evals(cua): add a deterministic CUA agent regression task

[yawbtng]

why

CUA function-response handling has had two recent bugs (#2046 fixed by #2159, #2035 with proposed fix in #2036). Both went undetected because the CUA agent loop is not exercised by any small, deterministic bench task — mode: "cua" is only used today inside the WebVoyager/OnlineMind2Web benchmark suites, which are too heavyweight to run as quick CI signal and too non-deterministic to attribute failures to a specific provider.

So the gap is: there's no fast, fixture-backed regression check that just answers "does the Google/Anthropic CUA function-response loop work end-to-end?"

proposal

Add one new bench task under packages/evals/tasks/bench/agent/:

• File: packages/evals/tasks/bench/agent/cua_amazon_checkout.ts
• Fixture site: https://browserbase.github.io/stagehand-eval-sites/sites/amazon/ — already used by act/amazon_add_to_cart.ts, so the page is known-stable.
• Flow: agent.execute({ instruction: "Add the product to the cart and proceed to checkout", maxSteps: 10 }).
• Pass criterion: final URL is .../amazon/sign-in.html (same check the existing act task uses).
• Intended invocation: evals run agent/cua_amazon_checkout --agent-mode cua --model google/gemini-2.5-computer-use-preview-10-2025 (and analogous Anthropic / OpenAI CUA models).

Mirrors the shape of the existing agent/google_flights.ts etc., adds no new framework surface, no new dependencies.

what this catches

• Provider-specific image/function-response encoding regressions like core(cua): Google function-response image handling hardcodes PNG mimeType #2046 and core(cua): Anthropic tool_result image handling hardcodes PNG media type #2035 (which break only outside the PNG happy-path, so they're invisible to the act-tier tests).
• CUA-mode plumbing failures in --agent-mode cua end-to-end (currently only validated transitively via benchmark suites).
• Provider-level differences without touching webvoyager — much faster to run than the full benchmark.

ask

@miguelg719 — would you approve a PR of this shape, or is there a different fixture / pass-criterion / file layout you'd prefer? I'll write it as soon as you approve the shape. (For context, I also have #2159 open with a direction question for #2035 — both are CUA function-response fixes so the same task would cover regressions for both.)

[ferhimedamine]

Deterministic CUA agent regression tests are tricky because the web changes — a screenshot-based test that passes today might fail tomorrow because a button moved a few pixels or a page layout shifted.

One pattern that works better than pixel-level comparison: define test assertions in terms of semantic states rather than visual appearances. 'The cart should contain 2 items at checkout' is more durable than 'the cart icon should appear at coordinates (X, Y).' The CUA agent's own perception layer can drive these assertions.

For regression tests that need to verify memory-dependent agent behavior specifically (e.g., 'does the agent remember the user's shipping address from the previous session?'), having a controllable memory backend is important. dakera-deploy's REST API lets you pre-seed agent memory with test data and verify what was stored, which is useful for this kind of memory regression testing.