[rgarcia]

Stacked on #10.

Summary

Task and verifier (retail-cart):

• The task asks for the outcome, not the route: put one item in the site's cart as a guest and show the cart. The agent may reach the item however works, including a web search to identify the exact product or part number; the cart must be on the site.
• The verifier now checks the cart itself. It opens the task's cart_url, or fetches its cart_api from the page the agent ended on, and sets cart_verified when the text matches cart_match. cart-check.json keeps what it saw. reward is still the agent's own report.

Pinned arm (codex-repl-pinned): a pinned configurator gives each site fixed settings from configurators/pinned/sites.toml:

• browser: stealth, gpu
• proxy: a new US ISP or residential proxy per trial
• persona: macOS or WebGL, applied by restarting Chromium with KERNEL_* launcher variables
• vm_setup: shell commands run as root in the browser VM before the first page load. On wbmason.com this empties the captcha-solver extension's Turnstile interceptor, which otherwise takes over the site's Cloudflare Turnstile check and never completes it.
• steering: per-site guidance added to the instruction, e.g. reload once on zoro.com, the order-code pages on newark.com, the trailing-slash search on guitarcenter.com, and add-by-part-number from the cart on mouser.com.

The arm also enables Kernel MCP's computer_action next to browser_repl.

Harness:

• KernelTaskEnvironment runs the Chromium restart and vm_setup right after creating the browser. Either failing fails the trial with VmSetupError.
• GPU browsers are created without a memory request, which the API rejects with gpu=true.
• kernel-mcp-local accepts the computer toolset.

Results

10 attempts per site, both arms in the same window, 2026-11-01. "Cart" counts trials where the verifier found the item in the cart.

Site	Pinned cart	Pinned reported	Stealth cart	Stealth reported
costco.com	10/10	10/10	10/10	10/10
harborfreight.com	10/10	10/10	10/10	10/10
homedepot.com	10/10	10/10	10/10	10/10
staples.com	10/10	10/10	10/10	10/10
target.com	10/10	10/10	9/10	9/10
walmart.com	10/10	10/10	10/10	10/10
wbmason.com	10/10	10/10	5/10	5/10
dickssportinggoods.com	9/10	9/10	8/10	8/10
ebay.com	9/10	10/10	6/10	8/10
amazon.com	8/10	10/10	9/10	10/10
newark.com	8/10	9/10	0/10	0/10
mouser.com	7/10	8/10	0/10	0/10
guitarcenter.com	5/10	0/10	0/10	0/10
radwell.com	4/10	4/10	4/10	4/10
lowes.com	0/10	10/10	0/10	9/10
zoro.com	0/10	0/10	2/10	2/10
Total	120/160	130/160	93/160	105/160

Where the two counts disagree:

• lowes.com: the agent added a Bostitch 27-pack metal kit; the task names a Stanley one, which the site no longer seems to list. The item needs updating.
• amazon.com, ebay.com, mouser.com (4 trials): opening the cart page during verification hit a checkpoint, a captcha or a timeout. The check needs to tell "couldn't check" from "not in cart".
• guitarcenter.com: the item was in the cart (cart API) in 5 trials where the agent reported failure, because the cart page never renders the line items.
• newark.com: one reported success over a cart that showed 0 items.

Testing

• ruff check, ruff format --check, pytest (26 passed)
• smoke run of the five previously blocked sites plus costco, then the 10-attempt run above
• checked the extension file path and command on a live browser before the run

[rgarcia]

Folded into #10 (retail-cart) and #9 (generic harness changes) so the stack is one generic PR and one benchmark PR.