observable per your QA card's scenario contract (qa.md). A flaky UI assert is real noise → quarantine per eval-harness (rerun, report FAIL only when it reproduces) before counting it.

Defer

Report in your QA card's PASS/FAIL + two-artifacts + smallest-repro format (qa.md); take all non-determinism / reliability-metric rules from eval-harness. This skill changes only how you reach the element and the verdict, not how you list scenarios or report them.